On Thu, 17 Aug 2023, Tony Josi via Gcc-help wrote:
> Thanks a lot Alexander for the detailed explanation
>
> > Worse, the attribute is documented
> > to reduce struct alignment down to one byte, and 'ch1' has the same
> > alignment as the whole struct.
>
> This is indeed new information to me, I always thought packed structs had
> alignment of 1 byte.
That's correct.
> And non packed (unpacked) structs have alignment equal to the largest
> alignment of its fields.(I believe my understanding in the case of non
> packed (unpacked) is correct)
Also correct.
> So in this case *struct Struct_test *has an alignment of 10 bytes
No, it has size 10, alignment 1.
> as it's packed, causing it to place *ch1* on a 10 byte aligned address which
No, ch1 has alignment 1 (non-power-of-two alignment doesn't make sense)
> causes inefficient alignment for *ch1*. Similarly for *ch2* but not for *ch3*
> as it falls on (10 + 4 + 2 = 16) 4 byte aligned address (its natural
> alignment) hence no warning for *ch3.*
>
> Is my understanding correct?
The warning is (quite cryptically) telling you about the fields that remain
aligned relative to the struct (which itself may be at an odd address).
Consider
struct __attribute__((packed)) Struct_test
{
uint32_t ch1;
uint16_t ch2;
uint32_t ch3;
uint16_t ch4;
uint32_t ch5;
};
All fields except ch3 are diagnosed.
Alexander
