Hi Mike,
The problem with changing initialization order is that you can never
enumerate all the possible initialization paths. I don't see anything
problematic with the move in relation to Thread.currentThread() or
ThreadGroup.add(). But the call to setJavaLangAccess requires
initialization of SharedSecrets which in turn requires initialization of
Unsafe which in turn initializes sun.reflect.Reflection which in turn
initializes a bunch of Collection classes - and suddenly we may have a
whole swag of classes now being initialized before the VM appears to be
booted.
If you then throw into the mix the possibility of different system
properties affecting initialization actions, or the existence of an
agent, then who knows what might actually happen.
Cheers,
David
-----
On 28/02/2012 1:57 PM, Mike Duigou wrote:
Hello all;
This issue is a patch for review that I have split out of a larger issue I'll
be submitting later.
WEBREV @ http://cr.openjdk.java.net/~mduigou/7149320/0/webrev/
sun.misc.VM.booted() is currently called before setJavaLangAccess(). For code
which uses the JavaLangAccess shared secrets it's convenient to be able to
check whether the secrets are initialized without having to call
sun.misc.SharedSecrets.getJavaLangAccess() every time the secrets are used and
checking for null. In particular with the static inner class holder idiom it
would be desirable to do :
static class Holder {
final sun.misc.JavaLangAccess ACCESS =
sun.misc.SharedSecrets.getJavaLangAccess();
}
...
if(sun.misc.VM.isBooted()&& Holder.ACCESS...
In my research on this issue I was unable to determine why sun.misc.VM.booted()
isn't the currently the last activity in System.initSystemClass(). Neither of
the two tasks which currently follow it depend upon booted state in any way
that I could determine. I am tempted, thinking about it, to add a comment to
System.initSystemClass before the call to sun.misc.VM.booted() asking future
maintainers to leave boot() as the last activity or explain why not.
I have done JTReg runs on linux x86, linux x64 and Solaris 64 which
incorporated this change without any problems encountered. It's rather
difficult to write a unit test for this issue as it requires a class that is
initialized after java.lang.System but before
java.lang.System.initSystemClass() and uses JavaLangAccess.
Thanks,
Mike
