Re: bug in GCC or C++ standard ?

[Lawrence Crowl <crowl at google dot com>]

On 11/12/08, James Dennett <james.dennett@gmail.com> wrote:
> On Wed, Nov 12, 2008 at 11:29 AM, Joe Buck <Joe.Buck@synopsys.com> wrote:
>  > On Tue, Nov 11, 2008 at 08:43:40PM -0800, James Dennett wrote:
>  >> (There are secondary uses of unions for type punning.  Most such uses
>  >> are not valid portable C++, but g++ supports them because they're so
>  >> common in real code.)
>  >
>  > On the contrary: the uses of unions for type-punning, while not portable,
>  > are valid C++, and while not portable, the behavior is
>  > implementation-defined (because it exposes the bitwise implementation of C
>  > and C++ types).
>
>
> Interesting that you call them "valid C++" when the standard leaves
>  them as undefined behavior (not just implementation-defined).  As I
>  implied, there are some permitted uses of type punning in unions, but
>  they're quite restricted.
>
>  Let's be concrete:
>
>  union {
>   int i;
>   double d;
>  } u;
>
>  u.i = 1;
>  u.d = 2.0;
>
>  printf("%d\n", u.i);
>
>  AIUI, the printf() statement above has undefined behavior, and might
>  quite possibly print 1, if the compiler uses the knowledge that reads
>  from u.i cannot be affected (in code with defined behavior) by writes
>  to u.d.  Indeed, the write to u.d above might reasonably be eliminated
>  by the optimizer if u.d is never read subsequently.
>
>  Placing most types inside a union doesn't make it any more valid to
>  alias them than if reinterpret_cast<int*>(&d) were used on a double d.

Now would be a good time to mention Google's bit_cast<>.

>
>
>  > Without this feature, it would not be possible to
>  > implement a function like ldexp or frexp in C or C++, where we want
>  > to consider the same storage either as a double or as a sequence of
>  > bytes.
>
>
> It's possible to do that without using unions; pointers to character
>  types are explicitly permitted to alias any storage.
>
>
>  -- James
>


-- 
Lawrence Crowl