Currently, when an exception is thrown from native code called from Java, such as:

const DWORD EX_CODE = 42;

__declspec(dllexport) void throw_C_exception() {
    RaiseException(EX_CODE, EXCEPTION_NONCONTINUABLE, 0, NULL);
}

The stack is unwound until we get back to Java frames, at which point topLevelExceptionFilter in os_windows.cpp is invoked (which acts as the exception handler for the entire code cache).

Since we have no explicit handling for this exception code, we eventually end up in Handle_FLT_Exception which checks if MxCsr is modified, restores it, and then returns EXCEPTION_CONTINUE_EXECUTION. However, this is done without making sure that the exception code corresponds to an FLT* exception.

As a result, we end up entering topLevelExceptionFilter over and over, until the process finally abruptly terminates due to a stack overflow error.

We should more carefully try to filter for FLT* exception codes in Handle_FLT_Exception, so that unrecognized exceptions are handled through the default VM error reporting mechanism.