runtime: benchmark for bulk write barriers

This adds a benchmark of typedslicecopy and its bulk write barriers.

For #22460.

Change-Id: I439ca3b130bb22944468095f8f18b464e5bb43ca
Reviewed-on: https://go-review.googlesource.com/74051
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>

src/runtime/gc_test.go

@@ -517,6 +517,32 @@ func TestUserForcedGC(t *testing.T) {
 	}
 }
 
+func writeBarrierBenchmark(b *testing.B, f func()) {
+	runtime.GC()
+	var ms runtime.MemStats
+	runtime.ReadMemStats(&ms)
+	//b.Logf("heap size: %d MB", ms.HeapAlloc>>20)
+
+	// Keep GC running continuously during the benchmark, which in
+	// turn keeps the write barrier on continuously.
+	var stop uint32
+	done := make(chan bool)
+	go func() {
+		for atomic.LoadUint32(&stop) == 0 {
+			runtime.GC()
+		}
+		close(done)
+	}()
+	defer func() {
+		atomic.StoreUint32(&stop, 1)
+		<-done
+	}()
+
+	b.ResetTimer()
+	f()
+	b.StopTimer()
+}
+
 func BenchmarkWriteBarrier(b *testing.B) {
 	if runtime.GOMAXPROCS(-1) < 2 {
 		// We don't want GC to take our time.
@@ -546,53 +572,70 @@ func BenchmarkWriteBarrier(b *testing.B) {
 	const depth = 22 // 64 MB
 	root := mkTree(22)
 
-	runtime.GC()
-	var ms runtime.MemStats
-	runtime.ReadMemStats(&ms)
-	//b.Logf("heap size: %d MB", ms.HeapAlloc>>20)
+	writeBarrierBenchmark(b, func() {
+		var stack [depth]*node
+		tos := -1
 
-	// Keep GC running continuously during the benchmark.
-	var stop uint32
-	done := make(chan bool)
-	go func() {
-		for atomic.LoadUint32(&stop) == 0 {
-			runtime.GC()
+		// There are two write barriers per iteration, so i+=2.
+		for i := 0; i < b.N; i += 2 {
+			if tos == -1 {
+				stack[0] = root
+				tos = 0
+			}
+
+			// Perform one step of reversing the tree.
+			n := stack[tos]
+			if n.l == nil {
+				tos--
+			} else {
+				n.l, n.r = n.r, n.l
+				stack[tos] = n.l
+				stack[tos+1] = n.r
+				tos++
+			}
+
+			if i%(1<<12) == 0 {
+				// Avoid non-preemptible loops (see issue #10958).
+				runtime.Gosched()
+			}
 		}
-		close(done)
-	}()
-
-	b.ResetTimer()
-
-	var stack [depth]*node
-	tos := -1
-
-	// There are two write barriers per iteration, so i+=2.
-	for i := 0; i < b.N; i += 2 {
-		if tos == -1 {
-			stack[0] = root
-			tos = 0
-		}
-
-		// Perform one step of reversing the tree.
-		n := stack[tos]
-		if n.l == nil {
-			tos--
-		} else {
-			n.l, n.r = n.r, n.l
-			stack[tos] = n.l
-			stack[tos+1] = n.r
-			tos++
-		}
-
-		if i%(1<<12) == 0 {
-			// Avoid non-preemptible loops (see issue #10958).
-			runtime.Gosched()
-		}
-	}
-
-	b.StopTimer()
-	atomic.StoreUint32(&stop, 1)
-	<-done
+	})
 
 	runtime.KeepAlive(wbRoots)
 }
+
+func BenchmarkBulkWriteBarrier(b *testing.B) {
+	if runtime.GOMAXPROCS(-1) < 2 {
+		// We don't want GC to take our time.
+		b.Skip("need GOMAXPROCS >= 2")
+	}
+
+	// Construct a large set of objects we can copy around.
+	const heapSize = 64 << 20
+	type obj [16]*byte
+	ptrs := make([]*obj, heapSize/unsafe.Sizeof(obj{}))
+	for i := range ptrs {
+		ptrs[i] = new(obj)
+	}
+
+	writeBarrierBenchmark(b, func() {
+		const blockSize = 1024
+		var pos int
+		for i := 0; i < b.N; i += blockSize {
+			// Rotate block.
+			block := ptrs[pos : pos+blockSize]
+			first := block[0]
+			copy(block, block[1:])
+			block[blockSize-1] = first
+
+			pos += blockSize
+			if pos+blockSize > len(ptrs) {
+				pos = 0
+			}
+
+			runtime.Gosched()
+		}
+	})
+
+	runtime.KeepAlive(ptrs)
+}