[llvm] [SLP]Improve minbitwidth analysis. (PR #78976)

[Nikita Popov via llvm-commits]

nikic wrote:

It looks like this change causes large compile-time regressions in some cases: http://llvm-compile-time-tracker.com/compare.php?from=1b1aea79194117d8f1729ef9c8f80454aea381fe&to=083d8aa03aca55b88098a91e41e41a8e321a5721&stat=instructions%3Au

For example, mode_decision.c from lencod regresses by 10% with `-O3` (without codegen change).

https://github.com/llvm/llvm-project/pull/78976