|
Assembly languages also have execution models, the same as any other language. Such an execution model is implemented by a CPU micro-architecture. For example, both a 5 stage in-order pipeline and a large out of order CPU implement the same assembly language execution model. The execution model is the definition of the behavior, so all implementations, whether in-order or out-of-order or interpreted or JIT'd etc.. must all give the exact same result, and that result is defined by the execution model.
== Parallel Execution Models ==
In the modern age, parallel programming is an increasingly important topic. Parallel execution models tend to be complex because they involve multiple timelines. Parallel execution models necessarily include the behavior of [[Synchronization_construct| synchronization constructs]]. A synchronization construct has the effect of establishing an ordering between activities in one timeline relative to activities in another timeline.
For example, a common synchronization construct is the lock. Consider one timeline. The timeline has a point at which it executes the "gain ownership of the lock" synchronization construct. In Posix threads this would be pthread_mutex_lock(&myMutex). In Java this would be lock.lock(). In both cases, the timeline is called a thread. The C and Java execution models are sequential, and they state that the timeline has activities that come before the call to "gain ownership of the lock", and activities that come after the call. Likewise there is a "give up ownership of the lock" operation. In C this would be pthread_mutex_unlock(&myMutex). In Java this would be lock.unlock(). Again, the execution models C and Java define that one group of statements is executed before ownership of the lock is given up, and another group of statements is executed after ownership of the lock is given up.
Now, consider the case of two timelines, also known as two threads. One thread, call it thread A, executes some statements, call them A-pre-gain-lock statements. Then thread A executes "gain ownership of the lock", then thread A executes A-post-gain-lock statements, which come after A gains ownership of the lock. Finally, thread A performs "give up ownership of the lock". Then thread A performs A-post-giveup-lock statements.
A second thread, call it thread B, executes some statements, call them B-pre-lock statements. Then thread B executes "gain ownership of the lock", then thread B executes B-post-lock statements, which come after B gains ownership of the lock.
Now, we can say the parallel execution model of the "gain ownership of lock" and "give up ownership of lock" synchronization construct. The execution model is this:
"In the case that ownership of the lock goes from thread A to thread B, A-post-gain-lock statements come before B-post-gain-lock statements."
And that's it.
Seems simple, right? The complication comes from the fact that the execution model does not have any means for the execution of "give up ownership of the lock" to have any influence over which execution of "gain ownership of the lock" in some other timeline (thread) follows. Very often, only certain handoffs give valid results. Thus, the programmer must think of all possible combinations of one thread giving up a lock and another thread getting it next, and make sure their code only allows valid combinations.
Note that the only effect is that A-post-gain-lock statements come before B-post-gain-lock statements. No other effect happens, and no other relative ordering can be relied upon. Specifically, A-post-give-up-lock and B-post-gain-lock have ''no relative ordering'' defined, which surprises many people. But thread A may have been swapped out after giving up ownership, so A-post-give-up-lock statements may happen long after many B-post-gain-lock statements have finished. That is one of the possibilities that must be thought about when designing locks, and illustrates why multi-threaded programming is difficult.
Note that modern parallel languages have much easier to use execution models. The thread model was one of the original parallel execution models, which may account for why it has persisted despite being difficult to use.
== See also ==
| 