x86 / x64架构是否能够抵御不安全的发布？

Question

I have observed, that unsafe publication of objects in JVM 1.5+ is very unlikely to cause any trouble, regardless of JVM specs stating that it is not guaranteed that such objects will be visible between threads.

While looking around on the internet I have found this: http://forum.springsource.org/archive/index.php/t-60676.html Person with nick " al0 " claims, that "[...] it is very unlikely to meet such behaviour on x86/x64 based machines, but on HP PA-RISC or IBM Power... based computers (for example AS400) it is much more likely".

Are x86/x64 architectures resistant to unsafe publication? How come?

Answer 1

x86 processors are somewhat resistant to unsafe publication. In particular, as long as one thread is only writing to shared memory and another is only reading, the processor treats all loads and stores as if the memory locations had Java volatile semantics. Writes are never reordered past writes, and reads are never reordered past reads, so the reading thread always sees writes in the correct order

However:

• writes can be moved after a read that comes later in the program. In this sense, volatile is still stronger than what the x86 promises

• this only covers what the processor can do to your code. The virtual machine can still reorder the code as it wishes. For example, it could rewrite xb = 4; ya = 5; xb = 4; ya = 5; to ya = 4; xb = 5 ya = 4; xb = 5 .

Such a decision can be taken based on many different factors: choosing which parts of the code will be JIT-compiled, inlining, scheduling... So even on processors with strong memory ordering, unsafe publication remains unsafe.

Answer 2

I don't want to speculate if the problem is more or less likely to occur on x86/x64 CPUs. Claiming that it is very unlikely to see this behaviour on x86/x64 CPUs is wrong .

Niklas Schlimm has written an article about this in DZone with a working example to demonstrate the incorrect behaviour if a variable with multi-threaded access is not declared volatile. He writes in the article that he can only reproduce the behaviour with the server VM, but for me, the test fails with the client VM as well (Oracle Java 7, Intel x64 Mobile CPU). It is important to remember that it depends on at least the exact version of the Java VM and the CPU model if the behaviour is reproducable at all or perhaps even only sporadic.