gc: make merely referencing an outer variable in a closure not force heapallocation.

before: runtime_test.BenchmarkCallClosure1       20000000              135 ns/op
after:  runtime_test.BenchmarkCallClosure1      500000000                6 ns/op

R=rsc
CC=golang-dev
https://golang.org/cl/4527091

src/cmd/gc/closure.c

@@ -75,7 +75,7 @@ closurebody(NodeList *body)
 }
 
 void
-typecheckclosure(Node *func)
+typecheckclosure(Node *func, int top)
 {
 	Node *oldfn;
 	NodeList *l;
@@ -106,6 +106,10 @@ typecheckclosure(Node *func)
 			v->op = 0;
 			continue;
 		}
+		// For a closure that is called in place, but not
+		// inside a go statement, avoid moving variables to the heap.
+		if ((top & (Ecall|Eproc)) == Ecall)
+			v->heapaddr->etype = 1;
 		typecheck(&v->heapaddr, Erv);
 		func->enter = list(func->enter, v->heapaddr);
 		v->heapaddr = N;

src/cmd/gc/dcl.c

@@ -188,6 +188,7 @@ declare(Node *n, int ctxt)
 		else if(n->op == ONAME)
 			gen = ++vargen;
 		pushdcl(s);
+		n->curfn = curfn;
 	}
 	if(ctxt == PAUTO)
 		n->xoffset = BADWIDTH;

src/cmd/gc/go.h

@@ -254,6 +254,7 @@ struct	Node
 	Node*	ntype;
 	Node*	defn;
 	Node*	pack;	// real package for import . names
+	Node*	curfn;	// function for local variables
 
 	// ONAME func param with PHEAP
 	Node*	heapaddr;	// temp holding heap address of param
@@ -526,6 +527,7 @@ enum
 	Easgn = 1<<7,		// assigning to expression
 	Eindir = 1<<8,		// indirecting through expression
 	Eaddr = 1<<9,		// taking address of expression
+	Eproc = 1<<10,		// inside a go statement
 };
 
 #define	BITS	5
@@ -815,7 +817,7 @@ int	bset(Bits a, uint n);
  */
 Node*	closurebody(NodeList *body);
 void	closurehdr(Node *ntype);
-void	typecheckclosure(Node *func);
+void	typecheckclosure(Node *func, int top);
 Node*	walkclosure(Node *func, NodeList **init);
 void	walkcallclosure(Node *n, NodeList **init);
 

src/cmd/gc/typecheck.c

@@ -780,7 +780,7 @@ reswitch:
 			n = r;
 			goto reswitch;
 		}
-		typecheck(&n->left, Erv | Etype | Ecall);
+		typecheck(&n->left, Erv | Etype | Ecall |(top&Eproc));
 		l = n->left;
 		if(l->op == ONAME && l->etype != 0) {
 			if(n->isddd && l->etype != OAPPEND)
@@ -1217,7 +1217,7 @@ reswitch:
 
 	case OCLOSURE:
 		ok |= Erv;
-		typecheckclosure(n);
+		typecheckclosure(n, top);
 		if(n->type == T)
 			goto error;
 		goto ret;
@@ -1246,11 +1246,15 @@ reswitch:
 		goto ret;
 
 	case ODEFER:
-	case OPROC:
 		ok |= Etop;
 		typecheck(&n->left, Etop);
 		goto ret;
 
+	case OPROC:
+		ok |= Etop;
+		typecheck(&n->left, Etop|Eproc);
+		goto ret;
+
 	case OFOR:
 		ok |= Etop;
 		typechecklist(n->ninit, Etop);
@@ -2165,7 +2169,9 @@ addrescapes(Node *n)
 		if(n->noescape)
 			break;
 		switch(n->class) {
-		case PAUTO:
+		case PPARAMREF:
+			addrescapes(n->defn);
+			break;
 		case PPARAM:
 		case PPARAMOUT:
 			// if func param, need separate temporary
@@ -2173,7 +2179,7 @@ addrescapes(Node *n)
 			// the function type has already been checked
 			// (we're in the function body)
 			// so the param already has a valid xoffset.
-			if(n->class == PPARAM || n->class == PPARAMOUT) {
+
 			// expression to refer to stack copy
 			n->stackparam = nod(OPARAM, n, N);
 			n->stackparam->type = n->type;
@@ -2182,7 +2188,8 @@ addrescapes(Node *n)
 				fatal("addrescapes before param assignment");
 			n->stackparam->xoffset = n->xoffset;
 			n->xoffset = 0;
-			}
+			// fallthrough
+		case PAUTO:
 
 			n->class |= PHEAP;
 			n->addable = 0;
@@ -2195,7 +2202,9 @@ addrescapes(Node *n)
 			snprint(buf, sizeof buf, "&%S", n->sym);
 			n->heapaddr->sym = lookup(buf);
 			n->heapaddr->class = PHEAP-1;	// defer tempname to allocparams
-			curfn->dcl = list(curfn->dcl, n->heapaddr);
+			n->heapaddr->ullman = 1;
+			n->curfn->dcl = list(n->curfn->dcl, n->heapaddr);
+
 			break;
 		}
 		break;

src/pkg/runtime/closure_test.go

@@ -19,3 +19,35 @@ func BenchmarkCallClosure1(b *testing.B) {
 		s += func(ii int) int { return 2*ii + j }(i)
 	}
 }
+
+var ss *int
+
+func BenchmarkCallClosure2(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		j := i
+		s += func() int {
+			ss = &j
+			return 2
+		}()
+	}
+}
+
+func addr1(x int) *int {
+	return func() *int { return &x }()
+}
+
+func BenchmarkCallClosure3(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		ss = addr1(i)
+	}
+}
+
+func addr2() (x int, p *int) {
+	return 0, func() *int { return &x }()
+}
+
+func BenchmarkCallClosure4(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		_, ss = addr2()
+	}
+}