perf(port_arm_m): issue DSB in yield_cpu

Despite the addition of an additional instruction, this may hypotheti-
cally confer a performance benefit: In `wait_until_woken_up`, if the
pend operation didn't complete before the subsequent `enter_cpu_lock`
operation, `wait_until_woken_up` would run extra iterations before
PendSV is actually taken, thereby incurring performance penalty. The DSB
operation prevents this from happening.

The measured changes in the execution times of the benchmark suite lie
across the ranges of -3–+5 cycles (STM32F401, Cortex-M4F) and
-158–0 cycles (RP2040, Cortex-M0+).

Author: yvt

src/r3_port_arm_m/src/threading/imp.rs

@@ -171,6 +171,10 @@ impl State {
     pub unsafe fn yield_cpu<Traits: PortInstance>(&'static self) {
         // Safety: See `use_port!`
         cortex_m::peripheral::SCB::set_pendsv();
+
+        // Technically this DSB isn't required for correctness, but ensures
+        // PendSV is taken before the next operation.
+        cortex_m::asm::dsb();
     }
 
     pub unsafe fn exit_and_dispatch<Traits: PortInstance>(