[Perf] Investigate BenchI/Ackerman x86 Regression between release/2.0.0 and release/2.1

[michellemcdaniel]

We are regressed in x86.

[AndyAyersMS]

2.0 and 2.1 codegen is identical for both Acker and Bench.

This test is notoriously sensitive to method alignment. Let's look at the longer-term stability in master:

So there seem to be 4 stable levels: 695, 720, 770, and 790.

2.0 seems to have settled in on 695:

2.1 started there but has jumped up to 720:

So I assume this perf difference is an artifact of method entry alignment. On x86 we end up aligning most jitted method entries to 8 byte boundaries, so it would appear that there is some 32-byte sensitivity here.

At some point we might want to rethink alignment, either to stabilize cases like this, or as an enabler for loop top alignment (#11607), or perhaps in general. I have opened dotnet/coreclr#16873 to investigate method alignment impact.

I don't think there is anything actionable here for 2.1 so will close this issue.