There is no VP intrinsic version of llvm.stepvector?

[wangpc-pp]

I think my memory has been messed up and I was thinking there should be an equal VP intrinsic for llvm.stepvector, but it seems not.

Remind me if I missed something. Why is there no such intrinsic? And do we need it?

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Pengcheng Wang (wangpc-pp)

I think my memory has been messed up and I was thinking there should be an equal VP intrinsic for [`llvm.stepvector`](https://llvm.org/docs/LangRef.html#llvm-stepvector-intrinsic), but it seems not.

Remind me if I missed something. Why is there no such intrinsic? And do we need it?

[lukel97]

It only recently got promoted out of experimental IIRC, maybe that's why we don't already have a VP intrinsic

I don't see any reason why we shouldn't have one anyway, did you have a specific use case in mind for it?

[topperc]

We haven't found a use for it yet in our downstream EVL vectorizer. I think I've only seen stepvector in the loop preheader where is no EVL yet. On some microarchitectures it might be profitable to use a vp.stepvector in the preheader if the total number of elements to process is less than VLMAX. Meaning the vector loop is a single iteration. If the microarcitecture dynamically reduces LMUL based on VL.