[llvm] [Analysis] Enchance alignment propagation in computeKnownBits. (PR #171828)

[Nikita Popov via llvm-commits]

nikic wrote:

I'm not sure it makes sense to extend InferAlign beyond what it handles now. Once we go beyond ptrtoint + and, there's a lot of possible patterns to handle, and hardcoding them all is not scalable.

On the other hand, doing this in ValueTracking in a way that has little compile-time impact will also be unreliable, as we'll have to severely restrict the number of visited uses.

Probably the ideal solution here is to have something like `align` attribute/metadata on `ptrtoint` so we can infer that once and then computeKnownBits can use the computed value. But that's a more complex change...

https://github.com/llvm/llvm-project/pull/171828