math, runtime: use a NaN that matches gcc's

our old choice is not working properly at least on VFPv2 in
ARM1136JF-S (it's not preserved across float64->float32 conversions).

        Fixes #3745.

R=dave, rsc
CC=golang-dev
https://golang.org/cl/6344078

src/pkg/math/all_test.go

@@ -1693,6 +1693,17 @@ func alike(a, b float64) bool {
 	return false
 }
 
+func TestNaN(t *testing.T) {
+	f64 := NaN()
+	if f64 == f64 {
+		t.Fatalf("NaN() returns %g, expected NaN", f64)
+	}
+	f32 := float32(f64)
+	if f32 == f32 {
+		t.Fatalf("float32(NaN()) is %g, expected NaN", f32)
+	}
+}
+
 func TestAcos(t *testing.T) {
 	for i := 0; i < len(vf); i++ {
 		a := vf[i] / 10

src/pkg/math/bits.go

@@ -5,7 +5,7 @@
 package math
 
 const (
-	uvnan    = 0x7FF0000000000001
+	uvnan    = 0x7FF8000000000001
 	uvinf    = 0x7FF0000000000000
 	uvneginf = 0xFFF0000000000000
 	mask     = 0x7FF

src/pkg/runtime/float.c

@@ -4,7 +4,7 @@
 
 #include "runtime.h"
 
-static	uint64	uvnan		= 0x7FF0000000000001ULL;
+static	uint64	uvnan		= 0x7FF8000000000001ULL;
 static	uint64	uvinf		= 0x7FF0000000000000ULL;
 static	uint64	uvneginf	= 0xFFF0000000000000ULL;