[llvm] [SLSR] Prove disjoint instruction reuse using an annotation-free IR (PR #218308)

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

But ignoring the correctness issue above, assuming that is an easy fix, the bigger problem with the patch is its approach to clone the whole function to do its analysis, I think. Not only do we clone it, but we then also do a full walk and use different data structures. This all looks very compile time and memory expensive if there is for example just 1 disjoint or in a large function. This approach looks like a showstopper to me. The question I have is: is there a different approach?

https://github.com/llvm/llvm-project/pull/218308