[vectorIntrinsics+fp16] RFR: Merge vectorIntrinsics

[Jatin Bhateja]

On Wed, 20 Aug 2025 15:57:18 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

> Merge vectorIntrinsics into vectorIntrinsics+fp16 as per the proposal on panama-dev mailing list[1]. This was a major merge after a gap of 2+ years.
> 
> Summary of changes:-
>    - Added a new package-private field CTYPE (carrier type) along with existing ETYPE (element type) to differentiate b/w the storage type from the vector lane element type.
>    - Pass an explicit operType argument to inline expanders to ease inferencing of Float16 vector IR. Idea: leverage existing backend implementation added along with JDK-8346236[2] and [3]
>    - Appropriate modifications in existing inline expansion entry points and VectorSupport class.
>    - Replaced existing Halffloat class references with jdk.incubator.vector.Float16, along with appropriate API adjustments.
>   
> What is remaining:-
>   -    Use the newly passed operType to infer Float16 Vector IR.
>   -    Functional validation
>   - Performance validation
>  
> [1] https://mail.openjdk.org/pipermail/panama-dev/2025-August/021100.html
> [2] https://github.com/openjdk/jdk/pull/22755
> [3] https://github.com/openjdk/jdk/pull/22754

With the latest bug-fix commit, all VectorAPI jtreg tests, including new HalffloatVectorTest are passing on AVX512 and AVX2 targets. HalffloatVectorTests are currently running at compilation tier=3, we will lift this limitation along with JDK-https://github.com/openjdk/panama-vector/pull/231.

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

PR Comment: https://git.openjdk.org/panama-vector/pull/230#issuecomment-3231171981