9.16 Use of Address Clauses for Memory-Mapped I/O

A common pattern is to use an address clause to map an atomic variable to a location in memory that corresponds to a memory-mapped I/O operation or operations, for example:

    type Mem_Word is record
       A,B,C,D : Byte;
    end record;
    pragma Atomic (Mem_Word);
    for Mem_Word_Size use 32;
    
    Mem : Mem_Word;
    for Mem'Address use some-address;
    ...
    Temp := Mem;
    Temp.A := 32;
    Mem := Temp;

For a full access (reference or modification) of the variable (Mem) in this case, as in the above examples, GNAT guarantees that the entire atomic word will be accessed. It is not clear whether the RM requires this. For example in the above, can the compiler reference only the Mem.A field as an optimization? Whatever the answer to this question is, GNAT makes the guarantee that for such a reference, the entire word is read or written.

A problem arises with a component access such as:

    Mem.A := 32;

Note that the component A is not declared as atomic. This means that it is not clear what this assignment means. It could correspond to full word read and write as given in the first example, or on architectures that supported such an operation it might be a single byte store instruction. The RM does not have anything to say in this situation, and GNAT does not make any guarantee. The code generated may vary from target to target. GNAT will issue a warning in such a case:

    Mem.A := 32;
    |
    >>> warning: access to non-atomic component of atomic array,
        may cause unexpected accesses to atomic object

It is best to be explicit in this situation, by either declaring the components to be atomic if you want the byte store, or explicitly writing the full word access sequence if that is what the hardware requires.