Is the modify method atomic?

[K4LCIFER]

The docs state the following:

the modify method performs a single read-modify-write operation that involves one read (LDR) to the register, modifying the value and then a single write (STR) of the modified value to the register.

I interpret this to mean that it isn't atomic as there is one read operation and one write operation.

---

Side note, I apologize if this isn't the right repo for this question — it's not immediately clear to me what repos related to avr-hal are explicitly responsible for what.

[Rahix]

Well, the answer is actually not that straight forward :)

I interpret this to mean that it isn't atomic as there is one read operation and one write operation.

Basically, this is correct. The calls to .modify() get lowered into instructions to first read the register, modify the value, and then write it back. Similar to the instructions that the typical |= or &= statements in AVR C code generate.

So the resulting update of the register is not atomic. If you need it to be, you have to wrap the .modify() call in a critical section using avr_device::interrupt::free(). Or in certain cases use .write() instead which is atomic by writing a whole new value in a single instruction (but only for 8-bit registers).

But there is some more nuance. For certain kinds of .modify() calls, the resulting register update is atomic anyway. If only a single known bit is set in the .modify() call or a single known bit is cleared, the compiler will lower this .modify() call into an sbi/cbi instruction which is atomic. This only works on low registers and when the aforementioned constraint is met. But it can be very useful. For example, we exploit this trick in the GPIO pin handling code to atomically update pin values.

[K4LCIFER]

Very speedy and informative answer. Thank you!

If only a single known bit is set in the .modify() call or a single known bit is cleared, the compiler will lower this .modify() call into an sbi/cbi instruction which is atomic. This only works on low registers and when the aforementioned constraint is met.

Presumably this is only the case when compiler optimizations are enabled (which I think is generally the case with --release)?

[Rahix]

Yes, true. But building without optimization on AVR is mostly impossible because the binary size will immediately surpass the available flash space ;)