Re: strict aliasing

[Andrew Haley <aph at redhat dot com>]

Ingo Rohloff wrote:

> I am sorry that I ask about this topic,
> even if it has been discussed so often, but I am really at a loss:
> 
> First of all is there any website, link or docu (whatever) which lists some
> examples of what is legal (under strict aliasing) and what is not ?

Yes.  Any version of the C standard will tell you.
http://www.open-std.org/JTC1/SC22/WG14/www/standards

> Let me try to explain what I understood:
> 
> With the following variable definitions,
> 
> unsigned int *ptr1;
> unsigned short *ptr2;
> 
> "strict aliasing" means that the compiler is allowed to assume that
> "ptr1" and "ptr2" will NOT acceess to the same address (object).

Right.

> For optimization this means that if someone writes
> 
> ptr2[0]=0x1111
> ptr1[0]=0x22222222;
> 
> then the compiler can assume that ptr2[0] still is "0x1111", because "ptr1"
> must not access to memory which is accessed through ptr2.
> 
> Is this correct so far ?

Yes.

> Now I am writing code for an embedded application (no OS).
> 
> Because I know the architecture I want to do the following
> ---
> unsigned int value;
> unsigned char *ptr2;
> 
> *((unsigned int *)ptr2) = value;
> ---
> 
> The idea here is that you write a 32 bit value in big endian order to a
> byte (unsigned char) array.

That's perfectly OK.  Standard C has special rules for character types, which
alias everything.

> (The used architecture is big endian; I know that on some architectures
> that means that the pointer must point to an address which is aligned...)
> 
> 
> Now my understanding was that this code breaks strict aliasing rules,
> because I cast a "unsigned char *" to an "unsigned int *".

No, because of the special rules for character types.

> BUT: The g++ 4.4.2 compiler does not complain with "-Wstrict-aliasing".
> 
> Is the above statement OK or not ?
> 
> Or is it, because "ptr2" is an "unsigned char *", that ptr2 is assumed to
> be able to alias anything ? (Because it is a char pointer).

That's right.

> Does these "char pointer can alias anything" really only refer to "char *"
> or also to "unsigned char *".

All character types.

But don't believe me: go to http://www.open-std.org/JTC1/SC22/WG14/www/docs/n1124.pdf
and read 6.3.2.3 Paragraph 7.

Andrew.