[Spice-devel] [PATCH] demarshall: Add safe method for unaligned byte access.

[Frediano Ziglio]

> 
> On Fri, Nov 27, 2015 at 07:04:55AM -0500, Frediano Ziglio wrote:
> > > 
> > > On Wed, Nov 25, 2015 at 09:04:58AM -0500, Frediano Ziglio wrote:
> > > > > 
> > > > > Usefull for ARM. Enable with "USE_MEMCPY" define.
> > > > >  Should be better to add "--use-memcpy" to configure.ac
> > > > > 
> > > > > Signed-off-by: Anton D. Kachalov <mouse at yandex-team.ru>
> > > > 
> > > > Are you sure there are not other way except memcpy?
> > > 
> > > An alternate way which I proposed was to do
> > > uint16_t my_int = unaligned_int[0] | unaligned_int[1] << 8
> > > but Anton said it was likely to be slower than a memcpy. I don't know
> > > how these unaligned accesses are usually handled in ARM code.
> > > I found
> > > https://github.com/yura-pakhuchiy/snappy-c/commit/a33077391874b36fb4c430ff9ed5f6e437b65f6f
> > > doinng something similar, but that's about it.
> > > 
> > > Christophe
> > > 
> > 
> > Being architecture dependent usually the compiler is the best that know
> > how to do it.
> > 
> > In the gcc word speaking the trick is to create a packed structure,
> > turn off aliasing and packing, something like
> > 
> > struct unaligned_u16
> > {
> >     uint16_t u16;
> > } __attribute__((__may_alias__)) __attribute__((packed));
> > 
> > static inline uint16_t read_u16ua(const void *p)
> > {
> >    return ((struct unaligned_u16 *)p)->u16;
> > }
> 
> Are you sure this is doing the right thing? This is not very far from
> what we currently have (struct { uint16_t field1; uint16_t field2; }
> __attribute__((packed));
> with field1 and field2 ending up unaligned. gcc is apparently not doing
> the right thing here, and unaligned memory access traps are triggered,
> causing this code to be slow.
> 
> Christohpe
> 

The theory is this! And is working for me.

"unaligned memory access traps are triggered". Architecture?
Version? Debugging? Watching at hardware counters? gcc version?
compiler flags?

"causing this code to be slow". Unaligned access are usually slower but
implemented in hardware are faster than doing it manually.
Or in a particular architecture gcc compile with unaligned instructions
and this is triggered in kernel and emulated? In this case perhaps you should
change architecture target (for instance ARM version), or the gcc version
is too old and is better to use some macros/functions.

Frediano