runtime: measure speed of procyield and osyield

These are delay primitives for lock2. If a mutex isn't immediately
available, we can use procyield to tell the processor to wait for a
moment, or osyield to allow the OS to run a different process or thread
if one is waiting. We expect a processor-level yield to be faster than
an os-level yield, and for both of them to be fast relative to entering
a full sleep (via futexsleep or semasleep).

Each architecture has its own way of hinting to the processor that it's
in a spin-wait loop, so procyield presents an architecture-independent
interface for use in lock_futex.go and lock_sema.go.

Measure the (single-threaded) speed of these to confirm.

For #68578

Change-Id: I90cd46ea553f2990395aceb048206285558c877e
Reviewed-on: https://go-review.googlesource.com/c/go/+/601396
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Dmitri Shuralyov <[email protected]>
Reviewed-by: Michael Knyszek <[email protected]>

src/runtime/export_test.go

@@ -1223,6 +1223,9 @@ func PageCachePagesLeaked() (leaked uintptr) {
 	return
 }
 
+var ProcYield = procyield
+var OSYield = osyield
+
 type Mutex = mutex
 
 var Lock = lock

src/runtime/runtime_test.go

@@ -539,3 +539,25 @@ func TestTimediv(t *testing.T) {
 		})
 	}
 }
+
+func BenchmarkProcYield(b *testing.B) {
+	benchN := func(n uint32) func(*testing.B) {
+		return func(b *testing.B) {
+			for i := 0; i < b.N; i++ {
+				ProcYield(n)
+			}
+		}
+	}
+
+	b.Run("1", benchN(1))
+	b.Run("10", benchN(10))
+	b.Run("30", benchN(30)) // active_spin_cnt in lock_sema.go and lock_futex.go
+	b.Run("100", benchN(100))
+	b.Run("1000", benchN(1000))
+}
+
+func BenchmarkOSYield(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		OSYield()
+	}
+}