Re: [RFA/JDWP] ReferenceTypeCommandSet.java

[Bryce McKinlay <mckinlay at redhat dot com>]

Keith Seitz wrote:

> > There's also a couple places I use classes from java.lang.reflect
> > This is a bit of a problem as it's been pointed out if in the instance
> > of a command like 
> > http://java.sun.com/j2se/1.5.0/docs/guide/jpda/jdwp/jdwp-protocol.html#JDWP_ReferenceType_GetValues
> > Using the Field object will cause the Class to initialize which would be
> > changing the execution of the program. Tom Tromey suggested we could
> > replace the java.lang.relfect classes we use by our own classes later on
> > that reach into the JVM and avoid this problem.
> >    
>
> The question is, though, is this really a problem?
>
> In order to use ReferenceType commands, the debugger must have already
> requested a reference ID for the given class. AFAICT the only way to get
> a reference type ID for a class is to register for some event. Eclipse,
> for example, uses class prepare events.
>
> Anyway, the existence of a reference type ID would mean that the class
> must have been loaded/initialized at some time, no?
>  

Class preparation is distinct from initialization. Preparation is 
effectively linking the class: laying out its fields, initializing class 
data structures in the VM, etc. Initialization involves actually running 
Java code to initialize the classes static fields, and any "static" code 
blocks the class may have. The point at which initialization occurs is 
strictly defined by the JLS, and its pretty easy to write a program that 
will behave differently if initializations are not performed in the 
"right" order.

That means that application could be behave differently when it is 
debugged. So, I think ultimately we will need to address this issue. 
Whether it matters in practice for the majority of debug sessions, I 
don't know.

> At the very least, as Tom suggests, we could certainly use this a first
> attempt, replacing when/if necessary.
>  

That sounds good to me.

Bryce