[vectorIntrinsics] RFR: Load store memory segment

[Maurizio Cimadamore]

On Tue, 29 Mar 2022 21:47:02 GMT, Paul Sandoz <psandoz at openjdk.org> wrote:

> This is the initial implementation for loading/storing vectors from/to memory segments.
> 
> Some of the changes are related to `int` -> `long` for the offset. I had to change the load/store intrinsic signatures to accept an offset as a `long`. This does not affect the C2 implementation, only the fallback lambdas. So in some cases there is a downcast to `int`, which is safe because the original input was an `int` e.g. the offset for primitive array access. 
> 
> Some changes related to module dependency and memory segment proxies will go away when the Memory Segment API is integrated into `java.base` as a preview API.
>  
> There is still more work to do but i wanted to get this in now and iterate:
> - I will remove the `byte[]` and `ByteBuffer` accepting load/store methods in a following PR.
> - some additional negative tests are required, specifically if a scope of segment is not alive, or the segment covers a heap array other than `byte[]`.

Looks good

src/java.base/share/classes/jdk/internal/misc/X-ScopedMemoryAccess.java.template line 674:

> 672:     private static
> 673:     <V extends VectorSupport.Vector<E>, E>
> 674:     void storeIntoMemorySegmentScoped(ScopedMemoryAccess.Scope scope,

I believe the rest of the class uses "internal" as a suffix for the scoped method

src/jdk.incubator.vector/share/classes/jdk/incubator/vector/Vector.java line 3527:

> 3525:      * {@link #intoMemorySegment(MemorySegment,long,ByteOrder,VectorMask)
> 3526:      * intoMemorySegment()} as follows:
> 3527:      * <pre>{@code

you might want to revisit later to use code snippets

src/jdk.incubator.vector/share/classes/jdk/incubator/vector/Vector.java line 3546:

> 3544:      *         or if access occurs from a thread other than the thread owning the session.
> 3545:      */
> 3546:     public abstract void intoMemorySegment(MemorySegment ms, long offset, ByteOrder bo);

Just a note - in the memory segment API we used to have API points taking an explicit byte order (e.g. dereference methods) and it turned out these were just specific instances of a more general API method accepting a full layout. I sense the same might apply here - e.g. if you take a segment, an offset and a layout, then you might be able to generalize the API to more than byte[]-backed memory segment. But for now it's fine. Or - maybe, in this case the vector instance should have a lane layout somewhere, which can be used here.

test/jdk/jdk/incubator/vector/Byte128VectorLoadStoreTests.java line 970:

> 968: 
> 969:         try {
> 970:             SPECIES.zero().intoMemorySegment(a, 0, bo);

I think these can be written more succinctly with `expectThrows(UnsupportedOperationException.class, () -> { ... })`

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

Marked as reviewed by mcimadamore (Committer).

PR: https://git.openjdk.java.net/panama-vector/pull/187