[llvm] [AArch64] Fix clmul.nxv1i128 lowering crash (PR #201835)

[Matthew Devereau via llvm-commits]

MDevereau wrote:

> This feels like an ABI implementation that we've not committed to yet. There's work happening in the background for how best to handle `<vscale x 1 x ?>` in general, and I would not be surprised if we decide to make `<vscale x 1 x i128>` type legal one day. That day is not today though.
> 
> Normally I'd suggest writing tests that bypass unsupported ABI mechanisms (e.g. by using memory) but as you'll see from https://godbolt.org/z/fPYPveYTW we have bigger gaps in `<vscale x 1 x i128>` support than just clmul. I recommend ensuring the cost model returns an invalid cost for this operation and leaving it at that.

Thanks, I'll abandon this and have a look at the cost model.

https://github.com/llvm/llvm-project/pull/201835