并发编程网 - ifeve.com

最近再次翻netty和disrupt的源码, 发现一些地方使用AtomicXXX.lazySet()/unsafe.putOrderedXXX系列, 以前一直没有注意lazySet这个方法, 仔细研究一下发现很有意思。我们拿AtomicReferenceFieldUpdater的set()和lazySet()作比较, 其他AtomicXXX类和这个类似。

[code lang=”java”]
public void set(T obj, V newValue) {
// …
unsafe.putObjectVolatile(obj, offset, newValue);
}

public void lazySet(T obj, V newValue) {
// …
unsafe.putOrderedObject(obj, offset, newValue);
}
[/code]

1.首先set()是对volatile变量的一个写操作, 我们知道volatile的write为了保证对其他线程的可见性会追加以下两个Fence(内存屏障)
1)StoreStore // 在intel cpu中, 不存在[写写]重排序, 这个可以直接省略了
2)StoreLoad // 这个是所有内存屏障里最耗性能的
注: 内存屏障相关参考Doug Lea大大的cookbook (http://g.oswego.edu/dl/jmm/cookbook.html)

2.Doug Lea大大又说了, lazySet()省去了StoreLoad屏障, 只留下StoreStore
在这里 http://bugs.java.com/bugdatabase/view_bug.do?bug_id=6275329
把最耗性能的StoreLoad拿掉, 性能必然会提高不少(虽然不能禁止写读的重排序了保证不了可见性, 但给其他应用场景提供了更好的选择, 比如上边连接中Doug Lea举例的场景)。

但是但是, 在好奇心驱使下我翻了下JDK的源码(unsafe.cpp):

[code lang=”java”]
// 这是unsafe.putObjectVolatile()
UNSAFE_ENTRY(void, Unsafe_SetObjectVolatile(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jobject x_h))
UnsafeWrapper("Unsafe_SetObjectVolatile");
oop x = JNIHandles::resolve(x_h);
oop p = JNIHandles::resolve(obj);
void* addr = index_oop_from_field_offset_long(p, offset);
OrderAccess::release();
if (UseCompressedOops) {
oop_store((narrowOop*)addr, x);
} else {
oop_store((oop*)addr, x);
}
OrderAccess::fence();
UNSAFE_END

// 这是unsafe.putOrderedObject()
UNSAFE_ENTRY(void, Unsafe_SetOrderedObject(JNIEnv *env, jobject unsafe, jobject obj, jlong offset, jobject x_h))
UnsafeWrapper("Unsafe_SetOrderedObject");
oop x = JNIHandles::resolve(x_h);
oop p = JNIHandles::resolve(obj);
void* addr = index_oop_from_field_offset_long(p, offset);
OrderAccess::release();
if (UseCompressedOops) {
oop_store((narrowOop*)addr, x);
} else {
oop_store((oop*)addr, x);
}
OrderAccess::fence();
UNSAFE_END
[/code]

仔细看代码是不是有种被骗的感觉, 他喵的一毛一样啊. 难道是JIT做了手脚?生成汇编看看

生成assembly code需要hsdis插件

mac平台从这里下载 https://kenai.com/projects/base-hsdis/downloads/directory/gnu-versions

linux和windows可以从R大的[高级语言虚拟机圈子]下载 http://hllvm.group.iteye.com/

为了测试代码简单, 使用AtomicLong来测:

[code lang=”java”]
// set()
public class LazySetTest {
private static final AtomicLong a = new AtomicLong();

public static void main(String[] args) {
for (int i = 0; i < 100000000; i++) {
a.set(i);
}
}
}

// lazySet()
public class LazySetTest {
private static final AtomicLong a = new AtomicLong();

public static void main(String[] args) {
for (int i = 0; i < 100000000; i++) {
a.lazySet(i);
}
}
}
[/code]

分别执行以下命令:

[code lang=”java”]
1.export LD_LIBRARY_PATH=~/hsdis插件路径/
2.javac LazySetTest.java && java -XX:+UnlockDiagnosticVMOptions -XX:+PrintAssembly LazySetTest

