long bit-fields with g++ 4.4.1

long bit-fields with g++ 4.4.1

[Vineet Soni]

The integral promotion behavior of long bit-fields has changed with g++ 4.4.1 on
x86_64-linux, and appears broken.  Consider:

#include <stdio.h>
#define T unsigned long
int main()
{
  struct { T f : 33; } s = { 1UL << 32 };
  printf("%lx %lx %lx\n", (T)s.f, (T)(s.f << 1), (T)((s.f << 1) >> 1));

  struct { T f : 16; } t = { 1UL << 15 };
  printf("%lx %lx %lx\n", (T)t.f, (T)(t.f << 17), (T)((t.f << 17) >> 17));

  return 0;
}

Based on my interpretation of the C++ standard, the expected output is:

    100000000 200000000 100000000
    8000 100000000 8000

The actual output is:

    $ g++-4.4.1 -Wall -pedantic x.c; ./a.out
    100000000 200000000 100000000
    8000 0 0

Is this a bug?

Note that 4.0.2 matches the expected output:

    $ g++-4.0.2 -Wall -pedantic x.c; ./a.out
    100000000 200000000 100000000
    8000 100000000 8000

GCC versions used:

$ gcc-4.0.2 -v
Using built-in specs.
Target: x86_64-unknown-linux-gnu
Configured with: ../../src/gcc-4.0.2/configure
--prefix=/home/global/Linux-x86_64-rhel3/encap/gcc-4.0.2 --disable-libgcj
--enable-languages=c,c++
Thread model: posix
gcc version 4.0.2

$ gcc-4.4.1 -v
Using built-in specs.
Target: x86_64-unknown-linux-gnu
Configured with: ./configure
--prefix=/home/global/Linux-x86_64-rhel5/encap/gcc-4.4.1 --disable-libgcj
--enable-languages=c,c++
--with-mpfr=/home/global/Linux-x86_64-rhel5/encap/mpfr-2.4.1
Thread model: posix
gcc version 4.4.1 (GCC)


-- Vineet

Re: long bit-fields with g++ 4.4.1

[Cedric Roux]

Vineet Soni wrote:
> The integral promotion behavior of long bit-fields has changed with g++ 4.4.1 on
> x86_64-linux, and appears broken.  Consider:
> 
> #include <stdio.h>
> #define T unsigned long
> int main()
> {
>   struct { T f : 33; } s = { 1UL << 32 };
>   printf("%lx %lx %lx\n", (T)s.f, (T)(s.f << 1), (T)((s.f << 1) >> 1));
> 
>   struct { T f : 16; } t = { 1UL << 15 };
>   printf("%lx %lx %lx\n", (T)t.f, (T)(t.f << 17), (T)((t.f << 17) >> 17));
> 
>   return 0;
> }
> 
> Based on my interpretation of the C++ standard, the expected output is:
> 
>     100000000 200000000 100000000
>     8000 100000000 8000
> 
> The actual output is:
> 
>     $ g++-4.4.1 -Wall -pedantic x.c; ./a.out
>     100000000 200000000 100000000
>     8000 0 0
> 
> Is this a bug?
> 
> Note that 4.0.2 matches the expected output:
> 
>     $ g++-4.0.2 -Wall -pedantic x.c; ./a.out
>     100000000 200000000 100000000
>     8000 100000000 8000

If that is what you refer to:
3 An  rvalue for an integral bit-field (_class.bit_) can be converted to
  an rvalue of type int if int can represent all the values of the  bit-
  field;  otherwise, it can be converted to unsigned int if unsigned int
  can represent all the values of the bit-field.  If  the  bit-field  is
  larger yet, no integral promotion applies to it.  If the bit-field has
  an enumerated type, it is treated as any other value of that type  for
  promotion purposes.
(found on the internet, I don't have the standard)

we can understand that "t" may be promoted to "int" (which is 32b) because
it occupies 16 bits. And the << arithmetic operator asks for the
promotion.

Maybe that's why you have 0.
Maybe that's the way it has to be and gcc 4.0.2 was wrong.

Note that doing:
   printf("%lx %lx %lx\n", (T)t.f, (T)((unsigned long)t.f << 17), (T)(((unsigned long)t.f << 17) >> 17));
prints what you wait (not 0)
(with a g++ 4.3.2, not even 4.4.1)

(Note that I am not involved in g++ or gcc development nor
C++ standardization, so all what I say may be totally wrong.)

Re: long bit-fields with g++ 4.4.1

[Vineet Soni]

Cedric Roux wrote:
> 
> If that is what you refer to:
> 3 An  rvalue for an integral bit-field (_class.bit_) can be converted to
>  an rvalue of type int if int can represent all the values of the  bit-
>  field;  otherwise, it can be converted to unsigned int if unsigned int
>  can represent all the values of the bit-field.  If  the  bit-field  is
>  larger yet, no integral promotion applies to it.  If the bit-field has
>  an enumerated type, it is treated as any other value of that type  for
>  promotion purposes.
> (found on the internet, I don't have the standard)
> 
> we can understand that "t" may be promoted to "int" (which is 32b) because
> it occupies 16 bits. And the << arithmetic operator asks for the
> promotion.
> 
> Maybe that's why you have 0.
> Maybe that's the way it has to be and gcc 4.0.2 was wrong.

Your citation applies, but my interpretation differs:

In C++ a bit-field has the declared type; the number of bits is not part of its
type [1][2].  The C++ integer promotion rules apply [your citation above]; in
particular the clause: "If the bit-field is larger yet, no integral promotion
applies to it".  A bit-field expression of declared type long is unchanged by
integer promotions, since (the number of bits in the bit-field being irrelevant)
long is larger than int and unsigned int.  So the expected output is:

    100000000 200000000 100000000
    8000 100000000 8000

[1] ISO/IEC 14882-1998 [class.bit] The bit-field attribute is not part of the
type of the class member.

[2] C++ Standard Core Language Issue #303
(http://www.open-std.org/JTC1/SC22/WG21/docs/cwg_closed.html#303)

> 
> Note that doing:
>   printf("%lx %lx %lx\n", (T)t.f, (T)((unsigned long)t.f << 17),
> (T)(((unsigned long)t.f << 17) >> 17));
> prints what you wait (not 0)
> (with a g++ 4.3.2, not even 4.4.1)

Yes, we can cast, but the question is what is the correct behavior in the absence of
a cast.  I don't want to find all the places in existing code where a cast might
be necessary, if 4.4.1 behavior is a bug.

-- Vineet

Re: long bit-fields with g++ 4.4.1

[Ian Lance Taylor]

Vineet Soni <vsoni@tilera.com> writes:

> The integral promotion behavior of long bit-fields has changed with g++ 4.4.1 on
> x86_64-linux, and appears broken.

That looks broken to me too.  Please file a bug report following the
directions at http://gcc.gnu.org/bugs/ .  Thanks.

Ian

Re: long bit-fields with g++ 4.4.1

[Vineet Soni]

Ian Lance Taylor wrote:
> That looks broken to me too.  Please file a bug report following the
> directions at http://gcc.gnu.org/bugs/ .  Thanks.

Filed as:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43393

Thanks.

Vineet