Why use sequential consistent atomic?

[MichaelSuen-thePointer]

-- This is not an official python repo so I put discussion here. --

I've read your blog. Using biased atomic ref counting is a genius idea. But in the code, I've noticed that you intentionally use sequential consistent atomic operations (seq_cst), in all of the atomic wrappers (std, gcc, msvc). Could you explain the reason behind the choice?

https://github.com/colesbury/nogil-3.12/blob/nogil-3.12/Include/pyatomic_std.h#L21

[colesbury]

• It's always safe to use seq_cst in place of weaker memory orders
• It's simpler and less error prone, especially because we have to support MSVC, which does not support C11 atomic currently
• There's no significant performance impact on the platforms we care about

[MichaelSuen-thePointer]

There's no significant performance impact on the platforms we care about

Ever tested on ARM platforms? From what I know, sequential consistent atomics are much heavier than acquire/release atomics in weak order CPUS.

[colesbury]

From what I know, sequential consistent atomics are much heavier than acquire/release atomics in weak order CPUS.

No. On 64-bit ARM, you will see the same machine code generated for every (or nearly every) atomic operation we use for seq_cst vs. acq/rel.