Re: [lwip-devel] [PATCH] Fix aliasing violation.

[address@hidden]

David Woodhouse wrote:
> On Fri, 2008-10-03 at 16:43 +0200, address@hidden wrote:
>   
> > In 1.3.0, we converted all calls to 'memcpy' to either 'MEMCPY' or 
> > 'SMEMCPY' (s for short). The goal is optimization: SMEMCPY is defined to 
> > memcpy (which the compiler may inline e.g. if size is known at compile 
> > time and small); whil MEMCPY can be overridden to provide a better copy 
> > mechanism (on my platform, for example, copying non-aligned data can be 
> > solved much better when loading full words and shifting in registers).
> >
> > Therfore, using 'memcpy' directly should be avoided: when you know your 
> > compiler can't optimize it, you can redefine SMEMCPY to inlining yourself.
> >     
>
> I'm inclined to suggest that the correct answer in that case is to _fix_
> the compiler, not work around it. But OK...
This isn't a question of broken compilers: It's a code size vs. speed 
question (an optimized memcpy is much bigger than a simple for-loop that 
does a byte copy). If the compiler decides not to inline, it would call 
'memcpy'. And since on most platforms, memcpy is a generic 
implementation of a C-library which is NOT optimized for the actual 
hardware, an assembly-routine can be much faster.

SMEMCPY was chosen to explicitly state this is a short memcpy and we 
(lwIP developers) don't want to call the fast routine for it since the 
compiler might optimize it.

Simon