gc: elide call to runtime.closure for function literals called in-place.

before:
runtime_test.BenchmarkCallClosure        5000000               499 ns/op
runtime_test.BenchmarkCallClosure1       5000000               681 ns/op

after:
runtime_test.BenchmarkCallClosure       500000000                5 ns/op
runtime_test.BenchmarkCallClosure1       10000000              160 ns/op

R=rsc
CC=golang-dev
https://golang.org/cl/4515167

src/cmd/gc/closure.c

@@ -199,3 +199,41 @@ walkclosure(Node *func, NodeList **init)
 	walkexpr(&call, init);
 	return call;
 }
+
+// Special case for closures that get called in place.
+// Optimize runtime.closure(X, __func__xxxx_, .... ) away
+// to __func__xxxx_(Y ....).
+// On entry, expect n->op == OCALL, n->left->op == OCLOSURE.
+void
+walkcallclosure(Node *n, NodeList **init)
+{
+	Node *z;
+	NodeList *ll, *cargs;
+
+	walkexpr(&n->left, init);
+	cargs =	n->left    // FUNC runtime.closure
+		->list     // arguments
+		->next     // skip first
+		->next;    // skip second
+
+	n->left = n->left  // FUNC runtime.closure
+		->list     // arguments
+		->next     // skip first
+		->n        // AS (to indreg) 
+		->right;   // argument  == the generated function 
+
+	// New arg list for n. First the closure-args, stolen from
+	// runtime.closure's 3rd and following,
+	ll = nil;
+	for (; cargs; cargs = cargs->next)
+		ll = list(ll, cargs->n->right);  // cargs->n is the OAS(INDREG, arg)
+
+	// then an extra zero, to fill the dummy return pointer slot,
+	z = nod(OXXX, N, N);
+	nodconst(z, types[TUINTPTR], 0);
+	z->typecheck = 1;
+	ll = list(ll, z);
+
+	// and finally the original parameter list.
+	n->list = concat(ll, n->list);
+}

src/cmd/gc/go.h

@@ -817,6 +817,7 @@ Node*	closurebody(NodeList *body);
 void	closurehdr(Node *ntype);
 void	typecheckclosure(Node *func);
 Node*	walkclosure(Node *func, NodeList **init);
+void	walkcallclosure(Node *n, NodeList **init);
 
 /*
  *	const.c

src/cmd/gc/walk.c

@@ -770,8 +770,15 @@ walkexpr(Node **np, NodeList **init)
 		t = n->left->type;
 		if(n->list && n->list->n->op == OAS)
 			goto ret;
+
+		if(n->left->op == OCLOSURE) {
+			walkcallclosure(n, init);
+			t = n->left->type;
+		} else
 			walkexpr(&n->left, init);
+
 		walkexprlist(n->list, init);
+
 		ll = ascompatte(n->op, n->isddd, getinarg(t), n->list, 0, init);
 		n->list = reorder1(ll);
 		if(isselect(n)) {

src/pkg/runtime/closure_test.go

+// Copyright 2011 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+package runtime_test
+
+import "testing"
+
+var s int
+
+func BenchmarkCallClosure(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		s += func(ii int) int { return 2 * ii }(i)
+	}
+}
+
+func BenchmarkCallClosure1(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		j := i
+		s += func(ii int) int { return 2*ii + j }(i)
+	}
+}