[vector api] RFR: Implement Vector API andAll/orAll/xorAll for AArch64 NEON

[John Rose]

Sandhya and the other Intel engineers are currently working on
the problem of AD file size.  They report they can reduce the
number of instructions by having one instruction support
many match rules, in the common case where one instruction
format performs a range of similar operations.  This does
not immediately reduce the size of the AD file, because
the adlc preprocessor internally duplicates such instructions,
once per match rule.  But some additional adlc work should
be able to refactor the generated code to be less repetition,
and therefore smaller.  I hope this work comes to the light
of day soon, so ARM can make use of it also.

On another note, you've probably noticed the rapid set
of pushes to the vector branch of panama.  We are still
hammering on the API design, of course, but it seems to
be cooling down a bit.  Now would be a great time for
vector experts on ARM to look at the javadoc and see if
there's anything in there that (accidentally) has semantics
that is too x86-specific, or is otherwise non-portable.

Here's a recent javadoc from Kishor Kharbas:

http://cr.openjdk.java.net/~kkharbas/vector-api/javadoc/VectorApi-JavaDoc.05/jdk.incubator.vector/jdk/incubator/vector/package-summary.html

http://cr.openjdk.java.net/~kkharbas/vector-api/javadoc/

It's good reading!  After we all work it over it will be even better,
of course.  There will also be a formal CSR review in the not-too-distant
future, since that is part of the JEP process.

Best wishes,

— John

On Jan 28, 2019, at 2:13 AM, Andrew Haley <aph at redhat.com> wrote:
> 
> On 1/28/19 2:44 AM, Yang Zhang (Arm Technology China) wrote:
>> I have a patch which implements Vector API andAll/orAll/xorAll for AArch64 NEON. Could you please help to review it?
>> http://cr.openjdk.java.net/~yzhang/vectorapi.andall/webrev.00/
> 
> OK, but:
> 
> You're doing a great deal of repetitive stuff in here. This makes the
> patch hard to maintain and review. Instead, I suggest that you have a
> look at aarch64_ad.m4 to see how it's possible to generate all of
> these patterns from a script: 487 lines of that script generates 1970
> lines of patterns.
> 
> -- 
> Andrew Haley
> Java Platform Lead Engineer
> Red Hat UK Ltd. <https://www.redhat.com>
> EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671