RFR: 8264762: ByteBuffer.byteOrder(BIG_ENDIAN).asXBuffer.put(Xarray) and ByteBuffer.byteOrder(nativeOrder()).asXBuffer.put(Xarray) are slow

[Alan Bateman]

On Fri, 23 Apr 2021 19:23:13 GMT, Brian Burkhalter <bpb at openjdk.org> wrote:

> Please consider this request to accelerate absolute and relative bulk array transfer on views of heap byte buffers where the element size is greater than one. What currently happens is that the transfer devolves to a “loopy” element-by-element copy such as
> 
>         int end = offset + length;
>         for (int i = offset, j = index; i < end; i++, j++)
>             dst[i] = get(j);
> 
> for `get()`, and
> 
>         int end = offset + length;
>         for (int i = offset, j = index; i < end; i++, j++)
>             this.put(j, src[i]);
> 
> for `put()`. This is of course relatively slow.
> 
> The change proposed hoists the accelerated versions of these methods using the `ScopedMemoryAccess` methods `copyMemory()` and `copySwapMemory()` from `Direct-X-Buffer` to `X-Buffer`. The array bulk transfer methods are removed from `Direct-X-Buffer` itself. The number of lines of code in the templates decreases by 87.
> 
> With this change the throughput of array bulk `put()` and `get()` for heap view buffers is increased by a factor of 6 to 11 compared with the current code. The performance of direct view buffers does not appear to be affected.
> 
> No tests are modified or added as existing tests already cover these methods. All tests in CI tiers 1-3 passed on all platforms.

The hoisting looks right but I think I need to checkout the branch and look at the generated code for the direct buffer and the one or two views to make sure.

In passing, I see we have an inconsistency with the ordering of the index and whether is read-only buffer, we might want to adjust that at some point so that have consistency behavior when both conditions arise.

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

PR: https://git.openjdk.java.net/jdk/pull/3660