Convert HELPER_METHOD_FRAME to QCalls (4/N)

[jkotas]

No description provided.

[hughbe]

Jan - is there an issue/PR somewhere that explains the rationale behind these changes (perf, correctness, other?)? just curious!

[DragosDanielBoia]

@dotnet-policy-service rerun

[am11]

@hughbe, perhaps there is no tracking issue, but I think this is a general goodness to have fewer FCalls as possible https://github.com/dotnet/runtime/blob/b345e2dbd6936eb281fadb7d70473358be578d24/docs/design/coreclr/botr/corelib.md#choosing-between-fcall-qcall-pinvoke-and-writing-in-managed-code. The ultimate goal is to gradually get rid of HELPER_METHOD_FRAME (dotnet/corert#3784 (comment)), and thereby reduce/remove the dependency on libunwind from Unix.

@jkotas, are we strictly converting FCalls to QCalls in this series, or to managed as well (when there is an opportunity)?

[jkotas]

Jan - is there an issue/PR somewhere that explains the rationale behind these changes (perf, correctness, other?)?

Good point. I have opened #95695. It is something that myself, @AaronRobinsonMSFT, @davidwrighton discussed over lunch as a general goodness a few weeks ago. I have decided to work on it as my side-project.

[jkotas]

@jkotas, are we strictly converting FCalls to QCalls in this series, or to managed as well (when there is an opportunity)?

Yes, if there is an easy opportunity. There are a few cases like that in this PR:

• ArgIterator.GetRemainingCount - move to managed
• StubHelpers - move a call of managed pointer to managed
• Moved argument checking to managed in some places

[davidwrighton]

This looks good although I might instead of converting the trailing byte code into a QCall instead produce a QCall/FCall that can produce the SyncBlock and then do the rest in managed. However, we could tackle that later.

[jkotas]

I might instead of converting the trailing byte code into a QCall instead produce a QCall/FCall that can produce the SyncBlock and then do the rest in managed. However, we could tackle that later.

Yes, it is a trade-off in how many data type definition to mirror between managed and unmanaged sides. I have been leaning towards keeping this mirroring more minimal.