runtime: faster chan creation on Linux/FreeBSD/Plan9

The change removes chan finalizer (Lock destructor)
if it is not required on the platform.

benchmark                    old ns/op    new ns/op    delta
BenchmarkChanCreation          1132.00       381.00  -66.34%
BenchmarkChanCreation-2        1215.00       243.00  -80.00%
BenchmarkChanCreation-4        1084.00       186.00  -82.84%
BenchmarkChanCreation-8        1415.00       154.00  -89.12%
BenchmarkChanCreation-16       1386.00       144.00  -89.61%
(on 2 x Intel Xeon E5620, 8 HT cores, 2.4 GHz, Linux)

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/4841041

9 files changed, 34 insertions, 26 deletions

diff --git a/src/pkg/runtime/chan.c b/src/pkg/runtime/chan.c
index b77e51b60d..65feacb78b 100644
--- a/src/pkg/runtime/chan.c
+++ b/src/pkg/runtime/chan.c
@@ -85,7 +85,6 @@ runtime·makechan_c(Type *elem, int64 hint)
 {
 	Hchan *c;
 	int32 n;
-	byte *by;
 
 	if(hint < 0 || (int32)hint != hint || (elem->size > 0 && hint > ((uintptr)-1) / elem->size))
 		runtime·panicstring("makechan: size out of range");
@@ -101,10 +100,9 @@ runtime·makechan_c(Type *elem, int64 hint)
 		n++;
 
 	// allocate memory in one call
-	by = runtime·mal(n + hint*elem->size);
-
-	c = (Hchan*)by;
-	runtime·addfinalizer(c, destroychan, 0);
+	c = (Hchan*)runtime·mal(n + hint*elem->size);
+	if(runtime·destroylock)
+		runtime·addfinalizer(c, destroychan, 0);
 
 	c->elemsize = elem->size;
 	c->elemalg = &runtime·algarray[elem->alg];
diff --git a/src/pkg/runtime/chan_test.go b/src/pkg/runtime/chan_test.go
index c5ffe93acc..71c9e2fd78 100644
--- a/src/pkg/runtime/chan_test.go
+++ b/src/pkg/runtime/chan_test.go
@@ -265,3 +265,25 @@ func BenchmarkChanProdConsWork10(b *testing.B) {
 func BenchmarkChanProdConsWork100(b *testing.B) {
 	benchmarkChanProdCons(b, 100, 100)
 }
+
+func BenchmarkChanCreation(b *testing.B) {
+	const CallsPerSched = 1000
+	procs := runtime.GOMAXPROCS(-1)
+	N := int32(b.N / CallsPerSched)
+	c := make(chan bool, procs)
+	for p := 0; p < procs; p++ {
+		go func() {
+			for atomic.AddInt32(&N, -1) >= 0 {
+				for g := 0; g < CallsPerSched; g++ {
+					myc := make(chan int, 1)
+					myc <- 0
+					<-myc
+				}
+			}
+			c <- true
+		}()
+	}
+	for p := 0; p < procs; p++ {
+		<-c
+	}
+}
diff --git a/src/pkg/runtime/darwin/thread.c b/src/pkg/runtime/darwin/thread.c
index 235d69abfc..6733e815e8 100644
--- a/src/pkg/runtime/darwin/thread.c
+++ b/src/pkg/runtime/darwin/thread.c
@@ -82,8 +82,8 @@ runtime·unlock(Lock *l)
 	}
 }
 
-void
-runtime·destroylock(Lock *l)
+static void
+destroylock(Lock *l)
 {
 	if(l->sema != 0) {
 		runtime·mach_semdestroy(l->sema);
@@ -147,6 +147,7 @@ runtime·osinit(void)
 	// to let the C pthread libary install its own thread-creation callback.
 	if(!runtime·iscgo)
 		runtime·bsdthread_register();
+	runtime·destroylock = destroylock;
 }
 
 void
diff --git a/src/pkg/runtime/freebsd/thread.c b/src/pkg/runtime/freebsd/thread.c
index 569098aa21..f8c550f578 100644
--- a/src/pkg/runtime/freebsd/thread.c
+++ b/src/pkg/runtime/freebsd/thread.c
@@ -102,11 +102,6 @@ runtime·unlock(Lock *l)
 	umtx_unlock(l);
 }
 
