[vectorIntrinsics] RFR: RFC: Vector API masking support proposal for Arm SVE [v3]

[Jatin Bhateja]

On Mon, 15 Mar 2021 02:55:45 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:

>> Thanks for your comment about this. `VectorLoadMask` and `VectorStoreMask` are vector nodes that have the `TypeVect` type, while `Vector2Mask` is a mask node that have the `TypeVMask` type.  For SVE, the two new added nodes are used for conversions between vector register and predicate register, e.g. `Mask2Vector` for `predicate -> vector` and `Vector2Mask` for `vector -> predicate`. Yes, it seems `VectorLoadMask` and `VectorStoreMask` can match the same requirements like what have been done in the AVX-512 patch. However, in our whole proposal, we defined different IRs to separate vector and mask operations. I think the main conflict is the IR part for mask. And it's better to get a final solution to the different operations together for SVE and AVX-512.
>
> Note that `MaskToVector` can be used for API `VectorMask.toVector()` in future which is implemented with `blend` currently.

> Thanks for your comment about this. `VectorLoadMask` and `VectorStoreMask` are vector nodes that have the `TypeVect` type, while `Vector2Mask` is a mask node that have the `TypeVMask` type. For SVE, the two new added nodes are used for conversions between vector register and predicate register, e.g. `Mask2Vector` for `predicate -> vector` and `Vector2Mask` for `vector -> predicate`. Yes, it seems `VectorLoadMask` and `VectorStoreMask` can match the same requirements like what have been done in the AVX-512 patch. However, in our whole proposal, we defined different IRs to separate vector and mask operations. I think the main conflict is the IR part for mask. And it's better to get a final solution to the different operations together for SVE and AVX-512.

Hi @XiaohongGong , 
Yes currently all the ideal nodes have fixed types , but technically there is nothing incorrect in having nodes (VectorLoadMask/VectorStoreMask) with types based on the target features, e.g. VectorLoadMask may have TypeVect for targets which do-not support predicate registers and a predicate vector type (TypeVectMask/TypeVMask) otherwise. 
By taking this approach we may be able to save this synthetic duplication for masked flavors of vector operation nodes.

-------------

PR: https://git.openjdk.java.net/panama-vector/pull/40