Last modified on 21 March 2013, at 09:28

Ada Programming/Pragmas/Atomic:3

It is almost always incorrect to use atomic or volatile variables for tasking.[1] When an object is atomic it just means that it will be read from or written to memory atomically. The compiler will not generate atomic instructions or memory barriers when accessing to that object, it will just:

  • check that the architecture guarantees atomic memory loads and stores,
  • disallow some compiler optimizations, like reordering or suppressing redundant accesses to the object.

For example, the following code, where A is an atomic object can be misunderstood:

A := A + 1;  -- Not an atomic increment!

The compiler will not (and is not allowed by the Standard to) generate an atomic increment instruction to directly increment and update from memory the variable A.[2] This is the code generated by the compiler:

  A := A + 1;
804969f:        a1 04 95 05 08          mov    0x8059504,%eax
80496a4:        40                      inc    %eax
80496a5:        a3 04 95 05 08          mov    %eax,0x8059504

As can be seen, no atomic increment instruction or test-and-set opcode will be generated. Like in other programming languages, if these specific instructions are required in the program they must be written explicitly using machine code insertions.[3]

The above code snippet is equivalent to the following code (both code sequences generates exactly the same object code), where T is a (non-atomic) temporary variable:

T := A;      -- A is copied atomically to local variable T
T := T + 1;  -- local variable T is incremented
A := T;      -- A is stored atomically

Thus it is incorrect to modify an atomic variable at the same time from multiple tasks. For example, two tasks incrementing a counter in parallel. Even in an uniprocessor, other Ada tasking features like a protected object should be used instead. In multiprocessors, depending on the memory consistency model, using various atomic or volatile variables for task communication can have surprising consequences.[2][4] Therefore, extreme care should be taken when using atomic objects for task data sharing or synchronization, specially in a multiprocessor.


ReferencesEdit

  1. Arch Robison (2007-11-30). "Volatile: Almost Useless for Multi-Threaded Programming". Intel Software Network. http://softwareblogs.intel.com/2007/11/30/volatile-almost-useless-for-multi-threaded-programming/. Retrieved 2008-05-30. "There is a widespread notion that the keyword volatile is good for multi-threaded programming (...) volatile is almost useless for multi-threaded programming." 
  2. a b Ian Lance Taylor (2008-03-05). "Volatile". http://www.airs.com/blog/archives/154. Retrieved 2008-05-28. "Using [the C/C++ qualifier] volatile does not mean that the variable is accessed atomically; no locks are used. Using volatile does not mean that other cores in a multi-core system will see the memory accesses; no cache flushes are used. (...) Using volatile does not imply any sort of memory barrier; the processor can and will rearrange volatile memory accesses. (...) You should not use more than one such variable to communicate between any pair of threads, as there is no guarantee that the different threads will see the accesses in the same order." 
  3. Laurent Guerby (January 1995). "C.5 Shared Variable Control". Ada 95 Rationale. Intermetrics. "A need to access specific machine instructions arises sometimes (...). Examples include instructions that perform compound operations atomically on shared memory, such as test-and-set and compare-and-swap (...)" 
  4. Sarita V. Adve, Kourosh Gharachorloo (December 1996). "Shared Memory Consistency Models: A Tutorial". IEEE Computer 29 (12): 66–76. http://www.hpl.hp.com/techreports/Compaq-DEC/WRL-95-7.pdf. Retrieved 2008-05-28.