Problem: When classes are imported in Clojure, the class is loaded using Class.forName(), which causes its static initialisers to be executed. This differs from Java where compilation does not cause classes to be loaded.
Motivation: In many cases when those classes are normally loaded by Java code during execution of a framework of some kind (IntelliJ in my case, and RoboVM is another culprit mentioned in that thread) the initialiser expects some infrastructure to be in place and will fail when it's not. This means that it's impossible to AOT compile namespaces importing these classes, which is a fairly serious limitation.
Approach: Modify ImportExpr to call RT.classForNameNonLoading() instead of Class.forName(), which will load the class but not initialise it. This change causes the Clojure test suite to fail, since clojure.test-clojure.genclass imports a gen-class'ed class which no longer loads its namespace on initialisation. I'm not sure if this is considered an incorrect use of such a class (IIRC with records it's required to import the class and require its namespace), but given that it's in the Clojure test case it's reasonable to assume that this fix would be a breaking change for more code out there. This test failure is also corrected in the attached patch.
Alternative: This patch enables the change unconditionally, but depending on the extent of breakage it causes, it might need to be enabled with a configuration flag. I propose we make it unconditional in an early 1.7 beta and monitor the fall-out.
Background: This issue has been discussed in the following threads