> > Are your structures marked as packed?  If not, what compilation flags are 
> > being used?
> 
> Hm. They are. But: Still all variables are 32 bit aligned! And: Using gcc 2.95.2
> under Linux or RiscOS I do not get this strange behaviour. Why does NetBSD
> still use egcs?
> 
> One example structure looks like this:
> 
> #pragma pack(1)
> 
> struct test {
>   int a,b;
> };
> 
> #pragma pack()

#prgama pack(1) means that this structure can appear at *any* address 
(that's part of the definition of a packed structure).  That includes odd 
addresses, so the code must be compiled as a set of byte loads and the 
final word built up from the bits.  I'm not sure why linux is doing 
differently, but it may well be a bug if it is.

From tm.texi (in the gcc documentation):

  The pack pragma specifies the maximum alignment (in bytes) of fields
  within a structure, in much the same way as the @samp{__aligned__} and
  @samp{__packed__} @code{__attribute__}s do.  A pack value of zero resets
  the behaviour to the default.


> If I now do access the content of this structure, the ints are accessed by 4
> LDRBs and 3 ORRs, which is absolutely not necessary.
> 
> Is there a chance that NetBSD/arm32 will soon use gcc 2.95.2?

Possibly, but it won't change anything, since the code generated is 
correct for the supplied input.

Richard.