RE: [PATCH v6 1/6] wifi: mac80211: Use struct instead of macro for PREQ frame

[Ping-Ke Shih <pkshih@realtek.com>]

Johannes Berg <johannes@sipsolutions.net> wrote:
> The point is that __packed forces the compiler to emit code that doesn't
> rely on alignment.

I was not aware that __packed is a hint of unalignment to compiler.
I learned. :)

> 
> So say e.g. you have
> 
>         __le32 *ptr = ...;
> 
>         u32 val = cpu_to_le32(*ptr);
> 
> In this case, the compiler can emit a load instruction that assumes
> alignment, so if 'ptr' can be unaligned, we need to use
> 
>         u32 val = get_unaligned_le32(ptr);

Got it.
With the assignment to pointer, the hint disappeared. 

> 
> instead.
> 
> However, if we have
> 
>         struct ieee80211_mesh_hwmp_preq_top *ptr = ...;
> 
>         u32 val = cpu_to_le32(ptr->preq_id);
> 
> then the compiler _cannot_ emit a load instruction that assumes
> alignment because of the __packed, and so the compiler has to emit a
> (perhaps sequence of) instruction(s) that load the 32-bit value without
> relying on alignment. As a consequence, we don't have to explicitly
> write it the more complicated way and can just write it the more natural
> way.

I'll remember this is better style in practice. Compiler can help.


In my tests with arm-gcc compiler, I did a special case: 

struct foo {
	int a;
	char b;
} __packed;

int bar(struct foo *foo)
{
	return foo->a;
}

It is obviously aligned (offset = 0), so I guessed arm-gcc can generate
a single load instruction, but actually it doesn't (even with -O3).
But this is not the course of this thread. :)

Thank you
Ping-Ke