This section provides a summary of Sather's features, with particular attention to features that are not found in the most common object oriented languages.
Data structures in Sather are constructed from objects, each of which has a specific concrete type that determines the operations that may be performed on it. Abstract types specify a set of operations without providing an implementation and correspond to sets of concrete types. The implementation of concrete types is defined by textual units called classes; abstract types are specified by textual units called abstract classes. Sather programs consist of classes and abstract class specifications. Each Sather variable has a declared type which determines the types of objects it may hold.
Classes define the following features: attributes which make up the internal state of objects, shareds and constants which are shared by all objects of a type, and methods which may be either routines or iterators. Any features are by default public, but may be declared private to allow only the class in which it appears access to it. An attribute or shared may instead be declared readonly to allow only the class in which it appears to modify it. Accessor routines are automatically defined for reading or writing attributes, shareds, and constants. The set of non-private methods in a class defines the interface of the corresponding type. Method definitions consist of statements; for their construction expressions are used. There are special literal expressions for boolean, character, string, integer, and floating point objects.
Certain conditions are described as fatal errors. These conditions should never occur in correct programs and all implementations of Sather must be able to detect them. For efficiency reasons, however, implementations may provide the option of disabling checking for certain conditions.
Like many object-oriented languages, Sather is garbage collected, so programmers never have to free memory explicitly. The runtime system does this automatically when it is safe to do so. Idiomatic Sather applications generate far less garbage than typical Smalltalk or Lisp programs, so the cost of collecting tends to be lower. Sather does allow the programmer to manually deallocate objects, letting the garbage collector handle the remainder. With checking compiled in, the system will catch dangling references from manual deallocation before any harm can be done.
More generally, when checking options have been turned on by compiler flags, the resulting program cannot crash disastrously or mysteriously. All sources of errors that cause crashes are either eliminated at compile-time or funneled into a few situations (such as accessing beyond array bounds) that are found at run-time precisely at the source of the error.
Sather does as little as possible behind the user's back at runtime. There are no implicitly constructed temporary objects, and therefore no rules to learn or circumvent. This extends to class constructors: all calls that can construct an object are explicitly written by the programmer. In Sather, constructors are ordinary routines distinguished only by a convenient but optional calling syntax (page 107). With garbage collection there is no need for destructors; however, explicit finalization is available when desired (page 143).
Sather never converts types implicitly, such as from integer to character, integer to floating point, single to double precision, or subclass to superclass. With neither implicit construction nor conversion, Sather resolves routine overloading (choosing one of several similarly named operations based on argument types) much more clearly than C++. The programmer can easily deduce which routine will be called. (See unnamedlink for details)
In Sather, the redefinition of operators is orthogonal to the rest of the language. (See unnamedlink) There is "syntactic sugar" for standard infix mathematical symbols such as '+' and '^' as calls to otherwise ordinary routines with names 'plus' and 'pow'. 'a+b' is just another way of writing 'a.plus(b)'. Similarly, 'a[i]' translates to 'a.aget(i)' when used in an expression. An assignment 'a[i] := expr' translates into 'a.aset(i,expr)'.
In many object-oriented languages, the term 'inheritance' is used to mean two things simultaneously. One is subtyping, which is the requirement that a class provide implementations for the abstract methods in a supertype. The other is code inheritance (called code inclusion in Sather parlance) which allows a class to reuse a portion of the implementation of another class. In many languages it is not possible to include code without subtyping or vice versa.
Sather provides separate mechanisms for these two concepts. Abstract classes represent interfaces: sets of signatures that subtypes of the abstract class must provide. Other kinds of classes provide implementation. Classes may include implementation from other classes using a special 'include' clause; this does not affect the subtyping relationship between classes. Separating these two concepts simplifies the language considerably and makes it easier to understand code. Because it is only possible to subtype from abstract classes, and abstract classes only specify an interface without code, sometimes in Sather one factors what would be a single class in C++ into two classes: an abstract class specifying the interface and a code class specifying code to be included. This often leads to cleaner designs.
Issues surrounding the decision to explicitly separate subtyping and code inclusion in Sather are discussed in the ICSI technical report TR 93-064: "Engineering a Programming Language: The Type and Class System of Sather,"[1]. It is available at the Sather WWW page.
