Number of merged contexts in SPMI collections differs significantly between platforms

[amanasifkhalid]

For the current (GUID: cc0e7adf-e397-40b6-9d14-a7149815c991) iteration, benchmarks.run_pgo has 95,310 contexts, of which 45,556 are FullOpts, on win-x64. On win-x86, this same collection has 124,430 collections, of which 81,492 are FullOpts. While some nominal variation in total context count is expected due to platform-specific implementation details, having almost twice as many optimized contexts is odd. @AndyAyersMS clarified that this discrepancy is likely due to how different contexts for the same method are consolidated when we have PGO data (notice how this discrepancy isn't nearly as dramatic for collections without tiering).

mcs currently identifies duplicate contexts to remove by comparing their hashes (see src/coreclr/tools/superpmi/mcs/removedup.cpp). This means mcs isn't all that nuanced in deciding whether to keep duplicate contexts with different profiles: Methods with identical profile counts are removed, but trivial differences in counts are enough to discern between two contexts. Thus, slight differences in profile count precision between platforms can significantly influence the number of duplicate contexts in the final collection.

This complicates analyzing diffs. Ideally, collections across platforms would be as similar as possible to facilitate triage, but removing profile data from the hash computation entirely such that no duplicates are removed would probably be impractical for collection sizes. It might be worth trying something more sophisticated for differentiating contexts based on their profiles; for example, maybe the profile counts should have to be different enough such that the likeliest branch out of a block changes in order for the profiles to be considered unique. However, it isn't clear to me how we would evaluate the efficacy of such a change.

For now, I think we all ought to be aware of this next time we encounter different amounts of churn across platforms. cc @dotnet/jit-contrib

[dotnet-policy-service]

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

[jakobbotsch]

Another cause of more tier 1 methods in x86 is that it lacks support for OSR, so many methods intentionally skip tier 0. It should be possible to check in the win-x64 collection how many of the tier 0 contexts would have been jitted in tier 1 on x86.

[AndyAyersMS]

I believe there was also a big discrepancy between x64 on linux/windows.

[amanasifkhalid]

I believe there was also a big discrepancy between x64 on linux/windows.

Yes, this is even more pronounced across OSes. On linux-x64, benchmarks.run_pgo has 176,455 contexts, of which 117,322 are FullOpts.