[llvm] [InstCombine] Fold `umax/umin(nuw_shl(base, x), nuw_shl(base, y)) -> nuw_shl(base, umax/umin(x, y))` (PR #131076)

[Nikita Popov via llvm-commits]

nikic wrote:

> > Can we handle this in foldIntrinsicUsingDistributiveLaws instead? It currently assumes commutative ops, but if we drop that assumption we can handle the shift case there as well.
> 
> I'm not quite sure how to modify that function to achieve this, as it not only assumes the inner ops are commutative but also tries to swap their operands.

What I'd expect is to basically skip the swapping logic if the ops are not commutative.

https://github.com/llvm/llvm-project/pull/131076