[llvm] [SLP]Improve minbitwidth analysis for operands of IToFP and ICmp instructions. (PR #85966)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Hi @alexey-bataev
> 
> I've bisected a miscompile back to this patch. Reproduce with
> 
> ```
> opt -passes=slp-vectorizer bbi-94784.ll -S -o - -mtriple=aarch64 -slp-threshold=-10 -slp-vectorize-hor=0
> ```
> 
> [bbi-94784.ll.gz](https://github.com/llvm/llvm-project/files/15064102/bbi-94784.ll.gz)
> 
> For input %in1=0x3 and %in2=ffff the input function returns 0, but after slp-vectorizer with this patch is returns 2. I think the problem is that after vectorization the "mul" operates on i16 values but it should really be on i64 otherwise the compares with 196605 will go wrong.

Thanks for the reproducer, will double check

https://github.com/llvm/llvm-project/pull/85966