/cc @jsotuyod re #817 (comment)

I've tried really hard (6 hours) to understand the issue and produce a minimal example, can't do better than this. Sorry for the long-ass description it may be TMI in some places.

Affects PMD Version: 6.0.0

Rule: UnusedPrivateMethod

Description: The attached project gives two false positive violations for the two used methods. After hours of investigation (went pretty deep into the rabbit hole sun.misc.URLClassPath) I figured out that the .class files were added to the classpath in the wrong way... I wasn't expecting the classpath to influence the results, because I thought PMD was a source analyzer, but I guess since the type resolution was introduced, the actual classes are required.

Anyway, net.sourceforge.pmd.lang.java.symboltable.ClassScope#findVariableHere@hasAuxclasspath is too much an important variable. It decides whether the method resolution is by name+argcount or by full type. I think there's a better way: as far as I see now the boolean hasAuxclasspath argument to the matchMethodDeclaration can be removed and replaced with a calculated value. parameterTypes and argumentTypes both contain SimpleTypedNameDeclaration objects, which, in case of missing .classes have null values in type field. My proposal is to define

hasAuxclasspath = all index in params/args (when count is the same) {
      (args[i].type != null && params[i].type != null) // best scenario, full type information available
     || ((args[i].type != null || params[i].type != null) && args[i].image == params[i].image) // partial info
     // if both null, or the image is not exactly the same,
     // don't try to guess, just accept the failure and match by name and arg number
}

Oh, and back to how this fixes UnusedPrivateMethod: it simply will fill in the scope based on the best available information, and hence the missing .class files will trigger a more lenient resolution thus preventing false positives.

Code Sample demonstrating the issue:
(see attached .zip file)

 void load(EngineKey key) {
    logWithTimeAndKey("", key);
    loadFromActiveResources(key, true);
  }

  private static void logWithTimeAndKey(String log, Key key) {
  }

  private void loadFromActiveResources(Key key, boolean isMemoryCacheable) {
  }

Running PMD through: [CLI | Ant | Maven | Gradle | Designer | Other] Gradle

Debugging
Here's how I debugged the issue:

1. Extract attached project

2. Import Gradle project into AS/IntellJ IDEA
   If Sync fails with

   Error:The modules ['buildSrc', 'pmd-classpath'] point to the same directory in the file system. Each module must have a unique path.

   you can fix it by going to View > Tool Windows > Gradle and using the refresh button there.
   It doesn't matter once it's synced, even if it says needs to resync.

3. Go to Run/Debug Configurations, and + a Remote run config

4. Set [x] Socket [x] Attach Port: 5005, module: root

5. From Terminal (in IDEA or external), run

   gradlew --stop && gradlew --no-daemon --debug -Dorg.gradle.debug=true cleanPmd pmd`

6. notice it stops at "> Starting Daemon"
   Observe: "daemon out: Listening for transport dt_socket at address: 5005"

7. Jump to class net.sourceforge.pmd.lang.java.symboltable.ClassScope

8. Put a breakpoint at line 244 (for > if) in matchMethodDeclaration

9. Debug the just-created run configuration,
   (wait a few seconds while Gradle build progresses to :pmd task)

• UnusedPrivateMethodRule.visit: occs.isEmpty() == true -> violation
  Why is it empty?
• The ClassScope's getMethodDeclarations has the method, but there are no NameOccurrences for it
• AbstractScope.nameDeclarations doesn't have values for MethodNameDeclaration
• ClassScope.addNameOccurrence is not executed for Engine.loadFromCache
• That method is called from Search which is used by OccurrenceFinder
  Why isn't Search traversing the load() method?
  OccurrenceFinder stops on "loadFromCache".equals(occ.getImage()), and that is in the "load" method: occ.location.getFirstParentOfType(ASTMethodDeclaration.class).findChildrenOfType(ASTMethodDeclarator.aclass).get(0)
• Search traverses the scopes via searchUpward, I expect it to stop and match on ClassScope.
• ClassScope.contains -> findVariableHere(image="loadFromCache", isMethodOrConstructorInvocation=true)
• getMethodDeclarations() contains the image
• ClassScope.matchMethodDeclaration(): there must be a mismatch here
  based on hacking around empirically EngineKey implement Key will mismatch somehow.
• name matches, count matches, parameterTypes.equals(argumentTypes) false because one is EngineKey and the other is com.bumptech.glide.load.Key; neither SimpleTypedNameDeclaration has .type nor .next.
• getEnclosingScope(SourceFileScope.class).resolveType -> TypeSet.findClass doesn't find them.
• pmdClassLoader.loadClass("com.bumptech.glide.load.Key") can't find it even though the ClassLoader.path has the file:/.../Key.class inside
• URLClassPath.path has all the class files, but URLClassPath.loaders only has 16 loaders, all them are JarLoader.
• sun.misc.URLClassPath#getLoader(java.net.URL) lead me to solution:

  task pmd {
  -    classpath += fileTree("${project.buildDir}/intermediates/classes/debug/")
  +    classpath += files("${project.buildDir}/intermediates/classes/debug/")
  }