Packed bit fields in c structures - GCC

Question

I am working with structs in c on linux. I started using bit fields and the "packed" attribute and I came across a wierd behavior:

struct __attribute__((packed)) {
    int a:12;
    int b:32;
    int c:4;
} t1;

struct __attribute__((packed))  {
    int a:12;
    int b;
    int c:4;
}t2;

void main()
{
    printf("%d\n",sizeof(t1)); //output - 6
    printf("%d\n",sizeof(t2)); //output - 7
}

How come both structures - that are exactly the same - take diffrent number of bytes?

Solution 1

Your structures are not "exactly the same". Your first one has three consecutive bit-fields, the second has one bit-field, an (non bit-field) int, and then a second bit-field.

This is significant: consecutive (non-zero width) bit-fields are merged into a single memory location, while a bit-field followed by a non-bit-field are distinct memory locations.

Your first structure has a single memory location, your second has three. You can take the address of the b member in your second struct, not in your first. Accesses to the b member don't race with accesses the a or c in your second struct, but they do in your first.

Having a non-bit-field (or a zero-length bit-field) right after a bit-field member "closes" it in a sense, what follows will be a different/independent memory location/object. The compiler cannot "pack" your b member inside the bit-field like it does in the first struct.

Solution 2

struct t1 // 6 bytes
{
    int a:12; // 0:11
    int b:32; // 12:43
    int c:4;  // 44:47
}__attribute__((packed));

struct t1 // 7 bytes
{
    int a:12; // 0:11
    int b;    // 16:47
    int c:4;  // 48:51
}__attribute__((packed));

The regular int b must be aligned to a byte boundary. So there is padding before it. If you put c right next to a this padding will no longer be necessary. You should probably do this, as accessing non-byte-aligned integers like int b:32 is slow.

Comments

[Kerrek SB]

Because t2::b is guaranteed to be a distinct memory location? Think of data races.

[phuclv]

the result of sizeof must be printed using %zu

[Alexis]

as accessing non-byte-aligned integers like int b:32 is slow source? define slow. Doesn't it just shift and mask the memory value?

[John Zwinck]

It depends, but it will never be as fast as byte alignment. How slow (or even how we define slow) depends on the application, the compiler, etc.

[Alexis]

Understood, my guess is it's just a matter of fast shift+mask.

[Matthew M.]

It may just 'shift+mask' if the entire value fits into a single register. But there's a lot of maybes. Maybe it doesn't fit into a single register (platform-dependent) so it's 2 shifts, 2 masks, and an or between 2 registers. Maybe this structure is involved in an operation that could've been vectorized before but can't now (or can be but w/ unaligned loads). Lots of maybes.