c - Is gcc's __attribute__((packed)) / #pragma pack unsafe?

In C, the compiler will lay out members of a struct in the order in which they're declared, with possible padding bytes inserted between members, or after the last member, to ensure that each member is aligned properly.

gcc provides a language extension, __attribute__((packed)), which tells the compiler not to insert padding, allowing struct members to be misaligned. For example, if the system normally requires all int objects to have 4-byte alignment, __attribute__((packed)) can cause int struct members to be allocated at odd offsets.

Quoting the gcc documentation:

The `packed' attribute specifies that a variable or structure field
should have the smallest possible alignment--one byte for a variable,
and one bit for a field, unless you specify a larger value with the
`aligned' attribute.

Obviously the use of this extension can result in smaller data requirements but slower code, as the compiler must (on some platforms) generate code to access a misaligned member a byte at a time.

But are there any cases where this is unsafe? Does the compiler always generate correct (though slower) code to access misaligned members of packed structs? Is it even possible for it to do so in all cases?

Answer

Yes, __attribute__((packed)) is potentially unsafe on some systems. The symptom probably won't show up on an x86, which just makes the problem more insidious; testing on x86 systems won't reveal the problem. (On the x86, misaligned accesses are handled in hardware; if you dereference an int* pointer that points to an odd address, it will be a little slower than if it were properly aligned, but you'll get the correct result.)

On some other systems, such as SPARC, attempting to access a misaligned int object causes a bus error, crashing the program.

There have also been systems where a misaligned access quietly ignores the low-order bits of the address, causing it to access the wrong chunk of memory.

Consider the following program:

#include 
#include 
int main(void)
{
    struct foo {
        char c;
        int x;
    } __attribute__((packed));
    struct foo arr[2] = { { 'a', 10 }, {'b', 20 } };

    int *p0 = &arr[0].x;
    int *p1 = &arr[1].x;
    printf("sizeof(struct foo)      = %d\n", (int)sizeof(struct foo));
    printf("offsetof(struct foo, c) = %d\n", (int)offsetof(struct foo, c));
    printf("offsetof(struct foo, x) = %d\n", (int)offsetof(struct foo, x));
    printf("arr[0].x = %d\n", arr[0].x);
    printf("arr[1].x = %d\n", arr[1].x);
    printf("p0 = %p\n", (void*)p0);
    printf("p1 = %p\n", (void*)p1);
    printf("*p0 = %d\n", *p0);

    printf("*p1 = %d\n", *p1);
    return 0;
}

On x86 Ubuntu with gcc 4.5.2, it produces the following output:

sizeof(struct foo)      = 5
offsetof(struct foo, c) = 0
offsetof(struct foo, x) = 1

arr[0].x = 10
arr[1].x = 20
p0 = 0xbffc104f
p1 = 0xbffc1054
*p0 = 10
*p1 = 20

On SPARC Solaris 9 with gcc 4.5.1, it produces the following:

sizeof(struct foo)      = 5
offsetof(struct foo, c) = 0
offsetof(struct foo, x) = 1
arr[0].x = 10
arr[1].x = 20
p0 = ffbff317
p1 = ffbff31c
Bus error

In both cases, the program is compiled with no extra options, just gcc packed.c -o packed.

(A program that uses a single struct rather than array doesn't reliably exhibit the problem, since the compiler can allocate the struct on an odd address so the x member is properly aligned. With an array of two struct foo objects, at least one or the other will have a misaligned x member.)

(In this case, p0 points to a misaligned address, because it points to a packed int member following a char member. p1 happens to be correctly aligned, since it points to the same member in the second element of the array, so there are two char objects preceding it -- and on SPARC Solaris the array arr appears to be allocated at an address that is even, but not a multiple of 4.)

When referring to the member x of a struct foo by name, the compiler knows that x is potentially misaligned, and will generate additional code to access it correctly.

Once the address of arr[0].x or arr[1].x has been stored in a pointer object, neither the compiler nor the running program knows that it points to a misaligned int object. It just assumes that it's properly aligned, resulting (on some systems) in a bus error or similar other failure.

Fixing this in gcc would, I believe, be impractical. A general solution would require, for each attempt to dereference a pointer to any type with non-trivial alignment requirements either (a) proving at compile time that the pointer doesn't point to a misaligned member of a packed struct, or (b) generating bulkier and slower code that can handle either aligned or misaligned objects.

I've submitted a gcc bug report. As I said, I don't believe it's practical to fix it, but the documentation should mention it (it currently doesn't).

UPDATE: As of 2018-12-20, this bug is marked as FIXED. The patch will appear in gcc 9 with the addition of a new -Waddress-of-packed-member option, enabled by default.

When address of packed member of struct or union is taken, it may
result in an unaligned pointer value. This patch adds
-Waddress-of-packed-member to check alignment at pointer assignment and warn unaligned address as well as unaligned pointer

I've just built that version of gcc from source. For the above program, it produces these diagnostics:

c.c: In function ‘main’:
c.c:10:15: warning: taking address of packed member of ‘struct foo’ may result in an unaligned pointer value [-Waddress-of-packed-member]
   10 |     int *p0 = &arr[0].x;
      |               ^~~~~~~~~
c.c:11:15: warning: taking address of packed member of ‘struct foo’ may result in an unaligned pointer value [-Waddress-of-packed-member]
   11 |     int *p1 = &arr[1].x;

      |               ^~~~~~~~~