The least requirement for the atomic operation

Question

I just wanna collect some evidence-based info about the requirement for the atomic operation.

eg (maybe not right, just some example)

1. at x86_64, when you write or read on the entry sizeof long and aligned by long, the operation is aotmic，because balabalabala ...

2. at x86_64, when you write or read bytes use gcc atomic primitive , your operation on these bytes will keep atomic.

Hope to receive and help :)

Answer 1

Quote from Professional C++ :

A simple assignment to a simple scalar variable typically requires no synchronization. For example, on the x86, a 32-bit aligned value is handled atomically at the hardware level and a single transaction requires no explicit synchronization. For example, setting a Boolean value to false to stop a thread requires no synchronization, but an operation like ++, -- or op= for any given op requires synchronization.

This doesn't answer all your questions, but is this part of what you were looking for?

Answer 2

If you're looking for the chip-level guarantees you will need to consult the technical manuals for those chips. On Intel check out the S oftware Developer Manuals which has the compelete information on memory ordering. The definition of atomic in this case includes a memory ordering guarantee. The basic guarantees are listing in "Volume 3: 8.2 Memory Ordering", where additionally some functions are marked as behaving in an "atomic" fashion.

Note the complete details are complex and require a detailed understanding of the chip. This is why C and C++ now wrap them in higher level guarantees. You can of course consult those respective standards to see their guarantees.