Jv_FindClass throws a ClassNotFoundException in case the requested class cannot be resolved (regardless of whether the bytecode missing or a further dependency cannot be resolved). This causes problems because in case of serious bytecode errors (eg. the class of a field is missing) where a NoClassDefFoundError should reach the user code. I found code in several places in gcj where the caller of Jv_FindClass expected NULL to be returned. This affects Jv_FindClassFromSignature too because it depends on the former function. Obviously such code will not be executed (unless the Jv_FindClass takes a special codepath where it call Jv_FindClassInCache). - link.cc: around line 1000, look for: if (target_class == NULL) throw new java::lang::NoClassDefFoundError (_Jv_NewStringUTF (sym.class_name->chars())); - verify.cc: around 387, look for data.klass = _Jv_FindClassFromSignature (data.name->chars(), loader); if (data.klass == NULL) throw new java::lang::NoClassDefFoundError(data.name->toString()); I want to hear what the experts have to say about that. My suggestion would be to introduce a special variant of Jv_FindClass which does not throw a ClassNotFoundException (by swallowing it) and then carefully adjusting the code location where the old variant was used.
By working on PR 24616 I found out that this belongs together.
The patch for PR 24616 contains variants of the methods I was talking here that do not throw any exception and instead make the caller responsible for throwing the correct exception or error: http://gcc.gnu.org/ml/java-patches/2006-q1/msg00124.html