[LUCENE-9117] RamUsageEstimator hangs with AOT compilation

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 9.0
• Lucene Fields: New
• Assignee: Dawid Weiss
• Reporter: Dawid Weiss
• Created: 07/Jan/20 07:56
• Updated: 28/Aug/22 15:55
• Resolved: 15/Jan/20 08:54

Description

Mailing list report by Cleber Muramoto.

After generating a pre-compiled image lucene-core (8.3.0)  with jaotc (JDK
13.0.1), RamUsageEstimator class is never loaded - it fails to complete the
static initializer.

Steps to reproduce:

 1)Generate the image with

>jaotc --info --ignore-errors --jar ~/.m2/repository/org/apache/lucene/lucene-core/8.3.0/lucene-core-8.3.0.jar --output lucene-core.so

2)Create a simple test class to trigger class loading

 import java.io.IOException;

import java.nio.file.Path;
import java.nio.file.Paths;

import org.apache.lucene.index.IndexWriter;
import org.apache.lucene.index.IndexWriterConfig;
import org.apache.lucene.store.MMapDirectory;

public class TestAOT {

public static void main(String...args){
 run();
 System.out.println("Done");
 }
static void run(){
 try {
 var iw = new IndexWriter(MMapDirectory.open(Paths.get("dir")), new
IndexWriterConfig());
 iw.close();
 } catch (IOException e) {
 e.printStackTrace();
 }
 }
}

3)Run the class with the pre-compiled image

>java -XX:AOTLibrary=./lucene-core.so -XX:+PrintAOT  -cp ~/.m2/repository/org/apache/lucene/lucene-core/8.3.0/lucene-core-8.3.0.jar TestAOT.java

4)The program never completes. After inspecting it with jstack <pid> it
shows that it's stuck in line 195 of RamUsageEstimator

"main" #1 prio=5 os_prio=0 cpu=174827,04ms elapsed=173,91s
tid=0x00007fe984018800 nid=0x2daf runnable [0x00007fe98bc3c000]
  java.lang.Thread.State: RUNNABLE
    at
org.apache.lucene.util.RamUsageEstimator.<clinit>(RamUsageEstimator.java:195)

    at org.apache.lucene.util.ArrayUtil.<clinit>(ArrayUtil.java:31)
    at org.apache.lucene.index.IndexWriter.<clinit>(IndexWriter.java:276)
    at TestAOT.run(TestAOT.java:20)
    at TestAOT.main(TestAOT.java:14)
    at
jdk.internal.reflect.NativeMethodAccessorImpl.invoke0(java.base@13.0.1/Native
Method)
    at jdk.internal.reflect.NativeMethodAccessorImpl.invoke(java.base@13.0.1/NativeMethodAccessorImpl.java:62)

    at
jdk.internal.reflect.DelegatingMethodAccessorImpl.invoke(java.base@13.0.1/DelegatingMethodAccessorImpl.java:43)

    at java.lang.reflect.Method.invoke(java.base@13.0.1/Method.java:567)
    at com.sun.tools.javac.launcher.Main.execute(jdk.compiler@13.0.1/Main.java:415)

    at com.sun.tools.javac.launcher.Main.run(jdk.compiler@13.0.1/Main.java:192)

    at com.sun.tools.javac.launcher.Main.main(jdk.compiler@13.0.1
/Main.java:132)

My guess is that the AOT compiler aggressive optimizations like
inlining/scalar replacing calls to Long.valueOf, are working against the
code's desired semantics.

Perhaps the logic to determine LONG_CACHE_[MIN/MAX]_VALUE could be
conditioned to VM version/vendor in Constants class. From [Open|Oracle]JDK
7 onwards (don't know about older versions) Long cache has a fixed range of
[-128,127], so there's no need to loop until a non-cached boxed value shows
up.

I know this compiler bug isn't Lucene's fault and as a workaround one could
use --compile-for-tiered or exclude RamUsageEstimator's methods from jaotc,
however, it would be nice to have a faster, allocation/loop-free
initializer for RamUsageEstimator nevertheless.