最近看了Advanced Design and Implementation of Virtual Machines，里面说了Native code的异常是通过保存在线程里面的变量实现的，真正的异常实现源码今天才看，顺便记录下来。
异常实现的源码在/hotspot/src/share/vm/utilities/exceptions.hpp中，一段简介:
// This file provides the basic support for exception handling in the VM.
// Note: We do not use C++ exceptions to avoid compiler dependencies and
// unpredictable performance.
//
// Scheme: Exceptions are stored with the thread. There is never more
// than one pending exception per thread. All functions that can throw
// an exception carry a THREAD argument (usually the last argument and
// declared with the TRAPS macro). Throwing an exception means setting
// a pending exception in the thread. Upon return from a function that
// can throw an exception, we must check if an exception is pending.
// The CHECK macros do this in a convenient way. Carrying around the
// thread provides also convenient access to it (e.g. for Handle
// creation, w/o the need for recomputation).
不使用c++的异常机制是为了减少编译器的依赖和性能的考虑。异常被保存在线程的变量中，一个线程同时最多只能有一个异常。所有带有 THREAD 参数的函数都有可能抛异常， THREAD 是关于线程的宏，下面会讲到。抛出一个异常意味着把这个异常放到线程的变量中，在函数返回的时候我们可以通过 CHECK 宏检查有没有抛异常。
class ThreadShadow: public CHeapObj {
friend class VMStructs;
friend class JVMCIVMStructs;
protected:
oop _pending_exception; // Thread has gc actions.
const char* _exception_file; // file information for exception (debugging only)
int _exception_line; // line information for exception (debugging only)
friend void check_ThreadShadow(); // checks _pending_exception offset
// The following virtual exists only to force creation of a vtable.
// We need ThreadShadow to have a vtable, even in product builds,
// so that its layout will start at an offset of zero relative to Thread.
// Some C++ compilers are so "clever" that they put the ThreadShadow
// base class at offset 4 in Thread (after Thread's vtable), if they
// notice that Thread has a vtable but ThreadShadow does not.
virtual void unused_initial_virtual() { }
public:
oop pending_exception() const { return _pending_exception; }
bool has_pending_exception() const { return _pending_exception != NULL; }
const char* exception_file() const { return _exception_file; }
int exception_line() const { return _exception_line; }
// Code generation support
static ByteSize pending_exception_offset() { return byte_offset_of(ThreadShadow, _pending_exception); }
// use THROW whenever possible!
void set_pending_exception(oop exception, const char* file, int line);
// use CLEAR_PENDING_EXCEPTION whenever possible!
void clear_pending_exception();
ThreadShadow() : _pending_exception(NULL),
_exception_file(NULL), _exception_line(0) {}
};
ThreadShadow 是 Thread 的父类，_pending_exception 保存了线程在运行过程中抛出的异常。
void ThreadShadow::set_pending_exception(oop exception, const char* file, int line) {
assert(exception != NULL && exception->is_oop(), "invalid exception oop");
_pending_exception = exception;
_exception_file = file;
_exception_line = line;
}
void ThreadShadow::clear_pending_exception() {
if (_pending_exception != NULL && log_is_enabled(Debug, exceptions)) {
ResourceMark rm;
outputStream* logst = Log(exceptions)::debug_stream();
logst->print("Thread::clear_pending_exception: cleared exception:");
_pending_exception->print_on(logst);
}
_pending_exception = NULL;
_exception_file = NULL;
_exception_line = 0;
}
抛异常的时候会调用 set_pending_exception，Exceptions类提供了很多很方便抛异常的宏，比如：
#define THROW_HANDLE(e) { Exceptions::_throw(THREAD_AND_LOCATION, e); return; }
去掉无用代码后，可以看到 _throw 对当前线程设置了异常
void Exceptions::_throw(Thread* thread, const char* file, int line, Handle h_exception, const char* message) {
// Check for special boot-strapping/vm-thread handling
if (special_exception(thread, file, line, h_exception)) {
return;
}
if (h_exception->is_a(SystemDictionary::OutOfMemoryError_klass())) {
count_out_of_memory_exceptions(h_exception);
}
// set the pending exception
thread->set_pending_exception(h_exception(), file, line);
}
Hotspot也提供了很多有用的宏帮助我们使用这些异常，比如：
// The THREAD & TRAPS macros facilitate the declaration of functions that throw exceptions.
// Convention: Use the TRAPS macro as the last argument of such a function; e.g.:
//
// int this_function_may_trap(int x, float y, TRAPS)
可以使用 TRAPS 宏来帮助我们友好的抛异常，使用 CHECK 宏来帮助我们检查异常。
#define THREAD __the_thread__
#define TRAPS Thread* THREAD
#define PENDING_EXCEPTION (((ThreadShadow*)THREAD)->pending_exception())
#define HAS_PENDING_EXCEPTION (((ThreadShadow*)THREAD)->has_pending_exception())
#define CLEAR_PENDING_EXCEPTION (((ThreadShadow*)THREAD)->clear_pending_exception())
#define CHECK THREAD); if (HAS_PENDING_EXCEPTION) return ; (void)(0
#define CHECK_(result) THREAD); if (HAS_PENDING_EXCEPTION) return result; (void)(0
#define CHECK_0 CHECK_(0)
#define CHECK_NH CHECK_(Handle())
#define CHECK_NULL CHECK_(NULL)
#define CHECK_false CHECK_(false)
#define CHECK_JNI_ERR CHECK_(JNI_ERR)
#define CHECK_AND_CLEAR THREAD); if (HAS_PENDING_EXCEPTION) { CLEAR_PENDING_EXCEPTION; return; } (void)(0
#define CHECK_AND_CLEAR_(result) THREAD); if (HAS_PENDING_EXCEPTION) { CLEAR_PENDING_EXCEPTION; return result; } (void)(0
#define CHECK_AND_CLEAR_0 CHECK_AND_CLEAR_(0)
#define CHECK_AND_CLEAR_NH CHECK_AND_CLEAR_(Handle())
#define CHECK_AND_CLEAR_NULL CHECK_AND_CLEAR_(NULL)
#define CHECK_AND_CLEAR_false CHECK_AND_CLEAR_(false)
举个例子:
int result = this_function_may_trap(x_arg, y_arg, CHECK_0);
把CHECK_0 替换为 CHECK_(0) 后展开：
int result = this_function_may_trap(x_arg, y_arg, THREAD);
if (HAS_PENDING_EXCEPTION) return 0; (void)(0);
可以很清楚的看到如果有异常不会执行接下来的代码，而是返回0。
以上就是整个 Native code 异常的实现，异常不会抛出而是存在线程的变量中，在调用方法的时候去检查这个变量，保证了整个异常机制的正常运作。