[1] C. Szyperski, S. Omohundro, S. Murer. "Engineering a programming language: The type and class system of Sather," In Jurg Gutknecht, ed., Programming Languages and System Architectures, p. 208-227. Springer Verlag, Lecture Notes in Computer Science 782, November 1993. Available at the Sather WWW page.
Early versions of Sather used a conventional 'until...loop...end' statement much like other languages. This made Sather susceptible to bugs that afflict looping constructs. Code which controls loop iteration is known for tricky "fencepost errors" (incorrect initialization or termination). Traditional iteration constructs also require the internal implementation details of data structures to be exposed when iterating over their elements.
Simple looping constructs are more powerful when combined with heavy use of cursor objects (sometimes called 'iterators' in other languages, although Sather uses that term for something else entirely) to iterate through the contents of container objects. Cursor objects can be found in most C++ libraries, and they allow useful iteration abstraction. However, they have a number of problems. They must be explicitly initialized, incremented, and tested in the loop. Cursor objects require maintaining a parallel cursor object hierarchy alongside each container class hierarchy. Since creation is explicit, cursors aren't elegant for describing nested or recursive control structures. They can also prevent a number of important optimizations in inner loops.
An important language improvement in Sather 1.0 over earlier versions was the addition of iterators. Iterators are methods that encapsulate user defined looping control structures just as routines do for algorithms. Code using iterators is more concise, yet more readable than code using the cursor objects needed in C++. It is also safer, because the creation, increment, and termination check are bound together inviolably at one point. Each class may define many sorts of iterators, whereas a traditional approach requires a different yet intimately coupled class for each kind of iteration over the major class. Sather iterators are part of the class interface just like routines.
Iterators act as a lingua-franca for operating on collections of items. Matrices define iterators to yield rows and columns; tree classes have recursive iters to traverse the nodes in pre-order, in-order, and post-order; graph classes have iters to traverse vertices or edges breadth-first and depth-first. Other container classes such as hash tables, queues, etc. all provide iters to yield and sometimes to set elements. Arbitrary iterators may be used together in loops with other code.
The rationale of the Sather iterator construct and comparisons with related constructs in other languages can be found in the ICSI technical report TR 93-045: "Sather Iters: Object-Oriented Iteration Abstraction,"[2]. It is available at the Sather WWW page.
[2] S. Murer, S. Omohundro, D. Stoutamire, C. Szyperski, "Iteration abstraction in Sather", Transactions on Programming Languages and Systems, Vol. 18, No. 1, Jan 1996 p. 1-15. Available at the Sather WWW page.
Sather provides higher-order functions through method closures, which are similar to closures and function pointers in other languages. These allow binding some or all arguments to arbitrary routines and iterators but defer the remaining arguments and execution until a later time. They support writing code in an applicative style, although iterators eliminate much of the motivation for programming that way. They are also useful for building control structures at run-time, for example, registering call-backs with a windowing system. Like other Sather methods, method closures follow static typing and behave with contravariant conformance.
Sather distinguishes between reference objects and immutable objects. Imutable objects never change once they are created. When one wishes to modify an immutable object, one is compelled to create a whole new object that reflects the modification.
Experienced C programmers immediately understand the difference when told about the internal representation the ICSI compiler uses: immutable types are implemented with stack or register allocated C 'struct's while reference types are pointers to the heap. Because of that difference, reference objects can be referred to from more than one variable (aliased), but immutable objects never appear to be. Many of the built-in types (integers, characters, floating point) are immutable classes. There are a handful of other differences between reference and immutable types; for example, reference objects must be explicitly allocated, but immutable objects 'just are'.
Immutable types can have several performance advantages over reference types. Immutable types have no heap management overhead, they don't reserve space to store a type tag, and the absence of aliasing makes more compiler optimizations possible. For a small class like 'CPX' (complex number), all these factors combine to give a significant win over a reference class implementation. Balanced against these positive factors in using an immutable object is the overhead that some C compilers introduce in passing the entire object on the stack. This problem is worse in immutable classes with many attributes. Unfortunately the efficiency of an immutable class is directly tied to how smart the C compiler is; at this time 'gcc' is not very bright in this respect, although other compilers are.
