Fix Windows atomic ops to return pre-operation value (match Linux/C11 semantics)

[Copilot]

ATOMIC_INC/ATOMIC_DEC/ATOMIC_ADD/ATOMIC_SUB return the new value on Windows (InterlockedIncrement/InterlockedDecrement/InterlockedAdd) but the old value on Linux (__sync_fetch_and_add) and C11 (atomic_fetch_add). This breaks cross-platform code like reference counting:

// Works on Linux (returns old value), broken on Windows (returns new value)
if (ATOMIC_DEC(&ref->refcnt) == 1) {
    release_cb(ref);
}

• Reorder #ifdef checks so GCC-based compilers (including MinGW) use __sync_fetch_and_add builtins via a direct __GNUC__ version check, instead of the broken __GNUC_PREREQ macro (which always evaluates to 0 on MinGW since it lacks glibc's <features.h>)
• For pure MSVC (non-GCC), replace InterlockedAdd/InterlockedIncrement/InterlockedDecrement with InterlockedExchangeAdd, which returns the original value
• Fix ATOMIC_SUB macro parenthesization: -(n) instead of -n

Original prompt

---

This section details on the original issue you should resolve

<issue_title>hatomic.h windows下原子操作错误</issue_title>
<issue_description>不同长度的整型使用这些原子操作接口时得到的结果不正确。
比如：atomic_flag_test_and_set函数（内部代码：InterlockedCompareExchange((LONG*)&p->_Value, 1, 0)）调用后并不能让atomic_flag置1。

还有这些原子操作接口在不同平台下的返回值不同。比如我用来实现引用计数的代码：

if (ATOMIC_DEC(&ref->refcnt) == 1)
{
    release_cb(ref);
}

在linux下，ATOMIC_DEC返回的减少之前的值，而在windows下返回的是减少之后的值，这导致这段代码在Windows下不正确。</issue_description>

<agent_instructions>Please analyze and respond to this issue.</agent_instructions>

Comments on the Issue (you are @copilot in this section)

@ithewei libhv里确实没有用到ATOMIC_INC、ATOMIC_DEC的返回值，这里windows下可以考虑使用InterlockedAdd，而不是InterlockedIncrement/InterlockedDecrement，应该就可以了吧

• Fixes hatomic.h windows下原子操作错误 #581

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[Copilot]

Pull request overview

Aligns Windows atomic operation return values with Linux (__sync_fetch_and_*) and C11 (atomic_fetch_*) semantics by making ATOMIC_* return the pre-operation value, fixing cross-platform correctness for patterns like reference counting.

Changes:

• Switch Windows ATOMIC_ADD/ATOMIC_SUB to InterlockedExchangeAdd (returns original value).
• Implement Windows ATOMIC_INC/ATOMIC_DEC via InterlockedExchangeAdd with +1/-1 (returns original value).
• Fix ATOMIC_SUB negation parenthesization (-(n)).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This change intentionally alters Windows return semantics for ATOMIC_ADD/SUB/INC/DEC to match fetch_* (returning the pre-operation value). There’s a unit test for hatomic, but it only prints values and doesn’t assert these semantics; please add an explicit regression test that verifies both the returned value and the stored counter value for inc/dec/add/sub (especially on Windows) so this doesn’t regress again.

[ithewei]

@copilot InterlockedExchangeAdd 在mingw平台上有吗，担心影响mingw编译

[Copilot]

@copilot InterlockedExchangeAdd 在mingw平台上有吗，担心影响mingw编译

InterlockedExchangeAdd is available on MinGW (it's been in Windows API since XP, actually more compatible than the previous InterlockedAdd). But to fully avoid the concern, in 0f3ec91 I reordered the #ifdef checks: GCC-based compilers (including MinGW) now use __sync_fetch_and_add builtins via a direct __GNUC__ version check instead of the broken __GNUC_PREREQ macro (which always evaluates to 0 on MinGW since it lacks glibc's <features.h>). Only pure MSVC falls through to the InterlockedExchangeAdd path.