Rationale For International Standard - Programming Languages - C
Rationale For International Standard - Programming Languages - C
International Standard—
Programming Languages—
C
Revision 2
20 October 1999
0. Introduction .............................................................................................................................. 1
0.1 Organization of the document ............................................................................................ 5
1. Scope ........................................................................................................................................ 6
4. Conformance ............................................................................................................................ 7
5. Environment ............................................................................................................................. 9
5.1 Conceptual models ............................................................................................................. 9
5.1.1 Translation environment.............................................................................................. 9
5.1.2 Execution environments ............................................................................................ 11
5.2 Environmental considerations .......................................................................................... 13
5.2.1 Character sets ............................................................................................................ 13
5.2.2 Character display semantics ...................................................................................... 17
5.2.3 Signals and interrupts ................................................................................................ 18
5.2.4 Environmental limits ................................................................................................. 18
6. Language ................................................................................................................................ 23
6.2 Concepts........................................................................................................................... 23
6.2.1 Scopes of identifiers .................................................................................................. 23
6.2.2 Linkages of identifiers............................................................................................... 24
6.2.3 Name spaces of identifiers ........................................................................................ 26
6.2.4 Storage durations of objects ...................................................................................... 26
6.2.5 Types ......................................................................................................................... 28
6.2.6 Representations of types............................................................................................ 30
6.2.7 Compatible type and composite type......................................................................... 31
6.3 Conversions...................................................................................................................... 31
6.3.1 Arithmetic operands .................................................................................................. 31
6.3.2 Other operands .......................................................................................................... 35
6.4 Lexical Elements .............................................................................................................. 37
6.4.1 Keywords................................................................................................................... 37
6.4.2 Identifiers................................................................................................................... 37
6.4.3 Universal character names......................................................................................... 38
6.4.4 Constants ................................................................................................................... 38
6.4.5 String literals ............................................................................................................. 41
6.4.6 Punctuators ................................................................................................................ 43
6.4.7 Header names ............................................................................................................ 44
6.4.8 Preprocessing numbers.............................................................................................. 44
6.4.9 Comments.................................................................................................................. 44
6.5 Expressions ...................................................................................................................... 45
6.5.1 Primary expressions .................................................................................................. 48
6.5.2 Postfix operators........................................................................................................ 48
6.5.3 Unary operators ......................................................................................................... 51
6.5.4 Cast operators ............................................................................................................ 53
i
CONTENTS
iii
CONTENTS
v
C9X RATIONALE WG14/N897 J11/99-032
0. Introduction
This Rationale summarizes the deliberations of NCITS J11 (formerly X3J11) and SC22 WG14,
5 respectively the ANSI Technical Committee and ISO/IEC JTC 1 Working Group, charged with
revising the International Standard for the C programming language; and it retains much of the text
of the Rationale for the original ANSI Standard (ANSI X3.159-1989, the so-called “C89”). This
document has been published along with the draft Standard to assist the process of formal public
review.
10
There have been several changes to the Standard already. C89 was quickly adopted as an
International Standard (ISO/IEC 9899:1990, commonly called “C90”), with changes to clause and
subclause numbering to conform to ISO practices. Since then, there have been two Technical
Corrigenda and one Amendment, AMD1; and those three documents, together with C90 itself,
15 compose the current International Standard, (“C95”). The draft Standard is often called “C9X.”
J11 represents a cross-section of the C community in the United States: it consists of about twenty
or thirty members representing hardware manufacturers, vendors of compilers and other software
development tools, software designers, consultants, academics, authors, applications programmers,
20 and others. WG14’s participants are representatives of national standards bodies such as AFNOR,
ANSI, BSI, DIN and DS. In this Rationale, the unqualified “Committee” refers to J11 and WG14
working together to create C9X.
Upon publication of the new Standard, the primary role of the Committee will be to offer
25 interpretations of the Standard. It will consider and respond to all correspondence it receives.
The Committee’s overall goal was to develop a clear, consistent, and unambiguous Standard for the
C programming language which codifies the common, existing definition of C and which promotes
the portability of user programs across C language environments.
30
The original X3J11 charter clearly mandated codifying common existing practice, and the C89
Committee held fast to precedent wherever that was clear and unambiguous. The vast majority of
the language defined by C89 was precisely the same as defined in Appendix A of the first edition of
The C Programming Language by Brian Kernighan and Dennis Ritchie, and as was implemented in
35 almost all C translators of the time. (This document is hereinafter referred to as K&R.)
K&R was not the only source of “existing practice.” Much work had been done over the years to
improve the C language by addressing its weaknesses, and the C89 Committee formalized
enhancements of proven value which had become part of the various dialects of C. This practice
40 has continued in the present Committee.
Existing practice, however, has not always been consistent. Various dialects of C have approached
problems in different and sometimes diametrically opposed ways. This divergence has happened
for several reasons. First, K&R, which once served as the language specification for almost all C
45 translators, is imprecise in some areas (thereby allowing divergent interpretations), and it does not
1
WG14/N897 J11/99-032 C9X RATIONALE
address some issues (such as a complete specification of a library) important for code portability.
Second, as the language has matured over the years, various extensions have been added in
different dialects to address limitations and weaknesses of the language; but these extensions have
not been consistent across dialects.
5
One of the C89 Committee’s goals was to consider such areas of divergence and to establish a set
of clear, unambiguous rules consistent with the rest of the language. This effort included the
consideration of extensions made in various C dialects, the specification of a complete set of
required library functions, and the development of a complete, correct syntax for C.
10
Much of the Committee’s work has always been in large part a balancing act. The C89 Committee
tried to improve portability while retaining the definition of certain features of C as machine-
dependent, it attempted to incorporate valuable new ideas without disrupting the basic structure and
fabric of the language, and it tried to develop a clear and consistent language without invalidating
15 existing programs. All of the goals were important and each decision was weighed in the light of
sometimes contradictory requirements in an attempt to reach a workable compromise.
In specifying a standard language, the C89 Committee used several principles which continue to
guide our deliberations today. The most important of these are:
20
Existing code is important, existing implementations are not. A large body of C code exists of
considerable commercial value. Every attempt has been made to ensure that the bulk of this code
will be acceptable to any implementation conforming to the Standard. The C89 Committee did not
want to force most programmers to modify their C programs just to have them accepted by a
25 conforming translator.
On the other hand, no one implementation was held up as the exemplar by which to define C. It
was assumed that all existing implementations must change somewhat to conform to the Standard.
30 C code can be portable. Although the C language was originally born with the UNIX operating
system on the DEC PDP-11, it has since been implemented on a wide variety of computers and
operating systems. It has also seen considerable use in cross-compilation of code for embedded
systems to be executed in a free-standing environment. The C89 Committee attempted to specify
the language and the library to be as widely implementable as possible, while recognizing that a
35 system must meet certain minimum criteria to be considered a viable host or target for the language.
C code can be non-portable. Although it strove to give programmers the opportunity to write
truly portable programs, the C89 Committee did not want to force programmers into writing
portably, to preclude the use of C as a “high-level assembler:” the ability to write machine-specific
40 code is one of the strengths of C. It is this principle which largely motivates drawing the distinction
between strictly conforming program and conforming program (§4).
Avoid “quiet changes.” Any change to widespread practice altering the meaning of existing code
causes problems. Changes that cause code to be so ill-formed as to require diagnostic messages are
45 at least easy to detect. As much as seemed possible consistent with its other goals, the C89
Committee avoided changes that quietly alter one valid program to another with different
semantics, that cause a working program to work differently without notice. In important places
2
C9X RATIONALE WG14/N897 J11/99-032
where this principle is violated, both the C89 Rationale and this Rationale point out a QUIET
CHANGE.
A standard is a treaty between implementor and programmer. Some numerical limits were
5 added to the Standard to give both implementors and programmers a better understanding of what
must be provided by an implementation, of what can be expected and depended upon to exist.
These limits were, and still are, presented as minimum maxima (that is, lower limits placed on the
values of upper limits specified by an implementation) with the understanding that any implementor
is at liberty to provide higher limits than the Standard mandates. Any program that takes advantage
10 of these more tolerant limits is not strictly conforming, however, since other implementations are at
liberty to enforce the mandated limits.
Keep the spirit of C. The C89 Committee kept as a major goal to preserve the traditional spirit of
C. There are many facets of the spirit of C, but the essence is a community sentiment of the
15 underlying principles upon which the C language is based. Some of the facets of the spirit of C can
be summarized in phrases like
The last proverb needs a little explanation. The potential for efficient code generation is one of the
25 most important strengths of C. To help ensure that no code explosion occurs for what appears to be
a very simple operation, many operations are defined to be how the target machine’s hardware does
it rather than by a general abstract rule. An example of this willingness to live with what the
machine does can be seen in the rules that govern the widening of char objects for use in
expressions: whether the values of char objects widen to signed or unsigned quantities typically
30 depends on which byte operation is more efficient on the target machine.
One of the goals of the C89 Committee was to avoid interfering with the ability of translators to
generate compact, efficient code. In several cases the C89 Committee introduced features to
improve the possible efficiency of the generated code; for instance, floating point operations may be
35 performed in single-precision if both operands are float rather than double.
At the WG14 meeting in Tokyo, Japan, in July 1994, the original principles were re-endorsed and
the following new ones were added:
40 Support international programming. During the initial standardization process, support for
internationalization was something of an afterthought. Now that internationalization has become an
important topic, it should have equal visibility. As a result, all revision proposals shall be reviewed
with regard to their impact on internationalization as well as for other technical merit.
45 Codify existing practice to address evident deficiencies. Only those concepts that have some
prior art should be accepted. (Prior art may come from implementations of languages other than C.)
3
WG14/N897 J11/99-032 C9X RATIONALE
Unless some proposed new feature addresses an evident deficiency that is actually felt by more
than a few C programmers, no new inventions should be entertained.
Minimize incompatibilities with C90 (ISO/IEC 9899:1990). It should be possible for existing C
5 implementations to gradually migrate to future conformance, rather than requiring a replacement of
the environment. It should also be possible for the vast majority of existing conforming programs
to run unchanged.
Minimize incompatibilities with C++. The Committee recognizes the need for a clear and
10 defensible plan for addressing the compatibility issue with C++. The Committee endorses the
principle of maintaining the largest common subset clearly and from the outset. Such a principle
should satisfy the requirement to maximize overlap of the languages while maintaining a distinction
between them and allowing them to evolve separately.
15 The Committee is content to let C++ be the big and ambitious language. While some features of
C++ may well be embraced, it is not the Committee’s intention that C become C++.
Maintain conceptual simplicity. The Committee prefers an economy of concepts that do the job.
Members should identify the issues and prescribe the minimal amount of machinery that will solve
20 the problems. The Committee recognizes the importance of being able to describe and teach new
concepts in a straightforward and concise manner.
During the revision process, it was important to consider the following observations:
25 • Regarding the 11 principles, there is a tradeoff between them—none is absolute. However, the
more the Committee deviates from them, the more rationale will be needed to explain the
deviation.
• There had been a very positive reception of the standard from both the user and vendor
30 communities.
• The standard was not considered to be broken. Rather, the revision was needed to track
emerging and/or changing technologies and internationalization requirements.
35 • Most users of C view it as a general-purpose high-level language. While higher level constructs
can be added, they should be done so only if they don’t contradict the basic principles.
• There are a good number of useful suggestions to be found from the public comments and
defect report processing.
40
Areas to which the Committee looked when revising the C Standard included:
• Incorporate AMD1.
4
C9X RATIONALE WG14/N897 J11/99-032
This Rationale focuses primarily on additions, clarifications, and changes made to the C language.
It is not a rationale for the C language as a whole: the C89 Committee was charged with codifying
25 an existing language, not designing a new one. No attempt is made in this Rationale to defend the
pre-existing syntax of the language, such as the syntax of declarations or the binding of operators.
The Standard is contrived as carefully as possible to permit a broad range of implementations, from
direct interpreters to highly optimizing compilers with separate linkers, from ROM-based
embedded microcomputers to multi-user multi-processing host systems. A certain amount of
30 specialized terminology has therefore been chosen to minimize the bias toward compiler
implementations shown in K&R.
This Rationale discusses some language or library features which were not adopted into the
Standard. These are usually features which are popular in some C implementations, so that a user
35 of those implementations might question why they do not appear in the Standard.
This Rationale is organized to parallel the Standard as closely as possible to facilitate finding
40 relevant discussions. Some subclauses of the Standard are absent from this Rationale: this
indicates that the Committee thought no special comment was necessary. Where a given discussion
touches on several areas, attempts have been made to include cross references within the text. Such
references, unless they specify the Standard or the Rationale, are deliberately ambiguous.
45 This document has one appendix called MSE which brings together information on the Multibyte
Support Extensions (MSE) that were added to C90 by AMD1. This is essentially the Rationale for
5
WG14/N897 J11/99-032 C9X RATIONALE
AMD1; and it was kept largely unchanged because it was thought that it would be clearer to have
the MSE rationale in one place, as opposed to scattered throughout the document.
Just as the Standard proper excludes all examples, footnotes, references, and informative annexes,
5 this Rationale is not part of the Standard. The C language is defined by the Standard alone. If any
part of this Rationale is not in accord with that definition, the Committee would very much like to
be so informed.
1. Scope
10
2. Normative References
15 The definitions of object, bit, byte, and alignment reflect a strong consensus, reached after
considerable discussion, about the fundamental nature of the memory organization of a C
environment:
(Thus, for instance, on a machine with 36-bit words, a byte can be defined to consist of 9, 12, 18, or
25 36 bits, these numbers being all the exact divisors of 36 which are not less than 8.) These strictures
codify the widespread presumption that any object can be treated as an array of characters, the size
of which is given by the sizeof operator with that object’s type as its operand.
These definitions do not preclude “holes” in struct objects. Such holes are in fact often
30 mandated by alignment and packing requirements. The holes simply do not participate in
representing the composite value of an object.
The definition of object does not employ the notion of type. Thus an object has no type in and of
itself. However, since an object may only be designated by an lvalue (see §6.3.2.1), the phrase “the
35 type of an object” is taken to mean, here and in the Standard, “the type of the lvalue designating this
object,” and “the value of an object” means “the contents of the object interpreted as a value of the
type of the lvalue designating the object.”
The concepts of multibyte character, wide character, universal character, and extended character
40 have been added to C to support very large character sets (see §5.2.1 and §MSE.1).
The terms unspecified behavior, undefined behavior, and implementation-defined behavior are used
6
C9X RATIONALE WG14/N897 J11/99-032
to categorize the result of writing programs whose properties the Standard does not, or cannot,
completely describe. The goal of adopting this categorization is to allow a certain variety among
implementations which permits quality of implementation to be an active force in the marketplace
as well as to allow certain popular extensions, without removing the cachet of conformance to the
5 Standard. Informative Annex G of the Standard catalogs those behaviors which fall into one of
these three categories.
Unspecified behavior gives the implementor some latitude in translating programs. This latitude
does not extend as far as failing to translate the program, however, because all possible behaviors
10 are “correct” in the sense that they don’t cause undefined behavior in any implementation.
Undefined behavior gives the implementor license not to catch certain program errors that are
difficult to diagnose. It also identifies areas of possible conforming language extension: the
implementor may augment the language by providing a definition of the officially undefined
15 behavior.
25 A new feature of C9X: While responding to a Defect Report filed against C89, the Committee
came to realize that the term, “implementation-defined,” was sometimes being used in the sense of
“implementation must document” when dealing with locales. The term, “locale-specific behavior,”
already in C89, was then used extensively in C95 to distinguish those properties of locales which
can appear in a strictly conforming program. Because the presence or absence of a specific locale
30 is, with two exceptions, implementation-defined, some users of the Standard were confused as to
whether locales could be used at all in strictly conforming programs.
A successful call to setlocale has side effects, known informally as “setting the contents of the
current locale,” which can alter the subsequent output of the program. A program whose output is
35 altered only by such side effects—for example, because the decimal point character has changed—
is still strictly conforming.
A program whose output is affected by the value returned by a call to setlocale might not be
strictly conforming. If the only way in which the result affects the final output is by determining,
40 directly or indirectly, whether to make another call to setlocale, then the program remains
strictly conforming; but if the result affects the output in some other way, then it does not.
4. Conformance
45 The three-fold definition of conformance is used to broaden the population of conforming programs
7
WG14/N897 J11/99-032 C9X RATIONALE
and distinguish between conforming programs using a single implementation and portable
conforming programs.
A strictly conforming program is another term for a maximally portable program. The goal is to
5 give the programmer a fighting chance to make powerful C programs that are also highly portable,
without seeming to demean perfectly useful C programs that happen not to be portable, thus the
adverb strictly.
By defining conforming implementations in terms of the programs they accept, the Standard leaves
10 open the door for a broad class of extensions as part of a conforming implementation. By defining
both conforming hosted and conforming freestanding implementations, the Standard recognizes the
use of C to write such programs as operating systems and ROM-based applications, as well as more
conventional hosted applications. Beyond this two-level scheme, no additional subsetting is
defined for C, since the C89 Committee felt strongly that too many levels dilutes the effectiveness
15 of a standard.
Conforming program is thus the most tolerant of all categories, since only one conforming
implementation need accept it. The primary limitation on this license is §5.1.1.3.
20 Diverse parts of the Standard comprise the “treaty” between programmers and implementors
regarding various name spaces: if the programmer follows the rules of the Standard the
implementation will not impose any further restrictions or surprises:
• A strictly conforming program can use only a restricted subset of the identifiers that
25 begin with underscore (§7.1.3). Identifiers and keywords are distinct (§6.4.1).
Otherwise, programmers can use whatever internal names they wish; a conforming
implementation is guaranteed not to use conflicting names of the form reserved for
the programmer. (Note, however, the class of identifiers which are identified in
§7.26 as possible future library names.)
30
• The external functions defined in, or called within, a portable program can be named
whatever the programmer wishes, as long as these names are distinct from the
external names defined by the Standard library (§7). External names in a maximally
portable program must be distinct within the first 31 characters (in C95, the first 6
35 characters mapped into one case) (see §5.2.4.1 and §6.4.2).
40 • Each function called within a maximally portable program must either be defined
within some source file of the program or else be a function in the Standard library.
One proposal long entertained by the C89 Committee was to mandate that each implementation
have a translation-time switch for turning off extensions and making a pure Standard-conforming
45 implementation. It was pointed out, however, that virtually every translation-time switch setting
effectively creates a different “implementation,” however close may be the effect of translating with
two different switch settings. Whether an implementor chooses to offer a family of conforming
8
C9X RATIONALE WG14/N897 J11/99-032
15 5. Environment
Because C has seen widespread use as a cross-compiled cross-compilation language, a clear
distinction must be made between translation and execution environments. The C89 preprocessor,
for instance, is permitted to evaluate the expression in a #if directive using the long integer or
20 unsigned long integer arithmetic native to the translation environment: these integers must
comprise at least 32 bits, but need not match the number of bits in the execution environment. In
C9X, this arithmetic must be done in intmax_t or uintmax_t, which must comprise at least 64
bits and must match the execution environment. Other translation time arithmetic, however, such
as type casting and floating point arithmetic, must more closely model the execution environment
25 regardless of translation environment.
The as if principle is invoked repeatedly in this Rationale. The C89 Committee found that
30 describing various aspects of the C language, library, and environment in terms of concrete models
best serves discussion and presentation. Every attempt has been made to craft the models so that
implementors are constrained only insofar as they must bring about the same result, as if they had
implemented the presentation model; often enough the clearest model would make for the worst
implementation.
35
5.1.1 Translation environment
40 The terms source file, external linkage, linked, libraries, and executable program all imply a
conventional compiler/linker combination. All of these concepts have shaped the semantics of C,
however, and are inescapable even in an interpreted environment. Thus, while implementations are
not required to support separate compilation and linking with libraries, in some ways they must
behave as if they do.
9
WG14/N897 J11/99-032 C9X RATIONALE
Perhaps the greatest undesirable diversity among pre-C89 implementations can be found in
5 preprocessing. Admittedly a distinct and primitive language superimposed upon C, the
preprocessing commands accreted over time, with little central direction, and with even less
precision in their documentation. This evolution has resulted in a variety of local features, each
with its ardent adherents: K&R offers little clear basis for choosing one over the other.
10 The consensus of the C89 Committee is that preprocessing should be simple and overt, that it
should sacrifice power for clarity. For instance, the macro invocation f(a,b) should assuredly
have two actual arguments, even if b expands to c,d; and the formal definition of f must call for
exactly two arguments. Above all, the preprocessing sub-language should be specified precisely
enough to minimize or eliminate dialect formation. To clarify the nature of preprocessing, the
15 translation from source text to tokens is spelled out as a number of separate phases. The separate
phases need not actually be present in the translator, but the net effect must be as if they were. The
phases need not be performed in a separate preprocessor, although the definition certainly permits
this common practice. Since the preprocessor need not know anything about the specific properties
of the target, a machine-independent implementation is permissible. The C89 Committee deemed
20 that it was outside the scope of its mandate to require that the output of the preprocessing phases be
available as a separate translator output file.
The phases of translation are spelled out to resolve questions raised about the precedence of
different parses. Can a #define begin a comment? (No.) Is backslash/new-line permitted within
25 a trigraph? (No.) Must a comment be contained within one #include file? (Yes.) And so on.
The Rationale on preprocessing (§6.10) discusses the reasons for many of the decisions that shaped
the specification of the phases of translation.
A backslash immediately before a newline has long been used to continue string literals, as well as
30 preprocessing command lines. In the interest of easing machine generation of C, and of
transporting code to machines with restrictive physical line lengths, the C89 Committee generalized
this mechanism to permit any token to be continued by interposing a backslash/newline sequence.
In translation phase 4, the syntactic category preprocessing-file applies to each included file
35 separately from the file it is included into. Thus an included file cannot contain, for example,
unbalanced #else or #elif directives.
5.1.1.3 Diagnostics
40 By mandating some form of diagnostic message for any program containing a syntax error or
constraint violation, the Standard performs two important services. First, it gives teeth to the
concept of erroneous program, since a conforming implementation must distinguish such a
program from a valid one. Second, it severely constrains the nature of extensions permissible to a
conforming implementation.
45
The Standard says nothing about the nature of the diagnostic message, which could simply be
10
C9X RATIONALE WG14/N897 J11/99-032
“syntax error”, with no hint of where the error occurs. (An implementation must, of course,
describe what translator output constitutes a diagnostic message, so that the user can recognize it as
such.) The C89 Committee ultimately decided that any diagnostic activity beyond this level is an
issue of quality of implementation, and that market forces would encourage more useful
5 diagnostics. Nevertheless, the C89 Committee felt that at least some significant class of errors must
be diagnosed, and the class specified should be recognizable by all translators.
The Standard does not forbid extensions provided that they do not invalidate strictly conforming
programs, and the translator must allow extensions to be disabled as discussed in Rationale §4.
10 Otherwise, extensions to a conforming implementation lie in such realms as defining semantics for
syntax to which no semantics is ascribed by the Standard, or giving meaning to undefined behavior.
15 The definition of program startup in the Standard is designed to permit initialization of static
storage by executable code, as well as by data translated into the program image.
20 As little as possible is said about freestanding environments, since little is served by constraining
them.
25 The properties required of a hosted environment are spelled out in a fair amount of detail in order to
give programmers a reasonable chance of writing programs which are portable among such
environments.
35 The specification of argc and argv as arguments to main recognizes extensive prior practice.
argv[argc] is required to be a null pointer to provide a redundant check for the end of the list,
also on the basis of common practice.
main is the only function that may portably be declared either with zero or two arguments. (The
40 number of other functions’ arguments must match exactly between invocation and definition.) This
special case simply recognizes the widespread practice of leaving off the arguments to main when
the program does not access the program argument strings. While many implementations support
more than two arguments to main, such practice is neither blessed nor forbidden by the Standard; a
program that defines main with three arguments is not strictly conforming (see §K.5.1.).
45
Command line I/O redirection is not mandated by the Standard, as this was deemed to be a feature
11
WG14/N897 J11/99-032 C9X RATIONALE
5 Because C expressions can contain side effects, issues of sequencing are important in expression
evaluation (see §6.5 and Annexes C and D). Most operators impose no sequencing requirements,
but a few operators impose sequence points upon their evaluation: comma, logical-AND, logical-
OR, and conditional. In the expression (i = 1, a[i] = 0), for example, the side effect
(alteration to storage) specified by i = 1 must be completed before the expression a[i] = 0 is
10 evaluated.
Other sequence points are imposed by statement execution and completion of evaluation of a full
expression (see §6.8). Thus in fn(++a), the incrementation of a must be completed before fn is
called. In i = 1; a[i] = 0; the side effect of i = 1 must be complete before a[i] = 0 is
15 evaluated.
The notion of agreement has to do with the relationship between the abstract machine defining the
semantics and an actual implementation. An agreement point for some object or class of objects is
a sequence point at which the value of the object(s) in the real implementation must agree with the
20 value prescribed by the abstract semantics.
For example, compilers that hold variables in registers can sometimes drastically reduce execution
times. In a loop like
25 sum = 0;
for (i = 0; i < N; ++i)
sum += a[i];
both sum and i might be profitably kept in registers during the execution of the loop. Thus, the
30 actual memory objects designated by sum and i would not change state during the loop.
Such behavior is, of course, too loose for hardware-oriented applications such as device drivers and
memory-mapped I/O. The following loop looks almost identical to the previous example, but the
specification of volatile ensures that each assignment to *ttyport takes place in the same
35 sequence, and with the same values, as the abstract machine would have done.
12
C9X RATIONALE WG14/N897 J11/99-032
evaluation of the subexpression mask1 & mask2 could be performed prior to the loop in the real
implementation, assuming that neither mask1 nor mask2 appear as an operand of the address-of
5 (&) operator anywhere in the function. In the abstract machine, of course, this subexpression is
reevaluated at each loop iteration, but the real implementation is not required to mimic this
repetitiveness, because the variables mask1 and mask2 are not volatile and the same results
are obtained either way.
10 The previous example shows that a subexpression can be precomputed in the real implementation.
A question sometimes asked regarding optimization is, “Is the rearrangement still conforming if the
precomputed expression might raise a signal (such as division by zero)?” Fortunately for
optimizers, the answer is “Yes,” because any evaluation that raises a computational signal has fallen
into an undefined behavior (§6.5), for which any action is allowable.
15
Behavior is described in terms of an abstract machine to underscore, once again, that the Standard
mandates results as if certain mechanisms are used, without requiring those actual mechanisms in
the implementation. The Standard specifies agreement points at which the value of an object or
class of objects in an implementation must agree with the value ascribed by the abstract semantics.
20
Annex C to the Standard lists the sequence points specified in the body of the Standard.
The class of interactive devices is intended to include at least asynchronous terminals, or paired
display screens and keyboards. An implementation may extend the definition to include other input
25 and output devices, or even network inter-program connections, provided they obey the Standard’s
characterization of interactivity.
The C89 Committee ultimately came to remarkable unanimity on the subject of character set
requirements. There was strong sentiment that C should not be tied to ASCII, despite its heritage
and despite the precedent of Ada being defined in terms of ASCII. Rather, an implementation is
35 required to provide a unique character code for each of the printable graphics used by C, and for
each of the control codes representable by an escape sequence. (No particular graphic
representation for any character is prescribed; thus the common Japanese practice of using the glyph
“¥” for the C character “\” is perfectly legitimate.) Translation and execution environments may
have different character sets, but each must meet this requirement in its own way. The goal is to
40 ensure that a conforming implementation can translate a C translator written in C.
For this reason, and for economy of description, source code is described as if it undergoes the
same translation as text that is input by the standard library I/O routines: each line is terminated by
some newline character regardless of its external representation.
45
A new feature of C9X: C9X adds the concept of universal character name (UCN) (see §6.4.3) in
13
WG14/N897 J11/99-032 C9X RATIONALE
order to allow the use of any character in a C source, not just English characters. The primary goal
of the Committee was to enable the use of any “native” character in identifiers, string literals and
character constants, while retaining the portability objective of C.
5 With the concept of multibyte characters, “native” characters could be used in string literals and
character constants, but this use was very dependent on the implementation and did not usually
work in heterogenous environments. Also, this did not encompass identifiers.
Both the C and C++ Committees studied this situation, and the adopted solution was to introduce a
10 new notation for UCNs. Its general forms are \unnnn and \Unnnnnnnn, to designate a given
character according to its short name as described by ISO/IEC 10646. Thus, \unnnn can be used to
designate a Unicode character. This way, programs that must be fully portable may use virtually
any character from any script used in the world and still be portable, provided of course that if it
prints the character, the execution character set has representation for it.
15
Of course the notation \unnnn, like trigraphs, is not very easy to use in everyday programming; so
there is a mapping that links UCN and multibyte characters to enable source programs to stay
readable by users while maintaining portability. Given the current state of multibyte encodings, this
mapping is specified to be implementation-defined; but an implementation can provide the users
20 with utility programs that do the conversion from UCNs to “native” multibytes or vice versa, thus
providing a way to exchange source files between implementations using the UCN notation.
UCN models
25 Once this was adopted, there was still one problem, how to specify UCNs in the Standard. Both the
C and C++ Committees studied this situation and the available solutions, and drafted three models:
A. Convert everything to UCNs in basic source characters as soon as possible, that is, in translation
phase 1.
30
B. Use native encodings where possible, UCNs otherwise.
C. Convert everything to wide characters as soon as possible using an internal encoding that
encompasses the entire source character set and all UCNs.
35
Furthermore, in any place where a program could tell which model was being used, the standard
should try to label those corner cases as undefined behavior.
The C++ committee defined its Standard in terms of model A, just because that was the clearest to
40 specify (used the fewest hypothetical constructs) because the basic source character set is a well-
defined finite set.
The situation is not the same for C given the already existing text for the standard, which allows
multibyte characters to appear almost anywhere (the most notable exception being in identifiers),
45 and given the more low-level (or “close to the metal”) nature of some uses of the language.
14
C9X RATIONALE WG14/N897 J11/99-032
Therefore, the C Committee agreed in general that model B, keeping UCNs and native characters
until as late as possible, is more in the “spirit of C” and, while probably more difficult to specify, is
more able to encompass the existing diversity. The advantage of model B is also that it might
encompass more programs and users’ intents than the two others, particularly if shift states are
5 significant in the source text as is often the case in East Asia.
In any case, translation phase 1 begins with an implementation-defined mapping; and such mapping
can choose to implement model A or C (but the implementation must document it). As a by-
product, a strictly conforming program cannot rely on the specifics handled differently by the three
10 models: examples of non-strict conformance include handling of shift states inside strings and calls
like fopen("\\ubeda\\file.txt","r") and #include "sys\udefault.h". Shift
states are guaranteed to be handled correctly, however, as long as the implementation performs no
mapping at the beginning of phase 1; and the two specific examples given above can be made much
more portable by rewriting these as fopen("\\" "ubeda\\file.txt", "r") and
15 #include "sys/udefault.h".
Trigraph sequences were introduced in C89 as alternate spellings of some characters to allow the
20 implementation of C in character sets which do not provide a sufficient number of non-alphabetic
graphics.
Implementations are required to support these alternate spellings, even if the character set in use is
ASCII, in order to allow transportation of code from systems which must use the trigraphs. AMD1
25 also added digraphs (see §6.4.6 and §MSE.4).
The C89 Committee faced a serious problem in trying to define a character set for C. Not all of the
character sets in general use have the right number of characters, nor do they support the graphical
symbols that C users expect to see. For instance, many character sets for languages other than
30 English resemble ASCII except that codes used for graphic characters in ASCII are instead used for
alphabetic characters or diacritical marks. C relies upon a richer set of graphic characters than most
other programming languages, so the representation of programs in character sets other than ASCII
is a greater problem than for most other programming languages.
35 ISO (the International Organization for Standardization) uses three technical terms to describe
character sets: repertoire, collating sequence, and codeset. The repertoire is the set of distinct
printable characters. The term abstracts the notion of printable character from any particular
representation; the glyphs R, R, R, R, R, R, and R, all represent the same element of the repertoire,
“upper-case-R”, which is distinct from “lower-case-r”. Having decided on the repertoire to be used
40 (C needs a repertoire of 91 characters plus whitespace), one can then pick a collating sequence
which corresponds to the internal representation in a computer. The repertoire and collating
sequence together form the codeset.
What is needed for C is to determine the necessary repertoire, ignore the collating sequence
45 altogether (it is of no importance to the language), and then find ways of expressing the repertoire in
a way that should give no problems with currently popular codesets.
15
WG14/N897 J11/99-032 C9X RATIONALE
C derived its repertoire from the ASCII codeset. Unfortunately, the ASCII repertoire is not a subset
of all other commonly used character sets; and widespread practice in Europe is not to implement
all of ASCII either, but to use some parts of its collating sequence for special national characters.
5
The solution is an internationally agreed-upon repertoire in terms of which an international
representation of C can be defined. ISO has defined such a standard, ISO/IEC 646, which describes
an invariant subset of ASCII.
10 The characters in the ASCII repertoire used by C and absent from the ISO/IEC 646 invariant
repertoire are:
#[]{}\|~^
15 Given this repertoire, the C89 Committee faced the problem of defining representations for the
absent characters. The obvious idea of defining two-character escape sequences fails because C
uses all the characters which are in the ISO/IEC 646 repertoire, so no single escape character is
available. The best that can be done is to use a trigraph: an escape digraph followed by a
distinguishing character.
20
?? was selected as the escape digraph because it is not used anywhere else in C except as noted
below; it suggests that something unusual is going on. The third character was chosen with an eye
to graphical similarity to the character being represented.
25 The sequence ?? cannot occur in a valid pre-C89 program except in strings, character constants,
comments, or header names. The character escape sequence '\?' (see §6.4.4.4) was introduced to
allow two adjacent question marks in such contexts to be represented as ?\?, a form distinct from
the escape digraph. The Committee makes no claims that a program written using trigraphs looks
attractive. As a matter of style, it may be wise to surround trigraphs with white space, so that they
30 stand out better in program text. Some users may wish to define preprocessing macros for some or
all of the trigraph sequences.
16
C9X RATIONALE WG14/N897 J11/99-032
between these large codes and the various byte-oriented media is necessary.
5 • Multibyte encodings of large character sets are necessary in I/O operations, in source
text comments, in source text string and character literals, and beginning with C9X,
in native language identifiers.
• The null character ('\0') may not be used as part of a multibyte encoding, except
for the one-byte null character itself. This allows existing functions which
manipulate strings to work transparently with multibyte sequences.
15
• Shift encodings (which interpret byte sequences in part on the basis of some state
information) must start out in a known (default) shift state under certain
circumstances such as the start of string literals.
The Standard defines a number of internal character codes for specifying “format-effecting actions
on display devices,” and provides printable escape sequences for each of them. These character
codes are clearly modeled after ASCII control codes, and the mnemonic letters used to specify their
25 escape sequences reflect this heritage. Nevertheless, they are internal codes for specifying the
format of a display in an environment-independent manner; they must be written to a text file to
effect formatting on a display device. The Standard states quite clearly that the external
representation of a text file (or data stream) may well differ from the internal form, both in character
codes and number of characters needed to represent a single internal code.
30
The distinction between internal and external codes most needs emphasis with respect to new-line.
NCITS L2, Codes and Character Sets (and now also ISO/IEC JTC 1/SC2/WG1, 8 Bit Character
Sets), uses the term to refer to an external code used for information interchange whose display
semantics specify a move to the next line. Although ISO/IEC 646 deprecates the combination of
35 the motion to the next line with a motion to the initial position on the line, the C Standard uses
new-line to designate the end-of-line internal code represented by the escape sequence '\n'.
While this ambiguity is perhaps unfortunate, use of the term in the latter sense is nearly universal
within the C community. But the knowledge that this internal code has numerous external
representations depending upon operating system and medium is equally widespread.
40
The alert sequence ('\a') was added by popular demand to replace, for instance, the ASCII BEL
code explicitly coded as '\007'.
Proposals to add '\e' for ASCII ESC ('\033') were not adopted because other popular
45 character sets have no obvious equivalent (see §6.4.4.4.)
17
WG14/N897 J11/99-032 C9X RATIONALE
The vertical tab sequence ('\v') was added since many existing implementations support it, and
since it is convenient to have a designation within the language for all the defined white space
characters.
5 The semantics of the motion control escape sequences carefully avoid the Western language
assumptions that printing advances left-to-right and top-to-bottom.
To avoid the issue of whether an implementation conforms if it cannot properly effect vertical tabs
(for instance), the Standard emphasizes that the semantics merely describe intent.
10
5.2.3 Signals and interrupts
Signals are difficult to specify in a system-independent way. The C89 Committee concluded that
about the only thing a strictly conforming program can do in a signal handler is to assign a value to
15 a volatile static variable which can be written uninterruptedly and promptly return. (The
header <signal.h> specifies a type sig_atomic_t which can be so written.) It is further
guaranteed that a signal handler will not corrupt the automatic storage of an instantiation of any
executing function, even if that function is called within the signal handler. No such guarantees can
be extended to library functions, with the explicit exceptions of longjmp (§7.13.2.1) and signal
20 (§7.14.1.1), since the library functions may be arbitrarily interrelated and since some of them have
profound effect on the environment.
Calls to longjmp are problematic, despite the assurances of §7.13.2.1. The signal could have
occurred during the execution of some library function which was in the process of updating
25 external state and/or static variables.
A second signal for the same handler could occur before the first is processed, and the Standard
makes no guarantees as to what happens to the second signal.
The C89 Committee agreed that the Standard must say something about certain capacities and
limitations, but just how to enforce these treaty points was the topic of considerable debate.
The Standard requires that an implementation be able to translate and execute some program that
meets each of the stated limits. This criterion was felt to give a useful latitude to the implementor
in meeting these limits. While a deficient implementation could probably contrive a program that
40 meets this requirement, yet still succeed in being useless, the C89 Committee felt that such
ingenuity would probably require more work than making something useful. The sense of both the
C89 and C9X Committees was that implementors should not construe the translation limits as the
values of hard-wired parameters, but rather as a set of criteria by which an implementation will be
judged.
45
Some of the limits chosen represent interesting compromises. The goal was to allow reasonably
18
C9X RATIONALE WG14/N897 J11/99-032
large portable programs to be written, without placing excessive burdens on reasonably small
implementations, some of which might run on machines with only 64K of memory. In C9X, the
minimum amount of memory for the target machine was raised to 512K. In addition, the
Committee recognized that smaller machines rarely serve as a host for a C compiler: programs for
5 embedded systems or small machines are almost always developed using a cross compiler running
on a personal computer or workstation. This allows for a great increase in some of the translation
limits.
C89’s minimum maximum limit of 257 cases in a switch statement allows coding of lexical
10 routines which can branch on any character (one of at least 256 values) or on the value EOF. This
has been extended to 1023 cases in C9X.
The requirement that a conforming implementation be able to translate and execute at least one
program that reaches each of the stated limits is not meant to excuse the implementation from doing
15 the best it can to translate and execute other programs. It was deemed infeasible to require
successful translation and execution of all programs not exceeding those limits. Many of these
limits require resources such as memory that a reasonable implementation might allocate from a
shared pool; so there is no requirement that all the limits be attained simultaneously. Requiring just
one acceptable program that attains each limit is simply meant to ensure conformance with these
20 requirements.
The C9X Committee reviewed several proposed changes to strengthen or clarify the wording on
conformance, especially with respect to translation limits. The belief was that it is simply not
practical to provide a specification which is strong enough to be useful, but which still allows for
25 real-world problems such as bugs. The Committee therefore chose to consider the matter a quality-
of-implementation issue, and to leave translation limits in the standard to give guidance.
Such a large body of C code has been developed for 8-bit byte machines that the integer sizes in
such environments must be considered normative. The prescribed limits are minima: an
implementation on a machine with 9-bit bytes can be conforming, as can an implementation that
35 defines int to be the same width as long. The negative limits have been chosen to accommodate
one’s-complement or sign-magnitude implementations, as well as the more usual two’s-
complement. The limits for the maxima and minima of unsigned types are specified as unsigned
constants (e.g., 65535u) to avoid surprising widening of expressions involving these extrema.
40 The macro CHAR_BIT makes available the number of bits in a char object. The C89 Committee
saw little utility in adding such macros for other data types.
The names associated with the short int types (SHRT_MIN, etc., rather than SHORT_MIN, etc.)
reflect prior art rather than obsessive abbreviation on the C89 Committee’s part.
45
5.2.4.2.2 Characteristics of floating types <float.h>
19
WG14/N897 J11/99-032 C9X RATIONALE
The characterization of floating point follows, with minor changes, that of the Fortran
standardization committee. The C89 Committee chose to follow the Fortran model in some part
out of a concern for Fortran-to-C translation, and in large part out of deference to the Fortran
5 committee’s greater experience with fine points of floating point usage. Note that the floating point
model adopted permits all common representations, including sign-magnitude and two’s-
complement, but precludes a logarithmic implementation.
The C89 Committee also endeavored to accommodate the IEEE 754 floating point standard by not
10 adopting any constraints on floating point which were contrary to that standard. IEEE 754 is now
an international standard, IEC 60559; and that is how it is referred to in C9X.
The term FLT_MANT_DIG stands for “float mantissa digits.” The Standard now uses the more
precise term significand rather than mantissa.
15
In C9X, all values except FLT_ROUNDS in <float.h> must be usable as static initializers.
The overflow and/or underflow thresholds may not be the same for all arithmetic operations. For
example, there is at least one machine where the overflow threshold for addition is twice as big as
20 for multiplication. Another implementation uses a pair of doubles to represent a long double.
In that implementation, the next representable long double value after 1.0L is 1.0L +
LDBL_MIN, yet, the difference between those two numbers (LDBL_MIN) is not b(1-p), otherwise
known as LDBL_EPSILON. Because of anomalies like these, there are few hard requirements on
the <float.h> values. But, the values in <float.h> should be in terms of the hardware
25 representation used to store floating point values in memory, not in terms of the effective accuracy
of operations, nor in terms of registers, and should apply to all operations. The representation
stored in memory may have padding bits and/or bytes that do not contribute to the value. The
padding should not be included in the <float.h> values.
30 Because of the practical difficulty involved in defining a uniform metric that all vendors would be
willing to follow (just computing the accuracy reliably could be a significant, and because the
importance of floating point accuracy differs greatly among users, the standard allows a great deal
of latitude in how an implementation documents the accuracy of the real and complex floating point
operations and functions.
35
Here are some ways that an implementation might address the need to define the accuracy:
digits correct
digits wrong
40 maximum Units in the Last Place (ULPs) error
maximum absolute error
maximum relative error
1
See X3J3 working document S8-112.
20
C9X RATIONALE WG14/N897 J11/99-032
There are two usages of the term ULP. One is in the context of differences between two numbers,
that is, f(x) differs from F(x) by 3 ULPs. The other is the value of the ULP of a number, that is, an
ULP of the value 1.0 is DBL_EPSILON. For this discussion, we are interested in the former; the
10 difference between the computed value and the infinitely precise value.
The error between two floating-point numbers in ULPs depends on the radix and the precision used
in representing the number, but not the exponent. With a decimal radix and 3 digits of precision,
the computed value 0.314e+1 differs from the value 0.31416e+1 by 0.16 ULPs. If both
15 numbers are scaled by the same power of the radix, for example, 0.314e+49 and
0.31416e+49, they still differ by 0.16 ULPs.
When the two numbers being compared span a power of the radix, the two possible ULP error
calculations differ by a factor of the radix. For a decimal radix and 3 digits of precision, consider
20 the two values 9.99e2 and 1.01e3. These are the two values adjacent to the value 1.00e3, a
power of the radix, in this number system. If 999 is the correct value and 1010 is the computed
value, the error is 11 ULPs; but, if 1010 is the correct value and 999 is the computed value, then the
error is 1.1 ULPs.
25 Some math functions such as those that do argument reduction modulo an approximation of π have
good accuracy for small arguments, but poor accuracy for large arguments. It is not unusual for an
implementation of the trigonometric functions to have zero bits correct in the computed result for
large arguments. For cases like this, an implementation might break the domain of the function into
disjoint regions and specify the accuracy in each region.
30
If an implementation documents worst case error, there is no requirement that it be the minimum
worst case error. That is, if a vendor believes that the worst case error for a function is around 5
ULPs, they could document it as 7 ULPs to be safe.
35 The Committee could not agree on upper limits on accuracy that all conforming implementations
must meet, for example, “addition is no worse than 2 ULPs for all implementations.” This is a
quality of implementation issue.
Implementations that conform to IEC 60559 have one half ULP accuracy in round-to-nearest mode,
40 and one ULP accuracy in the other three rounding modes, for the basic arithmetic operations and
square root. For other floating point arithmetics, it is a rare implementation that has worse than one
ULP accuracy for the basic arithmetic operations.
The accuracy of decimal-to-binary conversions and format conversions are discussed elsewhere in
45 the Standard.
21
WG14/N897 J11/99-032 C9X RATIONALE
For the math library functions, fast, correctly rounded 0.5 ULP accuracy remains a research
problem. Some implementations provide two math libraries, one being faster but less accurate than
the other.
5 The C9X Committee discussed the idea of allowing the programmer to find out the accuracy of
floating point operations and math functions during compilation (say, via macros) or during
execution (with a function call), but neither got enough support to warrant the change to the
Standard. The use of macros would require over one hundred symbols to name every math
function, for example, ULP_SINF, ULP_SIN, and ULP_SINL just for the real-valued sin
10 function. One possible function implementation might be a function that takes the name of the
operation or math function as a string, ulp_err("sin") for example, that would return a
double such as 3.5 to indicate the worst case error, with –1.0 indicating unknown error. But such
a simple scheme would likely be of very limited use given that so many functions have accuracies
that differ significantly across their domains. Constrained to worst case error across the entire
15 domain, most implementations would wind up reporting either unknown error or else a uselessly
large error for a very large percentage of functions. This would be useless because most programs
that care about accuracy are written in the first place to try to compensate for accuracy problems
that typically arise when pushing domain boundaries; and implementing something more useful like
the worst case error for a user-specified partition of the domain would be excessively difficult.
20
NaNs
C9X does not define the behavior of signaling NaNs, nor does it specify the interpretation of
NaN significands.
25
The IEC 60559 floating-point standard specifies quiet and signaling NaNs, but these terms can be
applied for some non-IEC 60559 implementations as well. For example, the VAX reserved
operand and the CRAY indefinite qualify as signaling NaNs. In IEC 60559 standard arithmetic,
operations that trigger a signaling NaN argument generally return a quiet NaN result provided no
30 trap is taken. Full support for signaling NaNs implies restartable traps, such as the optional traps
specified in the IEC 60559 floating-point standard.
The primary utility of quiet NaNs, as stated in IEC 60559, “to handle otherwise intractable
situations, such as providing a default value for 0.0/0.0,” is supported by this specification.
35
Other applications of NaNs may prove useful. Available parts of NaNs have been used to encode
auxiliary information, for example about the NaN’s origin. Signaling NaNs might be candidates
for filling uninitialized storage; and their available parts could distinguish uninitialized floating
objects. IEC 60559 signaling NaNs and trap handlers potentially provide hooks for maintaining
40 diagnostic information or for implementing special arithmetics.
However, C support for signaling NaNs, or for auxiliary information that could be encoded in
NaNs, is problematic. Trap handling varies widely among implementations. Implementation
mechanisms may trigger signaling NaNs, or fail to, in mysterious ways. The IEC 60559 floating-
45 point standard recommends that NaNs propagate; but it does not require this and not all
implementations do. And the floating-point standard fails to specify the contents of NaNs
22
C9X RATIONALE WG14/N897 J11/99-032
5 A draft version of the NCEG floating-point specification included signaling NaNs. It could serve
as a guide for implementation extensions in support of signaling NaNs.
6. Language
10 While more formal methods of language definition were explored, the C89 Committee decided
early on to employ the style of K&R: Backus-Naur Form for the syntax and prose for the
constraints and semantics. Anything more ambitious was considered to be likely to delay the
Standard, and to make it less accessible to its audience.
15 6.2 Concepts
C89 separated from the overloaded keywords for storage classes the various concepts of scope,
20 linkage, name space, and storage duration (see §6.2.2, §6.2.3 and §6.2.4.). This has traditionally
been a major area of confusion.
One source of dispute was whether identifiers with external linkage should have file scope even
when introduced within a block. K&R was vague on this point, and has been interpreted differently
25 by different pre-C89 implementations. For example, the following fragment would be valid in the
file scope scheme, while invalid in the block scope scheme:
30 first()
{
extern d_struct func();
/* ... */
}
35
second()
{
d_struct n = func();
}
40
While it was generally agreed that it is poor practice to take advantage of an external declaration
once it had gone out of scope, some argued that a translator had to remember the declaration for
checking anyway, so why not acknowledge this? The compromise adopted was to decree
essentially that block scope rules apply, but that a conforming implementation need not diagnose a
45 failure to redeclare an external identifier that had gone out of scope (undefined behavior).
23
WG14/N897 J11/99-032 C9X RATIONALE
A program relying on file scope rules may be valid under block scope rules but
5 behave differently, for instance, if d_struct were defined as type float rather
than struct data in the example above.
Although the scope of an identifier in a function prototype begins at its declaration and ends at the
end of that function’s declarator, this scope is ignored by the preprocessor. Thus an identifier in a
10 prototype having the same name as that of an existing macro is treated as an invocation of that
macro. For example:
#define status 23
void exit(int status);
15
generates an error, since the prototype after preprocessing becomes
#define status []
which is syntactically correct but semantically quite different from the intent.
Pre-C89 practice in this area was inconsistent. Some implementations avoided the renaming
45 problem simply by restricting internal linkage names by the same rules as the ones used for external
linkage. Others have disallowed a static declaration followed later by a defining instance, even
24
C9X RATIONALE WG14/N897 J11/99-032
though such constructs are necessary to declare mutually-recursive static functions. The
requirements adopted in C89 called for changes in some existing programs, but allowed for
maximum flexibility.
5 The definition model to be used for objects with external linkage was a major C89 standardization
issue. The basic problem was to decide which declarations of an object define storage for the
object, and which merely reference an existing object. A related problem was whether multiple
definitions of storage are allowed, or only one is acceptable. Pre-C89 implementations exhibit at
least four different models, listed here in order of increasing restrictiveness:
10
Common Every object declaration with external linkage, regardless of whether the keyword
extern appears in the declaration, creates a definition of storage. When all of the modules are
combined together, each definition with the same name is located at the same address in memory.
(The name is derived from common storage in Fortran.) This model was the intent of the original
15 designer of C, Dennis Ritchie.
Relaxed Ref/Def The appearance of the keyword extern in a declaration, regardless of whether
it is used inside or outside of the scope of a function, indicates a pure reference (ref), which does
not define storage. Somewhere in all of the translation units, at least one definition (def) of the
20 object must exist. An external definition is indicated by an object declaration in file scope
containing no storage class indication. A reference without a corresponding definition is an error.
Some implementations also will not generate a reference for items which are declared with the
extern keyword but are never used in the code. The UNIX operating system C compiler and
linker implement this model, which is recognized as a common extension to the C language (see
25 §K.5.11). UNIX C programs which take advantage of this model are standard conforming in their
environment, but are not maximally portable (not strictly conforming).
Strict Ref/Def This is the same as the relaxed ref/def model, save that only one definition is
allowed. Again, some implementations may decide not to put out references to items that are not
30 used. This is the model specified in K&R.
Initialization This model requires an explicit initialization to define storage. All other
declarations are references.
35 Figure 6.1 demonstrates the differences between the models. The intent is that Figure 6.1 shows
working programs in which the symbol i is neither undefined nor multiply defined.
The Standard model is a combination of features of the strict ref/def model and the initialization
model. As in the strict ref/def model, only a single translation unit contains the definition of a given
40 object because many environments cannot effectively or efficiently support the “distributed
definition” inherent in the common or relaxed ref/def approaches. However, either an initialization,
or an appropriate declaration without storage class specifier (see §6.9), serves as the external
definition. This composite approach was chosen to accommodate as wide a range of environments
and existing implementations as possible.
45
Figure 6.1: Comparison of identifier linkage models
25
WG14/N897 J11/99-032 C9X RATIONALE
5 Pre-C89 implementations varied considerably in the number of separate name spaces maintained.
The position adopted in the Standard is to permit as many separate name spaces as can be
distinguished by context, except that all tags (struct, union, and enum) comprise a single name
space.
It was necessary to clarify the effect on automatic storage of jumping into a block that declares local
storage (see §6.8.2.). While many implementations could traditionally allocate the maximum depth
of automatic storage upon entry to a function, the addition to C9X of the variable length array
15 feature (§6.7.5.2) forces the implementation to allocate some objects when the declaration is
encountered.
A new feature of C9X: C89 requires all declarations in a block to occur before any statements. On
the other hand, many languages similar to C (such as Algol 68 and C++) permit declarations and
20 statements to be mixed in an arbitrary manner. This feature has been found to be useful and has
been added to C9X.
Declarations that initialize variables can contain complex expressions and have arbitrary side-
26
C9X RATIONALE WG14/N897 J11/99-032
effects, and it is necessary to define when these take place, particularly when the flow of control
involves arbitrary jumps. There is a simple rule of thumb: the variable declared is created with an
unspecified value when the block is entered, but the initializer is evaluated and the value placed in
the variable when the declaration is reached in the normal course of execution. Thus a jump
5 forward past a declaration leaves it uninitialized, while a jump backwards will cause it to be
initialized more than once. If the declaration does not initialize the variable, it sets it to an
unspecified value even if this is not the first time the declaration has been reached.
The scope of a variable starts at its declaration. Therefore, although the variable exists as soon as
10 the block is entered, it cannot be referred to by name until its declaration is reached.
Example:
int j = 42;
15 {
int i = 0;
loop:
printf("I = %4d, ", i);
20 printf("J1 = %4d, ", ++j);
int j = i;
printf("J2 = %4d, ", ++j);
int k;
printf("K1 = %4d, ", k);
25 k = i * 10;
printf("K2 = %4d, ", k);
if (i % 2 == 0) goto skip;
int m = i * 5;
skip:
30 printf("M = %4d\n", m);
35 will output:
where “????” indicates an indeterminate value (and any use of an indeterminate value is undefined
behavior).
45
These rules have to be modified slightly for variable length arrays. The implementation will not
know how much space is required for the array until its declaration is reached, and so cannot create
27
WG14/N897 J11/99-032 C9X RATIONALE
A jump to a point after the declaration of a VLA is forbidden, because it would be possible to refer
to the VLA without creating it. Such a jump requires a diagnostic.
5
A jump to a point before the declaration of a VLA destroys the VLA.
A number of other approaches were considered, but there were problems with all of them. In
particular, this choice of rules ensures that VLAs can always be destroyed in the reverse order of
10 their creation, which is essential if they are placed on the stack.
To effect true reentrancy for functions in the presence of signals raised asynchronously (see §5.2.3),
an implementation must assure that the storage for function return values has automatic duration.
For example, the caller could allocate automatic storage for the return value and communicate its
15 location to the called function. (The typical case of return registers for small-sized types conforms
to this requirement: the calling convention of the implementation implicitly communicates the
return location to the called function.)
6.2.5 Types
void
void*
signed char
25 unsigned char
unsigned short
unsigned long
long double
_Bool
long long
40 unsigned long long
float _Imaginary
float _Complex
double _Imaginary
double _Complex
45 long dobule _Imaginary
long double _Complex
28
C9X RATIONALE WG14/N897 J11/99-032
C9X also allows extended integer types (see §7.8, <inttypes.h>, and §7.18, <stdint.h>)
and a boolean type (see §7.16, <stdbool.h>).
void is used primarily as the typemark for a function that returns no result. It may also be used as
5 the cast (void) to indicate explicitly that the value of an expression is to be discarded while
retaining the expression’s side effects. Finally, a function prototype list that has no arguments is
written as f(void), because f() retains its old meaning that nothing is said about the arguments.
Note that there is no such thing as a “void object.”
10 A “pointer to void,” void*, is a generic pointer capable of pointing to any object (except for bit-
fields and objects declared with the register storage class) without loss of information. A
pointer to void must have the same representation and alignment as a pointer to char; the intent
of this rule is to allow existing programs that call library functions such as memcpy and free to
continue to work. A pointer to void cannot be dereferenced, although such a pointer can be
15 converted to a normal pointer type which can be dereferenced. Pointers to other types coerce
silently to and from void* in assignments, function prototypes, comparisons, and conditional
expressions, whereas other pointer type clashes are invalid. It is undefined what will happen if a
pointer of some type is converted to void*, and then the void* pointer is converted to a type
with a stricter alignment requirement. Three types of char are specified: signed, plain, and
20 unsigned. A plain char may be represented as either signed or unsigned depending upon the
implementation, as in prior practice. The type signed char was introduced in C89 to make
available a one-byte signed integer type on those systems which implement plain char as
unsigned char. For reasons of symmetry, the keyword signed is allowed as part of the type
name of other integer types. Two varieties of the integer types are specified: signed and
25 unsigned. If neither specifier is used, signed is assumed. The only unsigned type in K&R is
unsigned int.
29
WG14/N897 J11/99-032 C9X RATIONALE
To avoid confusion, long float as a synonym for double was retired in C89.
5 Floating types of different widths (double wider than float and long double wider than
double) facilitate porting code that, intentionally or not, depends on differences in type widths.
Many results are exact or correctly rounded when computed with twice the number of digits of
precision as the data. For example, the calculation
10 float d, x, y, z, w;
/* ... */
d = (double) x * y - (double) z * w;
yields a correctly rounded determinant if double has twice the precision of float and the
15 individual operations are correctly rounded. (The casts to double are unnecessary if the
minimum evaluation format is double or long double.)
A new feature of C9X: Complex types were added to C as part of the effort to make C suitable
and attractive for general numerical programming. Complex arithmetic is used heavily in certain
20 important application areas.
The underlying implementation of the complex types is Cartesian, rather than polar, for overall
efficiency and consistency with other programming languages. The implementation is explicitly
stated so that characteristics and behaviors can be defined simply and unambiguously.
25
Enumerations permit the declaration of named constants in a more convenient and structured
fashion than does #define. Both enumeration constants and variables behave like integer types
for the sake of type checking, however.
40 The C89 Committee adopted the second alternative on the grounds that this approach most clearly
reflects common practice. Doing away with enumerations altogether would invalidate a fair
amount of existing code; stronger typing than integer creates problems, for example, with arrays
indexed by enumerations.
30
C9X RATIONALE WG14/N897 J11/99-032
The C89 Committee explicitly required binary representation of integers on the grounds that this
stricture was implicit in any case:
5
• Bit fields are specified by a number of bits, with no mention of “invalid integer”
representation. The only reasonable encoding for such bit fields is binary.
• The integer formats for printf suggest no provision for “invalid integer” values,
10 implying that any result of bitwise manipulation produces an integer result which
can be printed by printf.
The restriction to binary numeration systems rules out such curiosities as Gray code and makes
possible arithmetic definitions of the bitwise operators on unsigned types.
20
6.2.7 Compatible type and composite type
The concepts of compatible type and composite type were introduced to allow C89 to discuss those
situations in which type declarations need not be identical. These terms are especially useful in
25 explaining the relationship between an incomplete type and a completed type. With the addition of
variable length arrays (§6.7.5.2) in C9X, array type compatibility was extended so that variable
length arrays are compatible with both an array of known constant size and an array with an
incomplete type.
30 Structure, union, or enumeration type declarations in two different translation units do not formally
declare the same type, even if the text of these declarations come from the same header file, since
the translation units are themselves disjoint. The Standard thus specifies additional compatibility
rules for such types so that two such declarations are compatible if they are sufficiently similar.
35 6.3 Conversions
31
WG14/N897 J11/99-032 C9X RATIONALE
The unsigned preserving approach calls for promoting the two smaller unsigned types to
unsigned int. This is a simple rule, and yields a type which is independent of execution
5 environment.
The value preserving approach calls for promoting those types to signed int if that type can
properly represent all the values of the original type, and otherwise for promoting those types to
unsigned int. Thus, if the execution environment represents short as something smaller than
10 int, unsigned short becomes int; otherwise it becomes unsigned int.
Both schemes give the same answer in the vast majority of cases, and both give the same effective
result in even more cases in implementations with two’s-complement arithmetic and quiet
wraparound on signed overflow—that is, in most current implementations. In such
15 implementations, differences between the two only appear when these two conditions are both true:
2. The result of the preceding expression is used in a context in which its signedness is
significant:
In such circumstances a genuine ambiguity of interpretation arises. The result must be dubbed
questionably signed, since a case can be made for either the signed or unsigned interpretation.
35 Exactly the same ambiguity arises whenever an unsigned int confronts a signed int across
an operator, and the signed int has a negative value. Neither scheme does any better, or any
worse, in resolving the ambiguity of this confrontation. Suddenly, the negative signed int
becomes a very large unsigned int, which may be surprising, or it may be exactly what is
desired by a knowledgeable programmer. Of course, all of these ambiguities can be avoided by a
40 judicious use of casts.
One of the important outcomes of exploring this problem is the understanding that high-quality
compilers might do well to look for such questionable code and offer (optional) diagnostics, and
that conscientious instructors might do well to warn programmers of the problems of implicit type
45 conversions.
32
C9X RATIONALE WG14/N897 J11/99-032
The unsigned preserving rules greatly increase the number of situations where unsigned int
confronts signed int to yield a questionably signed result, whereas the value preserving rules
minimize such confrontations. Thus, the value preserving rules were considered to be safer for the
novice, or unwary, programmer. After much discussion, the C89 Committee decided in favor of
5 value preserving rules, despite the fact that the UNIX C compilers had evolved in the direction of
unsigned preserving.
15 The Standard clarifies that the integer promotion rules also apply to bit fields.
Note that, although _Bool is technically an integer type, conversion to _Bool does not always
20 work the same as conversion to other integer types. Consider, for example, that the expression
(_Bool)0.5 evaluates to 1, whereas (int)0.5 evaluates to 0. The first result is correct: it
simply says that 0.5 is non-zero; but it may be somewhat counter-intuitive unless a _Bool is
thought of as a “truth value” rather than as a one-bit integer.
Precise rules are now provided for converting to and from unsigned integers. On a two’s-
complement machine, the operation is still virtual (no change of representation is required), but the
rules are now stated independent of representation.
30
6.3.1.4 Real floating and integer
There was strong agreement in the C89 Committee that floating values should truncate toward zero
when converted to an integer type, the specification adopted in the Standard. Although K&R
35 permitted negative floating values to truncate away from zero, no C89 Committee member knew of
an implementation that functioned in such a manner.2
Note that conversion from integer to floating may indeed require rounding if the integer is too
wide to represent exactly in the floating-point format.
40
6.3.1.5 Real floating types
C89, unlike K&R, did not require rounding in the double to float conversion. Some widely
2
The Committee has since learned of one such implementation.
33
WG14/N897 J11/99-032 C9X RATIONALE
used IEC 60559 floating point processor chips control floating to integer conversion with the same
mode bits as for double-precision to single-precision conversion. Since truncation-toward-zero is
the appropriate setting for C in the former case, it would be expensive to require such
implementations to round to float. In C9X, §F.7.3 requires round-to-nearest for conversions
5 between floating formats as required by IEC 60559.
The rules in the Standard for these conversions are slight modifications of those in K&R: the
10 modifications accommodate the added types and the value preserving rules. Explicit license was
added to perform calculations in a “wider” type than absolutely necessary, since this can sometimes
produce smaller and faster code, not to mention the correct answer more often. Calculations can
also be performed in a “narrower” type by the as if rule so long as the same end result is obtained.
Explicit casting can always be used to obtain a value in a desired type.
15
The C9X Committee relaxed the requirement that float operands be converted to double. An
implementation may still choose to convert.
Real and imaginary operands are not converted to complex because doing so would require extra
25 computation, while producing undesirable results in certain cases involving infinities, NaNs and
signed zeros. For example, with automatic conversion to complex,
35 The following example illustrates the problem with signed zeros. With automatic conversion to
complex,
The problems illustrated in the examples above have counterparts for imaginary operands. The
mathematical product i2.0 × (∞ + i3.0) should yield –6.0 + i∞, but with automatic conversion to
45 complex,
34
C9X RATIONALE WG14/N897 J11/99-032
This also demonstrates the need for imaginary types specified in Annex G. Without them, i2.0
5 would have to be represented as 0.0 + i2.0, implying that NaN + i∞ would be the semantically
correct result regardless of conversion rules; and optimizers for implementations with infinities
would not be able to eliminate the operations with the zero real part.
A difference of opinion within the C community centered around the meaning of lvalue, one group
considering an lvalue to be any kind of object locator, another group holding that an lvalue is
15 meaningful on the left side of an assigning operator. The C89 Committee adopted the definition of
lvalue as an object locator. The term modifiable lvalue is used for the second of the above
concepts.
The role of array objects has been a classic source of confusion in C, in large part because of the
20 numerous contexts in which an array reference is converted to a pointer to its first element. While
this conversion neatly handles the semantics of subscripting, the fact that a[i] is a modifiable
lvalue while a is not has puzzled many students of the language. A more precise description was
incorporated in C89 in the hope of combatting this confusion.
25 6.3.2.2 void
The description of operators and expressions is simplified by saying that void yields a value, with
the understanding that the value has no representation, and hence requires no storage.
30 6.3.2.3 Pointers
C has now been implemented on a wide range of architectures. While some of these architectures
feature uniform pointers which are the size of some integer type, maximally portable code cannot
assume any necessary correspondence between different pointer types and the integer types. On
35 some implementations, pointers can even be wider than any integer type.
The use of void* (“pointer to void”) as a generic object pointer type is an invention of the C89
Committee. Adoption of this type was stimulated by the desire to specify function prototype
arguments that either quietly convert arbitrary pointers (as in fread) or complain if the argument
40 type does not exactly match (as in strcmp). Nothing is said about pointers to functions, which
may be incommensurate with object pointers and/or integers.
Since pointers and integers are now considered incommensurate, the only integer value that can be
safely converted to a pointer is a constant expression with the value 0. The result of converting any
45 other integer value, including a non-constant expression with the value 0, to a pointer is
implementation-defined.
35
WG14/N897 J11/99-032 C9X RATIONALE
Implicit in the Standard is the notion of invalid pointers. In discussing pointers, the Standard
typically refers to “a pointer to an object” or “a pointer to a function” or “a null pointer.” A special
25 case in address arithmetic allows for a pointer to just past the end of an array. Any other pointer is
invalid.
An invalid pointer might be created in several ways. An arbitrary value can be assigned (via a cast)
to a pointer variable. (This could even create a valid pointer, depending on the value.) A pointer to
30 an object becomes invalid if the memory containing the object is deallocated or moved by
realloc. Pointer arithmetic can produce pointers outside the range of an array.
Regardless how an invalid pointer is created, any use of it yields undefined behavior. Even
assignment, comparison with a null pointer constant, or comparison with itself, might on some
35 systems result in an exception.
36
C9X RATIONALE WG14/N897 J11/99-032
The Standard endeavors to bring preprocessing more closely into line with the token orientation of
the language proper. To do so requires that at least some information about white space be retained
5 through the early phases of translation (see §5.1.1.2). It also requires that an inverse mapping be
defined from tokens back to source characters (see §6.10.3).
6.4.1 Keywords
10 Several keywords were added in C89: const, enum, signed, void and volatile. New in
C9X are the keywords inline, restrict, _Bool, _Complex and _Imaginary.
Where possible, however, new features have been added by overloading existing keywords, as, for
example, long double instead of extended. It is recognized that each added keyword will
15 require some existing code that used it as an identifier to be rewritten. No meaningful programs are
known to be quietly changed by adding the new keywords.
The keywords entry, fortran, and asm have not been included since they were either never
used, or are not portable. Uses of fortran and asm as keywords are noted as common
20 extensions.
_Complex and _Imaginary, not complex and imaginary, are keywords in order that
freestanding implementations are not required to support complex. Old code using the names
complex or imaginary will still work (assuming <complex.h> is not included), and
25 combined C/C++ implementations will not have to finesse C-only public keywords.
6.4.2 Identifiers
6.4.2.1 General
30
Because of the linkers available at the time, the C89 Committee made the decision to restrict
significance of identifiers with external linkage to six case-insensitive characters. This limit is
increased in C9X to 31 case-sensitive characters.
35 While an implementation is not obliged to remember more than the first 63 characters of an
identifier with internal linkage, or the first 31 characters of an identifier with external linkage, the
programmer is effectively prohibited from intentionally creating two different identifiers that are the
same within the appropriate length. Implementations may therefore store the full identifier; they are
not obliged to truncate to 63 or 31.
40
QUIET CHANGE
A program that depends on identifiers matching only in the first few characters
may change to one with distinct objects for each variant spelling of the identifier.
45
37
WG14/N897 J11/99-032 C9X RATIONALE
A new feature of C9X: C9X introduces predefined identifiers, which have block scope (as distinct
from predefined macros which have file scope), and one such predefined identifier, _ _func_ _,
5 which allows the function name to be used at execution time.
A new feature of C9X: Note that, to allow for Universal Character Names (UCNs), a new
10 production has been added to the grammar that encompasses all forms of identifier elements (basic
letter, UCN, or extended character). There was some discussion about the need to require an
implementation to handle all digits, Arabic or otherwise, in a similar way. The general feeling was
that detecting the “extended digits” might be an undesirable burden for many implementations and
should be avoided if possible.
15
Note that a strictly conforming program may use in identifiers only the extended characters listed in
Annex I, and may not begin an identifier with an extended digit.
6.4.4 Constants
20
In folding and converting constants, an implementation must use at least as much precision as is
provided by the target environment. However, it is not required to use exactly the same precision as
the target, since this would require a cross compiler to simulate target arithmetic at translation time.
25 The C89 Committee considered the introduction of structure constants. Although it agreed that
structure literals would occasionally be useful, its policy was not to invent new features unless a
strong need exists. Since then, such structure constants have been shown to be quite useful, so C9X
introduces compound literals (see §6.5.2.5).
The C90 rule that the default type of a decimal integer constant is either int, long, or
unsigned long, depending on which type is large enough to hold the value without overflow,
simplifies the use of constants. The choices in C9X are int, long and long long.
35
C89 added the suffixes U and u to specify unsigned numbers. C9X adds LL to specify long
long.
Unlike decimal constants, octal and hexadecimal constants too large to be ints are typed as
40 unsigned int if within range of that type, since it is more likely that they represent bit patterns
or masks, which are generally best treated as unsigned, rather than “real” numbers.
Little support was expressed for the old practice of permitting the digits 8 and 9 in an octal
constant, so it was dropped in C89.
45
38
C9X RATIONALE WG14/N897 J11/99-032
A proposal to add binary constants was rejected due to lack of precedent and insufficient utility.
Despite a concern that a “lower-case-l” could be taken for the numeral one at the end of a numeric
literal, the C89 Committee rejected proposals to remove this usage, primarily on the grounds of
5 sanctioning existing practice.
The rules given for typing integer constants were carefully worked out in accordance with the C89
Committee’s deliberations on integer promotion rules. In C9X, this is clarified and extended with
the notion of “rank” (see §6.3.1.1).
10
QUIET CHANGE IN C89
Unsuffixed integer constants may have different types. In K&R, unsuffixed decimal
constants greater than INT_MAX, and unsuffixed octal or hexadecimal constants
15 greater than UINT_MAX are of type long.
20 Unsuffixed integer constants may have different types in C9X than. Such constants
greater than LONG_MAX are of type unsigned long in C89, but are of type
long long in C9X.
35 Note that the run-time selection of the decimal point character by setlocale (§7.11.1.1) has no
effect on the syntax of C source text: the decimal point character is always period.
Since floating constants are converted to appropriate internal representations at translation time,
default rounding direction and precision will be in effect and execution-time exceptions will not
40 be raised, even under the effect of an enabling FENV_ACCESS pragma. Library functions such
as strtod provide execution-time conversion of decimal strings.
A new feature of C9X: C9X adds hexadecimal notation because it more clearly expresses the
significance of floating constants. The binary-exponent part is required, instead of optional as it
45 is for decimal notation, to avoid ambiguity resulting from an f suffix being mistaken as a
hexadecimal digit.
39
WG14/N897 J11/99-032 C9X RATIONALE
Constants of long double type are not generally portable, even among IEC 60559
implementations.
5 Unlike integers, floating values cannot all be represented directly by hexadecimal constant
syntax. A sign can be prefixed for negative numbers and –0. Constant NaNs and infinities are
provided through macros in <math.h>. Note that 0x1.FFFFFEp128f, which might appear
to be an IEC 60559 single-format NaN, in fact overflows to an infinity in that format.
10 An alternate approach might have been to represent bit patterns. For example
This would have allowed representation of NaNs and infinities, however numerical values would
15 have been more obscure owing to bias in the exponent and the implicit significand bit, and NaN
representations would still not have been portable: even the determination of IEC 60559 quiet
NaN vs. signaling NaN is implementation-defined.
The straightforward approach of denoting octal constants by a 0 prefix would have been
20 inconsistent with allowing a leading 0 digit, a moot point as the need for octal floating constants
was deemed insufficient.
25 Whereas an enumeration variable may have any integer type that correctly represents all its values
when widened to int, an enumeration constant is only usable as the value of an expression. Hence
its type is simply int.
35 Hexadecimal escape sequences beginning with \x were adopted in C89, with precedent in several
existing implementations. There was little sentiment for providing \X as well. The escape
sequence extends to the first non-hex-digit character, thus providing the capability of expressing
any character constant no matter how large the type char is.
40 The C89 Committee chose to reserve all lower-case letters not currently used for future escape
sequences (undefined behavior). C9X adds \U from Java. All other characters with no current
meaning are left to the implementor for extensions (implementation-defined behavior). No portable
meaning is assigned to multi-character constants or ones containing other than the mandated source
character set (implementation-defined behavior).
45
The C89 Committee considered proposals to add the character constant '\e' to represent the
40
C9X RATIONALE WG14/N897 J11/99-032
ASCII ESC ('\033') character. This proposal was based upon the use of ESC as the initial
character of most control sequences in common terminal driving disciplines such as ANSI X3.64.
However, this usage has no obvious counterpart in other popular character codes such as EBCDIC.
5 A programmer merely wishing to avoid having to type "\033" to represent the ESC character in
an ANSI X3.64 (ISO/IEC 6429) environment, instead of
printf("\033[10;10h%d\n", somevalue);
10 may write
15 Notwithstanding the general rule that literal constants are non-negative3, a character constant
containing one character is effectively preceded with a (char) cast and hence may yield a
negative value if plain char is represented the same as signed char. This simply reflects
widespread past practice and was deemed too dangerous to change.
A constant of the form '\078' is valid, but now has different meaning. It now
denotes a character constant whose value is the (implementation-defined)
combination of the values of the two characters '\07' and '8'. In some
25 implementations the old meaning is the character whose code is 078 ≡ 0100 ≡ 64.
A constant of the form '\a' or '\x' now may have different meaning. The old
30 meaning, if any, was implementation dependent.
String literals are not required to be modifiable. This specification allows implementations to share
copies of strings with identical text, to place string literals in read-only memory, and to perform
3
Note that -4 is an expression: unary minus with operand 4.
41
WG14/N897 J11/99-032 C9X RATIONALE
certain optimizations. However, string literals do not have the type array of const char in order to
avoid the problems of pointer type checking, particularly with library functions, since assigning a
pointer to const char to a plain pointer to char is not valid. Those members of the C89 Committee
who insisted that string literals should be modifiable were content to have this practice designated a
5 common extension (see §K.5.5).
Existing code which modifies string literals can be made strictly conforming by replacing the string
literal with an initialized static character array. For instance,
A string can be continued across multiple lines by using the backslash–newline line continuation,
25 but this requires that the continuation of the string start in the first position of the next line. To
permit more flexible layout, and to solve some preprocessing problems (see §6.10.3), the C89
Committee introduced string literal concatenation. Two string literals in a row are pasted together,
with no null character in the middle, to make one combined string literal. This addition to the C
language allows a programmer to extend a string literal beyond the end of a physical line without
30 having to use the backslash–newline mechanism and thereby destroying the indentation scheme of
the program. An explicit concatenation operator was not introduced because the concatenation is a
lexical construct rather than a run-time operation.
A new feature of C9X: In C89, attempting to concatenate a character string literal and a wide string
35 literal resulted in undefined behavior, primarily because the C89 Committee saw little need to do
so. However, there are a number of macros defined by the standard as expanding into character
string literals which are frequently needed as wide strings instead (the format specifier macros in
<inttypes.h> are particularly notable examples, as are the predefined macros _ _FILE_ _,
_ _DATE_ _, and _ _TIME_ _). Rather than specifying two forms of each macro, one character
40 string literal and one wide string literal, the Committee decided to go ahead and define
concatenating a character string literal and a wide string literal as resulting in a wide string literal.
This solves the problem not only for the library and predefined macros, but for similar user-defined
macros as well.
45 Without concatenation:
42
C9X RATIONALE WG14/N897 J11/99-032
alpha = "abcdefghijklm\
nopqrstuvwxyz" ;
With concatenation:
5
// say the column is this wide
alpha = "abcdefghijklm"
"nopqrstuvwxyz";
10 String concatenation can be used to specify a hex-digit character following a hexadecimal escape
sequence:
20 A string of the form "\078" is valid, but now has different meaning.
It is neither required nor forbidden that identical string literals be represented by a single
35 copy of the string in memory; a program depending upon either scheme may behave
differently.
An L prefix distinguishes wide string literals. A prefix rather than a suffix notation was adopted so
that a translator can know at the start of the processing of a string literal whether it is dealing with
40 ordinary or wide characters.
6.4.6 Punctuators
C89 added the punctuator ... (ellipsis) to denote a variable number of trailing arguments in a
45 function prototype (see §6.7.5.3); and C9X extends this to function-like macros (see §6.10.3).
43
WG14/N897 J11/99-032 C9X RATIONALE
Header names in #include directives obey distinct tokenization rules; hence they are identified
as distinct tokens. Attempting to treat quote-enclosed header names as string literals creates a
5 contorted description of preprocessing, and the problems of treating angle bracket-enclosed header
names as a sequence of C tokens is even more severe.
10 The notion of preprocessing numbers was introduced to simplify the description of preprocessing.
It provides a means of talking about the tokenization of strings that look like numbers, or initial
substrings of numbers, prior to their semantic interpretation. In the interests of keeping the
description simple, occasional spurious forms are scanned as preprocessing numbers. For example,
0x123E+1 is a single token under the rules. The C89 Committee felt that it was better to tolerate
15 such anomalies than burden the preprocessor with a more exact, and exacting, lexical specification.
It felt that this anomaly was no worse than the principle under which the characters a+++++b are
tokenized as a ++ ++ + b (an invalid expression), even though the tokenization a ++ + ++ b
would yield a syntactically correct expression. In both cases, exercise of reasonable precaution in
coding style avoids surprises.
20
A new feature of C9X: C9X replaces nondigit with identifier-nondigit in the grammar to allow the
token pasting operator, ##, to work as expected. Given the code
#define mkident(s) s ## 1m
25 /* ... */
int mkident(int) = 0;
if an identifier is passed to the mkident macro, then 1m is parsed as a single pp-number, a valid
single identifier is produced by the ## operator, and nothing harmful happens. But consider a
30 similar construction that might appear using Greek script:
#define µk(p) p ## 1µ
µ
/* ... */
int µk(int) = 0;
35
µ must be parsed as only one pp-token. Restricting pp-numbers to only the
For this code to work, 1µ
basic letters would break this.
6.4.9 Comments
40
The C89 Committee considered proposals to allow comments to nest. The main argument for
nesting comments is that it would allow programmers to “comment out” code. The C89
Committee rejected this proposal on the grounds that comments should be used for adding
documentation to a program, and that preferable mechanisms already exist for source code
45 exclusion. For example,
44
C9X RATIONALE WG14/N897 J11/99-032
#if 0
/* code to be excluded */
#endif
5
Preprocessing directives such as this prevent the enclosed code from being scanned by later
translation phases. Bracketed material can include comments and other nested regions of bracketed
code.
if (0) {
/* code to be excluded */
}
15
Many modern compilers will generate no code for this if statement.
// comments were added for C9X due to their utility and widespread existing practice,
especially in dual C/C++ translators.
20
QUIET CHANGE IN C9X
In certain unusual situations, code could have different semantics for C90 and
C9X, for example
25
a = b //*divisor:*/ c
+ d;
35 a = b + d;
6.5 Expressions
Several closely related topics are involved in the precise specification of expression evaluation:
40 precedence, associativity, grouping, sequence points, agreement points, order of evaluation, and
interleaving.
The rules of precedence are encoded into the syntactic rules for each operator. For example, the
syntax for additive-expression includes the rule
45
additive-expression + multiplicative-expression
45
WG14/N897 J11/99-032 C9X RATIONALE
which implies that a+b*c parses as a+(b*c). The rules of associativity are similarly encoded
into the syntactic rules. For example, the syntax for assignment-expression includes the rule
With rules of precedence and associativity thus embodied in the syntax rules, the Standard
10 specifies, in general, the grouping (association of operands with operators) in an expression.
K&R’s rule imposes a high cost on translation of numerical code to C. Much numerical code is
25 written in Fortran, which does provide a no-regrouping guarantee; indeed, this is the normal
semantic interpretation in most high-level languages other than C. K&R’s advice, “rewrite using
explicit temporaries,” is burdensome to those with tens or hundreds of thousands of lines of code to
convert, a conversion which in most other respects could be done automatically.
30 Loss of the regrouping rule does not in fact prohibit much regrouping of integer expressions. The
bitwise logical operators can be arbitrarily regrouped since any regrouping gives the same result as
if the expression had not been regrouped. This is also true of integer addition and multiplication in
implementations with two’s-complement arithmetic and silent wraparound on overflow. Indeed, in
any implementation, regroupings which do not introduce overflows behave as if no regrouping had
35 occurred. (Results may also differ in such an implementation if the expression as written results in
overflows; but in such a case the behavior is undefined, so any regrouping couldn’t be any worse.)
The types of lvalues that may be used to access an object have been restricted so that an optimizer is
not required to make worst-case aliasing assumptions (see also §6.7.3.1)
40
In practice, aliasing arises with the use of pointers. A contrived example to illustrate the issues is
int a;
45 void f(int * b)
{
46
C9X RATIONALE WG14/N897 J11/99-032
a = 1;
*b = 2;
g(a);
}
5
It is tempting to generate the call to g as if the source expression were g(1), but b might point to
a, so this optimization is not safe. On the other hand, consider
int a;
10 void f( double * b )
{
a = 1;
*b = 2.0;
g(a);
15 }
Again the optimization is incorrect only if b points to a. However, this would only have come
about if the address of a were somewhere cast to double*. The C89 Committee has decided that
such dubious possibilities need not be allowed for.
20
In principle, then, aliasing only need be allowed for when the lvalues all have the same type. In
practice, the C89 Committee recognized certain prevalent exceptions:
• The lvalue types may differ in signedness. In the common range, a signed integer
25 type and its unsigned variant have the same representation; and it was felt that an
appreciable body of existing code is not “strictly typed” in this area.
• Character pointer types are often used in the bytewise manipulation of objects; a
byte stored through such a character pointer may well end up in an object of any
30 type.
• A qualified version of the object’s type, though formally a different type, provides
the same interpretation of the value of the object.
47
WG14/N897 J11/99-032 C9X RATIONALE
g(fip->i);
}
It is not safe to optimize the first call to g as g(2), or the second as g(1), since the call to f could
5 quite legitimately have been
struct fi x;
f( &x, &x.i );
An implementation that is able to multiply two double operands and produce a float result
in just one machine instruction might contract the multiplication and assignment in
15 float f;
double d1, d2;
/* ... */
f = d1 * d2;
20 Other examples of potential contraction operators include compound assignments (+=, -=, etc.),
ternary add (x + y + z), and multiply-add (xy + z).
Contractions can lead to subtle anomalies even while increasing accuracy. The value of C
expressions like a * b + c * d will depend on how the translator uses a contracted multiply-add.
25 The Intel i860’s multiply-add is slightly more problematic: since it keeps a wide but partial
product, a * b + z may differ from c * d + z even though the exact mathematical products a × b
and c × d are equal; the result depends not just on the mathematical result and the format, as
ordinarily expected for error analysis, but also on the particular values of the operands.
30 The programmer can control the use of fused multiply-adds by disabling use of contractions with
an FP_CONTRACT pragma and using the fma function where desired. The extra accuracy of a
fused multiply-add, which produces a result with just one rounding, can be exploited for simpler
and faster code.
A primary expression may be void (parenthesized call to a function returning void), a function
designator (identifier or parenthesized function designator), an lvalue (identifier or parenthesized
lvalue), or simply a value expression. Constraints ensure that a void primary expression is no part
40 of a further expression, except that a void expression may be cast to void, may be the second or
third operand of a conditional operator, or may be an operand of a comma operator.
48
C9X RATIONALE WG14/N897 J11/99-032
The C89 Committee found no reason to disallow the symmetry that permits a[i] to be written as
i[a].
The syntax and semantics of multidimensional arrays follow logically from the definition of arrays
5 and the subscripting operation. The material in the Standard on multidimensional arrays introduces
no new language features, but clarifies the C treatment of this important abstract data type.
10 Pointers to functions may be used either as (*pf)() or as pf(). The latter construct, not
sanctioned in K&R, appears in some present versions of C, is unambiguous, invalidates no old
code, and can be an important shorthand. The shorthand is useful for packages that present only
one external name, which designates a structure full of pointers to objects and functions: member
functions can be called as graphics.open(file) instead of (*graphics.open)(file).
15
The treatment of function designators can lead to some curious, but valid, syntactic forms. Given
the declarations
A new feature of C9X: The rule for implicit declaration of functions has been removed in C9X.
The effect is to guarantee the production of a diagnostic that will catch an additional category of
programming errors. After issuing the diagnostic, an implementation may choose to assume an
35 implicit declaration and continue translation in order to support existing programs that exploited
this feature.
For compatibility with past practice, all argument promotions occur as described in K&R in the
absence of a prototype declaration, including the not always desirable promotion of float to
40 double. A prototype gives the implementor explicit license to pass a float as a float rather
than a double, or a char as a char rather than an int, or an argument in a special register, etc.
If the definition of a function in the presence of a prototype would cause the function to expect
other than the default promotion types, then clearly the calls to this function must also be made in
the presence of a compatible prototype.
45
To clarify this and other relationships between function calls and function definitions, the Standard
49
WG14/N897 J11/99-032 C9X RATIONALE
describes the relationship between a function call or definition which does occur in the presence of
a prototype and one that does not.
Thus a prototyped function with no “narrow” types and no variable argument list must be callable
5 in the absence of a prototype, since the types actually passed in a call are equivalent to those in the
explicit function definition prototype. This constraint is necessary to retain compatibility with past
usage of library functions (see §7.1.4).
This provision constrains the latitude of an implementor because the parameter passing conventions
10 of prototype and non-prototype function calls must be the same for functions accepting a fixed
number of promoted arguments. Implementations in environments where efficient function calling
mechanisms are available must, in effect, use the efficient calling sequence either in all “fixed
argument list” calls or in none. Since efficient calling sequences often do not allow for variable
argument functions, the fixed part of a variable argument list may be passed in a completely
15 different fashion than in a fixed argument list with the same number and types of arguments.
The existing practice of omitting trailing parameters in a call if it is known that the parameters will
not be used has consistently been discouraged. Since omission of such parameters creates a
mismatch between the call and the declaration, the behavior in such cases is undefined, and a
20 maximally portable program will avoid this usage. Hence an implementation is free to implement a
function calling mechanism for fixed argument lists which would (perhaps fatally) fail if the wrong
number or type of arguments were to be provided.
Strictly speaking then, calls to printf are obliged to be in the scope of a prototype (as by
25 #include <stdio.h>), but implementations are not obliged to fail on such a lapse. (The
behavior is undefined.)
30 Since the language now permits structure parameters, structure assignment and functions returning
structures, the concept of a structure expression is now part of the C language. A structure value
can be produced by an assignment, by a function call, by a comma operator expression, by a
conditional operator expression, or by a compound literal:
35 s1 = (s2 = s3)
sf(x)
(x, s1)
x ? s1 : s2
40 Except for the case of the compound literal, the result is not an lvalue; hence it cannot be assigned
to nor can its address be taken.
Similarly, x.y is an lvalue only if x is an lvalue. Thus none of the following valid expressions are
lvalues.
45
sf(3).a
(s1=s2).a
50
C9X RATIONALE WG14/N897 J11/99-032
((i==6)?s1:s2).a
(x,s1).a
The Standard requires that an implementation diagnose a constraint error in the case that the
10 member of a structure or union designated by the identifier following a member selection operator
(. or ->) does not appear in the type of the structure or union designated by the first operand.
K&R is unclear on this point.
Increment and decrement operators are not defined for complex or imaginary types. Given the
20 regular definition, they would be surprising for imaginary types, as the operators would have no
effect. It is sometimes desirable to use the same source code with types being complex or
imaginary depending on the implementation. In this scenario, increment or decrement of the
complex 0+iy would differ from increment or decrement of the imaginary iy. Allowing
increment and decrement of complex but not imaginary objects would not be helpful here either.
25
6.5.2.5 Compound literals
A new feature of C9X: Compound literals provide a mechanism for specifying constants of
aggregate or union type. This eliminates the requirement for temporary variables when an
30 aggregate or union value will only be needed once.
Compound literals integrate easily into the C grammar and do not impose any additional run-time
overhead on a user’s program. They also combine well with designated initializers (§6.7.8) to form
an even more convenient aggregate or union constant notation. Their initial C implementation
35 appeared in a compiler by Ken Thompson at AT&T Bell Laboratories.
45 Some implementations have not allowed the & operator to be applied to an array or a function.
(The construct was permitted in early versions of C, then later made optional.) The C89 Committee
51
WG14/N897 J11/99-032 C9X RATIONALE
endorsed the construct since it is unambiguous, and since data abstraction is enhanced by allowing
the important & operator to apply uniformly to any addressable entity.
The bitwise complement operator ~, and the other bitwise operators, have now been defined
10 arithmetically for unsigned operands. Such operations are well-defined because of the restriction of
integer representations to “binary numeration systems.”
15 It is fundamental to the correct usage of functions such as malloc and fread that
sizeof(char) be exactly one. In practice, this means that a byte in C terms is the smallest unit
of storage, even if this unit is 36 bits wide; and all objects are composed of an integer number of
these smallest units.
20 C89, like K&R, defined the result of the sizeof operator to be a constant of an unsigned integer
type. Common implementations, and common usage, have often assumed that the resulting type is
int. Old code that depends on this behavior has never been portable to implementations that
define the result to be a type other than int. The C89 Committee did not feel it was proper to
change the language to protect incorrect code.
25
The type of sizeof, whatever it is, is published (in the library header <stddef.h>) as
size_t, since it is useful for the programmer to be able to refer to this type. This requirement
implicitly restricts size_t to be a synonym for an existing unsigned integer type. Note also that,
although size_t is an unsigned type, sizeof does not involve any arithmetic operations or
30 conversions that would result in modulus behavior if the size is too large to represent as a size_t,
thus quashing any notion that the largest declarable object might be too big to span even with an
unsigned long in C89 or uintmax_t in C9X. This also restricts the maximum number of
elements that may be declared in an array, since for any array a of N elements,
35 N == sizeof(a)/sizeof(a[0])
Thus size_t is also a convenient type for array sizes, and is so used in several library functions.
C89 specified that sizeof’s operand can be any value except a bit field, a void expression, or a
40 function designator. This generality allows for interesting environmental inquiries. Given the
declarations
52
C9X RATIONALE WG14/N897 J11/99-032
sizeof(vla) / sizeof(vla[0])
With the introduction of the long long and extended integer types, the sizeof
operator may yield a value that exceeds the range of a long.
20
6.5.4 Cast operators
A (void) cast is explicitly permitted, more for documentation than for utility.
25 Nothing portable can be said about casting integers to pointers, or vice versa, since the two are now
incommensurate.
The definition of these conversions adopted in the Standard resembles that in K&R, but with
several significant differences. K&R required that a pointer successfully converted to an integer
30 must be guaranteed to be convertible back to the same pointer. This integer-to-pointer conversion
is now specified as implementation-defined. While a high-quality implementation would preserve
the same address value whenever possible, it was considered impractical to require that the identical
representation be preserved. The C89 Committee noted that, on some current machine
implementations, identical representations are required for efficient code generation for pointer
35 comparisons and arithmetic operations.
The conversion of an integer constant expression with the value 0 to a pointer is defined similarly to
K&R. The resulting pointer must not address any object, must appear to be equal to an integer
value of 0, and may be assigned to or compared for equality with any other pointer. This definition
40 does not necessarily imply a representation by a bit pattern of all zeros: an implementation could,
for instance, use some address which causes a hardware trap when dereferenced.
The type char must have the least strict alignment of any type, so char* has often been used as a
portable type for representing arbitrary object pointers. This usage creates an unfortunate confusion
45 between the ideas of arbitrary pointer and character or string pointer. The new type void*,
which has the same representation as char*, is therefore preferable for arbitrary pointers.
53
WG14/N897 J11/99-032 C9X RATIONALE
It is possible to cast a pointer of some qualified type (§6.7.3) to an unqualified version of that type.
Since the qualifier defines some special access or aliasing property, however, any dereference of the
cast pointer results in undefined behavior.
5
The C89 Committee rejected extending the % operator to work on floating types as such usage
would duplicate the facility provided by fmod (see §7.12.10.1).
20
6.5.6 Additive operators
As with the sizeof operator (see §6.5.3.4), implementations have taken different approaches in
defining a type for the difference between two pointers. It is important that this type be signed in
25 order to obtain proper algebraic ordering when dealing with pointers within the same array.
However, the magnitude of a pointer difference can be as large as the size of the largest object that
can be declared; and since that is an unsigned type, the difference between two pointers can cause
an overflow on some implementations.
With the introduction of the long long and extended integer types, the subtraction
of pointers may return a value that exceeds the range of a long.
35 The C9X variable length array type (§6.7.5.2) does not affect the semantics of pointer difference.
Similarly, incrementing a pointer to a variable length array increments according to the number of
elements in the array just like a fixed length array.
void ptr_to_vla_incr(int n)
40 {
int a[2][n];
int (*p)[n] = a;
p++; // p == &a[1]
45
54
C9X RATIONALE WG14/N897 J11/99-032
// ...
5 If the declarations of a and p used an integer constant instead of the parameter n, then the
increment of pointer p still results in p pointing to the second row of a. That is, p is incremented
according to the number of elements in each row of a, and it doesn’t matter whether a is a variable
length array or a fixed length array. The expected behavior is preserved.
10 The type of pointer minus pointer is defined to be int in K&R. The Standard defines the result of
this operation to be a signed integer, the size of which is implementation-defined. The type is
published as ptrdiff_t, in the standard header <stddef.h>. Old code recompiled by a
conforming compiler may no longer work if the implementation defines the result of such an
operation to be a type other than int and if the program depended on the result to be of type int.
15 This behavior was considered by the C89 Committee to be correctable. Overflow was considered
not to break old code since it was undefined by K&R. Mismatch of types between actual and
formal argument declarations is correctable by including a properly defined function prototype in
the scope of the function invocation.
20 An important endorsement of widespread practice is the requirement that a pointer can always be
incremented to just past the end of an array, with no fear of overflow or wraparound:
SOMETYPE array[SPAN];
// ...
25 for (p = &array[0]; p < &array[SPAN]; p++)
This stipulation merely requires that every object be followed by one byte whose address is
representable. That byte can be the first byte of the next object declared for all but the last object
located in a contiguous segment of memory. (In the example, the address array + SPAN must
30 address a byte following the highest element of array.) Since the pointer expression p+1 need
not, and should not, be dereferenced, it is unnecessary to leave room for a complete object of size
sizeof(*p).
In the case of p-1, on the other hand, an entire object would have to be allocated prior to the array
35 of objects that p traverses, so decrement loops that run off the bottom of an array can fail. This
restriction allows segmented architectures, for instance, to place objects at the start of a range of
addressable memory.
55
WG14/N897 J11/99-032 C9X RATIONALE
The C89 Committee affirmed the freedom in implementation granted by K&R in not requiring the
signed right shift operation to sign extend, since such a requirement might slow down fast code and
5 since the usefulness of sign extended shifts is marginal. (Shifting a negative two’s-complement
integer arithmetically right one place is not the same as dividing by two!)
10 For an explanation of why the pointer comparison of the object pointer P with the pointer
expression P+1 is always safe, see Rationale §6.5.6.
Some mathematical practice would be supported by defining the relational operators for complex
operands so that z1 op z2 would be true if and only if both creal(z1) op creal(z2) and
15 cimag(z1) == cimag(z2). Believing such use to be uncommon, the C9X Committee voted
against including this specification.
20 The C89 Committee considered, on more than one occasion, permitting comparison of structures
for equality. Such proposals foundered on the problem of holes in structures. A byte-wise
comparison of two structures would require that the holes assuredly be set to zero so that all holes
would compare equal, a difficult task for automatic or dynamically allocated variables. The
possibility of union-type elements in a structure raises insuperable problems with this approach.
25 Without the assurance that all holes were set to zero, the implementation would have to be prepared
to break a structure comparison into an arbitrary number of member comparisons; a seemingly
simple expression could thus expand into a substantial stretch of code, which is contrary to the
spirit of C.
30 In pointer comparisons, one of the operands may be of type void*. In particular, this allows
NULL, which can be defined as (void*)0, to be compared to any object pointer.
35 The syntactic restrictions on the middle operand of the conditional operator have been relaxed to
include more than just logical-OR-expression: several extant implementations have adopted this
practice.
The type of a conditional operator expression can be void, a structure, or a union; most other
40 operators do not deal with such types. The rules for balancing type between pointer and integer
have, however, been tightened, since now only the constant 0 can portably be coerced to a pointer.
The Standard allows one of the second or third operands to be of type void*, if the other is a
pointer type. Since the result of such a conditional expression is void*, an appropriate cast must
45 be used.
56
C9X RATIONALE WG14/N897 J11/99-032
Certain syntactic forms of assignment operators have been discontinued, and others tightened up.
5
The storage assignment need not take place until the next sequence point. As a consequence, a
straightforward syntactic test for ambiguous expressions can be stated. Some definitions: A side
effect is a storage to any data object, or a read of a volatile object. An ambiguous expression is
one whose value depends upon the order in which side effects are evaluated. A pure function is one
10 with no side effects; an impure function is any other. A sequenced expression is one whose major
operator defines a sequence point: comma, &&, ||, or conditional operator; an unsequenced
expression is any other. We can then say that an unsequenced expression may be ambiguous if
more than one operand invokes any impure function, or if more than one operand contains an lvalue
referencing the same object and one or more operands specify a side-effect to that object. Further,
15 any expression containing an ambiguous sub-expression is ambiguous.
The optimization rules for factoring out assignments can also be stated. Let X(i,S) be an
expression which contains no impure functions or sequenced operators, and suppose that X contains
a storage S(i) to i which sets i to Snew(i) and returns Sval(i). The possible storage
20 expressions are
(T = i, i = Snew(i), X(T,Sval))
or
(T = X(i,Sval), i = Snew(i), T)
35
provided that neither i nor y have side effects themselves.
40 Structure assignment was added: its use was foreshadowed even in K&R, and many existing
implementations already support it.
The rules for type compatibility in assignment also apply to argument compatibility between actual
argument expressions and their corresponding parameter types in a function prototype.
45
An implementation need not correctly perform an assignment between overlapping operands.
57
WG14/N897 J11/99-032 C9X RATIONALE
Overlapping operands occur most naturally in a union, where assigning one field to another is often
desirable to effect a type conversion in place. The assignment may well work properly in all simple
cases, but it is not maximally portable. Maximally portable code should use a temporary variable as
an intermediate in such an assignment.
5
6.5.16.2 Compound assignment
The importance of requiring that the left operand lvalue be evaluated only once is not a question of
efficiency, although that is one compelling reason for using the compound assignment operators.
10 Rather, it is to assure that any side effects of evaluating the left operand are predictable.
Assignment operators of the form =+, described as old fashioned even in K&R, were dropped in
C89. The form += is now defined to be a single token, not two, so no white space is permitted
within it. No compelling case could be made for permitting such white space.
15
QUIET CHANGE IN C89
Expressions of the form x=-3 change meaning with the loss of the old-style
assignment operators.
20
6.5.17 Comma operator
The left operand of a comma operator may be void, since only the right hand operator is relevant
to the type of the expression.
25
6.6 Constant expressions
To clarify existing practice, several varieties of constant expression have been identified.
30 The expression following #if (§6.10.1) must expand to integer constants, character constants, the
special operator defined, and operators with no side effects. Environmental inquiries can be
made only using the macros defined in the standard headers, <limits.h>, <stdint.h>, etc.
Character constants, when evaluated in #if expressions, may be interpreted in the source character
35 set, the execution character set, or some other implementation-defined character set. This latitude
reflects the diversity of existing practice, especially in cross-compilers.
An integer constant expression must involve only numbers knowable at translation time, and
operators with no side effects. Casts and the sizeof operator whose operand does not have a
40 variable length array type (§6.7.5.2) may be used to interrogate the execution environment.
Static initializers include integer constant expressions, along with floating constants and simple
addressing expressions. An implementation must accept arbitrary expressions involving floating
and integer numbers and side effect-free operators in arithmetic initializers, but it is at liberty to turn
45 such initializers into executable code which is invoked prior to program startup. This scheme
58
C9X RATIONALE WG14/N897 J11/99-032
might impose some requirements on linkers or runtime library code in some implementations.
The translation environment must not produce a less accurate value for a floating-point initializer
than the execution environment, but it is at liberty to do better. Thus a static initializer may well be
5 slightly different from the same expression computed at execution time. However, while
implementations are certainly permitted to produce exactly the same result in translation and
execution environments, requiring this was deemed to be an intolerable burden on many cross-
compilers.
Due to the introduction of new types, the preprocessor arithmetic must be performed
using the semantics of either intmax_t or uintmax_t defined in
<stdint.h>. This is a quiet change for cross-compilation implementations
20 because C89 did not mandate that translation-time arithmetic have the properties of
the execution environment, but C9X does.
6.7 Declarations
25
The C89 Committee decided that empty declarations are invalid, except for a special case with
tags (see §6.7.2.3) and the case of enumerations such as enum {zero,one}; (see §6.7.2.2).
While many seemingly silly constructs are tolerated in other parts of the language in the interest of
facilitating the machine generation of C, empty declarations were considered sufficiently easy to
30 avoid.
Because the address of a register variable cannot be taken, objects of storage class register
35 effectively exist in a space distinct from other objects. (Functions occupy yet a third address
space.) This makes them candidates for optimal placement, the usual reason for declaring
registers; but it also makes them candidates for more aggressive optimization.
The practice of representing register variables as wider types (as when register char is
40 quietly changed to register int) is no longer acceptable.
Several new type specifiers were added to C89: signed, enum, and void. long float was
45 retired and long double was added, along with many integer types.
59
WG14/N897 J11/99-032 C9X RATIONALE
A new feature of C9X: several new type specifiers were added to C9X: _Bool, _Complex and
_Imaginary, along with the related types.
5 A new feature of C9X: C9X adds a new integer data type, long long, as consolidation of prior
art, whose impetus has been three hardware developments. First, disk density and capacity used
to double every 3 years, but after 1989 has quadrupled every 3 years, yielding low-cost,
physically small disks with large capacities. Although a fixed size for file pointers and file
system structures is necessary for efficiency, eventually it is overtaken by disk growth, and limits
10 need to be expanded. In the 1970s, 16-bit C (for the Digital PDP-11) first represented file
information with 16-bit integers, which were rapidly obsoleted by disk progress. People
switched to a 32-bit file system, first using int[2] constructs which were not only awkward,
but also not efficiently portable to 32-bit hardware.
15 To solve the problem, the long type was added to the language, even though this required C on
the PDP-11 to generate multiple operations to simulate 32-bit arithmetic. Even as 32-bit
minicomputers became available alongside 16-bit systems, people still used int for efficiency,
reserving long for cases where larger integers were truly needed, since long was noticeably
less efficient on 16-bit systems. Both short and long were added to C, making short
20 available for 16 bits, long for 32 bits, and int as convenient for performance. There was no
desire to lock the numbers 16 or 32 into the language, as there existed C compilers for at least
24- and 36-bit CPUs, but rather to provide names that could be used for 32 bits as needed.
PDP-11 C might have been re-implemented with int as 32-bits, thus avoiding the need for
25 long; but that would have made people change most uses of int to short or suffer serious
performance degradation on PDP-11s. In addition to the potential impact on source code, the
impact on existing object code and data files would have been worse, even in 1976. By the
1990s, with an immense installed base of software, and with widespread use of dynamic linked
libraries, the impact of changing the size of a common data object in an existing environment is
30 so high that few people would tolerate it, although it might be acceptable when creating a new
environment. Hence, many vendors, to avoid namespace conflicts, have added a 64-bit integer to
their 32-bit C environments using a new name, of which long long has been the most widely
used.
35 C9X has therefore adopted long long as the name of an integer type with at least 64 bits of
precision. People can and do argue about the particular choice of name, but it has been difficult
to pick a clearly better name early enough, and by now it is fairly common practice, and may be
viewed as one of the least bad choices.
40 To summarize this part: 32-bit CPUs are coming to need clean 64-bit integers, just as 16-bit
CPUs came to need 32-bit integers, and the need for wider integers happens irrespective of other
CPUs. Thus, 32-bit C has evolved from a common ILP32 model (int, long and pointers are
32 bits) to ILP32LL (ILP32 + 64-bit long long), and this still runs on 32-bit CPUs with
sequences to emulate 64-bit arithmetic.
45
In the second and third interrelated trends, DRAM memories continue to quadruple in size every
60
C9X RATIONALE WG14/N897 J11/99-032
3 years, and 64-bit microprocessors started to be widely used in 1992. By 1995, refrigerator-
sized, microprocessor-based servers were being sold with 8GB to 16GB of memory, which
required more than 32 bits for straightforward addressing. However, many 64-bit
microprocessors are actually used in video games, X-Terminals, network routers, and other
5 applications where pointer size is less important than performance for larger integers.
The memory trend encourages a C programming model in which pointers are enlarged to 64 bits
(called *P64), of which the consensus choice seems to be LP64 (long, pointers and long
long are 64 bits; int is 32 bits), with long long in some sense redundant, just as long was
10 on the 32-bit VAX. It is fairly difficult to mix this object code with the ILP32 model, and so it
is a new environment to which people must port code, but for which they receive noticeable
benefits: they can address large memories, and file pointers automatically are enlarged to 64 bits.
There do exist, of course, 32-bit CPUs with more-than-32-bit addressing, although C
environments become much more straightforward on 64-bit CPUs with simple, flat addressing.
15 In practice, people do not move from ILP32LL to LP64 unless they have no choice or gain
some clear benefit.
If people only consider LP64 in isolation, long is 64 bits, and there seems no need for long
long, just as the VAX 32-bit environment really did not need long. However, this view
20 ignores the difficulty of getting compilers, linkers, debuggers, libraries, etc., to exist for LP64.
In practice, these programs need to deal with 64-bit integers long before an LP64 environment
exists, in order to bootstrap, and later support, all these tools. Put another way, people must:
1. Using int[2], upgrade compilers and a minimal set of tools to compile and debug code that
25 uses long long.
2. Recode the compilers and all of the tools to actually use long long.
This ends up with a set of tools that run as ILP32LL, on existing 32-bit CPUs and new 64-bit
30 CPUs, and can compile code to either ILP32LL or LP64. This is yet another reason where
long long is important, not for the LP64 model, but for the tools that support that model.
Most 64-bit micros can, and for commercial reasons must, continue to run existing ILP32LL
object programs, alongside any new LP64 programs. For example, database server processes
35 often desire LP64 to access large memory pools, but the implementers prefer to leave the client
code as ILP32 so that it can run on existing 32-bit CPUs as well, and where LP64 provides no
obvious value.
In mixed environments, it is of course very useful for programs to share data structures, and
40 specifically for 32-bit programs to be able to cleanly describe aligned 64-bit integers, and in fact
for it to be easy to write structure definitions whose size and alignment are identical between
ILP32LL and LP64. This can be straightforwardly done using int and long long, just as it
was doable in the 1970s via short and long.
45 Finally, one more important case occurs, in which people want performance benefits of 64-bit
CPUs, while wishing to maintain source compatibility, but not necessarily binary compatibility,
61
WG14/N897 J11/99-032 C9X RATIONALE
with related 32-bit CPUs. In embedded control and consumer products, people have little
interest in 64-bit pointers, but they often like 64-bit integer performance for bit manipulation,
memory copies, encryption, and other tasks. They like ILP32LL, but with long long
compiled to use 64-bit registers, rather than being simulated via 32-bit registers. While this is
5 not binary-compatible with existing ILP32LL binaries, it is source-compatible; and it runs faster
and uses less space than LP64, both of which are important in these markets. It is worth noting
that of the many millions of 64-bit CPUs that exist, a very large majority are actually used in
such applications rather than traditional computer systems.
10 Thus, there are 3 choices, all of which have been done already, and different customers choose
different combinations:
25 A new integer data type is needed that can be used to express 64-bit integers efficiently and
portably among 32- and 64-bit systems. It must be a new name to avoid a disastrous set of
incompatibilities with existing 32-bit environments since one cannot safely change long to 64
bits and mix with existing object code. It is needed to deal with disk file size increases, but also
to help bootstrap to 64-bit environments, and then wider, so that many programs can be compiled
30 to exactly one binary that runs on both 32- and 64-bit CPUs.
While there is more argument about the specific syntax, nobody has seemed able to provide a
compellingly better syntax than long long, which at least avoided gratuitous namespace
pollution. Proposals like int64_t seem very awkward for 36-bit CPUs, for example.
35
Given the various complex interactions, long long seems a reasonable addition to C, as
existing practice has shown the need for a larger integer, and long long syntax seems one of
the least bad choices.
40 Implementation note: excluding the library, programs that do not use long long do not have
any overhead caused by the addition of long long to C9X; however, certain standard library
functions that are useful when the program does not otherwise use long long do have
overhead that an implementation may wish to eliminate. The primary examples of such library
functions are the printf and scanf families.
45
The only use of long long in some programs might be the support in printf to print long
62
C9X RATIONALE WG14/N897 J11/99-032
long values even though the program never actually prints such a value. On systems where
there is no hardware support for long long arithmetic, the long long support in printf
might cause runtime functions to be loaded to emulate long long arithmetic. For embedded
systems, where memory is precious, the code to support printing long long values and for
5 doing long long arithmetic is a great burden if that code is never actually used by the program.
(In contrast, large systems with shared libraries may have no overhead supporting long long
all of the time.)
The situation with long long is very simiular to the situration with floating point in C for the
10 PDP-11. Some models of the PDP-11 lacked floating point hardware and were limited to a 64K
address space. (A similar situation exists with MS-DOS-based implementations.) Many C
programs did not use floating point, but because of printf and scanf, wasted a significant
part of their address space on floating point emulation code. The PDP-11 C solution was simple
and effective: two versions of printf and scanf were provided. One version supported
15 printing or scanning floating point values; the other version did not. By linking with the proper
version of printf and scanf, the program could avoid the overhead of floating point code.
This technique also works for avoiding any overhead from long long.
In the simplest form, this technique can be purely manual and require no changes to compilers or
20 linkers: the user explicitly links the program with the proper versions of printf and scanf.
Automatic solutions are also possible by having the compiler inform the linker if the program
uses long long. A variety of sulutions are possible, and like the PDP-11 implementation,
many embedded systems have already solved the overhead problem for floating point, and may
find it useful to adopt a similar solution for long long.4
25
Since the technique discussed above allows an implementation to avoid the burden of linking in
long long support when it is not needed, the committee saw no reason to make support for
long long be optional.
In some environments such as LP64 and ILP64, long long and long are
equivalent. In the others (ILP32LL and LLP64), long long is larger than
long. If the system environment also changes standard definitions such as
35 ptrdiff_t to become long long or size_t to become unsigned long
long, then existing correct code can be broken. For example,
unsigned long x;
size_t y;
40 x = y;
silently truncates y.
4
An implementation may be able to optimize more cases by inspecting the argument types in calls to printf and scanf family functions, but
should be aware that vprintf and vscanf arguments types may not be available at compile time.
63
WG14/N897 J11/99-032 C9X RATIONALE
A new feature of C9X: In C89, all type specifiers could be omitted from the declaration specifiers
in a declaration. In such a case int was implied. The Committee decided that the inherent
danger of this feature outweighed its convenience, and so it was removed. The effect is to
5 guarantee the production of a diagnostic that will catch an additional category of programming
errors. After issuing the diagnostic, an implementation may choose to assume an implicit int
and continue to translate the program in order to support existing source code that exploited this
feature.
Three types of bit fields are now defined: plain int calls for implementation-defined signedness
(as in K&R), signed int calls for assuredly signed fields, and unsigned int calls for
unsigned fields. The old constraints on bit fields crossing word boundaries have been relaxed,
15 since so many properties of bit fields are implementation dependent anyway.
• if necessary, a hole is placed on the end to make the structure big enough to pack
tightly into arrays and maintain proper alignment.
25
Since some existing implementations, in the interest of enhanced access time, leave internal holes
larger than absolutely necessary, it is not clear that a portable deterministic method can be given
for traversing a structure member by member.
30 To clarify what is meant by the notion that “all the members of a union occupy the same storage,”
the Standard specifies that a pointer to a union, when suitably cast, points to each member (or, in
the case of a bit-field member, to the storage unit containing the bit field).
A new feature of C9X: There is a common idiom known as the “struct hack” for creating a
35 structure containing a variable-size array:
struct s
{
int n_items;
40 /* possibly other fields */
int items[1];
};
struct s *p;
size_t n, i;
45
/* code that sets n omitted */
p = malloc(sizeof(struct s) + (n - 1) * sizeof(int));
64
C9X RATIONALE WG14/N897 J11/99-032
The validity of this construct has always been questionable. In the response to one Defect Report,
The Committee decided that it was undefined behavior because the array p->items contains
only one item, irrespective of whether the space exists. An alternative construct was suggested:
10 make the array size larger than the largest possible case (for example, using
int items[INT_MAX];), but this approach is also undefined for other reasons.
The Committee felt that, although there was no way to implement the “struct hack” in C89, it was
nonetheless a useful facility. Therefore the new feature of “flexible array members” was
15 introduced. Apart from the empty brackets, and the removal of the “-1” in the malloc call, this
is used in the same way as the struct hack, but is now explicitly valid code.
There are a few restrictions on flexible array members that ensure that code using them makes
sense. For example, there must be at least one other member, and the flexible array must occur
20 last. Similarly, structures containing flexible arrays can’t occur in the middle of other structures or
in arrays. Finally, sizeof applied to the structure ignores the array but counts any padding
before it. This makes the malloc call as simple as possible.
30 6.7.2.3 Tags
As with all block-structured languages that also permit forward references, C has a problem with
structure and union tags. If one wants to declare, within a block, two mutually-referencing
structures, one must write something like
35
struct x { struct y *p; /*...*/ };
struct y { struct x *q; /*...*/ };
But if struct y is already defined in a containing block, the first field of struct x will refer to
40 the older declaration.
struct y;
45
which now hides the outer declaration of y, and “opens” a new instance in the current block.
65
WG14/N897 J11/99-032 C9X RATIONALE
The C89 Committee added to C two type qualifiers, const and volatile; and C9X adds a
10 third, restrict. Individually and in combination they specify the assumptions a compiler can
and must make when accessing an object through an lvalue.
The syntax and semantics of const were adapted from C++; the concept itself has appeared in
other languages. volatile and restrict are inventions of the Committee; and both follow
15 the syntactic model of const.
Type qualifiers were introduced in part to provide greater control over optimization. Several
important optimization techniques are based on the principle of “cacheing”: under certain
circumstances the compiler can remember the last value accessed (read or written) from a location,
20 and use this retained value the next time that location is read. (The memory, or “cache”, is
typically a hardware register.) If this memory is a machine register, for instance, the code can be
smaller and faster using the register rather than accessing external memory.
The basic qualifiers can be characterized by the restrictions they impose on access and cacheing:
25
const No writes through this lvalue. In the absence of this qualifier, writes may occur
through this lvalue.
volatile No cacheing through this lvalue: each operation in the abstract semantics must be
30 performed (that is, no cacheing assumptions may be made, since the location is not
guaranteed to contain any previous value). In the absence of this qualifier, the
contents of the designated location may be assumed to be unchanged except for
possible aliasing.
66
C9X RATIONALE WG14/N897 J11/99-032
A translator design with no cacheing optimizations can effectively ignore the type qualifiers,
except insofar as they affect assignment compatibility.
5 It would have been possible, of course, to specify nonconst instead of const, etc. The senses
of these concepts in the Standard were chosen to assure that the default, unqualified, case is the
most common, and that it corresponds most clearly to traditional practice in the use of lvalue
expressions.
10 Several combinations of the three qualifiers are possible and most define useful sets of lvalue
properties. The next several paragraphs describe typical uses of the const and volatile
qualifiers. The restrict qualifier is discussed in §6.7.3.1.
The translator may assume, for an unqualified lvalue, that it may read or write the referenced
15 object, that the value of this object cannot be changed except by explicitly programmed actions in
the current thread of control, but that other lvalue expressions could reference the same object.
const is specified in such a way that an implementation is at liberty to put const objects in
read-only storage, and is encouraged to diagnose obvious attempts to modify them, but is not
20 required to track down all the subtle ways that such checking can be subverted.
A static const volatile object appropriately models a memory-mapped input port, such
as a real-time clock. Similarly, a const volatile object models a variable which can be
altered by another process but not by this one.
35
Although the type qualifiers are formally treated as defining new types, they actually serve as
modifiers of declarators. Thus the declarations
declare x as a const object, but not y. The const property can be associated with the aggregate
type by means of a type definition:
67
WG14/N897 J11/99-032 C9X RATIONALE
stype y;
In these declarations the const property is associated with the declarator stype, so x and y are
both const objects.
5
The C89 Committee considered making const and volatile storage classes, but this would
have ruled out any number of desirable constructs, such as const members of structures and
variable pointers to const types.
10 A cast of a value to a qualified type has no effect; the qualification (volatile, say) can have no
effect on the access since it has occurred prior to the cast. If it is necessary to access a non-
volatile object using volatile semantics, the technique is to cast the address of the object
to the appropriate pointer-to-qualified type, then dereference that pointer.
A new feature of C9X: The restrict type qualifier allows programs to be written so that
translators can produce significantly faster executables. Anyone for whom this is not a concern
can safely ignore this feature of the language.
20
The problem that the restrict qualifier addresses is that potential aliasing can inhibit
optimizations. Specifically, if a translator cannot determine that two different pointers are being
used to reference different objects, then it cannot apply optimizations such as maintaining the
values of the objects in registers rather than in memory, or reordering loads and stores of these
25 values. This problem can have a significant effect on a program that, for example, performs
arithmetic calculations on large arrays of numbers. The effect can be measured by comparing a
program that uses pointers with a similar program that uses file scope arrays (or with a similar
Fortran program). The array version can run faster by a factor of ten or more on a system with
vector processors. Where such large performance gains are possible, implementations have of
30 course offered their own solutions, usually in the form of compiler directives that specify particular
optimizations. Differences in the spelling, scope, and precise meaning of these directives have
made them troublesome to use in a program that must run on many different systems. This was
the motivation for a standard solution.
35 The restrict qualifier was designed to express and extend two types of aliasing information
already specified in the language.
First, if a single pointer is directly assigned the return value from an invocation of malloc, then
that pointer is the sole initial means of access to the allocated object (that is, another pointer can
40 gain access to that object only by being assigned a value that is based on the value of the first
pointer). Declaring the pointer to be restrict-qualified expresses this information to a translator.
Furthermore, the qualifier can be used to extend a translator’s special treatment of such a pointer to
more general situations. For example, an invocation of malloc might be hidden from the
translator in another function, or a single invocation of malloc might be used to allocate several
45 objects, each referenced through its own pointer.
68
C9X RATIONALE WG14/N897 J11/99-032
Second, the library specifies two versions of an object copying function, because on many systems
a faster copy is possible if it is known that the source and target arrays do not overlap. The
restrict qualifier can be used to express the restriction on overlap in a new prototype that is
compatible with the original version:
5
void *memcpy(void * restrict s1, const void * restrict s2,
size_t n);
void *memmove(void * s1, const void * s2, size_t n);
The full specification is necessary to determine the precise meaning of a qualifier in any context,
and so must be understood by compiler implementors. Fortunately, most others will need to
understand only a few simple patterns of usage explained in the following examples.
25
A translator can assume that a file scope restrict-qualified pointer is the sole initial means of
access to an object, much as if it were the declared name of an array. This is useful for a
dynamically allocated array whose size is not known until run time. Note in the example how a
single block of storage is effectively subdivided into two disjoint objects.
30
float * restrict a1, * restrict a2;
void init(int n)
{
35 float * t = malloc(2 * n * sizeof(float));
a1 = t; // a1 refers to 1st half
a2 = t + n; // a2 refers to 2nd half
}
40 A translator can assume that a restrict-qualified pointer that is a function parameter is, at the
beginning of each execution of the function, the sole means of access to an object. Note that this
assumption expires with the end of each execution. In the following example, parameters a1 and
a2 can be assumed to refer to disjoint array objects because both are restrict-qualified. This
implies that each iteration of the loop is independent of the others, and so the loop can be
45 aggressively optimized.
69
WG14/N897 J11/99-032 C9X RATIONALE
A translator can assume that a restrict-qualified pointer declared with block scope is, during
10 each execution of the block, the sole initial means of access to an object. An invocation of the
macro shown in the following example is equivalent to an inline version of a call to the function
f1 above.
# define f2(N,A1,A2) \
15 { int n = (N); \
float * restrict a1 = (A1); \
float * restrict a2 = (A2); \
int i; \
for ( i = 0; i < n; i++ ) \
20 a1[i] += a2[i]; \
}
The restrict qualifier can be used in the declaration of a structure member. A translator can
assume, when an identifier is declared that provides a means of access to an object of that structure
25 type, that the member provides the sole initial means of access to an object of the type specified in
the member declaration. The duration of the assumption depends on the scope of the identifier,
not on the scope of the declaration of the structure. Thus a translator can assume that s1.a1 and
s1.a2 below are used to refer to disjoint objects for the duration of the whole program, but that
s2.a1 and s2.a2 are used to refer to disjoint objects only for the duration of each invocation of
30 the f3 function.
struct t {
int n;
float * restrict a1, * restrict a2;
35 };
struct t s1;
70
C9X RATIONALE WG14/N897 J11/99-032
the qualifier. In particular, a restrict-qualified pointer can be the actual argument for a
function parameter that is unqualified. On the other hand, it is easier for a translator to find
opportunities for optimization if as many as possible of the pointers in a program are restrict-
qualified.
5
6.7.4 Function specifiers
A new feature of C9X: The inline keyword, adapted from C++, is a function-specifier that can
be used only in function declarations. It is useful for program optimizations that require the
10 definition of a function to be visible at the site of a call. (Note that the Standard does not attempt
to specify the nature of these optimizations.)
Visibility is assured if the function has internal linkage, or if it has external linkage and the call is
in the same translation unit as the external definition. In these cases, the presence of the inline
15 keyword in a declaration or definition of the function has no effect beyond indicating a preference
that calls of that function should be optimized in preference to calls of other functions declared
without the inline keyword.
Visibility is a problem for a call of a function with external linkage where the call is in a different
20 translation unit from the function’s definition. In this case, the inline keyword allows the
translation unit containing the call to also contain a local, or inline, definition of the function.
A program can contain a translation unit with an external definition, a translation unit with an
inline definition, and a translation unit with a declaration but no definition for a function. Calls in
25 the latter translation unit will use the external definition as usual.
A conforming program must not rely on the implementation using the inline definition, nor may it
rely on the implementation using the external definition. The address of a function is always the
35 address corresponding to the external definition, but when this address is used to call the function,
the inline definition might be used. Therefore, the following example might not behave as
expected.
45 int compare_name(void)
{
71
WG14/N897 J11/99-032 C9X RATIONALE
Since the implementation might use the inline definition for one of the calls to saddr and use the
5 external definition for the other, the equality operation is not guaranteed to evaluate to 1 (true).
This shows that static objects defined within the inline definition are distinct from their
corresponding object in the external definition. This motivated the constraint against even
defining a non-const object of this type.
10 Inlining was added to the Standard in such a way that it can be implemented with existing linker
technology, and a subset of C9X inlining is compatible with C++. This was achieved by requiring
that exactly one translation unit containing the definition of an inline function be specified as the
one that provides the external definition for the function. Because that specification consists
simply of a declaration that either lacks the inline keyword, or contains both inline and
15 extern, it will also be accepted by a C++ translator.
Inlining in C9X does extend the C++ specification in two ways. First, if a function is declared
inline in one translation unit, it need not be declared inline in every other translation unit.
This allows, for example, a library function that is to be inlined within the library but available
20 only through an external definition elsewhere. The alternative of using a wrapper function for the
external function requires an additional name; and it may also adversely impact performance if a
translator does not actually do inline substitution.
Second, the requirement that all definitions of an inline function be “exactly the same” is replaced
25 by the requirement that the behavior of the program should not depend on whether a call is
implemented with a visible inline definition, or the external definition, of a function. This allows
an inline definition to be specialized for its use within a particular translation unit. For example,
the external definition of a library function might include some argument validation that is not
needed for calls made from other functions in the same library. These extensions do offer some
30 advantages; and programmers who are concerned about compatibility can simply abide by the
stricter C++ rules.
Note that it is not appropriate for implementations to provide inline definitions of standard library
functions in the standard headers because this can break some legacy code that redeclares standard
35 library functions after including their headers. The inline keyword is intended only to provide
users with a portable way to suggest inlining of functions. Because the standard headers need not
be portable, implementations have other options along the lines of:
6.7.5 Declarators
45 The function prototype syntax was adapted from C++ (see §6.5.2.2 and §6.7.5.3).
72
C9X RATIONALE WG14/N897 J11/99-032
Some pre-C89 implementations had a limit of six type modifiers (function returning, array of,
pointer to), the limit used in Ritchie’s original compiler. This limit was raised to twelve in C89
since the original limit has proven insufficient in some cases; in particular, it did not allow for
Fortran-to-C translation, since Fortran allows for seven subscripts. (Some users have reported
5 using nine or ten levels, particularly in machine-generated C code.)
A pointer declarator may have its own type qualifiers to specify the attributes of the pointer itself,
10 as opposed to those of the reference type. The construct is adapted from C++.
const int * means (variable) pointer to constant int, and int * const means constant
pointer to (variable) int, just as in C++. (And mutatis mutandis for the other type qualifiers.) As
with other aspects of C type declarators, judicious use of typedef statements can clarify the
15 code.
The concept of composite types (§6.2.7) was introduced to provide for the accretion of information
20 from incomplete declarations, such as array declarations with missing size, and function
declarations with missing prototype (argument declarations). Type declarators are therefore said to
specify compatible types if they agree except for the fact that one provides less information of this
sort than the other.
25 C9X adds a new array type called a variable length array type. The inability to declare arrays
whose size is known only at execution time was often cited as a primary deterrent to using C as a
numerical computing language. Adoption of some standard notion of execution time arrays was
considered crucial for C’s acceptance in the numerical computing world.
30 The number of elements specified in the declaration of a variable length array type is a runtime
expression. Before C9X, this size expression was required to be an integer constant expression.
C9X makes a distinction between variable length array types and variably modified types, for
example, a pointer to a variable length array. Variable length array types are a subset of all
35 possible variably modified types.
All variably modified types must be declared at either block scope or function prototype scope.
File scope identifiers cannot be declared with a variably modified type. Furthermore, array objects
declared with either the static or extern storage class specifiers cannot be declared with a
40 variable length array type, although block scope pointers declared with the static storage class
specifier can be declared as pointers to variable length array types. Finally, if the identifier that is
being declared has a variable length array type (as opposed to being a pointer to a variable length
array), then it must be an ordinary identifier. This eliminates structure and union members.
45 Restricting variable length array declarators to identifiers with automatic storage duration is natural
since “variableness” at file scope requires some notion of parameterized typing. There was
73
WG14/N897 J11/99-032 C9X RATIONALE
sentiment for allowing structure members to be variably modified; however allowing structure
members to have a variable length array type introduces a host of problems such as the treatment
when passing these objects, or even pointers to these objects, as parameters. In addition, the
semantics of the offsetof macro would need to be extended and runtime semantics added.
5 Finally, there was disagreement whether the size of a variable length array member could be
determined using one of the other members. The Committee decided to limit variable length array
types to declarations outside structures and unions.
Side effects in variable length array size expressions are guaranteed to be produced, except in one
10 context. If a size expression is part of the operand of a sizeof operator, and the result of that
sizeof operator does not depend on the value of the size expression, then it is unspecified
whether side effects are produced. In the following example:
{
15 int n = 5;
int m = 7;
size_t sz = sizeof(int (*)[n++]);
}
20 the value of the result of the sizeof operator is the same as in:
{
int n = 5;
int m = 7;
25 size_t sz = sizeof(int (*)[m++]);
}
Since the value stored in sz does not depend on the size expression, the side effect in n++ is not
guaranteed to occur. Requiring the side effrect introduced a burden on some implementations.
30 Since side effects in this context seemed to have limited utility and are not perceived to be a
desired coding style, the Committee decided to make it unspecified whether these size expressions
are actually evaluated.
A new feature of C9X: The static storage class specifier and any of the type-qualifiers,
35 restrict, const or volatile, can appear inside the [ and ] that are used to declare an
array type, but only in the outermost array type derivation of a function parameter.
The static keyword provides useful information about the intent of function parameters.
Consider:
40
void fadd(double *a, const double *b)
{
int i;
74
C9X RATIONALE WG14/N897 J11/99-032
a[i] += b[i];
}
5 return;
}
It would be a significant advantage on some systems for the translator to initiate, at the beginning
of the function, prefetches or loads of the arrays that will be referenced through the parameters.
10 There is no way in C89 for the user to provide information to the translator about how many
elements are guaranteed to be available.
a[i] += b[i];
}
25
return;
}
guarantees that both the pointers a and b provide access to the first element of an array containing
30 at least ten elements. The static keyword also guarantees that the pointer is not NULL and
points to an object of the appropriate effective type. It does not, however, guarantee that a and b
point to unique, non-overlapping objects. The restrict keyword is used for that purpose as in:
a[i] += b[i];
}
45
return;
}
75
WG14/N897 J11/99-032 C9X RATIONALE
This function definition specifies that the parameters a and b are restricted pointers. This is
information that an optimizer can use, for example, to unroll the loop and reorder the loads and
stores of the elements referenced through a and b.
5
The const keyword can be used to indicate that the pointer will always point to the same array
object. The function declaration:
15 There does not appear to be much value in using volatile to qualify an array function
parameter.
20 The function prototype mechanism is one of the most useful additions to the C language. The
feature, of course, has precedent in many of the Algol-derived languages of the past 25 years. The
particular form adopted in the Standard is based in large part upon C++.
In function calls not in the scope of a function prototype, integer arguments have the integer
30 promotions applied and float arguments are widened to double. It is not possible in such a
call to pass an unconverted char or float argument. Function prototypes give the programmer
explicit control over the function argument type conversions, so that the often inappropriate and
sometimes inefficient default widening rules for arguments can be suppressed by the
implementation.
35
Modifications of function interfaces are easier in cases where the actual arguments are still
assignment compatible with the new formal parameter type: only the function definition and its
prototype need to be rewritten in this case; no function calls need be rewritten. Allowing an
optional identifier to appear in a function prototype serves two purposes:
40
• the programmer can associate a meaningful name with each argument position for
documentation purposes.
• a function declarator and a function prototype can use the same syntax. The
45 consistent syntax makes it easier for new users of C to learn the language.
Automatic generation of function prototype declarators from function definitions is
76
C9X RATIONALE WG14/N897 J11/99-032
also facilitated.
The Standard requires that calls to functions taking a variable number of arguments must occur in
the presence of a prototype using the trailing ellipsis notation , (...). An implementation may
5 thus assume that all other functions are called with a fixed argument list, and may therefore use
possibly more efficient calling sequences. Programs using old-style headers in which the number
of arguments in the calls and the definition differ may not work in implementations which take
advantage of such optimizations. This is not a quiet change, strictly speaking, since the program
does not conform to the Standard. A word of warning is in order, however, since the style is not
10 uncommon in existing code, and since a conforming translator is not required to diagnose such
mismatches when they occur in separate translation units. Such trouble spots can be made
manifest (assuming an implementation provides reasonable diagnostics) by providing new-style
function declarations in the translation units with the non-matching calls. Programmers who
currently rely on being able to omit trailing arguments are advised to recode using the
15 <stdarg.h> paradigm.
A function prototype can have parameters that have variable length array types (§6.7.5.2) using a
30 special syntax as in
This is consistent with other C prototypes where the name of the parameter need not be specified.
35
There was considerable debate about whether to maintain the current lexical ordering rules for
variable length array parameters in function definitions. For example, the following old-style
declaration
void f(a, n)
int n;
double a[n][n];
45 {
// ...
}
77
WG14/N897 J11/99-032 C9X RATIONALE
Previously, programmers did not need to concern themselves with the order in which formal
10 parameters are specified, and one common programming style is to declare the most important
parameters first. With Standard C’s lexical ordering rules, the declaration of a would force n to be
undefined or captured by an outside declaration. The possibility of allowing the scope of
parameter n to extend to the beginning of the parameter-type-list was explored (relaxed lexical
ordering), which would allow the size of parameter a to be defined in terms of parameter n, and
15 could help convert a Fortran library routine into a C function. Such a change to the lexical
ordering rules is not considered to be in the “Spirit of C,” however. This is an unforeseen side
effect of Standard C prototype syntax.
The following example demonstrates how to declare parameters in any order and avoid lexical
20 ordering issues.
In this case, the parameter ap is assigned to a local pointer that is declared to be a pointer to a
30 variable length array. The function g can be called as in
{
double x[10][10];
g(&x[0][0], 10);
35 }
which allows the array address to be passed as the first argument. The strict lexical ordering rules
remain in place.
Empty parentheses within a type name are always taken as meaning function with unspecified
arguments and never as unnecessary parentheses around the elided identifier. This specification
avoids an ambiguity by fiat.
45
78
C9X RATIONALE WG14/N897 J11/99-032
A typedef may only be redeclared in an inner block with a declaration that explicitly contains a
type name. This rule avoids the ambiguity about whether to take the typedef as the type name
5 or the candidate for redeclaration.
Some pre-C89 implementations allowed type specifiers to be added to a type defined using
typedef. Thus
would give x the type unsigned short int. The C89 Committee decided that since this
interpretation may be difficult to provide in many implementations, and since it defeats much of
15 the utility of typedef as a data abstraction mechanism, such type modifications are invalid. This
decision is incorporated in the rules of §6.7.2.
20 In C89, a typedef could be redeclared in an inner block with a declaration that explicitly
contained a type name. This rule avoided the ambiguity about whether to take the typedef as
the type name or a candidate for redeclaration. In C9X, implicit int declarations are not allowed,
so this anbiguity is not possible and the rule is no longer necessary.
25 Using a typedef to declare a variable length array object (see §6.7.5.2) could have two possible
meanings. Either the size could be eagerly computed when the typedef is declared, or the size
could be lazily computed when the object is declared. For example
{
30 typedef VLA[n];
n++;
VLA object;
// ...
35 }
The question arises whether n should be evaluated at the time the type definition itself is
encountered or each time the type definition is used for some object declaration. The Committee
decided that if the evaluation were to take place each time the typedef name is used, then a
40 single type definition could yield variable length array types involving many different dimension
sizes. This possibility seemed to violate the spirit of type definitions. The decision was made to
force evaluation of the expression at the time the type definition itself is encountered.
6.7.8 Initialization
45
An implementation might conceivably have codes for floating zero and/or null pointer other than
79
WG14/N897 J11/99-032 C9X RATIONALE
all bits zero. In such a case, the implementation must fill out an incomplete initializer with the
various appropriate representations of zero; it may not just fill the area with zero bytes.
The C89 Committee considered proposals for permitting automatic aggregate initializers to consist
5 of a brace-enclosed series of arbitrary execution-time expressions, instead of just those usable for a
translation-time static initializer. However, cases like this were troubling:
10 Rather than determine a set of rules which would avoid pathological cases and yet not seem too
arbitrary, the C89 Committee elected to permit only static initializers. Consequently, an
implementation may choose to build a hidden static aggregate, using the same machinery as for
other aggregate initializers, then copy that aggregate to the automatic variable upon block entry.
15 A structure expression, such as a call to a function returning the appropriate structure type, is
permitted as an automatic structure initializer, since the usage seems unproblematic.
For programmer convenience, even though it is a minor irregularity in initializer semantics, the
trailing null character in a string literal need not initialize an array element, as in
20
char mesg[5] = "help!";
25 K&R allows a trailing comma in an initializer at the end of an initializer-list. The Standard has
retained this syntax, since it provides flexibility in adding or deleting members from an initializer
list, and simplifies machine generation of such lists.
Various implementations have parsed aggregate initializers with partially elided braces differently.
30
The Standard has reaffirmed the top-down parse described in K&R. Although the construct is
allowed, and its parse well defined, the C89 Committee urges programmers to avoid partially
elided initializers because such initializations can be quite confusing to read.
Code which relies on a bottom-up parse of aggregate initializers with partially elided
braces will not yield the expected initialized object.
40 The C89 Committee has adopted the rule (already used successfully in some implementations) that
the first member of the union is the candidate for initialization. Other notations for union
initialization were considered, but none seemed of sufficient merit to outweigh the lack of prior art.
This rule has a parallel with the initialization of structures. Members of structures are initialized in
45 the sequence in which they are declared. The same could be said of C89 unions, with the
significant difference that only one union member, the first, can be initialized.
80
C9X RATIONALE WG14/N897 J11/99-032
A new feature of C9X: Designated initializers provide a mechanism for initializing sparse arrays, a
practice common in numerical programming. They add useful functionality that already exists in
Fortran so that programmers migrating to C need not suffer the loss of a program-text-saving
notational feature.
5
This feature also allows initialization of sparse structures, common in systems programming, and
allows initialization of unions via any member, regardless of whether or not it is the first member.
Designated initializers integrate easily into the C grammar and do not impose any additional run-
10 time overhead on a user’s program. Their initial C implementation appeared in a compiler by Ken
Thompson at AT&T Bell Laboratories.
15 The C89 Committee considered proposals for forbidding a goto into a block from outside, since
such a restriction would make possible much easier flow optimization and would avoid the whole
issue of initializing auto storage; but it rejected such a ban out of fear of invalidating working code,
however undisciplined, and out of concern for those producing machine-generated C.
20 A new feature of C9X: A common coding practice is always to use compound statements for every
selection and iteration statement because this guards against inadvertent problems when changes
are made in the future. Because this can lead to surprising behavior in connection with certain uses
of compound literals (§6.5.2.5), the concept of a block has been expanded in C9X.
return *p;
}
40
it seemed surprising that just introducing compound statements also introduced undefined behavior:
81
WG14/N897 J11/99-032 C9X RATIONALE
{
if (*(q = (int[2]){i, j})) {
25 // *q is guaranteed to designate an object
fn(p = (int[5]){9, 8, 7, 6, 5}, q);
} else {
// *q is guaranteed to designate an object
fn(p = (int[5]){4, 3, 2, 1, 0}, q + 1);
30 }
}
If compound literals are defined in selection or iteration statements, their lifetimes are limited to
the implied enclosing block; therefore the definition of “block” has been moved to this section.
40 This change is compatible with similar C++ rules.
There are some pathological cases where program behavior changes quietly as demonstrated by
45 the following example.
82
C9X RATIONALE WG14/N897 J11/99-032
int different(void)
{
if (sizeof(enum {b, a}) != sizeof(int))
5 return a; // a == 1
return b; // which b?
}
10 In C89, the declaration enum {b, a} persists after the if statement terminates; but in C9X, the
implied block that encloses the entire if statement limits the scope of that declaration; therefore
the different function returns different values in C89 and C9X. The Committee views such
cases as unintended artifacts of allowing declarations as operands of cast and sizeof operators;
and this change is not viewed as a serious problem.
15
6.8.1 Labeled statements
Since label definition and label reference are syntactically distinctive contexts, labels are
established as a separate name space.
20
6.8.2 Compound statement
A new feature of C9X: declarations and statements may be mixed in an arbitrary manner (see
§6.2.4).
25
6.8.3 Expression and null statements
The void cast is not needed in an expression statement, since any value is always discarded. Some
compilers prefer this reassurance, however, for functions that return objects of types other than
30 void.
A new feature of C9X: Unlike in C89, all selection statements and their associated substatements
35 are blocks. See §6.8.2.
See §6.8.
40
6.8.4.2 The switch statement
The controlling expression of a switch statement may now have any integer type, even
unsigned long long. Floating types were rejected for switch statements since exact equality
45 in floating point is not portable.
83
WG14/N897 J11/99-032 C9X RATIONALE
case labels are first converted to the type of the controlling expression of the switch, then
checked for equality with other labels. No two may match after conversion.
5 Case ranges of the form, lo .. hi, were seriously considered, but ultimately not adopted in the
Standard on the grounds that it added no new capability, just a problematic coding convenience.
The construct seems to promise more than it could be mandated to deliver:
• A great deal of code or jump table space might be generated for an innocent-looking
10 case range such as 0 .. 65535.
• The range 'A' .. 'Z' would specify all the integers between the character code
for “upper-case-A” and that for “upper-case-Z”. In some common character sets
this range would include non-alphabetic characters, and in others it might not
15 include all the alphabetic characters, especially in non-English character sets.
No serious consideration was given to making switch more structured, as in Pascal, out of fear of
invalidating working code.
long expressions and constants in switch statements are no longer truncated to int.
To simplify the syntax, each loop is limited to a single declaration (though this can declare several
40 variables), and these must have auto or register storage class.
Example:
int i = 42;
45 for (int i = 5, j = 15; i < 10; i++, j--)
printf("Loop %d %d\n", i, j);
84
C9X RATIONALE WG14/N897 J11/99-032
will output:
5 Loop 5 15
Loop 6 14
Loop 7 13
Loop 8 12
Loop 9 11
10 I = 42
A new feature of C9X: In C89, for loops were defined in terms of a syntactic rewrite into while
loops. This introduced problems for the definition of the continue statement; and it also introduced
problems when the operands of cast and sizeof operators contain declarations as in:
20
enum {a, b};
{
int i, j = b;
25
for (i = a; i < j; i += sizeof(enum {b, a}))
j += b;
}
{
35 int i, j = b;
i = a;
while (i < j) {
j += b; // which b?
40 i += sizeof(enum {b, a}); // declaration of b moves
}
}
because a different b is used to increment i in each case. For this reason, the syntactic rewrite
45 has been replaced by words that describe the behavior.
85
WG14/N897 J11/99-032 C9X RATIONALE
5 With the combination of variable length arrays (see §6.7.5.2) and mixed code and declarations,
situations can arise where a variable length array definition is skipped. In the following example
{
int n = 1;
10 goto label;
int a[n];
label:
// ...
15 }
it is problematic to allocate the array a because the goto statement causes a jump past the
declaration. Therefore, it is forbidden to branch from outside the scope of a variably modified
declaration to a point that is inside the scope, although it is permitted to jump from inside the scope
20 to a point outside the scope. In the latter case the translator is expected to deallocate the memory
associated with the variable length array. In the following example
{
int n = 1;
25 label:
int a[n];
// ...
if (n++ < 10) goto label;
}
30
the goto statement causes the array a to be deallocated. It is reallocated with a new size that is the
value of n each time the declaration is encountered. Other automatic objects are not deallocated if a
goto causes them to go out of scope.
The C89 Committee rejected proposed enhancements to continue and break which would
40 allow specification of an iteration statement other than the immediately enclosing one on grounds
of insufficient prior art.
45 See §6.8.6.2.
86
C9X RATIONALE WG14/N897 J11/99-032
5 A function definition may have its old form and say nothing about arguments on calls, or it may be
introduced by a prototype which affects argument checking and coercion on subsequent calls.
To avoid a nasty ambiguity, the Standard bans the use of typedef names as formal parameters.
For instance, in translating the text
10
int f(size_t, a_t, b_t, c_t, d_t, e_t, f_t, g_t,
h_t, i_t, j_t, k_t, l_t, m_t, n_t, o_t,
p_t, q_t, r_t, s_t)
15 the translator determines that the construct can only be a prototype declaration as soon as it scans
the first size_t and following comma. In the absence of this rule, it might be necessary to see the
token following the right parenthesis that closes the parameter list, which would require a sizable
look-ahead, before deciding whether the text under scrutiny is a prototype declaration or an old-
style function header definition.
20
An argument list must be explicitly present in the declarator; it cannot be inherited from a
typedef (see §6.7.5.3). That is to say, given the definition:
p funk // weird
{ return q + r ; }
30
Some current implementations rewrite the type of, for instance, a char parameter as if it were
declared int, since the argument is known to be passed as an int in the absence of a prototype.
The Standard requires, however, that the received argument be converted as if by assignment upon
function entry. Type rewriting is thus no longer permissible.
35
QUIET CHANGE IN C89
Functions that depend on char or short parameter types being widened to int,
or float widened to double, may behave differently.
40
Notes for implementors: the assignment conversion for argument passing often requires no
executable code. In most two’s-complement machines, a short or char is a contiguous subset of
the bytes comprising the int actually passed for even the most unusual byte orderings, so that
assignment conversion can be effected by adjusting the address of the argument if necessary.
45
87
WG14/N897 J11/99-032 C9X RATIONALE
For an argument declared float, however, an explicit conversion must usually be performed from
the double actually passed to the float desired. Not many implementations can subset the
bytes of a double to get a float. Even those that apparently permit simple truncation often get
the wrong answer on certain negative numbers.
5
Some current implementations permit an argument to be masked by a declaration of the same
identifier in the outermost block of a function. This usage is almost always an erroneous attempt by
a novice C programmer to declare the argument; it is rarely the result of a deliberate attempt to
render the argument unreachable. The C89 Committee decided, therefore, that arguments are
10 effectively declared in the outermost block, and hence cannot be quietly redeclared in that block.
The C89 Committee considered it important that a function taking a variable number of arguments,
printf for example, be expressible portably in C. Hence, the C89 Committee devoted much time
to exploring methods of traversing variable argument lists. One proposal was to require arguments
15 to be passed as a “brick,” that is, a contiguous area of memory, the layout of which would be
sufficiently well specified that a portable method of traversing the brick could be determined.
Several diverse implementations, however, can implement argument passing more efficiently if the
arguments are not required to be contiguous. Thus, the C89 Committee decided to hide the
20 implementation details of determining the location of successive elements of an argument list
behind a standard set of macros (see §7.15).
The rule which caused undeclared parameters in an old-style function definition to be implicitly
declared int has been removed: undeclared parameters are now a constraint violation. The
25 effect is to guarantee production of a diagnostic that will catch an additional category of
programming errors. After issuing the diagnostic, an implementation may choose to assume an
implicit int declaration and continue translation in order to support existing programs that
exploited this feature.
Different implementations have had different notions about whether white space is permissible
before and/or after the # signalling a preprocessor line. The C89 Committee decided to allow any
white space before the #, and horizontal white space (spaces or tabs) between the # and the
35 directive, since the white space introduces no ambiguity, causes no particular processing problems,
and allows maximum flexibility in coding style. Note that similar considerations apply for
comments, which are reduced to white space early in the phases of translation (§5.1.1.2):
88
C9X RATIONALE WG14/N897 J11/99-032
5 The operator defined was added to C89 to make possible writing boolean combinations of
defined flags with one another and with other inclusion conditions. If the identifier defined were
to be defined as a macro, defined(X) would mean the macro expansion in C text proper and the
operator expression in a preprocessing directive (or else that the operator would no longer be
available). To avoid this problem, such a definition is not permitted (§6.10.8).
10
#elif was added to minimize the stacking of #endif directives in multi-way conditionals.
Processing of skipped material is defined such that an implementation need only examine a logical
line for the # and then for a directive name. Thus, assuming that xxx is undefined, in this
15 example:
# ifndef xxx
# define xxx "abc"
# elif xxx > 0
20 // ...
# endif
an implementation is not required to diagnose an error for the #elif directive, even though if it
were processed, a syntax error would be detected.
25
Various proposals were considered for permitting text other than comments at the end of directives,
particularly #endif and #else, presumably to label them for more easily matching their
corresponding #if directives. The C89 Committee rejected all such proposals because of the
difficulty of specifying exactly what would be permitted and how the translator would have to
30 process it.
Various proposals were considered for permitting additional unary expressions to be used for the
purpose of testing for the system type, testing for the presence of a file before #include, and
other extensions to the preprocessing language. These proposals were all rejected on the grounds of
35 insufficient prior art and/or insufficient utility.
Specification of the #include directive raises distinctive grammatical problems because the file
40 name is conventionally parsed quite differently from an “ordinary” token sequence:
• The double quotes do not delimit a string literal with all its defined escape
45 sequences (in some systems, backslash is a legitimate character in a filename); the
89
WG14/N897 J11/99-032 C9X RATIONALE
• White space or characters not in the C repertoire may be permissible and significant
within either or both forms.
5
These points in the description of phases of translation are of particular relevance to the parse of the
#include directive:
• The filename on the #include and #line directives, if it does not begin with "
or <, is macro-expanded prior to execution of the directive. Allowing macros in the
#include directive facilitates the parameterization of include file names, an
20 important issue in transportability.
The file search rules used for the filename in the #include directive were left as implementation-
defined. The Standard intends that the rules which are eventually provided by the implementor
correspond as closely as possible to the original K&R rules. The primary reason that explicit rules
25 were not included in the Standard is the infeasibility of describing a portable file system structure.
It was considered unacceptable to include UNIX-like directory rules due to significant differences
between this structure and other popular commercial file system structures.
Nested include files raise an issue of interpreting the file search rules. In UNIX C a #include
30 directive found within an included file entails a search for the named file relative to the file system
directory that holds the outer #include. Other implementations, including the earlier UNIX C
described in K&R, always search relative to the same current directory. The C89 Committee
decided in principle in favor of K&R approach, but was unable to provide explicit search rules as
explained above.
35
The Standard specifies a set of include file names which must map onto distinct host file names. In
the absence of such a requirement, it would be impossible to write portable programs using
included files.
40 Subclause §5.2.4.1 on translation limits contains the required number of nesting levels for included
files. The limits chosen were intended to reflect reasonable needs for users constrained by
reasonable system resources available to implementors.
By defining a failure to read an included file as a syntax error, the Standard requires that the failure
45 be diagnosed. More than one proposal was presented for some form of conditional include, or a
directive such as #ifincludable, but none were accepted by the Committee due to lack of prior
90
C9X RATIONALE WG14/N897 J11/99-032
art.
In C9X, the number of significant characters in header and source file names was raised from six to
eight, and digits were allowed, in the belief that all implementations could support this, and because
5 it could be of help to users.
The specification of macro definition and replacement in the Standard was based on these
10 principles:
• Define macro expansion such that it produces the same token sequence whether the
macro calls appear in open text, in macro arguments, or in macro definitions.
20
Preprocessing is specified in such a way that it can be implemented either as a separate text-to-text
prepass or as a token-oriented portion of the compiler itself. Thus, the preprocessing grammar is
specified in terms of tokens.
25 However the newline character must be a token during preprocessing because the preprocessing
grammar is line-oriented. The presence or absence of white space is also important in several
contexts, such as between the macro name and a following parenthesis in a #define directive. To
avoid overly constraining the implementation, the Standard allows both the preservation of each
white space character (which is easy for a text-to-text prepass) and the mapping of white space into
30 a single “white space” token (which is easier for token-oriented translators).
(in header1.h)
35
#define NBUFS 10
(in header2.h)
40 #define NBUFS 12
which are clearly invitations to serious bugs in a program. There remained, however, the question
of “benign redefinitions,” such as
45 (in header1.h)
91
WG14/N897 J11/99-032 C9X RATIONALE
(in header2.h)
The C89 Committee concluded that safe programming practice is better served by allowing benign
redefinition where the definitions are the same. This allows independent headers to specify their
understanding of the proper value for a symbol of interest to each, with diagnostics generated only
10 if the definitions differ.
The definitions are considered “the same” if the identifier-lists, token sequences, and occurrences of
white space (ignoring the spelling of white space) in the two definitions are identical.
25 The first case might be useful in moving extant code from an implementation in which plain char
is signed to one in which it is unsigned. The second case might be useful in adapting code which
assumes that the sizeof operator yields an int value. The redefinition of const could be
useful in retrofitting more modern C code to an older implementation.
30 As with any other powerful language feature, keyword redefinition is subject to abuse. Users
cannot expect any meaningful behavior to come about from source files starting with
A new feature of C9X: C89 introduced a standard mechanism for defining functions with variable
numbers of arguments, but did not allow any way of writing macros with the same property. For
40 example, there is no way to write a macro that looks like a call to printf.
This facility is now available. The macro definition uses an ellipsis in the same way to indicate a
variable argument list. However, since macro substitution is textual rather than run-time, a
different mechanism is used to indicate where to substitute the arguments: the identifier
45 _ _VA_ARGS_ _. This is replaced by all the arguments that match the ellipsis, including the
commas between them.
92
C9X RATIONALE WG14/N897 J11/99-032
#ifdef DEBUG
#define dfprintf(stream, ...) \
5 fprintf(stream, "DEBUG: " _ _VA_ARGS_ _)
#else
#define dfprintf(stream, ...) ((stream, _ _VA_ARGS_ _, 0))
#endif
For example,
or
25
((stderr, "X = %d\n", x, 0));
If DEBUG is true, this calls fprintf, but first catenating "DEBUG: " to the format (which must
therefore be a simple string). Otherwise it creates a comma expression (so that the arguments are
30 still evaluated) with the value zero.
There must be at least one argument to match the ellipsis. This requirement avoids the problems
that occur when the trailing arguments are included in a list of arguments to another macro or
function. For example, if dprintf had been defined as
35
#define dprintf(format,...) \
dfprintf(stderr, format, _ _VA_ARGS_ _)
and it were allowed for there to be only one argument, then there would be a trailing comma in the
40 expanded form. While some implementations have used various notations or conventions to work
around this problem, the Committee felt it better to avoid the problem altogether. Similarly, the
_ _VA_ARGS_ _ notation was preferred to other proposals for this syntax.
A new feature of C9X: Function-like macro invocations may also now have empty arguments, that
45 is, an argument may consist of no preprocessing tokens. In C89, any argument that consisted of no
preprocessing tokens had undefined behavior, but was noted as a common extension.
93
WG14/N897 J11/99-032 C9X RATIONALE
A function-like macro invocation f() has the form of either a call with no arguments or a call with
one empty argument. Which form it actually takes is determined by the definition of f, which
indicates the expected number of arguments.
5
The sequence
15 float f = xglue(TENTH,F) ;
double d = xglue(TENTH,D) ;
long double ld = xglue(TENTH,LD);
results in
20
float f = 0.1f ;
double d = 0.1 ;
long double ld = 0.1L;
30 Some pre-C89 implementations decided to replace identifiers found within a string literal if they
matched a macro argument name. The replacement text is a “stringized” form of the actual
argument token sequence. This practice appears to be contrary to K&R’s definition of
preprocessing in terms of token sequences. The C89 Committee declined to elaborate the syntax of
string literals to the point where this practice could be condoned; however, since the facility
35 provided by this mechanism seems to be widely used, the C89 Committee introduced a more
tractable mechanism of comparable power.
The # operator, which may be used only in a #define expansion, was introduced for stringizing.
It causes the formal parameter name following to be replaced by a string literal formed by
40 stringizing the actual argument token sequence. In conjunction with string literal concatenation
(see §6.4.5), use of this operator permits the construction of strings as effectively as by identifier
replacement within a string. An example in the Standard illustrates this feature.
One problem with defining the effect of stringizing is the treatment of white space occurring in
45 macro definitions. Where this could be discarded in the past, now upwards of one logical line may
have to be retained. As a compromise between token-based and character-based preprocessing
disciplines, the C89 Committee decided to permit white space to be retained as one bit of
94
C9X RATIONALE WG14/N897 J11/99-032
information: none or one. Arbitrary white space is replaced in the string by one space character.
The remaining problem with stringizing was to associate a “spelling” with each token. The
problem arises in token-based preprocessors that might, for instance, convert a numeric literal to a
5 canonical or internal representation, losing information about base, leading zeros, etc. In the
interest of simplicity, the C89 Committee decided that each token should expand to just those
characters used to specify it in the original source text.
45 • The token resulting from a paste operation is subject to further macro expansion.
95
WG14/N897 J11/99-032 C9X RATIONALE
These principles codify the essential features of prior art and are consistent with the specification of
the stringizing operator.
the expansion will to be either 2*f(9) or 2*9*g: there are no clear grounds for making a
25 decision whether the f(9) token string resulting from the initial expansion of f and the
examination of the rest of the source file should be considered as nested within the expansion of f
or not. The C89 Committee intentionally left this behavior ambiguous as it saw no useful purpose
in specifying all the quirks of preprocessing for such questionably useful constructs.
Some pre-C89 implementations maintained a stack of #define instances for each identifier, and
#undef simply popped the stack. The C89 Committee agreed that more than one level of
#define was more prone to error than utility.
35
It is explicitly permitted to #undef a macro that has no current definition. This capability is
exploited in conjunction with the standard library (see §7.1.4).
45 A new proposal for C9X to allow the #line directive to appear within macro invocations was
considered. The Committe decided to not allow any preprocessor directives to be recognized as
96
C9X RATIONALE WG14/N897 J11/99-032
5 The #error directive was introduced in C89 to provide an explicit mechanism for forcing
translation to fail under certain conditions. Formally, the Standard can require only that a
diagnostic be issued when the #error directive is processed. It is the intent of the Committee,
however, that translation cease immediately upon encountering this directive if this is feasible in the
implementation. Further diagnostics on text beyond the directive are apt to be of little value.
10
6.10.6 Pragma directive
The #pragma directive was added in C89 as the universal method for extending the space of
directives.
15
A new feature of C9X: Some #pragma directives have been standardized; and directives whose
first preprocessing token is STDC are reserved for standardized directives.
25 The rule that these macros may not be redefined or undefined reduces the complexity of the name
space that the programmer and implementor must understand; and it recognizes that these macros
have special built-in properties.
The macros _ _DATE_ _ and _ _TIME_ _ were added in C89 to make available the time of
30 translation. A particular format for the expansion of these macros was specified to aid in parsing
strings initialized by them.
The macros _ _LINE_ _ and _ _FILE_ _ were added in C89 to give programmers access to the
source line number and file name.
35
The macro _ _STDC_ _ allows for conditional translation on whether the translator claims to be
standard-conforming. It is defined as having the value 1. Future versions of the Standard could
define it as 2, 3, etc., to allow for conditional compilation on which version of the Standard a
translator conforms to. The C89 Committee felt that this macro would be of use in moving to a
40 conforming implementation.
A new feature of C9X: C9X adds two additional predefined macros: _ _STDC_IEC_559_ _ and
97
WG14/N897 J11/99-032 C9X RATIONALE
_ _STDC_IEC_559_COMPLEX_ _.
5 A new feature of C9X: As an alternative syntax for a #pragma directive, the _Pragma operator
has the advantage that it can be used in a macro replacement list. If a translator is directed to
produce a preprocessed version of the source file, then expressions involving the unary _Pragma
operator and #pragma directives should be treated consistently in whether they are preserved and
in whether macro invocations within them are expanded.
10
6.11 Future language directions
This subclause includes specific mention of the future direction in which the Committee intends to
extend and/or restrict the language. The contents of this subclause should be considered as quite
15 likely to become a part of the next version of the Standard. Implementors are advised that failure to
take heed of the points mentioned herein is considered undesirable for a conforming
implementation. Users are advised that failure to take heed of the points mentioned herein is
considered undesirable for a conforming program.
The practice of placing the storage class specifier other than first in a declaration was branded as
obsolescent. The Committee felt it desirable to rule out such constructs as
The characterization as obsolescent of the use of the “old style” function declarations and
definitions, that is, the traditional style not using prototypes, signals the Committee’s intent that the
35 new prototype style should eventually replace the old style.
The case for the prototype style is presented in §6.5.2.2 and §6.7.5.3. The gist of this case is that
the new syntax addresses some of the most glaring weaknesses of the language defined in K&R,
that the new style is superior to the old style on every count.
40
It was obviously out of the question to remove syntax used in the overwhelming majority of extant
C code, so the Standard specifies two ways of writing function declarations and function
definitions. Characterizing the old style as obsolescent is meant to discourage its use and to serve
as a strong endorsement by the Committee of the new style. It confidently expects that approval
45 and adoption of the prototype style will make it feasible for some future C Standard to remove the
98
C9X RATIONALE WG14/N897 J11/99-032
99
WG14/N897 J11/99-032 C9X RATIONALE
7. Library
7.1 Introduction
5
The base document for this part of the Standard was the 1984 /usr/group Standard. The /usr/group
document contains definitions of some facilities which were specific to the UNIX operating system
and not relevant to other operating environments, such as pipes, ioctl, file access permissions
and process control facilities. Those definitions were dropped from C89. Other functions were
10 excluded as well because they were non-portable or were ill-defined.
Other facilities not in the /usr/group library but present in many UNIX implementations, such as the
curses (terminal-independent screen handling) library were considered to be more complex and less
essential than the facilities of /usr/group; these functions were not added to the Standard.
15
The prototypes for several library routines were changed in C9X and they now contain the new
keyword restrict as part of some parameter declarations. The restrict keyword allows the
prototype to express what was previously expressed by words.
20 The definition of certain C library routines such as memcpy contain the words:
If copying takes place between objects that overlap, the behavior is undefined.
These words are present because copying between overlapping objects is quite rare, and this
25 allowed vendors to provide efficient implementations of these library routines. Now that
restrict allows users to express these same non-overlapping semantics, it is used in prototype
declarations to demonstrate the utility of the keyword, and to act as guidance to those wishing to
understand how to use it correctly.
35 and the restrict keywords tell the translator that the first two parameters, s1 and s2, are
pointers that point to disjoint data objects. Essentially, this keyword provides the same information
as the words that indicate copying between overlapping objects is not allowed.
Besides the library functions whose specifications state that copying between overlapping objects is
40 not allowed, several others have also had their prototype adorned with the restrict keyword.
For example:
45 A critical question that one asks when deciding if a pointer parameter should be restrict-
100
C9X RATIONALE WG14/N897 J11/99-032
qualified or not is, if copying takes place between overlapping objects, will the function behave as
expected. In the case of the printf function, unexpected behavior occurs in a call such as:
{
5 int *p = malloc(n * sizeof(int));
char *cp = (char *) p;
strcpy(cp, "%s %n %s\n");
Remember that the const qualifier in the printf prototype only guarantees that the parameter
pointing at the format string is read-only. Another alias, p, is allowed to modify the same format
20 string.
Since the implementation costs are high if vendors are forced to cater to this extremely rare case,
the restrict keyword is used to explicitly forbid situations like these.
30 Again, since a character pointer can be a potential alias with other pointers, restrict is used to
make it clear to the translator that parameter s is never an alias with parameter stream when the
fgets function is called in a strictly conforming program.
Finally, the prototypes of certain library functions are adorned with restrict only if the pointer
35 is used to access data. For example:
101
WG14/N897 J11/99-032 C9X RATIONALE
only used to track the location in the wide character string where the search terminated. Thus there
is no possibility of copying taking place between overlapping objects through the lower-level
pointer.
15 The decimal-point character is the single character used in the input or output of floating point
numbers, and may be changed by setlocale. This is a library construct; the decimal point in
numeric literals in C source text is always a period.
25 In many implementations the names of headers are the names of files in special directories. This
implementation technique is not required, however: the Standard makes no assumptions about the
form that a file name may take on any system. Headers may thus have a special status if an
implementation so chooses. Standard headers may even be built into a translator, provided that
their contents do not become “known” until after they are explicitly included. One purpose of
30 permitting these header “files” to be “built in” to the translator is to allow an implementation of the
C language as an interpreter in a free-standing environment where the only “file” support may be a
network interface.
The C89 Committee decided to make library headers “idempotent,” that is, they should be
35 includable any number of times, and includable in any order. This requirement, which reflects
widespread existing practice, may necessitate some protective wrappers within the headers to avoid,
for instance, redefinitions of typedefs. To ensure that such protective wrapping can be made to
work, and to ensure proper scoping of typedefs, standard headers may only be included outside
of any declaration.
40
A common way of providing this “protective wrapping” is
#ifndef _ _ERRNO_H
#define _ _ERRNO_H
45 /* body of <errno.h> */
102
C9X RATIONALE WG14/N897 J11/99-032
/* ... */
#endif
To give implementors maximum latitude in packing library functions into files, all external
identifiers defined by the library are reserved in a hosted environment. This means, in effect, that
no user-supplied external names may match library names, not even if the user function has the
25 same specification. Thus, for instance, strtod may be defined in the same object module as
printf, with no fear that link-time conflicts will occur. Equally, strtod may call printf, or
printf may call strtod, for whatever reason, with no fear that the wrong function will be
called.
30 Also reserved for the implementor are all external identifiers beginning with an underscore, and all
other identifiers beginning with an underscore followed by a capital letter or an underscore. This
gives a name space for writing the numerous behind-the-scenes non-external macros and functions
a library needs to do its job properly.
35 With these exceptions, the Standard assures the programmer that all other identifiers are available,
with no fear of unexpected collisions when moving programs from one implementation to another5.
Note, in particular, that part of the name space of internal identifiers beginning with underscore is
available to the user: translator implementors have not been the only ones to find use for “hidden”
names. C is such a portable language in many respects that the issue of “name space pollution” has
40 been and is one of the principal barriers to writing completely portable code. Therefore the
Standard assures that macro and typedef names are reserved only if the associated header is
explicitly included.
5
See §6.2.2.1 for a discussion of some of the precautions an implementor should take to keep this promise. Note also that any implementation-
defined member names in structures defined in <time.h> and <locale.h> must begin with an underscore, rather than following the pattern of
other names in those structures.
103
WG14/N897 J11/99-032 C9X RATIONALE
To make usage more uniform for both implementor and programmer, the Standard requires that
5 every library function, unless specifically noted otherwise, must be represented as an actual
function, in case a program wishes to pass its address as a parameter to another function. On the
other hand, every library function is now a candidate for redefinition in its associated header as a
macro, provided that the macro performs a “safe” evaluation of its arguments, that is, it evaluates
each of the arguments exactly once and parenthesizes them thoroughly; and provided that its top-
10 level operator is such that the execution of the macro is not interleaved with other expressions.
Two exceptions are the macros getc and putc, which may evaluate their arguments in an unsafe
manner (see §7.19.7.5 and §7.19.7.8).
If a program requires that a library facility be implemented as an actual function, not as a macro,
15 then the macro name, if any, may be erased by using the #undef preprocessing directive (see
§6.10.3.5).
All library prototypes are specified in terms of the “widened” types: an argument formerly declared
as char is now written as int. This ensures that most library functions can be called with or
20 without a prototype in scope, thus maintaining backwards compatibility with pre-C89 code. Note,
however, that since functions like printf and scanf use variable-length argument lists, they
must be called in the scope of a prototype.
The Standard contains an example showing how certain library functions may be “built in” in an
25 implementation that remains conforming.
Unlike in C89, some names are no longer unique in the first six characters. Such uniqueness is
unnecessary because C9X no longer allows that minimum translation limit.
To minimize the number of different methods for program termination, assert is now defined in
45 terms of the abort function.
104
C9X RATIONALE WG14/N897 J11/99-032
Note that defining the macro NDEBUG to disable assertions may change the behavior of a program
with no failing assertion if any argument expression to assert has side effects, because the
expression is no longer evaluated.
5
It is possible to turn assertions off and on in different functions within a translation unit by defining
or undefining NDEBUG and including <assert.h> again. The implementation of this behavior
in <assert.h> is simple: undefine any previous definition of assert before providing the new
one. Thus the header might look like
10
#undef assert
#ifdef NDEBUG
#define assert(ignore) ((void)0)
#else
15 extern void _ _gripe(char *_Expr, char *_File,
int _Line, const char *_Func);
#define assert(expr) \
((expr) ? (void)0 :\
_ _gripe(#expr, _ _FILE_ _, _ _LINE_ _, _ _func_ _))
20 #endif
Note that assert must expand to a void expression, so the more obvious if statement does not
suffice as a definition of assert. Note also the avoidance of names in a header that would
conflict with the user’s name space.
25
7.3 Complex arithmetic <complex.h>
30 The choice of I instead of i for the imaginary unit concedes to the widespread use of the
identifier i for other purposes. The programmer can use a different identifier, say j, for the
imaginary unit by following the inclusion of <complex.h> with
#undef I
35 #define j _Imaginary_I
40 The corresponding real type for the imaginary unit is float so that use of I for algorithmic or
notational convenience will not result in widening types.
On systems with imaginary types, the programmer has the ability to control whether use of the
macro I introduces an imaginary type, by explicitly defining I to be _Imaginary_I or
45 _Complex_I. Disallowing imaginary types is useful for some programs intended to run on
implementations without support for such types.
105
WG14/N897 J11/99-032 C9X RATIONALE
The macro _Imaginary_I provides a test for whether imaginary types are supported (whether
or not the implementation fully supports Annex G).
5 The cis function (cos(x) + I*sin(x)) was considered but rejected because its
implementation is easy and straightforward, even though some implementations could compute
sine and cosine more efficiently in tandem.
Two topologies are commonly used in complex mathematics: the complex plane with its
continuum of infinities, and the Riemann sphere with its single infinity. The complex plane is
15 better suited for transcendental functions, the Riemann sphere for algebraic functions. The
complex types with their multiplicity of infinities provide a useful (though imperfect) model for
the complex plane. The cproj function helps model the Riemann sphere by mapping all
infinities to one, and should be used just before any operation, especially comparisons, that might
give spurious results for any of the other infinities.
20
Note that a complex value with one infinite part and one NaN part is regarded as an infinity, not
a NaN, because if one part is infinite, the complex value is infinite independent of the value of
the other part. For the same reason, cabs returns an infinity if its argument has an infinite part
and a NaN part.
25
7.4 Character Handling <ctype.h>
Pains were taken to eliminate any ASCII dependencies from the definition of the character handling
functions. One notable result of this policy was the elimination of the function isascii, both
30 because of the name and because its function was hard to generalize. Nevertheless, the character
functions are often most clearly explained in concrete terms, so ASCII is used frequently to express
examples.
Since these functions are often used primarily as macros, their domain is restricted to the small
35 positive integers representable in an unsigned char, plus the value of EOF. EOF is traditionally
-1, but may be any negative integer, and hence distinguishable from any valid character code.
These macros may thus be efficiently implemented by using the argument as an index into a small
array of attributes.
40 §7.26.2 warns that names beginning with is and to, when these are followed by lower-case letters,
are subject to future use in adding items to <ctype.h>.
45 The definitions of printing character and control character have been generalized from ASCII.
106
C9X RATIONALE WG14/N897 J11/99-032
Note that none of these functions returns a nonzero value (true) for the argument value EOF.
A new feature of C9X: text processing applications often need to distinguish white space that can
20 occur within lines from white space that also separates lines (for example, see §6.10 regarding use
of whitespace in the preprocessor). This distinction is also a property of POSIX locale definition
files.
30 Pre-C89 libraries had almost equivalent macros, _tolower and _toupper, for these functions.
The Standard now permits any library function to be additionally implemented as a macro provided
that the underlying function must still be present. _toupper and _tolower are thus
unnecessary and were dropped as part of the general standardization of library macros.
<errno.h> is a header invented to encapsulate the error handling mechanism used by many of
the library routines in <math.h> and <stdlib.h>.6
40 The error reporting machinery centered about the setting of errno is generally regarded with
tolerance at best. It requires a “pathological coupling” between library functions and makes use of
6
In early drafts of C89, errno and related macros were defined in <stddef.h>. When the C89 Committee decided that the other definitions
in this header were of such general utility that they should be required even in freestanding environments, it created <errno.h>.
107
WG14/N897 J11/99-032 C9X RATIONALE
a static writable memory cell, which interferes with the construction of shareable libraries.
Nevertheless, the C89 Committee preferred to standardize this existing, however deficient,
machinery rather than invent something more ambitious. In C9X, errno need no longer be set by
math functions.
5
The definition of errno as an lvalue macro grants implementors the license to expand it to
something like *_ _errno_addr(), where the function returns a pointer to the current
modifiable copy of errno.
The floating-point environment as defined here includes only execution-time modes, not the
15 myriad of possible translation-time options that can affect a program’s results. Each such
option’s deviation from this specification should be well documented.
1. the programmer may have to defend against undesirable mode settings, which imposes
intellectual as well as time and space overhead.
25 2. the translator may not know which mode settings will be in effect or which functions change
them at execution time, which inhibits optimization.
C9X addresses these problems without changing the dynamic nature of the modes.
30 An alternate approach would have been to present a model of static modes with explicit
utterances to the translator about what mode settings would be in effect. This would have
avoided any uncertainty due to the global nature of dynamic modes or the dependency on
unenforced conventions. However, some essentially dynamic mechanism still would have been
needed in order to allow functions to inherit (honor) their caller’s modes. The IEC 60559
35 standard requires dynamic rounding direction modes. For the many architectures that maintain
these modes in control registers, implementation of the static model would be more costly. Also,
standard C has no facility, other than pragmas, for supporting static modes.
An implementation on an architecture that provides only static control of modes, for example
40 through opword encodings, still could support the dynamic model, by generating multiple code
streams with tests of a private global variable containing the mode setting. Only modules under
an enabling FENV_ACCESS pragma would need such special treatment.
Translation
45
An implementation is not required to provide a facility for altering the modes for translation-time
108
C9X RATIONALE WG14/N897 J11/99-032
arithmetic, or for making exception flags from the translation available to the executing program.
The language and library provide facilities to cause floating-point operations to be done at
execution time when they can be subjected to varying dynamic modes and their exceptions
detected. The need does not seem sufficient to require similar facilities for translation.
5
The fexcept_t type
fexcept_t does not have to be an integer type. Its values must be obtained by a call to
fegetexceptflag, and cannot be created by logical operations from the exception macros.
10 An implementation might simply implement fexcept_t as an int and use the representations
reflected by the exception macros, but isn’t required to: other representations might contain extra
information about the exceptions. fexcept_t might be a struct with a member for each
exception (that might hold the address of the first or last floating-point instruction that caused
that exception). C9X makes no claims about the internals of an fexcept_t, and so the user
15 cannot inspect it.
Unsupported macros are not defined in order to assure that their use results in a translation error.
20 A program might explicitly define such macros to allow translation of code (perhaps never
executed) containing the macros. An unsupported exception macro should be defined to be 0, for
example
#ifndef FE_INEXACT
25 #define FE_INEXACT 0
#endif
30 Exceptions
In previous drafts of this specification, several of the exception functions returned an int
indicating whether the excepts argument represented supported exceptions. This facility was
deemed unnecessary because excepts & ~FE_ALL_EXCEPT can be used to test invalidity of
35 the excepts argument.
Rounding precision
The IEC 60559 floating-point standard prescribes rounding precision modes (in addition to the
40 rounding direction modes covered by the functions in this section) as a means for systems whose
results are always double or extended to mimic systems that deliver results to narrower formats.
An implementation of C can meet this goal in any of the following ways:
109
WG14/N897 J11/99-032 C9X RATIONALE
or double precision.
3. By providing functions to dynamically set and get rounding precision modes which shorten
results by rounding to IEC 60559 single or double precision. Recommended are functions
5 fesetprec and fegetprec and macros FE_FLTPREC, FE_DBLPREC, and
FE_LDBLPREC, analogous to the functions and macros for the rounding direction modes.
This specification does not include a portable interface for precision control because the IEC
60559 floating-point standard is ambivalent on whether it intends for precision control to be
10 dynamic (like the rounding direction modes) or static. Indeed, some floating-point architectures
provide control modes suitable for a dynamic mechanism, and others rely on instructions to
deliver single- and double-format results suitable only for a static mechanism.
Raising overflow or underflow is allowed to also raise inexact because on some architectures the
30 only practical way to raise an exception is to execute an instruction that has the exception as a
side effect. Any IEC 60559 operation that raises either overflow or underflow raises inexact as
well.
The function is not restricted to accept only valid coincident expressions for atomic operations,
35 so the function can be used to raise exceptions accrued over several operations.
7.6.3 Rounding
7.6.4 Environment
45
110
C9X RATIONALE WG14/N897 J11/99-032
In previous drafts the function returned nonzero to indicate success. This was changed for
consistency with other C functions that return a status indicator.
5
feholdexcept should be effective on typical IEC 60559 implementations which have the
default non-stop mode and at least one other mode for trap handling or aborting. If the
implementation provides only the non-stop mode, then installing the non-stop mode is trivial.
fegetenv(envp);
fesetenv(FE_DFL_ENV);
<float.h> makes available to programmers a set of useful quantities for numerical analysis
20 (see §5.2.4.2.2). This set of quantities has seen widespread use for such analysis, in C and in
other languages, and was recommended by the numerical analysts on the C89 Committee. The
set was chosen so as not to prejudice an implementation’s selection of floating-point
representation. See also §7.10 for general remarks.
25 Regarding the FLT_DIG formula, note that the process of converting an arbitrary value in one
floating-point format to a second floating-point format and then back again so as not to change the
original value requires more precision than might be expected. In general, the formulas for number
of digits needed for base conversions for integers do not work when applied to floating-point
representations.
30
If the radix b is a power of 10, then it is obvious that all decimal numbers with p × log10b digits
convert exactly to a floating-point representation and then convert exactly back to the same decimal
number (as long as p × log10b is an integer).
35 When the radix b is not a power of 10, it can be difficult to find a case where a decimal number
with p × log10b digits fails. Consider a four-bit mantissa system (that is, base b=2 and precision
p=4) used to represent one-digit decimal numbers. While four bits are enough to represent one-
digit numbers, they are not enough to support the conversions of decimal to binary and back to
decimal in all cases (but they are enough for most cases). Consider a power of 2 that is just under
40 9.5e21, for example, 273 = 9.44e21. For this number, the three consecutive one-digit numbers near
that special value and their round-to-nearest representations are:
111
WG14/N897 J11/99-032 C9X RATIONALE
See §MSE.4.
40
7.10 Sizes of integer types <limits.h>
Both <float.h> and <limits.h> are inventions of the C89 Committee. Included in these
headers are various parameters of the execution environment which are potentially useful at
45 compile time, and which are difficult or impossible to determine by other means.