Immutable classes aren't strictly necessary; reference classes with immutable semantics work too. For example, the reference class 'INTI' implements immutable infinite precision integers and can be used like the built-in immutable class 'INT'. The standard string class 'STR' is also a reference type but behaves with immutable semantics. Explicitly declaring immutable classes allows the compiler to enforce immutable semantics and provides a hint for good code generation. Common immutable classes are defined in the standard libraries; defining a new immutable class is unusual.
Sather attempts to conform to the IEEE 754-1985 specification for its floating point types. Unfortunately, many platforms make it difficult to do so. For example, underflow is often improperly implemented to flush to zero rather than use IEEE's gradual underflow. This happens because gradual underflow is a special case and can be quite slow if implemented using traps. When benchmarks include simulations which cause many underflows, marketing pressures make flush-to-zero the default.
There are many other problems. Microsoft's C and C++ compilers defeat the purpose of the invalid flag by using it exclusively to detect floating-point stack overflows, so programmers cannot use it. There is no portable C interface to IEEE exception flags and their behavior with respect to 'setjmp' is suspect. Threads packages often fail to address proper handling of IEEE exceptions and rounding modes.
Correct IEEE support from various platforms was the single worst porting problem of the Sather 1.0 compiler. In 1.1, we give up and make full IEEE compliance optional. Sather implementations are expected to conform to the spirit, if not the letter, of IEEE 754, although proper exceptions, extended types, underflow handling, and correct handling of positive and negative zero are specifically not required.
The Sather treatment of NaNs is particularly tricky; IEEE wants NaN to be neither equal nor unequal to anything else, including other NaNs. Because Sather defines 'x /= y' as 'x.is_eq(y).not', to get the IEEE notion of unequal is necessary to write 'x=x and y=y and x/=y'. Other comparison operators present similar difficulties.
Parallel Sather (pSather) is a parallel extension of the language, developed and in use at ICSI. It extends serial Sather with threads, synchronization, and data distribution.
pSather differs from concurrent object-oriented languages that try to unify the notions of objects and processes by following the actors model[3]. There can be a grave performance impact for the implicit synchronization this model imposes on threads even when they do not conflict. While allowing for actors, pSather treats object-orientation and parallelism as orthogonal concepts, explicitly exposing the synchronization with new language constructs.
[3] G. Agha, "Actors: A Model of Concurrent Computation in Distributed Systems", The MIT Press, Cambridge, Massachusetts, 1986.
pSather follows the Sather philosophy of shielding programmers from common sources of bugs. One of the great difficulties of parallel programming is avoiding bugs introduced by incorrect synchronization. Such bugs cause completely erroneous values to be silently propagated, threads to be starved out of computational time, or programs to deadlock. They can be especially troublesome because they may only manifest themselves under timing conditions that rarely occur (race conditions) and may be sensitive enough that they don't appear when a program is instrumented for debugging (heisenbugs). pSather makes it easier to write deadlock and starvation free code by providing structured facilities for synchronization. A lock statement automatically performs unlocking when its body exits, even if this occurs under exceptional conditions. It automatically avoids deadlocks when multiple locks are used together. It also guarantees reasonable properties of fairness when several threads are contending for the same lock.
pSather allows the programmer to direct data placement. Machines do not need to have large latencies to make data placement important. Because processor speeds are outpacing memory speeds, attention to locality can have a profound effect on the performance of even ordinary serial programs. Some existing languages can make life difficult for the performance-minded programmer because they do not allow much leeway in expressing placement. For example, extensions allowing the programmer to describe array layout as block-cyclic is helpful for matrix-oriented code but of no use for general data structures.
Because high performance appears to require explicit human-directed placement, pSather implements a shared memory abstraction using the most efficient facilities of the target platform available, while allowing the programmer to provide placement directives for control and data (without requiring them). This decouples the performance-related placement from code correctness, making it easy to develop and maintain code enjoying the language benefits available to serial code. Parallel programs can be developed on simulators running on serial machines. A powerful object-oriented approach is to write both serial and parallel machine versions of the fundamental classes in such a way that a user's code remains unchanged when moving between them.