(at least for GNUC platforms) don't rely on nasty division-by-zero tricks to generate +-inf and nan. instead borrow the bitlevel twiddling javascript uses. I would imagine that this would work for all platforms that use IEEE floats (which is necessarily all platforms for which EF will work), but someone else needs to figure out if windows will work with this code.

ef/Utilities/General/FloatUtils.cpp

@@ -79,29 +79,37 @@ union dpun {
     Flt64 d;
 };
 
+union spun {
+  uint32 s;
+  Flt32 f;
+};
+
 struct DummyInit
 {
   DummyInit() {
-    union dpun u;
+    union dpun du;
+    union spun su;
 
-#if notyet    
-    floatPositiveInfinity = 1.0f/fZero;
-    floatNegativeInfinity = -1.0f/fZero;
-    floatNaN = fZero/fZero;
-#endif
+    su.s = 0x7f800000;
+    floatPositiveInfinity = su.f;
     
-	u.s.hi = DOUBLE_HI32_EXPMASK;
-	u.s.lo = 0x00000000;
-	doublePositiveInfinity = u.d;
+	du.s.hi = DOUBLE_HI32_EXPMASK;
+	du.s.lo = 0x00000000;
+	doublePositiveInfinity = du.d;
 
-	u.s.hi = DOUBLE_HI32_SIGNBIT | DOUBLE_HI32_EXPMASK;
-	u.s.lo = 0x00000000;
-	doubleNegativeInfinity = u.d;
+    su.s = 0xff800000;
+    floatNegativeInfinity = su.f;
     
+	du.s.hi = DOUBLE_HI32_SIGNBIT | DOUBLE_HI32_EXPMASK;
+	du.s.lo = 0x00000000;
+	doubleNegativeInfinity = du.d;
 
-	u.s.hi = DOUBLE_HI32_EXPMASK | DOUBLE_HI32_MANTMASK;
-	u.s.lo = 0xffffffff;
-	doubleNaN = u.d;
+    su.s = 0x7f900000;
+    floatNaN = su.f;
+
+	du.s.hi = DOUBLE_HI32_EXPMASK | DOUBLE_HI32_MANTMASK;
+	du.s.lo = 0xffffffff;
+	doubleNaN = du.d;
   }
 };