Re: ARM compiler misbehaves ?

[Doug Evans <dje at transmeta dot com>]

Richard Earnshaw writes:
 > 
 > On Thu, 2004-04-29 at 16:30, Grant Edwards wrote:
 > > On Thu, Apr 29, 2004 at 04:23:26PM +0100, Richard Earnshaw wrote:
 > > > On Thu, 2004-04-29 at 15:43, Bill Gatliff wrote:
 > > > 
 > > > > Manipulations through pointers won't work, that much should be clear by 
 > > > > now (as should be the underlying reason).  But _references_ to packed 
 > > > > structure members by name, even when misaligned, do work.  Gcc will 
 > > > > reconstruct the misaligned object in a properly-aligned one (a register, 
 > > > > or on the stack), when necessary.
 > > > 
 > > > Indeed.  There's no dispute about this.
 > > 
 > > Except that you said it didn't work in a previous post.
 > 
 > There seems to be some general confusion here.  The thing I was (trying)
 > to say doesn't work is placing the address of a packed object into a
 > pointer to an object of the same base type.  Hence
 > 
 > struct s {char a; short b;} __attribute__((packed));
 > 
 > struct s S;
 > 
 > short *p = &S.b;  
 > 
 > says that although S is a packed object, I'm now asserting that S.b is,
 > in fact, aligned as a short.
 > 
 > If it turns out that S.b wasn't aligned, then the code won't work,
 > because the compiler is entitled to assume that it was.

Indeed.

Also, I sense that trying to make this a problem with the C standard
isn't right.  We're already outside the scope of the C standard
with the use of `packed'. gcc introduced `packed', it should follow
through with a useful implementation of it (no claim is made that
it hasn't).

Maybe gcc should have an error checking mode that complains
about assigning an address of an unaligned value to an aligned
pointer (if it doesn't already).

Also, if one modified the type of `p' to be "pointer to packed short",
then memcpy should "just work".

------
Want more information?  See the CrossGCC FAQ, http://www.objsw.com/CrossGCC/
Want to unsubscribe? Send a note to crossgcc-unsubscribe@sources.redhat.com