[llvm] [Support] Assert that DomTree nodes share parent (PR #101198)

[Michael Kruse via llvm-commits]

Meinersbur wrote:

The more difficult problem turns out to be ScalarEvolution. It has the same problem in that Polly uses it across function boundaries, but also uses DominatorTree of which we need two (one for the original function, one for the outlined function). We also cannot just not create a DominatorTree/LoopInfo for the outline function (since it would be recomputed for the next pass anyway) because `SCEVExpander` uses them during generation.

In contrast to DomTree and LoopInfo, the separation on whether the SCEV is on the original code or the new code emitted by `SCEVExpander` is not as clear. That is, Polly subclasses it as the `ScopExpander` class and should be able to map to the outlined function. Stay tuned.

https://github.com/llvm/llvm-project/pull/101198