[llvm] [ValueTracking] Enhance alignment propagation in computeKnownBits. (PR #166935)

[Nikita Popov via llvm-commits]

nikic wrote:

I think that making the limit low enough to avoid the compile time impact will make this optimization quite unreliable. While any given load/store user will likely have the same alignment, it's likely that the one inspected user will either be a GEP or not dominate. And of course, the picked user depends on the use-list order.

Doing this in InferAlignment should be more reliable, with the downside that it runs late in the pipeline and other computeKnownBits users can't benefit.

https://github.com/llvm/llvm-project/pull/166935