Revert "Kill REG_PROFILER_RET_SCRATCH (#37679)"

[AndyAyersMS]

This reverts commit b179e19.

[AndyAyersMS]

Undoing change that I think might have lead to gc stress failures; see #38230 and #36971.

[AndyAyersMS]

Most of the stress runs are done (and all the linux arm ones) -- there are about 30 failures on linux arm amd ~100 total failures.

Arm linux failures are mostly in various HW Intrinsics tests. So it seems the change I've reverted above is the root cause for the spike in failures.

@BruceForstall @CarolEidt -- Seems like we should revert #37679.

[BruceForstall]

Agreed, although #37223 will regress and will need a different fix. It's optional that you'd want to disable the test there that will regress.

[AndyAyersMS]

What's the mechanism for excluding a libraries test?

[BruceForstall]

E.g.,

https://github.com/BruceForstall/runtime/blob/95d9bdb3a4f250b9335035cad28ca6eb32766844/src/libraries/System.Numerics.Vectors/tests/Matrix3x2Tests.cs#L689

[SkipOnCoreClr("https://github.com/dotnet/runtime/issues/36587", RuntimeTestModes.JitStress)]

or, e.g., #38580

[CarolEidt]

Can you articulate why reverting this change eliminates the issue? Codegen still uses r2, I believe, and it is codegen, not LSRA, that tracks the GC state.

[AndyAyersMS]

There are special tricks to extend gc liveness of r0 around the late profile hook and they won't work right if the return value is temporarily elsewhere. I'll add more details when I have a chance (might be a few days).

[AndyAyersMS]

@BruceForstall looks like we didn't specify which test hit this assert, so I don't know what to disable. It might affect more than one test. @CarolEidt do you recall?

[CarolEidt]

@AndyAyersMS - were you asking about which test hit the assert that I fixed with #37679? The issue was #37223 and it failed in System.Threading.Tasks.Dataflow.Tests, though it doesn't directly fail in any of the tests themselves, and I don't recall if I was able to narrow it down to a class or method within those tests.

[AndyAyersMS]

@CarolEidt thanks -- the method was Dictionary.Remove so not surprising if it ends up impacting more than one test or in the harness itself.

Let me see if the failure still repros and if so we can figure out whether to exclude it here or do something else.

[AndyAyersMS]

Hmm, can't get CI to run. Let me close this an open a new PR.