Fix LDouble.to_int

`long double` must be casted to `intnat` first, not to `uintnat`
directly. Otherwise the conversion is undefined behavior for nearly
all negative values, see:
http://c0x.coding-guidelines.com/6.3.1.4.html , Note 50. The code was
correct until
ocaml/ocaml@24c118d

src/ctypes/ldouble_stubs.c

@@ -213,7 +213,9 @@ CAMLprim value ctypes_ldouble_of_int(value a) {
 }
 CAMLprim value ctypes_ldouble_to_int(value a) {
   CAMLparam1(a);
-  CAMLreturn(Val_long(ldouble_custom_val(a)));
+  long double b = ldouble_custom_val(a);
+  intnat c = b;
+  CAMLreturn(Val_long(c));
 }
 
 #define OP2(OPNAME, OP)                                                               \

tests/test-ldouble/test_ldouble.ml

@@ -116,6 +116,18 @@ let test_conv _ =
   List.iter (fun a -> assert_bool "to/of_float" (a = LDouble.(to_float (of_float a)))) flts;
   assert_bool "to_int" (3 = LDouble.(to_int (of_float 3.45)));
   assert_bool "to_int" (-34 = LDouble.(to_int (of_float (-34.999))));
+  if Sys.word_size = 32 then (
+    let mi = float_of_int min_int in
+    let ma = float_of_int max_int in
+    assert_bool "to_int" (min_int = LDouble.(to_int (of_float mi)));
+    assert_bool "to_int" (max_int = LDouble.(to_int (of_float ma)));
+  )
+  else (
+    let mi = -9007199254740991. in
+    let ma =  9007199254740991. in
+    assert_bool "to_int" (Int64.to_int (-9007199254740991L) = LDouble.(to_int (of_float mi)));
+    assert_bool "to_int" (Int64.to_int 9007199254740991L = LDouble.(to_int (of_float ma)));
+  );
   assert_bool "of_string" (3.5 = LDouble.(to_float (of_string "3.5")));
   assert_bool "to_string" ("3.500000" = LDouble.(to_string (of_float 3.5)))