C++ aliasing problem (noticed with -O2 -march=athlon)

[Jakub Bogusz <qboosh at pld dot org dot pl>]

I have a question related to C++ aliasing rules - because I'm not sure
what's wrong - some program code or aliasing rules in gcc.

Some library+program (djvulibre) compiled by gcc 3.2 (3.2.1 too)
with "-O2 -march=athlon" segfaults. With other optimizations
(e.g. "-O1 -march=athlon", "-O2 -march=i686" etc.) it doesn't happen.
Also "-O2 -march=athlon -fno-strict-aliasing" optimization generates
good, working code (that's why I suppose it's aliasing-related problem
and athlon is only coincidence).

Offending djvulibre code is (declaration):

#v+
class GPBufferBase
{
public:
  GPBufferBase(void *&,const size_t n,const size_t t);
private:
  void *&ptr;
};

inline
GPBufferBase::GPBufferBase(void *&xptr,const size_t n,const size_t t) : ptr(xptr), num(n)
{
  xptr=(n*t)?(::operator new(n*t)):0;
}

template<class TYPE>
class GPBuffer : public GPBufferBase
{
public:
  GPBuffer(TYPE *&xptr,const size_t n=0) : GPBufferBase((void *&)xptr,n,sizeof(TYPE)) {}
}
#v-

...and then use:

#v+
  unsigned int *posn;
  GPBuffer<unsigned int> gposn(posn,blocksize);
  memset(posn, 0, sizeof(unsigned int)*size);
#v-      

Good x86 code, generated with e.g. "-O2 -march=i686" or
"-O2 -march=athlon -fno-sctrict-aliasing":

#v+
  leal 60(%esp),%esi
[...]
.Lxxx
  movl %eax,(%esi)
  leal 0(,%edx,4),%ecx
  movl 60(%esp),%edi
#v-

Wrong code, generated with "-O2 -march=athlon":

#v+
  leal 60(%esp),%esi
[...]
.Lxxx
  movl 60(%esp),%edi
  leal 0(,%edx,4),%ecx
  movl %eax,(%esi)
#v-

"void*" should be valid alias for "unsigned int*", so
isn't "void*&" vs "unsigned int*&" case similar?


TIA

-- 
Jakub Bogusz    http://www.cs.net.pl/~qboosh/
PLD Linux       http://www.pld.org.pl/