Programmers steeped in the posix tradition
often tend to think of mutex_lock/_unlock
as memory barriers of sorts. That might have
been one source/rationale for the restriction below?
-- ramki.
> Mime-Version: 1.0
> X-Sender: pugh@localhost (Unverified)
> Date: Mon, 22 Apr 2002 19:31:23 -0400
> To: "Jack, Paul" <pjack@sfaf.org>
> From: Bill Pugh <pugh@cs.umd.edu>
> Subject: Re: JavaMemoryModel: FWD: Question regarding nested synchronized
blocks
> Cc: javamemorymodel@cs.umd.edu
>
> Paul,
> Note that the Java memory model is currently being revised. Don't
> bother paying attention to the existing chapter 17 except as a
> historical curiosity.
>
>
> >[ Forwarded from "Jack, Paul" <pjack@sfaf.org>; who isn't subscribed
> >via that email address]
> >I read in a JavaWorld article that it's forbidden
> >for two synchronized blocks to be merged:
> >
> > sychronized (lock) {
> > <op1>;
> > <op2>;
> > }
> > synchronized (lock) {
> > <op3>;
> > <op4>;
> > }
> >
> >Now assuming there's no side-effects, an optimizing
> >compiler can legally reorder <op1> and <op2>, or
> >reorder <op3> and <op4>, but it can't, for instance,
> >put <op3> between <op1> and <op2>. This all makes
> >sense to me.
>
> Under the new memory model, you will be able to merge synchronization
> blocks (although there are fairness issues associated with doing so).
> If the synchronization blocks are fused so that no other thread can
> obtain the lock between the two synchronization blocks above, then
> <op1>, <op2>, <op3> and <op4> can be freely reordered.
>
>
> >
> >My question involves nested sychronized blocks.
> >According to the JLS (17.9), "Any association between
> >locks and variables is purely conventional. Locking
> >*any* lock conceptually flushes *all* variables from a
> >thread's working memory..."
> >
> >So given nested locks...
> >
> > synchronized (lock1) {
> > synchronized (lock2) {
> > <op1>;
> > <op2>;
> > }
> > <op3>;
> > <op4>;
> > <op5>;
> > synchronized (lock2) {
> > <op6>;
> > <op7>;
> > }
> > }
> >
> >...what reorderings are allowed? It seems that
> ><op2><op1> <op5><op4><op3> <op7><op6> would
> >be legal, but not <op4> <op2><op1> <op5>
> ><op7><op6> <op3>. Am I correct in this assumption,
> >and if I am, what's the rationale?
>
> You are correct with regards to the old memory model.
>
> However, under the new memory model, this restriction is being
> removed (because there was no rationale).
>
> Bill
> -------------------------------
> JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel
-------------------------------
JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel
This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:39 EDT