assigned to @rotateright

Note that this is similar to the existing @​clamp_check_for_no_infinite_loop3 in llvm/test/Transforms/InstCombine/minmax-fold.ll, but with some predicates flipped.

We could view this as a missing optimization.

The special-case of clamping to a range of only 2 values can be reduced to a single cmp+select:
https://alive2.llvm.org/ce/z/txD5Dk

I don't know why we don't hit the same inf-loop for the signed predicate example. I'll take a look.

The existing test looks similar, but it's not the same as the new problem case.

I added a test that shows we can hit the same infinite loop using a different combination of predicates here:
https://reviews.llvm.org/rGa6f79b56711e

The opposing transforms are within InstCombinerImpl::foldICmpWithDominatingICmp() and canonicalizeMinMaxWithConstant().

For intrinsics, I also added the min/max clamp optimization suggested in the previous comment::
https://reviews.llvm.org/rG025bb5290379

We could replicate that for cmp+select, but I'm not sure if that would guarantee that we don't inf-loop. Hopefully, we can switch over to the intrinsics soon and not have to deal with these conflicts as much within instcombine.