fix assertion merge for degenerate flow

[kunalspathak]

Sometimes, we have degenerate flow involving 2 blocks A and B such that A->jumpDst == B as well as A->bbNext == B. In such case we should merge the assertions of out edges from both flows into the B's in edge. Eventually we want to avoid generating such flows in redundant branch optimization. See #48609.

Thanks Andy for help identify the root cause.

Fixes: #47101

[kunalspathak]

@dotnet/jit-contrib , @AndyAyersMS

[kunalspathak]

https://dev.azure.com/dnceng/public/_build/results?buildId=1006551&view=results is green again with this change.

[briansull]

Looks Good

[AndyAyersMS]

Did you get a chance to look at other merges to see if any of them might be impacted by the same problem?

[AndyAyersMS]

This deserves a comment and some extra jit dump output, since it is an unusual case.

[kunalspathak]

Agree. Added.

[kunalspathak]

Did you get a chance to look at other merges to see if any of them might be impacted by the same problem?

Are you asking about CSE? Apart from AssertionProp, that's the only one that uses DataFlow and perform Merge. For CSE, I didn't see logic where we chose which edge to merge so I think it should be fine.

If you meant something else, can you please clarify other phases that does similar thing and I will double check.

[AndyAyersMS]

Did you get a chance to run diffs for this? Curious if it popped up anywhere else, and why we didn't notice it, if so.

[kunalspathak]

Did you get a chance to run diffs for this? Curious if it popped up anywhere else, and why we didn't notice it, if so.

Yes, I ran libraries.pmi, tests and benchmarks superpmi collection and found no asmdiff. The original issue was in R2R with jitstress flag. While investigating, I tried to repro using JIT, reduce jitstress flags, etc. but none of them helped in reproing so I believe that it took a unique combination of things to get to this, so "no asm diff" don't surprise me.