[llvm] [LV] Add vplan folds for (X - (X & Y)) -> (X & ~Y) (PR #226466)

[Luke Lau via llvm-commits]

lukel97 wrote:

> > Not sure what the fix for this is yet
> 
> That's strange because the vplan fold mirrors what instcombine would do anyway! You can see this by piping
> 
> ```
> define i64 @foo(i64 %a) {
>   %b = and i64 %a, 31
>   %c = sub i64 %a, %b
>   ret i64 %c
> }
> ```
> 
> into instcombine with the command `opt -p instcombine -S < foo.ll`

Yeah this is the bit that's making me scratch my head too. I think InstCombine must do some other combine utilising KnownBits before the `(X - (X & Y)) -> (X & ~Y)` combine. I wonder if we can utilise SCEV in VPlan when we know the trip remainder is a constant to mirror this?

https://github.com/llvm/llvm-project/pull/226466