Re: Invalid compilation without -fno-strict-aliasing

[Jakub Jelinek <jakub@redhat.com>]

On Wed, Feb 26, 2003 at 03:57:54PM -0500, Daniel Jacobowitz wrote:
> > > > > 	if((stream + event_len) < ends) {
> > > > > 		iwe->len = event_len;
> > > > > 		memcpy(stream, (char *) iwe, event_len);
> > > > > 		stream += event_len;
> > > > > 	}
> > > > > 	return stream;

> > > > The compiler is free to assume char *stream and struct iw_event *iwe
> > > > point to separate areas of memory, due to strict aliasing.

No.

> Stream is not the same storage as iwe, so this is hardly the issue. 
> Writes to stream don't affect iwe.  The problem was the assignment to
> iwe->len being moved after the access, according to the report.

GCC can do that with -fstrict-aliasing:
All calls to this inline function are done with constant event_len.
E.g. in case of event_len = IW_EV_UINT_LEN, memcpy macro expands to
__constant_memcpy which does:
                        *(unsigned long *)to = *(const unsigned long *)from;
                        *(1+(unsigned long *)to) = *(1+(const unsigned long *)from);
iwe->len is unsigned short and the code writes the value as
unsigned short (iwe->len = 8) and then reads the same memory as unsigned
long (something = *(const unsigned long *)(char *)iwe). But those two types
cannot alias and thus gcc can reorder them.

To fix that, __constant_memcpy would have to access the data through
union, ir you could as well forget about __constant_memcpy and use
__builtin_memcpy where gcc will take care about the constant copying.

	Jakub