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

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> LGTM - as far as I can tell - the AArch64/getelementptr.ll code explosion seems unfortunate though - have you confirmed it results in better codegen?

AArch64/getelementptr.ll is not a regression. It uses -slp-threshold=-7, which turns on the vectorization of previously non-profitable graph with this patch. Need to set -slp-threshold=-6 (or even -5) to return it back, but in this case the first test will be failed (unaffected by this patch). I'll split this test into 2 separate test with the different thresholds.

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