Re: [lwip-devel] Curious struct packing issue - is it GCC?

[address@hidden]

Bill Auerbach wrote:
> I’m trying to inline SMEMCPY. I see it’s used with lengths only of 4, 
> 6, 18, 20 and 28. By inlining it with byte copies, I see over a 5% 
> increase in outbound bandwidth (that’s all I’m focused on right now). 
> As has come up this week, the call to memcpy to copy 4 or 6 bytes is 
> very silly (IMO).

Shouldn't the compiler do that inlining job for you? Isn't there an 
option for gcc for that?

> I thought I would copy u16_t in half the copies since either 
> everything is u32_t aligned for my processor (NIOS II – GCC as 
> mentioned) or **should** be u16_t aligned for IP related items.

Why that? For which reason should stack-internal data be 16-bit aligned?

> Everything **is** u16_t aligned except one item – hwaddr in struct 
> netif. Netif does **not** include packing around its definition, but 
> curiously, it **does** include packed struct members.

The struct netif is an lwIP-internal struct and thus does not need to be 
packed. The only structs that need packing are those used in protocol 
headers/data, since their layout is determined by the network/protocol.

As such, I don't think you can rely on all data being u16_t aligned to 
achieve good performance: every struct member of an non-packed struct 
should be aligned as it is required - thus, a char array can always be 
non aligned! After all, you still have the alignment-problem if setting 
LWIP_NETIF_HOSTNAME to 1.

> Have I found a problem in that GCC carried the included packed struct 
> override through the remainder of the netif struct?

I'm not sure I understand you here: The packed struct members you are 
talking about are the IP-addresses, right? Is there anything unaligned 
after the 3 addresses? If so, that would be a bug, yes.

> If I delete hwaddr_len from struct netif and then replace the only 2 
> uses of it in dhcp.c (netif->hwaddr_len ) with ETHARP_HWADDR_LEN, the 
> remainder of netif is properly aligned.
>
> Since we use ETHARP_HWADDR_LEN everywhere else in lwIP, why keep it in 
> the netif, especially if it’s used only in DHCP 2 times?

Did I understand you correctly that you want to remove the u8_t to get 
hwaddr aligned? That might work now, but I don't think every developer 
changing that struct will remember such a subtle alignment issue...

Aside from that, we explicitly *don't* use ETHERP_HWADDR_LEN throughout 
the stack, only in places that are ethernet-related (etharp.c, 
ethernetif.c, example port ethernet drivers). This is necessary to use 
the stack in non-ethernet environments!

> |..] I think this can be changed to MEMCPY since I don’t think any 
> compiler will inline a 28 byte copy. And this call is only in the 
> icmp_dest_unreach function.

Hm, I guess that's up to the compiler: But since it doesn't know the 
alignemnt of src and dst, it probably won't... Is there any disadvantage 
in keeping the SMEMCPY?

To sum it up, I think the idea of that copy define is quite handy, but 
it should by all means be optional and keep working even if some 
developers change structs to bad alignment.

Simon