TypeBinding hashCode may be bad #3412
Comments
There where 2 commits that introduced similar workarounds: PTBKey (https://bugs.eclipse.org/bugs/show_bug.cgi?id=521438, ReferenceBindingSetWrapper (https://bugs.eclipse.org/bugs/show_bug.cgi?id=544921) for noticed performance problems. A clean solution would be to not require UnresolvedReferenceBindings have same hashCode as ReferenceBinding or to copy the hashCode from the unresolved to the resolved one. |
|
property is used in org.eclipse.jdt.internal.compiler.lookup.TypeSystem.updateCaches(UnresolvedReferenceBinding, ReferenceBinding) to update keys in a org.eclipse.jdt.internal.compiler.util.SimpleLookupTable TypeSystem.annotationTypes |
|
Maybe while at it some hashcode implementions could be updated in case they are not updated already (did not check) |
This is not the full story. Class TypeSystem was introduced in 2013 (Java 8). The design decision behind hashCode for ReferenceBinding dates from 2004 (Java 5). In some ways TypeSystem "inherited" the responsibilities of former fields like SimpleLookupTable uniqueParameterizedTypeBindings (in LookupEnvironment), but 20 years after the design was made, there may be many more locations relying on its guarantees. To get a full picture, all hash collections (incl. SimpleLookupTable) over TypeBinding must be considered. In all of them an initial UnresolvedReferenceBinding must occupy the same slot that eventually its resolved version will assume. I suggest that any changes in this area must at the very least be thoroughly reviewed by @srikanth-sankaran , ideally also by Philipe Mulet :) |
That would be a separate issue, because the quality/reviewing requirements for the compiler differ from other code areas. |
|
While it's great to try to understand the true intention behind this admittedly irregular design, I'm not sure there is an actual flaw to be corrected. |
If they don't have the same hashcode then "in place" resolution won't work. An UnresolvedReferenceBinding gets resolved automagically without perturbing the container semantics today. |
I vaguely recall some super complex scenario defects 15 years ago involving UnresolvedReferenceBinding where we would end up with two or more ReferenceBindings after resolution for the same URB - a design violation triggered by something to do with hashcode 😊 Here be dragons! |
|
Currently compiler.codegen.ObjectCache as used in ExceptionHandlingFlowContext constructor (and only there) uses ReferenceBinding.hashCode() but not equals(), it uses |
SuperTypeTest.test016(), SuperTypeTest.test017() failed when implementation of ReferenceBinding.hashCode() changed. The message was either "already defined by IChangeRulerColumn" or "already defined by IRevisionRulerColumn" depending on which Interface had the lower hashCode. relates to eclipse-jdt#3412
SuperTypeTest.test016(), SuperTypeTest.test017() failed when implementation of ReferenceBinding.hashCode() changed. The message was either "already defined by IChangeRulerColumn" or "already defined by IRevisionRulerColumn" depending on which Interface had the lower hashCode. Now using a LinkedHashMap to rely on the insertion order rather then the hashcode. relates to eclipse-jdt#3412
* avoids raw types * deletes custom ObjectCache ObjectCache implemented a HashMap using hashCode() but not equals(). Instead it used == identity. JDK offers a special Map for identity since 1.4, consistently using System.identityHashCode(). relates to eclipse-jdt#3412
* avoids raw types * deletes custom ObjectCache ObjectCache implemented a HashMap using hashCode() but not equals(). Instead it used == identity. JDK offers a special Map for identity since 1.4, consistently using System.identityHashCode(). relates to eclipse-jdt#3412 Also solved follow up * LambdaExpression "Redundant specification of type arguments"
Environment is only used during put relates to eclipse-jdt#3412
SuperTypeTest.test016(), SuperTypeTest.test017() failed when implementation of ReferenceBinding.hashCode() changed. The message was either "already defined by IChangeRulerColumn" or "already defined by IRevisionRulerColumn" depending on which Interface had the lower hashCode. Now using a LinkedHashMap to rely on the insertion order rather then the hashcode. relates to #3412
* avoids raw types * deletes custom ObjectCache ObjectCache implemented a HashMap using hashCode() but not equals(). Instead it used == identity. JDK offers a special Map for identity since 1.4, consistently using System.identityHashCode(). relates to #3412 Also solved follow up * LambdaExpression "Redundant specification of type arguments"
Especially UnresolvedReferenceBinding.hashCode()
See discussion around #3387 (comment)
It is not obvious if the method violates equals contract or is a good hashCode(), given that ReferenceBindingSetWrapper tries to work around it.
I will investigate.
The text was updated successfully, but these errors were encountered: