The libraries of the Scheme standard are organized according to projected use. Hence, the (rnrs base (6)) library exports procedures and syntactic abstractions that are likely to be useful for most Scheme programs and libraries. Conversely, each of the libraries relegated to the separate report on libraries is likely to be missing from the imports of a substantial number of programs and libraries. Naturally, the specific decisions about this organization and the separation of concerns of the libraries are debatable, and represent a best attempt of the editors.
A number of secondary criteria were also used in choosing the exports of the base library. In particular, macros transformers defined using the facilities of the base library are guaranteed to be hygienic; hygiene-breaking transformers are only available through the (rnrs syntax-case (6)) library.
Note that (rnrs base (6)) is not a “primitive library” in the sense that all other libraries of the Scheme standard can be implemented portably using only its exports. Moreover, the library organization is not layered from more primitive to more advanced libraries, even though some libraries can certainly be implemented in terms of others. Organizing the libraries according to such a layering would have meant imposing unnecessary constraints on implementations. Implementations differ in their sets of primitives, and thus any layering mandated by the report is unlikely to reflect the layering employed by any specific implementation. (The distinction between primitive and derived features was removed from the report for the same reasons.)
In library bodies and local bodies, all definitions must precede all expressions. R6RS treats bodies in top-level programs as a special case. Allowing definitions and expressions to be mixed in top-level programs has ugly semantics, and introduces a special case, but was allowed as a concession to convenience when constructing programs rapidly via cut and paste.
Definitions are not interchangeable with expressions, so definitions cannot be allowed to appear wherever expressions can appear, and expressions cannot be allowed to appear wherever definitions can appear. Composition of definitions with expressions will therefore be restricted in some way; the question is not whether there will be restrictions, but what those restrictions will be.
Historically, top-level definitions in Scheme have had a different semantics from definitions in bodies. In a body, definitions serve as syntactic sugar for the bindings of a letrec (or letrec* in R6RS) that is implicit at the head of every body.
That semantics can be stretched to cover top-level programs by converting expressions to definitions of ignored variables, but does not easily generalize to allow definitions to be placed anywhere within expressions. Different generalizations of definition placement are possible, however a survey of current Scheme code found surprisingly few places where such a generalization would be useful.
If such a generalization proposed, were adopted, programmers who are familiar with Java and similar languages might expect definitions to be allowed in the same kinds of contexts that allow declarations in Java. These programmers would have to be reminded that Scheme’s definitions have letrec* scope, while Java declarations (inside a method body) have let* scope and cannot be used to define recursive procedures; that Scheme’s begin expressions do not introduce a new scope, while Java’s curly braces do introduce a new scope; that flow analysis is nontrivial in higher order languages, while Java can use a trivial flow analysis to reject programs with undefined variables; that Scheme’s macro expander must locate all definitions, while Java has no macro system; and so on. Rather than explain how those facts justify restricting definitions to appear as top-level forms of a body, it is simpler to explain that definitions are just syntactic sugar for the bindings of an implicit letrec* at the head of each body, and to explain that the relaxation of that restriction for top-level bodies is (like several other features of top-level programs) an ad hoc special case.
The syntax-rules and identifier-syntax forms are used to create macro transformers and are thus needed only at expansion time, i.e., meta level 1.
The identifiers unquote, unquote-splicing, =>, and else serve as literals in the syntax of one or more syntactic forms; e.g., else serves as a literal in the syntax of cond and case. Bindings of these identifiers are exported from the base library so that they can be distinguished from other bindings of these identifiers or renamed on import. The identifiers ..., _, and set! serve as literals in the syntax of syntax-rules and identifier-syntax forms and are thus exported along with those forms with level 1.
The let-values and let-values* forms are compatible with SRFI 11 [16].
The above definition of eqv? allows implementations latitude in their treatment of procedures: implementations are free either to detect or to fail to detect that two procedures are equivalent to each other, and can decide whether or not to merge representations of equivalent procedures by using the same pointer or bit pattern to represent both. Moreover, they can use implementation techniques such as inlining and beta reduction that duplicate otherwise equivalent procedures.
The basic reason why the behavior of eqv? is not specified on NaNs is that the IEEE-754 standard does not say much about how the bits of a NaN are to be interpreted, and explicitly allows implementations of that standard to use most of a NaN’s bits to encode implementation-dependent semantics. The implementors of a Scheme system should therefore decide how eqv? should interpret those bits.
Arguably, R6RS should require
(let ((x <expression>)) (eqv? x x))to evaluate to true when <expression> evaluates to a number object; both R5RS and R6RS imply this for certain other types, and for most numbers objects, but not for NaNs. Since the IEEE-754 and draft IEEE-754R [18] both say that the interpretation of a NaN’s payload is left up to implementations, and implementations of Scheme often do not have much control over the implementation of IEEE arithmetic, it would be unwise for R6RS to insist upon the truth of
(let ((x <expression>))even though that expression is likely to evaluate to true in most systems. For example, a system with delayed boxing of inexact real number objects might box the two arguments to eqv? separately, the boxing process might involve a change of precision, and the two separate changes of precision may result in two different payloads.
When x and y are flonums represented in IEEE floating point or similar, it is reasonable to implement (eqv? x y) by a bitwise comparison of the floating-point representations. R6RS should not require this, however, because
R6RS does not require that flonums be represented by a floating-point representation,
the interpretation of a NaN’s payload is explicitly implementation-dependent according to both the IEEE-754 standard and the current draft of its proposed replacement, IEEE-754R, and
the semantics of Scheme should remain independent of bit-level representations.
For example, IEEE-754, IEEE-754R, and the draft R6RS all allow the external representation +nan.0 to be read as a NaN whose payload encodes the input port and position at which +nan.0 was read. This is no different from any other external representation such as (), #(), or 324. An implementation can have arbitrarily many bit-level representations of the empty vector, for example, and some do. That is why the behavior of the eq? and eqv? procedures on vectors cannot be defined by reference to bit-level representations, and must instead be defined explicitly.
It is usually possible to implement eq? much more efficiently than eqv?, for example, as a simple pointer comparison instead of as some more complicated operation. One reason is that it may not be possible to compute eqv? of two number objects in constant time, whereas eq? implemented as pointer comparison will always finish in constant time. The eq? predicate may be used like eqv? in applications using procedures to implement objects with state since it obeys the same constraints as eqv?.
R5RS did not require implementations to support the full numeric tower. Consequently, it was hard to write portable programs that perform substantial arithmetic, and it is unnecessarily difficult even to write programs whose arithmetic is portable between different implementations in the same category. The portability problems were most easily solved by requiring all implementations to support the full numerical tower.
As mentioned in chapter 3, the treatment of infinities, NaNs and -0.0, if present in a Scheme implementation, are in line with IEEE-754 [17] and IEEE-754R [18]. Analogously, the specification of branch cuts for certain transcendental functions have been changed from R5RS to conform to the IEEE standard.
The specification of the transcendental functions follows Steele [35], which in turn cites Penfield [24]; refer to these sources for more detailed discussion of branch cuts, boundary conditions, and implementation of these functions.
The domains of the finite?, infinite?, and nan? procedures could be expanded to include all number objects, or perhaps even all objects. However, R6RS restricts them to real number objects. Expanding nan? to complex number objects would involve at least some arbitrariness; not expanding its domain while expanding the domains of the other two would introduce an irregularity into the domains of these three procedures, which are likely to be used together; it is easier for programmers who wish to use these procedures with complex number objects to express their intent in terms of the real-only versions than it would be for the editors to guess their intent.
Scheme’s numerical types are the exactness types exact and inexact, the tower types integer, rational, real, complex, and number, and the Cartesian product of the exactness types with the tower types, where < t1, t2 >; is regarded as a subtype of both t1 and t2.
These types have an aesthetic symmetry to them, but they are not equally important. Judging by the number of R5RS procedures whose domain is restricted to values of some numerical type, the most important numerical types are the exact integer objects, the integer objects, the rational number objects, the real number objects, and the complex number objects.
Many programmers and implementors regard the focus of R5RS on those particular numerical types as something of a mistake. In practice, there is reason to believe that the most important numerical types are the exact integer objects, the exact rational number objects, the inexact real number objects, and the inexact complex number objects. This section explores one of the reasons those four types are so important in practice, and why real number objects have an exact zero as their imaginary part in R6RS (a change from R5RS).
Scheme’s types are not completely arbitrary. Each type corresponds to a set of values that turns up repeatedly as the natural domain or range of the functions that are computed by Scheme’s standard procedures. The reason these types turn up so often is that they are closed under certain sets of operations.
The exact integer objects, for example, are closed under the integral operations of addition, subtraction, and multiplication. The exact rational number objects are closed under the rational operations, which consist of the integral operations plus division (although division by zero is a special case). The real number objects (and inexact real number objects) are closed under some (often inexact) interpretation of rational and irrational operations such as exp and sin, but are not closed under operations such as log, sqrt, and expt. The complex (and inexact complex) number objects are closed under the largest set of operations.
A naive implementation of Scheme’s arithmetic operations is slow compared to the arithmetic operations of most other languages, mainly because most operations must perform some sort of case dispatch on the representation types of their arguments. The potential for this case dispatch arises when the type of an operation’s argument is represented by a union of two or more representation types, or because the operation must raise an exception when given an argument of an incorrect type. (The second reason can be regarded as a special case of the first.)
To make Scheme’s arithmetic more efficient, many implementations provide sets of operations whose domain is restricted to a single representation type, and which are not expected to raise an exception when given arguments of incorrect type.
Alternatively, or in addition, several compilers perform some kind of flow analysis that attempts to infer the representation types of expressions. When a single representation type can be inferred for each argument of an operation, and those types match the types expected by some representation-specific version of the operation, then the compiler can substitute the specific version for the more general version that was specified in the source code.
Flow analysis is performed by solving the type and interval constraints that arise from such things as:
the types of literal constants, e.g. 2 is an exact integer object that is known to be within the interval [2,2]
conditional control flow that is predicated on known inequalities, e.g. (if (< i n) <expression1> <expression2>)
conditional control flow that is predicated on known type predicates, e.g. (if (real? x) <expression1> <expression2>)
the closure properties of known operations (for example, (+ flonum1 flonum2) always evaluates to a flonum)
The purpose of flow analysis (as motivated in this section) is to infer a single representation type for each argument of an operation. That places a premium on predicates and closure properties from which a single representation type can be inferred.
In practice, the most important single representation types are fixnum, flonum, and compnum. (A compnum is a pair of flonums, representing an inexact complex number object.) These are the representation types for which a short sequence of machine code can be generated when the representation type is known, but for which considerably less efficient code will probably have to be generated when the representation type cannot be inferred.
The fixnum representation type is not closed under any operation of R5RS, so it is hard for flow analysis to infer the fixnum type from portable code. Sometimes the combination of a more general type (e.g. exact integer object) and an interval (e.g. [0,n), where n is known to be a fixnum) can imply the fixnum representation type. Adding fixnum-specific operations that map fixnums to fixnums greatly increases the number of fixnum representation types that a compiler can infer.
The flonum representation type is not closed under operations such as sqrt and expt, so flow analysis tends to break down in the presence of those operations. This is unfortunate, because those operations are normally used only with arguments for which the result is expected to be a flonum. Adding flonum-specific versions such as flsqrt and flexpt improves the effectiveness of flow analysis.
R5RS creates a more insidious problem by defining (real? z) to be true if and only if (zero? (imag-part z)) is true. This means, for example, that -2.5+0.0i is real. If -2.5+0.0i is represented as a compnum, then the compiler cannot rely on x being a flonum during the <real case> of (if (real? x) <expression1> <expression2>). This problem could be fixed by writing all of the arithmetic operations so that any compnum with a zero imaginary part is converted to a flonum before it is returned, but that merely creates an analogous problem for compnum arithmetic, as explained below. R6RS adopted a proposal by Brad Lucier to fix the problem: (real? z) is now true if and only if (imag-part z) is an exact zero.
The compnum representation type is closed under virtually all operations, provided no operation that accepts two compnums as its argument ever returns a flonum. To work around the problem described in the paragraph above, several implementations automatically convert compnums with a zero imaginary part to the flonum representation. This practice virtually destroys the effectiveness of flow analysis for inferring the compnum representation, so it is not a good workaround. To improve the effectiveness of flow analysis, it is better to change the definition of Scheme’s real number objects as described in the paragraph above.
Given arithmetic on exact integer objects of arbitrary precision, it is a trivial matter to derive signed and unsigned integer types of finite range from it by modular reduction. For example 32-bit signed two-complement arithmetic behaves like computing with the residue classes “mod 232”, where the set { - 231, ..., 231 - 1} has been chosen to represent the residue classes. Likewise, unsigned 32-bit arithmetic also behaves like computing “mod 232”, but with a different set of representatives {0, ..., 232 - 1}.
Unfortunately, the R5RS operations quotient, remainder, and modulo are not ideal for this purpose. In the following example, remainder fails to transport the additive group structure of the integers over to the residues modulo 3.
(remainder (+ -2 3) 3) ⇒ 1,In fact, modulo should have been used, producing residues in {0,1,2}. For modular reduction with symmetric residues, i.e. in { - 1,0,1} in the example, it is necessary to define a more complicated reduction altogether.
Therefore, quotient, remainder, and modulo have been replaced in R6RS by the div, mod, div0, and mod0 procedures, which are more useful when implementing modular reduction. The underlying mathematical functions div, mod, div0, and mod0 (see report section on “Integer division”) have been adapted from the div and mod operations by Egner et al. [11]. They differ in the representatives from the residue classes they return: div and mod always compute a non-negative residue, whereas div0 and mod0 compute a residue from a set centered on 0. The former can be used, for example, to implement unsigned fixed-width arithmetic, whereas the latter correspond to two’s-complement arithmetic.
These operations differ slightly from the div and mod operations from Egner et al. The latter make both operations available through a single pair of operations that distinguish between the two cases for residues by the sign of the divisor (as well as returning 0 for a zero divisor). Splitting the operations into two sets of procedures avoids potential confusion.
The procedures modulo, remainder, and quotient from R5RS can easily be defined in terms of div and mod.
The behavior of the numerical type predicates complex?, real?, rational?, and integer behavior is motivated by closure properties described in section 11.6.6. Conversely, the procedures real-valued?, rational-valued?, and integer-valued? test whether a given number object can be coerced to the specified type without loss of numerical accuracy.
The round procedure rounds to even for consistency with the default rounding mode specified by the IEEE floating-point standard.
The behavior of sqrt is consistent with the IEEE floating-point standard.
If z is an inexact number object represented using binary floating point, and the radix is 10, then the expression listed in the specification is normally satisfied by a result containing a decimal point. The unspecified case allows for infinities, NaNs, and representations other than binary floating-point.
Where R5RS specified characters and strings in terms of its own, limited character set, R6RS specifies characters and strings in terms of Unicode. The primary goal of the design change were to improve the portability of Scheme programs that manipulate text, while preserving a maximum of backward compatibility with R5RS.
R6RS defines characters to be representations of Unicode scalar values, and strings to be indexed sequences of characters. This is a different representation for Unicode text than the representations chosen by some other programming languages such as Java or C#, which use UTF-16 code units as the basis for the type of characters.
The representation of Unicode text corresponds to the lowest semantic level of the Unicode standard: The Unicode standard specifies most semantic properties in terms of Unicode scalar values. Thus, Unicode strings in Scheme allow the straightforward implementation of semantically sensitive algorithms on strings in terms of these scalar values.
In contrast, UTF-16 is a specific encoding for Unicode text, and performing semantic manipulation on UTF-16 representations of text is awkward. Choosing UTF-16 as the basis for the string representation would have meant that a character object potentially carries no semantic information at all, as surrogates have to be combined pairwise to yield the corresponding Unicode scalar value. (As a result, the APIs of Java for semantic operations on Unicode text often come in two overloadings, one for character objects and one for integers that are Unicode scalar values.)
The surrogates cover a numerical range deliberately omitted from the set of Unicode scalar values. Hence, surrogates have no representation as characters—they are merely an artefact of the design of UTF-16. Including surrogates in the set of characters introduces complications similar to the complications of using UTF-16 directly. In particular, most Unicode consortium standards and recommendations explicitly prohibit unpaired surrogates, including the UTF-8 encoding, the UTF-16 encoding, the UTF-32 encoding, and recommendations for implementing the ANSI C wchar_t type. Even UCS-4, which originally permitted a larger range of values that includes the surrogate range, has been redefined to match UTF-32 exactly. That is, the original UCS-4 range was shrunk and surrogates were excluded.
Arguably, a higher-level model for text could be used as the basis for Scheme’s character and string types, such as grapheme clusters. However, no design satisfying the goals stated above was available when the report was written.
Symbols have exactly the properties needed to represent identifiers in programs, and so most implementations of Scheme use them internally for that purpose. Symbols are useful for many other applications; for instance, they may be used the way enumerated values are used in C and Pascal.
A common use of call-with-current-continuation is for structured, non-local exits from loops or procedure bodies, but in fact call-with-current-continuation is useful for implementing a wide variety of advanced control structures.
Most programming languages incorporate one or more special-purpose escape constructs with names like exit, return, or even goto. In 1965, however, Peter Landin [22] invented a general purpose escape operator called the J-operator. John Reynolds [26] described a simpler but equally powerful construct in 1972. The catch special form described by Sussman and Steele in the 1975 report on Scheme is exactly the same as Reynolds’s construct, though its name came from a less general construct in MacLisp. Several Scheme implementors noticed that the full power of the catch construct could be provided by a procedure instead of by a special syntactic construct, and the name call-with-current-continuation was coined in 1982. This name is descriptive, but opinions differ on the merits of such a long name, and some people use the name call/cc instead.
The dynamic-wind procedure was added more recently in R5RS. It enables implementing a number of abstractions related to continuations, such as implementing a general dynamic environment, and making sure that finalization code runs when some dynamic extent expires. More generally, the dynamic-wind procedure provides a guarantee that
(dynamic-wind before thunk after)cannot return unless both before and after have been evaluated, in that order. They might be evaluated several times, but always in that order, and these evaluations are never nested. As this guarantee is crucial for enabling many of the uses of call-with-current-continuation and dynamic-wind, both are specified jointly.
Many computations conceptually return several results. Scheme expressions implementing such computations can return the results as several values using the values procedure. Of course, such expressions could alternatively return the results as a single compound value, such as a list, vector, or a record. However, values in programs usually represent conceptual wholes; in many cases, multiple results yielded by a computation lack this coherence. Moreover, this would be inefficient in many implementations, and a compiler would need to perform significant optimization to remove the boxing and unboxing inherent in packaging multiple results into a single values. Most importantly, the mechanism for multiple values in Scheme establishes a standard policy for returning several results from an expression, which makes constructing interfaces and using them easier.
R6RS does not specify the semantics of multiple values completely. In particular, it does not specify what happens when more than one value (or zero values) are returned to a continuation that implicitly accepts only one value. In particular:
((lambda (x) x) (values 1 2))Whether an implementation must raise an exception when evaluating such an expression, or should exhibit some other, non-exceptional behavior is a contentious issue. Variations of two different and fundamentally incompatible positions on this issue exist, each with its own merits:
Passing the wrong number of values to a continuation is typically an error, one that implementations ideally detect and report.
There is no such thing as returning the wrong number of values to a continuation. In particular, continuations not created by begin or call-with-values should ignore all but the first value, and treat zero values as one unspecified value.
R6RS allows an implementation to take either position. Moreover, it allows an implementation to let set!, vector-set!, and other effect-only operators to pass zero values to their continuations, preventing a program to make obscure use of the return value. This causes a potential compatibility problem with R5RS, which specifies that such expression return a single unspecified value, but the benefits of the change were deemed to outweigh the costs.
While the first subform of <srpattern> of a <syntax rule> in a syntax-rules form (see report section on “Macro transformers”) may be an identifier, the identifier is not involved in the matching and is not considered a pattern variable or literal identifier. This is actually important, as the identifier is most often the keyword used to identify the macro. The scope of the keyword is determined by the binding form or syntax definition that binds it to the associated macro transformer. If the keyword were a pattern variable or literal identifier, then the template that follows the pattern would be within its scope regardless of whether the keyword were bound by let-syntax, letrec-syntax, or define-syntax.