Bug ID: JDK-8302139 Speed up SuperWord reduction tests

JDK-8302139 : Speed up SuperWord reduction tests

Type: Enhancement
Component: hotspot
Sub-Component: compiler
Affected Version: 21

Priority: P4
Status: Closed
Resolution: Won't Fix

Submitted: 2023-02-09
Updated: 2025-04-24
Resolved: 2025-04-24

Related Reports

Blocks :	JDK-8302652 - [SuperWord] Reduction should happen after loop, when possible
Cloners :	JDK-8294715 - Add IR checks to the reduction vectorization tests
Cloners :	JDK-8302093 - Speed up compiler/loopopts/superword/Vec_MulAddS2I.java
Duplicate :	JDK-8302093 - Speed up compiler/loopopts/superword/Vec_MulAddS2I.java
Relates :	JDK-8295561 - Review running times of hotspot compiler tier1 tests
Relates :	JDK-8300865 - C2: product reduction in ProdRed_Double is not vectorized
Relates :	JDK-8302668 - [TESTBUG] Tests require feature sse4_1 which does not exist, should be sse4.1
Relates :	JDK-8302673 - [SuperWord] MaxReduction and MinReduction should vectorize for int
Relates :	JDK-8302665 - [SuperWord] Allow MulReductionVL not just for avx512dq
Relates :	JDK-8302662 - [SuperWord] Vectorize loop when value from last iteration is used after loop

Description

https://github.com/openjdk/jdk/blame/master/test/hotspot/jtreg/compiler/loopopts/superword/SumRed_Double.java
https://github.com/openjdk/jdk/blame/master/test/hotspot/jtreg/compiler/loopopts/superword/SumRedSqrt_Double.java

These are two of the slowest compiler tests in tier1, take about 30 and 25sec.

Restrict compilation, see if compile threashold is ok. Consider merging the two tests.

Similar issue in JDK-8302093. We should just speed up all of the reduction tests of superword.
Currently, [~dskantz] is working on JDK-8294715. Afterwards we can speed these tests up. Maybe we have to unify all into a single test-file.

Comments

Closing all tasks created for the test speedups under JDK-8295561. If this becomes a priority again, we should probably measure test performance of all tests, and prioritize by which tests are taking the most time now, and not based on data from 2023.

24-04-2025

Emanuel's work is nice. I have made some slight changes to it (add Vec_MulAddS2I, provisional support for platform features, some refactoring). I am not sure if we want to wait for JDK-8302673 and JDK-8280120 before proceeding. JDK-8280120 makes it possible to add all tests into one file by disabling some IR checks on 32-bit platforms.

17-04-2023

I had a patch. I'll just leave it here for later. https://github.com/openjdk/jdk/pull/12590

23-03-2023