ARM memcpy on unaligned pointers and -O2

[Richard Earnshaw]

> On Thu, Dec 13, 2001 at 09:53:01AM -0700, Shaun Jackman wrote:
> 
>  > I recognized that afterwords and cast the arguments to memcpy()
>  > 
>  > uint32
>  > ld32( const uint32* p)
>  > {
>  > 	uint32 ret;
>  > 	memcpy( &ret, (uint8*)p, sizeof( ret));
>  > 	return ret;
>  > }
>  > 
>  > void
>  > str32( uint32* p, uint32 val)
>  > {
>  > 	memcpy( (uint8*)p, &val, sizeof( *p));
>  > }
>  > 
>  > It seems the behaviour persists though. In fact, it has the exact same 
>  > behaviour if I change the parameters from uint32* to void*.
> 
> This behavior matches my experience with GCC, as well.  Other compilers
> (such as Digital Unix's) to honor the casts in this case, but for GCC,
> you have to actually do this:
> 
> 	char *cp = (char *) src;
> 	memcpy(&ret, cp, sizeof(ret));
> 
> I think GCC should honor the cast, like other compilers do.

I don't think even this can be expected to work.  From the moment you put 
something that isn't a uint32* into p, you've lied to the compiler.  If a 
uint32 * is an aligned type, then it must only contain aligned pointers; 
playing games with assignments like this is just trying to cover up your 
lie -- but eventually the compiler will find you out.

R.