diff --git a/src/cmd/gc/builtin.c b/src/cmd/gc/builtin.c
index 197255be4e..d993bc7fc8 100644
--- a/src/cmd/gc/builtin.c
+++ b/src/cmd/gc/builtin.c
@@ -102,7 +102,7 @@ char *runtimeimport =
 	"func @\"\".int64tofloat64(? int64) (? float64)\n"
 	"func @\"\".uint64tofloat64(? uint64) (? float64)\n"
 	"func @\"\".complex128div(@\"\".num complex128, @\"\".den complex128) (@\"\".quo complex128)\n"
-	"func @\"\".racefuncenter()\n"
+	"func @\"\".racefuncenter(? uintptr)\n"
 	"func @\"\".racefuncexit()\n"
 	"func @\"\".raceread(? uintptr)\n"
 	"func @\"\".racewrite(? uintptr)\n"
diff --git a/src/cmd/gc/racewalk.c b/src/cmd/gc/racewalk.c
index 3e0feffd5a..034a5a5a34 100644
--- a/src/cmd/gc/racewalk.c
+++ b/src/cmd/gc/racewalk.c
@@ -33,6 +33,7 @@ racewalk(Node *fn)
 {
 	int i;
 	Node *nd;
+	Node *nodpc;
 	char s[1024];
 
 	if(myimportpath) {
@@ -42,10 +43,14 @@ racewalk(Node *fn)
 		}
 	}
 
-	// TODO(dvyukov): ideally this should be:
-	// racefuncenter(getreturnaddress())
-	// because it's much more costly to obtain from runtime library.
-	nd = mkcall("racefuncenter", T, nil);
+	// nodpc is the PC of the caller as extracted by
+	// getcallerpc. We use -widthptr(FP) for x86.
+	// BUG: this will not work on arm.
+	nodpc = nod(OXXX, nil, nil);
+	*nodpc = *nodfp;
+	nodpc->type = types[TUINTPTR];
+	nodpc->xoffset = -widthptr;
+	nd = mkcall("racefuncenter", T, nil, nodpc);
 	fn->enter = list(fn->enter, nd);
 	nd = mkcall("racefuncexit", T, nil);
 	fn->exit = list(fn->exit, nd); // works fine if (!fn->exit)
diff --git a/src/cmd/gc/runtime.go b/src/cmd/gc/runtime.go
index cfabbb174b..b8204ebcf3 100644
--- a/src/cmd/gc/runtime.go
+++ b/src/cmd/gc/runtime.go
@@ -140,7 +140,7 @@ func uint64tofloat64(uint64) float64
 func complex128div(num complex128, den complex128) (quo complex128)
 
 // race detection
-func racefuncenter()
+func racefuncenter(uintptr)
 func racefuncexit()
 func raceread(uintptr)
 func racewrite(uintptr)
diff --git a/src/pkg/runtime/race.c b/src/pkg/runtime/race.c
index 97bfe6864e..bea16cc832 100644
--- a/src/pkg/runtime/race.c
+++ b/src/pkg/runtime/race.c
@@ -70,11 +70,13 @@ runtime·raceread(uintptr addr)
 
 // Called from instrumented code.
 void
-runtime·racefuncenter(void)
+runtime·racefuncenter(uintptr pc)
 {
-	uintptr pc;
+	// If the caller PC is lessstack, use slower runtime·callers
+	// to walk across the stack split to find the real caller.
+	if(pc == (uintptr)runtime·lessstack)
+		runtime·callers(2, &pc, 1);
 
-	runtime·callers(2, &pc, 1);
 	m->racecall = true;
 	runtime∕race·FuncEnter(g->goid-1, (void*)pc);
 	m->racecall = false;