This extends a similar pattern from D125500. If we know that operand 1 of a subtract is itself a non-overflowing subtract from operand 0, then the final/outer subtract is also non-overflowing:
https://alive2.llvm.org/ce/z/Bqan8v
InstCombine uses this analysis to trigger a narrowing optimization, so that is what the first changed test shows.
The last test models the motivating case from issue #48013. In that example, we determine 'nsw' on the first sub from the srem, then we determine that the 2nd sub can be narrowed, and that leads to eliminating both subtracts.
This works for unsigned sub too, but I left that out to keep the patch minimal. If this looks ok, I will follow up with that change. There are also several missing subtract narrowing optimizations demonstrated in the tests above the diffs shown here - those should be handled in InstCombine with another set of patches.