JIT: Stop calling morph from CSE

[jakobbotsch]

CSE cannot tolerate morph removing any defs, which many of morphs transformations can do. We have been slowly infecting morph with more and more checks to avoid this, but in reality CSE just should not reinvoke morph -- the benefits of doing this are minimal.

Fix #108600

[jakobbotsch]

/azp run runtime-coreclr superpmi-diffs

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jakobbotsch]

cc @dotnet/jit-contrib PTAL @EgorBo

Diffs. Regressions in cases where some morph optimizations no longer kick in after CSE, but I think they're small enough that we can accept them and see if something shows up in micro benchmarks before porting more transformations to lowering. Some TP improvements now that we avoid the morph call.

[EgorBo]

I am curious why we didn't do it before instead of fixing morph with these checks 🙂 I presume some parts of morph also conservatively use fgGlobalMorph check

[EgorBo]

LGTM, agree that diffs aren't too bad to fix a popular source of bugs

[jakobbotsch]

I am curious why we didn't do it before instead of fixing morph with these checks 🙂

Good question, I guess it was simpler to just keep adding those checks.

I presume some parts of morph also conservatively use fgGlobalMorph check

Ah yeah, some of those can probably be cleaned up. We still call morph from some other places, like assertion prop, but it's probably not as sensitive as CSE. I'll open an issue about investigating those.

I also realized that I missed removing checks for optValnumCSE_phase from morph. Let me push a commit that removes those too...