cmd/gc: don't copy string in range []byte(str)

Using benchmark from the issue:

benchmark                    old ns/op     new ns/op     delta
BenchmarkRangeStringCast     2162          1152          -46.72%

benchmark                    old allocs     new allocs     delta
BenchmarkRangeStringCast     1              0              -100.00%

Fixes #2204

Change-Id: I92c5edd2adca4a7b6fba00713a581bf49dc59afe
Reviewed-on: https://go-review.googlesource.com/3790Reviewed-by: Keith Randall <khr@golang.org>

src/cmd/gc/builtin.c

@@ -38,6 +38,7 @@ char *runtimeimport =
 	"func @\"\".slicebytetostringtmp (? []byte) (? string)\n"
 	"func @\"\".slicerunetostring (? []rune) (? string)\n"
 	"func @\"\".stringtoslicebyte (? string) (? []byte)\n"
+	"func @\"\".stringtoslicebytetmp (? string) (? []byte)\n"
 	"func @\"\".stringtoslicerune (? string) (? []rune)\n"
 	"func @\"\".stringiter (? string, ? int) (? int)\n"
 	"func @\"\".stringiter2 (? string, ? int) (@\"\".retk·1 int, @\"\".retv·2 rune)\n"

src/cmd/gc/go.h

@@ -467,6 +467,7 @@ enum
 	OARRAYBYTESTRTMP, // string(bytes) ephemeral
 	OARRAYRUNESTR,	// string(runes)
 	OSTRARRAYBYTE,	// []byte(s)
+	OSTRARRAYBYTETMP,	// []byte(s) ephemeral
 	OSTRARRAYRUNE,	// []rune(s)
 	OAS,	// x = y or x := y
 	OAS2,	// x, y, z = xx, yy, zz

src/cmd/gc/order.c

@@ -757,6 +757,10 @@ orderstmt(Node *n, Order *order)
 		default:
 			fatal("orderstmt range %T", n->type);
 		case TARRAY:
+			// Mark []byte(str) range expression to reuse string backing storage.
+			// It is safe because the storage cannot be mutated.
+			if(n->right->op == OSTRARRAYBYTE)
+				n->right->op = OSTRARRAYBYTETMP;
 			if(count(n->list) < 2 || isblank(n->list->next->n)) {
 				// for i := range x will only use x once, to compute len(x).
 				// No need to copy it.

src/cmd/gc/runtime.go

@@ -52,6 +52,7 @@ func slicebytetostring(*[32]byte, []byte) string
 func slicebytetostringtmp([]byte) string
 func slicerunetostring([]rune) string
 func stringtoslicebyte(string) []byte
+func stringtoslicebytetmp(string) []byte
 func stringtoslicerune(string) []rune
 func stringiter(string, int) int
 func stringiter2(string, int) (retk int, retv rune)

src/cmd/gc/walk.c

@@ -1406,6 +1406,11 @@ walkexpr(Node **np, NodeList **init)
 		n = mkcall("stringtoslicebyte", n->type, init, conv(n->left, types[TSTRING]));
 		goto ret;
 
+	case OSTRARRAYBYTETMP:
+		// stringtoslicebytetmp(string) []byte;
+		n = mkcall("stringtoslicebytetmp", n->type, init, conv(n->left, types[TSTRING]));
+		goto ret;
+
 	case OSTRARRAYRUNE:
 		// stringtoslicerune(string) []rune
 		n = mkcall("stringtoslicerune", n->type, init, n->left);

src/runtime/string.go

@@ -135,6 +135,18 @@ func stringtoslicebyte(s string) []byte {
 	return b
 }
 
+func stringtoslicebytetmp(s string) []byte {
+	// Return a slice referring to the actual string bytes.
+	// This is only for use by internal compiler optimizations
+	// that know that the slice won't be mutated.
+	// The only such case today is:
+	// for i, c := range []byte(str)
+
+	str := (*stringStruct)(unsafe.Pointer(&s))
+	ret := slice{array: (*byte)(str.str), len: uint(str.len), cap: uint(str.len)}
+	return *(*[]byte)(unsafe.Pointer(&ret))
+}
+
 func stringtoslicerune(s string) []rune {
 	// two passes.
 	// unlike slicerunetostring, no race because strings are immutable.

src/runtime/string_test.go

@@ -221,3 +221,17 @@ func TestIntStringAllocs(t *testing.T) {
 		t.Fatalf("want 0 allocs, got %v", n)
 	}
 }
+
+func TestRangeStringCast(t *testing.T) {
+	s := "abc"
+	n := testing.AllocsPerRun(1000, func() {
+		for i, c := range []byte(s) {
+			if c != s[i] {
+				t.Fatalf("want '%c' at pos %v, got '%c'", s[i], i, c)
+			}
+		}
+	})
+	if n != 0 {
+		t.Fatalf("want 0 allocs, got %v", n)
+	}
+}