[llvm-branch-commits] [llvm] [LoopInterchange] Disable LoopCacheAnalysis-based heuristic by default (PR #193478)

[Ryotaro Kasuga via llvm-branch-commits]

kasuga-fj wrote:

@sjoerdmeijer As I mentioned at EuroLLVM, I think we should merge this (and all the PRs it depends on) before enabling loop‑interchange by default. As stated in the PR description, the LoopCacheAnalysis‑based heuristic is an interesting one, but its behavior is not very clear and it sometimes produces odd decisions. Using this in the default pipeline may cause us pain in the future. To make the pass stable, I believe it would be better to adopt the simpler heuristic first.

A little bonus:
- LoopCacheAnalysis is somewhat expensive, so this change might mitigate compile-time degradation.
- We will be freed from the accident of forgetting to specify `cache-line-size`.

https://github.com/llvm/llvm-project/pull/193478