What is VC++ doing when packing bitfields?

Question

To clarify my question, let's start off with an example program:

#include <stdio.h>

#pragma pack(push,1)
struct cc {
    unsigned int a   :  3;  
    unsigned int b   : 16;
    unsigned int c   :  1;
    unsigned int d   :  1;
    unsigned int e   :  1;
    unsigned int f   :  1;
    unsigned int g   :  1;
    unsigned int h   :  1;
    unsigned int i   :  6;  
    unsigned int j   :  6;  
    unsigned int k   :  4;  
    unsigned int l   : 15;
};
#pragma pack(pop)

struct cc c;

int main(int argc, char **argv)

{   printf("%d\n",sizeof(c));
}

The output is "8", meaning that the 56 bits (7 bytes) I want to pack are being packed into 8 bytes, seemingly wasting a whole byte. Curious about how the compiler was laying these bits out in memory, I tried writing specific values to &c, e.g.:

int main(int argc, char **argv)

{
unsigned long long int* pint = &c;
*pint = 0xFFFFFFFF;
printf("c.a = %d", c.a);
...
printf("c.l = %d", c.l);
}

Predictably, on x86_64 using Visual Studio 2010, the following happens:

*pint = 0x00000000 000000FF :

c[0].a = 7
c[0].b = 1
c[0].c = 1
c[0].d = 1
c[0].e = 1
c[0].f = 1
c[0].g = 0
c[0].h = 0
c[0].i = 0
c[0].j = 0
c[0].k = 0
c[0].l = 0

*pint = 0x00000000 0000FF00 :

c[0].a = 0
c[0].b = 0
c[0].c = 0
c[0].d = 0
c[0].e = 0
c[0].f = 0
c[0].g = 1
c[0].h = 127
c[0].i = 0
c[0].j = 0
c[0].k = 0
c[0].l = 0


*pint = 0x00000000 00FF0000 :

c[0].a = 0
c[0].b = 0
c[0].c = 0
c[0].d = 0
c[0].e = 0
c[0].f = 0
c[0].g = 0
c[0].h = 32640
c[0].i = 0
c[0].j = 0
c[0].k = 0
c[0].l = 0

etc.

Forget portability for a moment and assume you care about one CPU, one compiler, and one runtime environment. Why can't VC++ pack this structure into 7 bytes? Is it a word-length thing? The MSDN docs on #pragma pack says "the alignment of a member will be on a boundary that is either a multiple of n [1 in my case] or a multiple of the size of the member, whichever is smaller." Can anyone give me some idea of why I get a sizeof 8 and not 7?

Answer 1

MSVC++ always allocates at least a unit of memory that corresponds to the type you used for your bit-field. You used unsigned int, meaning that a unsigned int is allocated initially, and another unsigned int is allocated when the first one is exhausted. There's no way to force MSVC++ to trim the unused portion of the second unsigned int.

Basically, MSVC++ interprets your unsigned int as a way to express the alignment requirements for the entire structure.

Use smaller types for your bit-fields (unsigned short and unsigned char) and regroup the bit-fields so that they fill the allocated unit entirely - that way you should be able to pack things as tightly as possible.

Answer 2

Bitfields are stored in the type that you define. Since you are using unsigned int, and it won't fit in a single unsigned int then the compiler must use a second integer and store the last 24 bits in that last integer.

Answer 3

Well you are using unsigned int which happens to be 32 Bit in this case. The next boundary (to fit in the bitfield) for unsigned int is 64 Bit => 8 Bytes.

Answer 4

pst is right. The members are aligned on 1-byte boundaries, (or smaller, since it's a bitfield). The overall structure has size 8, and is aligned on an 8-byte boundary. This complies with both the standard and the pack option. The docs never say there will be no padding at the end.

Answer 5

To give another interesting illustrates what's going on, consider the case where you want to pack a structure that crosses a type boundary. E.g.

struct state {
    unsigned int cost     : 24; 
    unsigned int back     : 21; 
    unsigned int a        :  1; 
    unsigned int b        :  1; 
    unsigned int c        :  1;
};

This structure can't be packed into 6 bytes using MSVC as far as I know. However, we can get the desired packing effect by breaking up the first two fields:

struct state_packed {
    unsigned short cost_1   : 16; 
    unsigned char  cost_2   :  8;
    unsigned short back_1   : 16; 
    unsigned char  back_2   :  5;
    unsigned char  a        :  1; 
    unsigned char  b        :  1; 
    unsigned char  c        :  1; 
};

This can indeed be packed into 6 bytes. However, accessing the original cost field is extremely awkward and ugly. One method is to cast a state_packed pointer to a specialized dummy struct:

struct state_cost {
    unsigned int cost     : 24;
    unsigned int junk     :  8; 
};

state_packed    sc;
state_packed *p_sc = ≻

sc.a = 1;
(*(struct state_cost *)p_sc).cost = 12345;
sc.b = 1;

If anyone knows a more elegant way of doing this, I would love to know!