Integrated: 8323641: Test compiler/loopopts/superword/TestAlignVectorFuzzer.java timed out
Emanuel Peter
epeter at openjdk.org
Mon Jan 15 10:47:30 UTC 2024
On Fri, 12 Jan 2024 08:22:54 GMT, Emanuel Peter <epeter at openjdk.org> wrote:
> It seems that allowing `90%` of the timeout-time was cutting it too close. Some individual tests can take more time occasionally, one even took more than `80 sec` (very rare).
>
> Why do these tests take so long?
> - Sometimes compilation can take quite a bit of time. `-XX:LoopUnrollLimit=250` already increases the number of nodes allowed in a loop-body, and with unrolling this increases significantly. SuperWord then has to work through all these nodes, and has a lot of quadratic and higher complexity loops.
> - The loop bodies are quite large (hand unrolled), and often lead to partial vectorization, with lots of scalar memory ops. This produces quite sub-optimal code, with a bit too many instructions in the loop body. Combined with lots of array copying, this probably takes quite a hit on caches.
>
> I will investigate SuperWord compilation time in the future, and lower the runtime complexity if neccessary/possible, it is part of my autovectorization plans.
>
> I now lowered the allowance down to `40%`, which is hopefully small enough to avoid timeout, while still allowing sufficient many run to get decent test coverage.
This pull request has now been integrated.
Changeset: cd0fe377
Author: Emanuel Peter <epeter at openjdk.org>
URL: https://git.openjdk.org/jdk/commit/cd0fe377417be65dbf1338d8b47da8817985c7d8
Stats: 18 lines in 1 file changed: 2 ins; 0 del; 16 mod
8323641: Test compiler/loopopts/superword/TestAlignVectorFuzzer.java timed out
Reviewed-by: chagedorn, kvn
-------------
PR: https://git.openjdk.org/jdk/pull/17389
More information about the hotspot-compiler-dev
mailing list