RFR 8078563 - add profitability tests for reductions

Berg, Michael C michael.c.berg at intel.com
Mon Apr 27 19:57:24 UTC 2015


No, there cannot be such a mapping, that would be scalar code with no vector solution where superword would fail to find a packset mapping.


From: Christian Thalinger [mailto:christian.thalinger at oracle.com]
Sent: Monday, April 27, 2015 9:49 AM
To: Berg, Michael C
Cc: hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>
Subject: Re: RFR 8078563 - add profitability tests for reductions

+       // Length 2 reductions of INT/LONG do not offer performance benefits

+       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {

I don’t know that code very well but can there be reductions with size == 1?

On Apr 23, 2015, at 5:53 PM, Berg, Michael C <michael.c.berg at intel.com<mailto:michael.c.berg at intel.com>> wrote:

Hi Folks,

We (Intel) would like to add profitability tests to superword to gate scenarios where reduction optimization overhead is roughly equal to the benefit gained by vectorization.
We would like to do this for all x86 enabled microarchitectures that support reductions and superword.  This new constraint was tested on SSE and AVX (1,2) enabled platforms.
The contribution as referenced by RFR 8078563 is defined by the information at the links below.

Please review this bug entry and its code and comment as needed:


And its code and test addition (this is a small patch):


Vladimir Kozlov has offered to sponsor this patch.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20150427/6f27de77/attachment.html>

More information about the hotspot-compiler-dev mailing list