[CI Problem] Flaky Test: tests/python/unittest/test_meta_schedule_schedule_rule_auto_inline.py::test_inline_into_cache

[junrushao]

Branch/PR Failing

Looks like this PR introduced a flaky test. Likely it's due to some concurrent merge.

Jenkins Link

https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/2417/pipeline/

Flakiness

Yes.

CC: @MasterJH5574

[junrushao]

CC @wrongtest @Hzfengsy

It looks like some randomness was introduced recently in GetBlockReadWriteRegion, possibly in this commit, but this particular robustness issue wasn't revealed until @MasterJH5574's PR:

[wrongtest-intellif]

cc @junrushao1994 @Hzfengsy

The fix trial is here #9958

I am sorry for introducing this issue. I believe it is introduced by #9372, but detected until subsequent change #9880 which touch more testcases.

Developers have already pointed out that with equations like xx + rx >= 1, the SolveInequalitiesToRange create undetermined results for the failed case. After a look through I found that actually the result of SolveInequalitiesToRange depend on the to-solve variables order we pass in. However, in the region analysis context, I used an unordered_map with ObjectPtrHash to record the variables.

[junrushao]

Thanks for acting quickly on this issue @MasterJH5574 @wrongtest! I tried myself last night but failed to find the root reason. Thank you for finding this out!