-void
-runtime·destroylock(Lock*)
-{
-}
-
 // Event notifications.
 void
 runtime·noteclear(Note *n)
diff --git a/src/pkg/runtime/linux/thread.c b/src/pkg/runtime/linux/thread.c
index 8efba2b98b..4878a00f25 100644
--- a/src/pkg/runtime/linux/thread.c
+++ b/src/pkg/runtime/linux/thread.c
@@ -199,11 +199,6 @@ runtime·unlock(Lock *l)
 	futexunlock(l);
 }
 
-void
-runtime·destroylock(Lock*)
-{
-}
-
 
 // One-time notifications.
 void
diff --git a/src/pkg/runtime/plan9/thread.c b/src/pkg/runtime/plan9/thread.c
index b091c59788..776989242d 100644
--- a/src/pkg/runtime/plan9/thread.c
+++ b/src/pkg/runtime/plan9/thread.c
@@ -114,12 +114,6 @@ runtime·unlock(Lock *l)
 }
 
 
-void 
-runtime·destroylock(Lock *l)
-{
-	// nothing
-}
-
 // User-level semaphore implementation:
 // try to do the operations in user space on u,
 // but when it's time to block, fall back on the kernel semaphore k.
diff --git a/src/pkg/runtime/runtime.c b/src/pkg/runtime/runtime.c
index c572897d2c..57c0873306 100644
--- a/src/pkg/runtime/runtime.c
+++ b/src/pkg/runtime/runtime.c
@@ -10,6 +10,7 @@ enum {
 };
 
 uint32	runtime·panicking;
+void	(*runtime·destroylock)(Lock*);
 
 /*
  * We assume that all architectures turn faults and the like
diff --git a/src/pkg/runtime/runtime.h b/src/pkg/runtime/runtime.h
index 434c82b95d..d2e4378b59 100644
--- a/src/pkg/runtime/runtime.h
+++ b/src/pkg/runtime/runtime.h
@@ -380,6 +380,7 @@ extern	uint32	runtime·panicking;
 extern	int32	runtime·gcwaiting;		// gc is waiting to run
 int8*	runtime·goos;
 extern	bool	runtime·iscgo;
+extern	void	(*runtime·destroylock)(Lock*);
 
 /*
  * common functions and data
@@ -515,7 +516,6 @@ void	runtime·starttheworld(void);
  */
 void	runtime·lock(Lock*);
 void	runtime·unlock(Lock*);
-void	runtime·destroylock(Lock*);
 
 /*
  * sleep and wakeup on one-time events.
diff --git a/src/pkg/runtime/windows/thread.c b/src/pkg/runtime/windows/thread.c
index 5644fd5dd7..4ab043e887 100644
--- a/src/pkg/runtime/windows/thread.c
+++ b/src/pkg/runtime/windows/thread.c
@@ -40,12 +40,14 @@ extern void *runtime·WaitForSingleObject;
 extern void *runtime·WriteFile;
 
 static int64 timerfreq;
+static void destroylock(Lock *l);
 
 void
 runtime·osinit(void)
 {
 	runtime·stdcall(runtime·QueryPerformanceFrequency, 1, &timerfreq);
 	runtime·stdcall(runtime·SetConsoleCtrlHandler, 2, runtime·ctrlhandler, (uintptr)1);
+	runtime·destroylock = destroylock;
 }
 
 void
@@ -154,8 +156,8 @@ runtime·unlock(Lock *l)
 	eventunlock(l);
 }
 
-void
-runtime·destroylock(Lock *l)
+static void
+destroylock(Lock *l)
 {
 	if(l->event != 0)
 		runtime·stdcall(runtime·CloseHandle, 1, l->event);