Bugzilla – Full Text Bug Listing
| Summary: | [compiler][1.5] Generics - JDT cannot compile hudson-core when javac can | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Product: | [Eclipse Project] JDT | Reporter: | Sébastien Vauclair <sebastien.vauclair> | ||||||||
| Component: | Core | Assignee: | Philipe Mulet <philippe_mulet> | ||||||||
| Status: | VERIFIED FIXED | QA Contact: | |||||||||
| Severity: | normal | ||||||||||
| Priority: | P3 | CC: | eric_jodet, kent_johnson, Olivier_Thomann | ||||||||
| Version: | 3.4 | Flags: | maxime_daniel:
review+
Olivier_Thomann: review+ |
||||||||
| Target Milestone: | 3.4 RC2 | ||||||||||
| Hardware: | PC | ||||||||||
| OS: | Windows XP | ||||||||||
| Whiteboard: | |||||||||||
| Attachments: |
|
||||||||||
|
Description
Sébastien Vauclair
Reproduced. I would agree this is our bug. Tentatively tagging for 3.4RC2, since Hudson libs are quite popular. Thanks for looking after this one! Note that Hudson seems to have been fixed in the meantime, according to: https://hudson.dev.java.net/servlets/BrowseList?listName=dev&from=1082374&to=1082374&count=7&by=thread&paged=false Problem does not come from inference, but from #getClass() invocation. Reduced testcase:
public class X extends AbstractProject<X, Object> {
public void testBug() {
filter(getClass());
}
public static void filter(Class<? extends AbstractProject<?, ?>> type) {
}
}
abstract class AbstractProject<P extends AbstractProject<P, R>, R> {
}
Re: comment 4 Thanks for the pointer, indeed this is really strange. An implicit message send fails, where explicit ones do succeed... e.g. public class X extends AbstractProject<X, Object> { public void testBug() { filter(getClass());//1 filter(this.getClass());//2 filter(X.class);//3 } public static void filter(Class<? extends AbstractProject<?, ?>> type) { } } abstract class AbstractProject<P extends AbstractProject<P, R>, R> { } only (1) does fail Problem found. Indeed implicit msg send to #getClass() did not get properly substituted for its return type: Class<?> into Class<? extends |receiverType|>. Added GenericTypeTest#test1333-1334. Created attachment 100718 [details]
Proposed patch
Reviewers - there was a code path (through Scope#getImplicitMethod(...) which did not process #getClass(). The processing code was already written elsewhere (see Scope#getMethod(...) line 2118 for instance), but not invoked.
+1 for 3.4RC2. Now the code is identical to the getMethod(...) path. Created attachment 100749 [details]
Patch for 3.3.x
Released in 3.3.x maintenance stream. Given OTT is off today, asking Maxime for review. Per the same token as Olivier (symmetry of code paths), I +1 the proposed change for 3.4 RC2. Given the fact that this code has to be invoked very often, I'd suggest that in 3.5 we consider reordering the terms of the boolean expression though, in both places, to make the most of the shortcut. I've not established full-blown statistics for this, but I would expect the two first terms to be true very often, the third to be false most of the time, and the fourth to be expansive to compute (which is a good reason to keep it where it is). Which probes for a comparison of the performances of the current ordering with (at least) one ordering that would bring the third term (or even a faster quick check of part of it) in first position. Adding Kent to the discussion for this specific point. Re: comment 14 The first 2 tests are identity tests, where the 3rd one is a char[] comparison. So slower, also message sends to as receiver type Object are not super frequent. if (receiverType.id != T_JavaLangObject && argumentTypes == Binding.NO_PARAMETERS && CharOperation.equals(selector, GETCLASS) && methodBinding.returnType.isParameterizedType()/*1.5*/) { Forgot to mark it as fixed. verified for 3.4RC2 using build I20080523-0100 |