[llvm] [polly] [SCEV] Introduce SDiv expressions (PR #216862)

[Nikita Popov via llvm-commits]

nikic wrote:

"It exists in IR so we should model it in SCEV" is not a valid argument -- we specifically want SCEV to cut off parts that are not relevant to the analysis it performs using SCEVUnknown. It reduces the size of the analyzed graph and avoids rematerialization problems.

https://github.com/llvm/llvm-project/pull/216862