// ——————————————————
// set()的assembly code片段:
0x000000010ccbfeb3: mov %r10,0x10(%r9)
0x000000010ccbfeb7: lock addl $0x0,(%rsp) ;*putfield value
; – java.util.concurrent.atomic.AtomicLong::set@2 (line 112)
; – LazySetTest::main@13 (line 13)
0x000000010ccbfebc: inc %ebp ;*iinc
; – LazySetTest::main@16 (line 12)
// ——————————————————
// lazySet()的assembly code片段:
0x0000000108766faf: mov %r10,0x10(%rcx) ;*invokevirtual putOrderedLong
; – java.util.concurrent.atomic.AtomicLong::lazySet@8 (line 122)
; – LazySetTest::main@13 (line 13)
0x0000000108766fb3: inc %ebp ;*iinc
; – LazySetTest::main@16 (line 12)
[/code]

好吧, set()生成的assembly code多了一个lock前缀的指令

查询IA32手册可知道, lock addl $0x0,(%rsp)其实就是StoreLoad屏障了, 而lazySet()确实没生成StoreLoad屏障

这里JIT除了将方法内联, 相同代码生成不同指令是怎么做到的?

http://hg.openjdk.java.net/jdk7u/jdk7u/hotspot/file/6e9aa487055f/src/share/vm/classfile/vmSymbols.hpp

查看如上代码, 812行和868行分别有如下代码:

[code lang=”java”]
do_intrinsic(_putObjectVolatile, sun_misc_Unsafe, putObjectVolatile_name, putObject_signature, F_RN)
do_intrinsic(_putOrderedObject, sun_misc_Unsafe, putOrderedObject_name, putOrderedObject_signature, F_RN)
[/code]

putObjectVolatile与putOrderedObject都在vmSymbols.hpp的宏定义中,jvm会根据instrinsics id生成特定的指令集 putObjectVolatile与putOrderedObject生成的汇编指令不同估计是源于这里了, 继续往下看 hotspot/src/share/vm/opto/libaray_call.cpp这个类:
首先看如下两行代码:

[code lang=”java”]
case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
[/code]

再看inline_unsafe_access()和inline_unsafe_ordered_store(), 不贴出全部代码了, 只贴出重要的部分:

[code lang=”java”]
bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
// This is another variant of inline_unsafe_access, differing in
// that it always issues store-store ("release") barrier and ensures
// store-atomicity (which only matters for "long").

// …
if (type == T_OBJECT) // reference stores need a store barrier.
store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
else {
store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
}
insert_mem_bar(Op_MemBarCPUOrder);
return true;
}

———————————————————————————————————

bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
// ….

if (is_volatile) {
if (!is_store)
insert_mem_bar(Op_MemBarAcquire);
else
insert_mem_bar(Op_MemBarVolatile);
}

if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);

return true;
}
[/code]

我们可以看到 inline_unsafe_access()方法中, 如果是is_volatile为true, 并且是store操作的话, 有这样的一句代码 insert_mem_bar(Op_MemBarVolatile), 而inline_unsafe_ordered_store没有插入这句代码

再继续看/hotspot/src/cpu/x86/vm/x86_64.ad的membar_volatile

[code lang=”java”]
instruct membar_volatile(rFlagsReg cr) %{
match(MemBarVolatile);
effect(KILL cr);
ins_cost(400);

format %{
$$template
if (os::is_MP()) {
$$emit$$"lock addl [rsp + #0], 0\t! membar_volatile"
} else {
$$emit$$"MEMBAR-volatile ! (empty encoding)"
}
%}
ins_encode %{
__ membar(Assembler::StoreLoad);
%}
ins_pipe(pipe_slow);
%}
[/code]

lock addl [rsp + #0], 0\t! membar_volatile指令原来来自这里

总结:
错过一些细节, 但在主流程上感觉是有一点点明白了, 有错误之处请指正

参考了以下资料:
1.http://g.oswego.edu/dl/jmm/cookbook.html
2.https://wikis.oracle.com/display/HotSpotInternals/PrintAssembly
3.http://www.quora.com/How-does-AtomicLong-lazySet-work
4.http://bad-concurrency.blogspot.ru/2012/10/talk-from-jax-london.html