[11u] RFR: JDK-8222074: Enhance auto vectorization for x86

[Andrew Haley]

On 24/06/2020 18:44, Hohensee, Paul wrote:

> We would like to backport aarch64 performance work because
> ARM64-based server are being rapidly adopted for high performance
> floating point applications. JDK 11 will be the primary Java
> execution platform for them for some time to come.

While I have some sympathy for this, please remember that JDK 11 is a
long-term maintenance version. The ratio of reward to risk must be
high, and therefore the benefits of large changes must be substantial.
Many people depend on us not to break anything in production.

AArch64 is, for most people, a new architecture and therefore it's
reasonable (to some extent) for it to be enhanced in JDK 11, if only
for parity with x86. The same cannot be said of x86, where the
ultimate stability is essential. Multi-thousand-line backports for
non-bugs should be very rare indeed. Is the gain really it? note
that, whatever the motivation, this patch must be justified for x86.

Also, although we really need to build confidence in the AArch64
platform, and there's nothing like regressions to destroy trust. Small
performance improvements help a little, but crashes are disastrous. Be
careful out there.

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671