[llvm] IndVarSimplify: fix high-cost-expand check (PR #125828)

[Ramkumar Ramachandra via llvm-commits]

artagnon wrote:

> Hm, it seems like costing IVLimit instead of ExitCount ends up having the opposite effect and we do more LFTR? Is IVLimit usually a simpler expression than ExitCount?

It's neither strictly more or strictly less complex. In some cases, we do more LFTR; in others, we do less LFTR. Considering that we're trying to phase out LFTR gradually, maybe I can add a FIXME high-cost-check with ExitCount as well, to avoid regressions?

https://github.com/llvm/llvm-project/pull/125828