runtime: implement procyield as ISB instruction on arm64

[rhysh]

When looking into #68578, I found that the implementation of runtime.procyield on GOARCH=arm64 uses the YIELD instruction, and that the YIELD instruction is in effect a (fast) NOP.

The current code is at

go/src/runtime/asm_arm64.s

Lines 917 to 923 in 9a4fe7e

	TEXT runtime·procyield(SB),NOSPLIT,$0-0
	MOVWU cycles+0(FP), R0
	again:
	YIELD
	SUBW $1, R0
	CBNZ R0, again
	RET

, and I added a benchmark, runtime.BenchmarkProcYield, at https://go.dev/cl/601396 .

The difference in delay between amd64 (slow, using PAUSE) and arm64 (fast, using YIELD) makes it hard to be confident in the tuning of the runtime.lock2 spin loop. Note that it's easy to tune the spin loop for the specific duration of a microbenchmark's critical section, which might not be the best tuning for Go overall.

It looks like Rust uses ISB SY, https://github.com/rust-lang/rust/blob/d6c8169c186ab16a3404cd0d0866674018e8a19e/library/core/src/hint.rs#L291-L295 , changed in rust-lang/rust@c064b65 . I've confirmed that using ISB results in a longer delay on the hardware most easily available to me (M1 MacBook Air), which I'd expect to be a benefit to runtime.lock2, both reducing the likelihood of acquiring the lock without a sleep and controlling the electrical energy used to do so.

What I'd most prefer is for Go 1.24 to include a fix for #68578 , and for that to be the only change to runtime.lock2 in the Go 1.24 cycle (so it's clear whether that change is to blame for any changes in mutex performance). But I'm opening this now so we at least don't lose track of it.

CC @golang/runtime @golang/arm

[gabyhelp]

Related Issues and Documentation

• runtime: procyield doesn't use yield instruction on ARM, unlike ARM64 #16663 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[prattmic]

Without truly understanding the details, at face value I am hesitant to use ISB SY over YIELD.

YIELD means exactly what we intend: "Software with a multithreading capability can use a YIELD instruction to indicate to the PE that it is performing a task, for example a spin-lock, that could be swapped out to improve overall system performance."

ISB SY is a sort of instruction fetch barrier: "Instruction Synchronization Barrier flushes the pipeline in the processor, so that all instructions following the ISB are fetched from cache or memory, after the instruction has been completed."

The purpose of ISB SY is for various page table operations (and perhaps self-modifying code?), which is not what we mean.

While I acknowledge that ISB SY takes longer to execute, it doesn't seem like a great fit, especially without understanding the full effects (at minimum it seems to affect branch predictor behavior).

Rather than picking another instruction simply because it takes longer, it seems to me we could instead adjust the number of iterations of the YIELD instruction to make it take longer rather than switching to a slower instruction [1].

I'm happy to be convinced otherwise if we have more concrete information available (such as recommendations from an ARM optimization manual e.g.).

(FWIW, from a quick glance, it looks like Linux uses YIELD: https://elixir.bootlin.com/linux/v6.10.7/source/arch/arm64/include/asm/vdso/processor.h)

[1] At a higher level, all of the various tuned busy/sleep waits in the runtime are sufficiently old that I am not convinced they are particularly well tuned for modern CPUs at all.