When optimizing a struct that is packed, the compiler tries to deduce the alignment of each field, to improve access. However, it is not always possible for the compiler to deduce the alignment of each field in a __packed struct. In contrast, when individual fields in a struct are declared as __packed, fast access is guaranteed to naturally aligned members within the struct. Therefore, when the use of a packed structure is required, it is recommended that you always pack individual fields of the structure, rather than the entire structure itself.

The differences between not packing a struct, packing an entire struct, and packing individual fields of a struct are illustrated by the three implementations of a struct shown in Table 4.9.

In the first implementation, the struct is not packed. In the second implementation, the entire structure mystruct is qualified as __packed. In the third implementation, the __packed attribute is removed from the mystruct structure, and individual non-aligned fields are declared as __packed.

struct foo
{
    char one;
    short two;
    char three;
    int four;
} c;

__packed struct foo
{
    char one;
    short two;
    char three;
    int four;
} c;

struct foo
{
    char one;
    __packed short two;
    char three;
    int four;
} c;

Table 4.10 shows the corresponding disassembly of the machine code produced by the compiler for each of the sample implementations of Table 4.9, where the C code for each implementation has been compiled using the option -O2.

; r0 contains address of c

LDRB    r1, [r0, #0]
LDRSH   r2, [r0, #2]
LDRB    r3, [r0, #4]
LDR     r12, [r0, #8]

; r0 contains address of c

; char one
LDRB  r1, [r0, #0]

; short two
LDRB  r2, [r0, #1]
LDRSB r12, [r0, #2]
ORR   r2, r12, r2, LSL #8

; char three
LDRB  r3, [r0, #3]

; int four
ADD   r0, r0, #4
BL    __aeabi_uread4

; r0 contains address of c

; char one
LDRB  r1, [r0, #0]

; short two
LDRB  r2, [r0, #1]
LDRSB r12, [r0, #2]
ORR   r2, r12, r2, LSL #8

; char three
LDRB  r3, [r0, #3]

; int four
LDR   r12, [r0, #4]

In the disassembly of the unpacked struct in Table 4.10, the compiler always accesses data on aligned word or halfword addresses. The compiler is able to do this because the struct is padded so that every member of the struct lies on its natural size boundary.

In the disassembly of the __packed struct in Table 4.10, the fields one and three are aligned on their natural size boundaries by default, and so the compiler makes aligned accesses. The compiler always carries out aligned word or halfword accesses for fields it can identify are aligned. For the unaligned field two, the compiler uses multiple aligned memory accesses (LDR/STR/LDM/STM), combined with fixed shifting and masking, to access the correct bytes in memory. The compiler calls the AEABI runtime routine __aeabi_uread4 for reading an unsigned word at an unknown alignment to access the field four, because it is not able to determine that the field lies on its natural size boundary.

In the disassembly of the struct with individually packed fields in Table 4.10, the fields one, two, and three are accessed just as they are in the case where the entire struct is qualified as __packed. In contrast to the situation where the entire struct is packed, however, the compiler makes a word-aligned access to the field four, because the presence of the __packed short within the structure helps the compiler to determine that the field four lies on its natural size boundary.