[llvm] [LLVM][LangRef] Remove "n > 0" restriction from get.active.lanes.mask. (PR #152140)

[Paul Walker via llvm-commits]

paulwalker-arm wrote:

> > Does this sounds believable?
> 
> We certainly rely on separate intrinsic to set the trip count but, IIRC, SCEV is used to calculate whether the two intrinsics agree on the expected behaviour. If this guarantee isn't meant, then we rely on SelectionDAG lowering, like everyone else, which happens [here](https://github.com/llvm/llvm-project/blob/5499a70c39bfea10a0139ed6e98a267b9473448d/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp#L8057).

I meant when analysing the original loop to determine whether the MVE transformation is possible.  As in, I'm asking whether `getSCEV(ActiveLaneMask.getOperand(1)).isKnownNonZero()` didn't work for the MVE use case so instead the definition of the intrinsic was changed?

https://github.com/llvm/llvm-project/pull/152140