RFR(S): 8200477: Integer dot product no longer autovectorised

Andrew Haley aph at redhat.com
Mon Apr 23 15:44:53 UTC 2018

On 04/10/2018 03:54 PM, Roland Westrelin wrote:
> http://cr.openjdk.java.net/~roland/8200477/webrev.00/
> Superword only allows reduction to have Phi uses but with loop strip
> mining, there's a safepoint use as well. In any case, the check for only
> Phi uses seems too strict to me. So I replaced it with a check for only
> uses out of loop.

I've tried to get this to work, both on AArch64 and x86, but without
any success.  No matter what I do I do not get any vector insns.
Please tell me what you do to run this JMH test and dee any vector
instructions.  An example of what you see would be nice.  Does it
need AVX, or some special level of SSE?

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

More information about the hotspot-compiler-dev mailing list