The C++ memory model and CAS loops

The C++ Memory Model

Goal: understand "it works on my machine" vs. "correct in theory" (provably correct)

---

Before C++11

• No standard concurrency model. pthread/WinAPI were de facto standards, but not part of C++.
• Compiler/optimizations could break your assumptions: Sharing objects across threads was unconditionally UB. Reordering, caching, dead store elimination.
• Portable code was hard: Code working on x86 might crash on ARM due to different memory ordering.

---

After C++11

• formal, portable concurrency model in the language standard.
• Defined exactly what the compiler/CPU can and cannot do with memory operations across threads.
• Made multi-threaded C++ first-class, not just a wrapper around OS APIs.
• Your code is only "correct" if it conforms to the memory model. If it works without conforming you're relying on stronger hardware guarantees.

Takeaway: The memory model constrains the implementation. This allows C++ users to reason about concurrent behavior.

---

What Is the C++ Memory Model?

The memory model is the foundation that types like std::thread, std::mutex, or std::atomic rely on.

C++ standard: https://eel.is/c++draft/basic.exec#intro.multithread

Important concepts:

1. threads of Execution
2. Reordering & Visibility
   • by the compiler
   • by the CPU/GPU (also cache coherence protocol)
3. sequenced before
4. synchronizes with
5. happens before
6. forward progress (consider what a program requires to not be useless)

Takeaway: Don't introduce a data race. 😀

---

volatile is something else

volatile is not a tool relevant to the happens before relation. It's a tool for I/O.

• volatile is a tool for communicating with things outside of your program.
• atomics are a tool for communicating within your program.

---

Memory Orderings

• https://en.cppreference.com/w/cpp/atomic/memory_order.html

---

CAS: Compare and Swap

• docs: https://en.cppreference.com/w/cpp/atomic/atomic/compare_exchange
• example: https://compiler-explorer.com/z/aW4shjabd