112
C9X RATIONALE WG14/N897 J11/99-032
The availability of this information in headers provides a portable way of tuning a program to
different environments. Requiring that preprocessing always yield the same results as run-time
arithmetic, however, would cause problems for portable compilers (themselves written in C) or for
5 cross-compilers, which would then be required to implement the target machine’s arithmetic on the
host machine.
10 C has become an international language. Users of the language outside the United States have been
forced to deal with the various Americanisms built into the standard library routines. Areas
affected by international considerations include:
Alphabet. The English language uses 26 letters derived from the Latin alphabet which suffice only
15 for English and Swahili; other living languages use either the Latin alphabet plus other
characters, or other non-Latin alphabets or syllabaries.
In English, each letter has an upper-case and lower-case form, but this is not generally the
case. The German “sharp S”, ß, for example, occurs only in lower case. European French
20 usually omits diacritics on upper-case letters. Some scripts do not have the concept of two
cases.
Collation. In both EBCDIC and ASCII the code for “z” is greater than the code for “a”, and so on
for other letters in the alphabet, so a “machine sort” gives not unreasonable results for
25 ordering strings. In contrast, most European languages use a codeset resembling ASCII in
which some of the codes used in ASCII for punctuation characters are used for alphabetic
characters (see §5.2.1). The ordering of these codes is not alphabetic. In some languages
letters with diacritics sort as separate letters; in others they should be collated just as the
unmarked form. In Spanish, “ll” sorts as a single letter following “l”; in German, “ß” sorts
30 like “ss”.
Formatting of numbers and currency amounts. In the United States the period is invariably
used for the decimal point, and this usage was built into the definitions of such functions as
printf and scanf. Prevalent practice in several major European countries is to use a
35 comma; a raised dot is employed in some locales. Similarly, in the United States a comma
is used to separate groups of three digits to the left of the decimal point; but a period is
common in Europe, and in some countries digits are not grouped by threes at all. In printing
currency amounts, the currency symbol (which may be more than one character) may
precede, follow, or be embedded in the digits. Note that the decimal point is a single
40 character, not a multibyte string.
Date and time. The standard function asctime returns a string which includes abbreviations for
month and weekday names, and returns the various elements in a format which might be
considered unusual even in its country of origin.
45
Various common date formats include
113
WG14/N897 J11/99-032 C9X RATIONALE
20 The C89 Committee introduced mechanisms into the C library to allow these and other issues to be
treated in the appropriate locale-specific manner.
25 English for C source. The C language proper is based on English. Keywords are based on
English words. A program which uses “national characters” in identifiers was not strictly
conforming through C95, but C9X allows identifiers to be written using the “universal
character names” (UCNs) of ISO/IEC 10646. (Use of national characters in comments has
always been strictly conforming, though what happens when such a program is printed in a
30 different locale is unspecified.) The decimal point must be a period in C source, and no
thousands delimiter may be used.
Function interface. The locale is changed by calling a function, thus allowing the implementation
40 to recognize the change, rather than by, say, changing a memory location that contains the
decimal point character.
Immediate effect. When a new locale is selected, affected functions reflect the change
immediately. (This is not meant to imply that, if a signal-handling function were to change
45 the selected locale and return to a library function, the return value from that library function
must be completely correct with respect to the new locale.)
114
C9X RATIONALE WG14/N897 J11/99-032
5 The setlocale function provides the mechanism for controlling locale-specific features of the
library. The category argument allows parts of the library to be localized as necessary without
changing the entire locale-specific environment. Specifying the locale argument as a string gives an
implementation maximum flexibility in providing a set of locales. For instance, an implementation
could map the argument string into the name of a file containing appropriate localization
10 parameters; and these files could then be added and modified without requiring any recompilation
of a localizable program.
The localeconv function gives a programmer access to information about how to format
monetary and non-monetary numeric quantities. This sort of interface was considered preferable to
defining conversion functions directly: even with a specified locale, the set of distinct formats that
20 can be constructed from these elements is large; and the ones desired are very application-
dependent.
A new feature of C9X: C9X extends the members in the lconv structure to cover long-standing
POSIX practice and to permit additional flexibility for internationally formatted monetary amounts.
25
7.12 Mathematics <math.h>
Before C9X, the math library was defined only for the floating type double. All the names
formed by appending f or l to a name in <math.h> were reserved to allow for the definition of
30 float and long double libraries; and C9X provides for all three versions of math functions.
The functions ecvt, fcvt, and gcvt have been dropped since their capability is available
through sprintf.
35 Before C89, HUGE_VAL was usually defined as a manifest constant that approximates the largest
representable double value. As an approximation to infinity it is problematic. As a function
return value indicating overflow, it can cause trouble if first assigned to a float before testing,
since a float may not necessarily hold all values representable in a double.
40 After considering several alternatives, the C89 Committee decided to generalize HUGE_VAL to a
positive expression of type double so that it could be expressed as an external identifier naming a
location initialized precisely with the proper bit pattern. It can even be a special encoding for
machine infinity on implementations that support such codes. It need not be representable as a
float however. C9X adds HUGE_VALF and HUGE_VALL.
45
115
WG14/N897 J11/99-032 C9X RATIONALE
Similarly, domain errors before C89 were typically indicated by a zero return, which is not
necessarily distinguishable from a valid result. The C89 Committee agreed to make the return
value for domain errors implementation-defined, so that special machine codes can be used to
advantage. This makes possible an implementation of the math library in accordance with the IEC
5 60559 proposal on floating point representation and arithmetic.
Whether underflow should be considered a range error and cause errno to be set is specified as
10 implementation-defined since detection of underflow is inefficient on some systems. In C9X,
errno is no longer required to be set to EDOM or ERANGE because that is an impediment to
optimization.
The Standard has been crafted to neither require nor preclude any popular floating-point
15 implementation. This principle affects the definition of domain error: an implementation may
define extra domain errors to deal with floating-point arguments such as infinity or “not-a-number”
(NaN).
The C89 Committee considered the adoption of the matherr capability from UNIX System V. In
20 this feature of that system’s math library, any error such as overflow or underflow results in a call
from the library function to a user-defined exception handler named matherr. The C89
Committee rejected this approach for several reasons:
• This style is incompatible with popular floating point implementations such as IEC
25 60559, with its special return codes, or that of VAX/VMS.
• It conflicts with the error-handling style of Fortran, thus making it more difficult to
translate useful bodies of mathematical code from that language to C.
30 • It requires the math library to be reentrant since math routines could be called from
matherr, which may complicate some implementations.
45 Passing an integer, complex, or other non-floating type to a classification macro yields undefined
116
C9X RATIONALE WG14/N897 J11/99-032
behavior.
x - (2*pi) * (int)(x/(2*pi))
15
are ill-advised.
20 The atan2 function is modeled after Fortran’s. It is described in terms of arctan (y/x) for
simplicity. The C89 Committee did not wish to complicate the descriptions by specifying in detail
how to determine the appropriate quadrant, since that should be obvious from normal mathematical
convention. atan2(y,x) is well-defined and finite, even when x is 0; the one ambiguity occurs
when both arguments are 0, because at that point any value in the range of the function could
25 logically be selected. Since valid reasons can be advanced for all the different choices that have
been made in this situation by various implementations, the Standard preserves the implementor’s
freedom to return an arbitrary well-defined value such as 0, to report a domain error, or to return a
NaN.
The tangent function has singularities at odd multiples of π/2, approaching positive infinity from
one side and negative infinity from the other. Implementations commonly perform argument
reduction using the best machine representation of π; and for arguments to tan sufficiently close to
35 a singularity, such reduction may yield a value on the wrong side of the singularity. In view of such
problems, the C89 Committee recognized that tan is an exception to the range error rule (see
§7.12.1) that an overflowing result produces HUGE_VAL properly signed.
The functions frexp, ldexp, and modf are primitives used by the remainder of the library.
There was some sentiment for dropping them for the same reasons that ecvt, fcvt, and gcvt
45 were dropped, but their adherents rescued them for general use. Their use is problematic: on non-
117
WG14/N897 J11/99-032 C9X RATIONALE
binary architectures, ldexp may lose precision and frexp may be inefficient.
5 See §7.12.6.4.
Whether log(0.0) is a domain error or a range error is arguable. The choice in the Standard,
10 range error, is for compatibility with IEC 60559. Some such implementations would represent the
result as –∞, in which case no error is raised.
15 See §7.12.6.7.
See §7.12.6.7.
20
7.12.6.10 The log2 functions
See §7.12.6.7.
The treatment of subnormal x follows the recommendation in IEEE 854, which differs from
IEEE 754 on this point. Even 754 implementations should follow this definition rather than the
one recommended (not required) by 754.
30
7.12.6.12 The modf functions
See §6.12.6.4.
In earlier versions of the specification, this function was called scalb. The name was changed
to avoid conflicting with the Single Unix scalb function whose second argument is double
instead of int. Single Unix’s scalb was not included in C9X as its specification of certain
40 special cases is inconsistent with the C9X approach and because the scalbn and scalbln
functions were considered sufficient.
scalbln, whose second parameter has type long int is provided because the factor required
to scale from the smallest positive floating-point value to the largest finite one, on many
45 implementations, is too large to represent in the minimum-width int format.
118
C9X RATIONALE WG14/N897 J11/99-032
Adding an absolute value operator was rejected by the C89 Committee. An implementation can
provide a built-in function for efficiency.
IEC 60559, unlike the Standard, requires sqrt(-0.) to return a negatively signed magnitude-
zero result. This is an issue on implementations that support a negative floating zero. The Standard
specifies that taking the square root of a negative number (in the mathematical sense of less than 0)
20 is a domain error which requires the function to return an implementation-defined value. This rule
permits implementations to support either the IEC 60559 or vendor-specific floating point
representations.
Since the mathematical gamma function increases in value so quickly (it is around 10306 for an
argument of only 170), the logarithm of gamma extends the useful domain. Also, for computing
30 combinations and permutations, it is the quotient of the (potentially large) gammas that is needed;
taking differences of the lgammas instead allows for calculations without overflow.
In Single Unix, a call to lgamma sets an external variable, signgam, to the sign of gamma(x),
which is -1 if x < 0 && remainder(floor(x), 2) != 0.
35
Note that this specification does not remove the external identifier signgam from the user’s
name space. An implementation that supports lgamma’s setting of signgam as an extension
must still protect the external identifier signgam if defined by the user.
In many other standards, the meaning of gamma has changed over the years. Originally, it
computed the logarithm of the absolute value of the mathematical gamma function, with an external
int, signgam, being set to the sign of the gamma function.
45
119
WG14/N897 J11/99-032 C9X RATIONALE
Then gamma was replaced with lgamma, and gamma was slated to be withdrawn. About that
time, NCEG changed gamma to compute the mathematical gamma function, and that is what was
adopted into C9X CD1; but it appears that the old meaning of gamma has not yet been withdrawn,
so there would have been a conflict between C9X and current industry practice. C9X therefore
5 changed the name in the FCD to tgamma, meaning “true gamma,” to avoid this conflict.
15 (double)(int)x
is ill-advised.
instead of
30 long lrint(double);
long long llrint(double);
long lround(double);
long long llround(double);
There were two changes here. First, the parameter type changed to double to match other
functions which, like these, return an integer-type result; this makes the interface style more
consistent. Second, the names changed to make way for f- and l-suffixed versions of the
40 functions, which become needed because of the first change (otherwise rinttoll could be
either the double version of the long long function or the long double version of the
long function).
For functions with a floating argument and an integer return type, the previous specification took
45 the approach of declaring the parameter to be long double. The rationale was to avoid
unnecessary multiple versions of the function in the interface. The implementation need not
120
C9X RATIONALE WG14/N897 J11/99-032
actually promote a float or double argument to long double, so any potential inefficiency
could be avoided.
With the previous interface, however, a programmer would be left to worry about the risk of
5 incurring a costly promotion to long double. Also, the current specification seems more
consistent with the rest of the interface where all the other functions come in three sizes. (A
programmer might initially be surprised not to find float and double versions.)
The fmod function is defined even if the quotient x/y is not representable. This function is properly
implemented by scaled subtraction rather than by division. The Standard defines the result in terms
15 of the formula x – n × y, where n is some integer. This integer need not be representable, and need
not even be explicitly computed. Thus implementations are advised not to compute the result
using code like
x - y * (int)(x/y)
20
Instead, the result can be computed in principle by subtracting ldexp(y,n) from x, for
appropriately chosen decreasing n, until the remainder is between 0 and x, although efficiency
considerations may dictate a different actual implementation.
25 The result of fmod(y, 0.0) is either a domain error or 0; the result always lies between 0 and y,
so specifying the non-erroneous result as 0 simply recognizes the limit case.
The C89 Committee considered a proposal to use the remainder operator % for this function; but it
was rejected because the operators in general correspond to hardware facilities, and fmod is not
30 supported in hardware on most machines.
The remquo functions are intended for implementing argument reductions which can exploit a
35 few low-order bits of the quotient. Note that x may be so large in magnitude relative to y that an
exact representation of the quotient is not practical.
copysign and signbit need not be consistent with each other if the arithmetic is not
consistent in its treatment of zeros. For example, the IBM S/370 has instructions to flip the sign
bit making it possible to create a negative zero, but ±0.0 × ±1.0 is always +0.0. In this case,
45 copysign will treat –0.0 as positive, while signbit will treat it as negative.
121
WG14/N897 J11/99-032 C9X RATIONALE
It is sometimes desirable to find the next representation after a value in the direction of a
5 previously computed value, maybe smaller, maybe larger. The nextafter functions have a
second floating argument so that the program will not have to include floating-point tests for
determining the direction in such situations. (On some machines, these tests may fail due to
overflow, underflow or roundoff.)
10 For the case x = y, IEC 60559 recommends that x be returned. This specification differs so that
nextafter(-0.0, +0.0) returns +0.0 and nextafter(+0.0, -0.0) returns –0.0.
The nextafter functions can be employed to obtain next values in a particular format. For
example, nextafterf(x, y) will return the next float value after (float)x in the direction
15 of (float)y regardless of the evaluation method.
1. nextafter and nexttoward conform to the usual rules for suffixes and type-generic
macros. Before, nextafterd and nextafter were exceptional on both counts.
25 2. Without the change, nextafterf is not the float version of nextafter, which is
potentially surprising.
3. It better matches prior art, which typically has a nextafter function with two double
parameters.
30
7.12.11.4 The nexttoward functions
The second parameter of the nexttoward function has type long double so that the
uncoerced value of the second argument can be used to determine the direction.
35
7.12.12 Maximum, minimum, and positive difference functions
The names for fmax, fmin and fdim have f prefixes to allow for extension integer versions
following the example of fabs and abs.
40
7.12.13 Floating multiply-add
45 In many cases, clever use of floating (fused) multiply-add leads to much improved code; but its
122
C9X RATIONALE WG14/N897 J11/99-032
unexpected use by the compiler can undermine carefully written code. The FP_CONTRACT
macro can be used to disallow use of floating multiply-add; and the fma function guarantees its
use where desired. Many current machines provide hardware floating multiply-add instructions;
software implementation can be used for others.
5
7.13 Nonlocal jumps <setjmp.h>
jmp_buf must be an array type for compatibility with existing practice: programs typically omit
the address operator before a jmp_buf argument, even though a pointer to the argument is desired,
10 not the value of the argument itself. Thus, a scalar or structure type is unsuitable. Note that a one-
element array of the appropriate type is a valid definition.
20 One proposed requirement on setjmp is that it be usable like any other function, that is, that it be
callable in any expression context, and that the expression evaluate correctly whether the return
from setjmp is direct or via a call to longjmp. Unfortunately, any implementation of setjmp
as a conventional called function cannot know enough about the calling environment to save any
temporary registers or dynamic stack locations used part way through an expression evaluation. (A
25 setjmp macro seems to help only if it expands to inline assembly code or a call to a special built-
in function.) The temporaries may be correct on the initial call to setjmp, but are not likely to be
on any return initiated by a corresponding call to longjmp. These considerations dictated the
constraint that setjmp be called only from within fairly simple expressions, ones not likely to
need temporary storage.
30
An alternative proposal considered by the C89 Committee was to require that implementations
recognize that calling setjmp is a special case7, and hence that they take whatever precautions are
necessary to restore the setjmp environment properly upon a longjmp call. This proposal was
rejected on grounds of consistency: implementations are currently allowed to implement library
35 functions specially, but no other situations require special treatment.
7
This proposal was considered prior to the adoption of the stricture that setjmp be a macro. It can be considered as equivalent to proposing that
the setjmp macro expand to a call to a special built-in compiler function.
123
WG14/N897 J11/99-032 C9X RATIONALE
environment fully, that is, that upon execution of longjmp, all local variables in the environment
of setjmp have the values they did at the time of the longjmp call. Register variables create
problems with this idea. Unfortunately, the best that many implementations attempt with register
variables is to save them in jmp_buf at the time of the initial setjmp call, then restore them to
5 that state on each return initiated by a longjmp call. Since compilers are certainly at liberty to
change register variables to automatic, it is not obvious that a register declaration will indeed be
rolled back. And since compilers are at liberty to change automatic variables to register if their
addresses are never taken, it is not obvious that an automatic declaration will not be rolled back,
hence the vague wording. In fact, the only reliable way to ensure that a local variable retain the
10 value it had at the time of the call to longjmp is to define it with the volatile attribute. Note
this does not apply to the floating-point environment (status flags and control modes) which is part
of the global state just as f is.
Some implementations leave a process in a special state while a signal is being handled. Explicit
15 reassurance must be given to the environment when the signal handler returns. To keep this job
manageable, the C89 Committee agreed to restrict longjmp to only one level of signal handling.
The longjmp function should not be called in an exit handler, that is, a function registered with
the atexit function (see §7.20.4.2), since it might jump to code that is no longer in scope.
20
7.14 Signal handling <signal.h>
This facility was retained from /usr/group since the C89 Committee felt it important to provide
some standard mechanism for dealing with exceptional program conditions. Thus a subset of the
25 signals defined in UNIX were retained in the Standard, along with the basic mechanisms of
declaring signal handlers and, with adaptations, raising signals (see §7.14.2.1). For a discussion of
the problems created by including signals, see §5.2.3.
The signal machinery contains many misnomers: SIGFPE, SIGILL, and SIGSEGV have their
30 roots in PDP-11 hardware terminology, but the names are too entrenched to change. The
occurrence of SIGFPE, for instance, does not necessarily indicate a floating-point error. A
conforming implementation is not required to field any hardware interrupts.
The C89 Committee has reserved the space of names beginning with SIG to permit
35 implementations to add local names to <signal.h>. This implies that such names should not be
otherwise used in a C source file which includes <signal.h>.
When a signal occurs, the normal flow of control of a program is interrupted. If a signal occurs that
is being trapped by a signal handler, that handler is invoked. When it is finished, execution
continues at the point at which the signal occurred. This arrangement could cause problems if the
45 signal handler invokes a library function that was being executed at the time of the signal. Since
124
C9X RATIONALE WG14/N897 J11/99-032
library functions are not guaranteed to be reentrant, they should not be called from a signal handler
that returns (see §5.2.3). A specific exception to this rule was granted for calls to signal from
within the signal handler; otherwise, the handler could not reliably reset the signal.
5 The specification that some signals may be effectively set to SIG_IGN instead of SIG_DFL at
program startup allows programs under UNIX systems to inherit this effective setting from parent
processes.
For performance reasons, UNIX does not reset SIGILL to default handling when the handler is
10 called (usually to emulate missing instructions). This treatment is sanctioned by specifying that
whether reset occurs for SIGILL is implementation-defined.
The raise function replaces /usr/group’s kill function. The latter has an extra argument which
refers to the “process ID” affected by the signal. Since the execution model of the Standard does
not deal with multi-processing, the C89 Committee deemed it preferable to introduce a function
20 which requires no process argument. The kill function has been standardized in the POSIX
specification.
These macros, modeled, after the UNIX <varargs.h> macros, have been added to enable the
portable implementation in C of library functions such as printf and scanf (see §7.19.6). Such
implementation could otherwise be difficult, considering newer machines that may pass arguments
30 in machine registers rather than using the more traditional stack-oriented methods.
The definitions of these macros in the Standard differ from their forebears: they have been
extended to support argument lists that have a fixed set of arguments preceding the variable list.
35 va_start and va_arg must exist as macros, since va_start uses an argument that is passed
by name and va_arg uses an argument which is the name of a data type. Using #undef on these
names leads to undefined behavior.
The va_list type is not necessarily assignable, however a function can pass a pointer to its
40 initialized argument list object as noted below. The wording has been changed in C9X to state
clearly that va_list is an object type.
125
WG14/N897 J11/99-032 C9X RATIONALE
Changing an arbitrary type name into a type name which is a pointer to that type could require
sophisticated rewriting. To allow the implementation of va_arg as a macro, va_arg need only
correctly handle those type names that can be transformed into the appropriate pointer type by
5 appending a *, which handles most simple cases. Typedefs can be defined to reduce more
complicated types to a tractable form. When using these macros, it is important to remember that
the type of an argument in a variable argument list will never be an integer type smaller than int,
nor will it ever be float (see §6.7.5.3).
10 va_arg can only be used to access the value of an argument, not to obtain its address.
A much simpler approach is to copy the va_list object used to represent processing of the
arguments. However, there is no safe way to do this in C89 because the object may include
pointers to memory allocated by the va_start macro and destroyed by the va_end macro. The
25 new va_copy macro provides this safe mechanism.
Calling the va_copy macro exactly duplicates the state of a va_list object; therefore an
identical call to the va_arg macro on the two objects will produce the same results, and both
objects must be cleaned up with separate calls to the va_end macro.
30
7.15.1.3 The va_end macro
va_end must also be called from within the body of the function having the variable argument list.
In many implementations, this is a do-nothing operation; but those implementations that need it
35 probably need it badly.
va_start must be called within the body of the function whose argument list is to be traversed.
40 That function can then pass a pointer to its va_list object to other functions to do the actual
traversal, or it can traverse the list itself.
The parmN argument to va_start was intended to be an aid to implementors writing the
definition of a conforming va_start macro entirely in C, even using pre-C89 compilers (for
45 example, by taking the address of the parameter). The restrictions on the declaration of the parmN
126
C9X RATIONALE WG14/N897 J11/99-032
parameter follow from the intent to allow this kind of implementation, as applying the & operator to
a parameter name might not produce the intended result if the parameter’s declaration did not meet
these restrictions.
5 In practice, many current implementations have “hidden machinery” that is used by the va_start
macro to diagnose incorrect usage (for example, to verify that parmN actually is the name of the
last fixed parameter) or to handle more complex argument passing mechanisms. Such machinery
would be capable of handling any kind of parameter without restriction, but the C89 Committee
saw no compelling reason to lift these restrictions, as that would require all implementations to
10 have such machinery.
Multiple va_list variables can be in use simultaneously in the same function; each requires its
own calls to va_start and va_end.
The offsetof macro was added to provide a portable means of determining the offset, in bytes,
35 of a member within its structure. This capability is useful in programs, such as are typical in
database implementations, which declare a large number of different data structures: it is desirable
to provide “generic” routines that work from descriptions of the structures, rather than from the
structure declarations themselves.8
8
Consider, for instance, a set of nodes (structures) which are to be dynamically allocated and garbage-collected, and which can contain pointers to
other such nodes. A possible implementation is to have the first field in each node point to a descriptor for that node. The descriptor includes a table
of the offsets of fields which are pointers to other nodes. A garbage-collector "mark" routine needs no further information about the content of the
node (except, of course, where to put the mark). New node types can be added to the program without requiring the mark routine to be rewritten or
even recomplied.
127
WG14/N897 J11/99-032 C9X RATIONALE
(size_t)&(((s_name*)0)->m_name)
5 or
(size_t)(char*)&(((s_name*)0)->m_name)
or, where X is some predeclared address (or 0) and A(Z) is defined as ((char*)&Z),
10
(size_t)(A((s_name*)X->m_name ) - A(X))
It was not feasible, however, to mandate any single one of these forms as a construct guaranteed to
be portable. Some implementations may choose to expand this macro as a call to a built-in function
15 that interrogates the translator’s symbol table.
Note that these can be implementation-defined types that are wider than long long.
Many implementations of the C runtime environment, most notably the UNIX operating system,
40 provide, aside from the standard I/O library’s fopen, fclose, fread, fwrite, and fseek, a
set of unbuffered I/O services, open, close, read, write, and lseek. The C89 Committee
decided not to standardize the latter set of functions.
Additional semantics for these functions may be found in the POSIX standard. The standard I/O
45 library functions use a file pointer for referring to the desired I/O stream. The unbuffered I/O
128
C9X RATIONALE WG14/N897 J11/99-032
services use a file descriptor (a small integer) to refer to the desired I/O stream.
Due to weak implementations of the standard I/O library, many implementors have assumed that
the standard I/O library was used for small records and that the unbuffered I/O library was used for
5 large records. However, a good implementation of the standard I/O library can match the
performance of the unbuffered services on large records. The user also has the capability of tuning
the performance of the standard I/O library (with setvbuf) to suit the application.
Some subtle differences between the two sets of services can make the implementation of the
10 unbuffered I/O services difficult:
• The model of a file used in the unbuffered I/O services is an array of characters.
Many C environments do not support this file model.
15 • Difficulties arise when handling the newline character. Many hosts use conventions
other than an in-stream newline character to mark the end of a line. The unbuffered
I/O services assume that no translation occurs between the program’s data and the
file data when performing I/O, so either the newline character translation would be
lost (which breaks programs) or the implementor must be aware of the newline
20 translation (which results in non-portable programs).
In summary, the C89 Committee chose not to standardize the unbuffered I/O services because
• The performance of the standard I/O services can be the same or better than the
unbuffered I/O services.
35 • The unbuffered I/O file model may not be appropriate for many C language
environments.
7.19.1 Introduction
40 The macros _IOFBF, _IOLBF, and _IONBF are enumerations of the third argument to
setvbuf, a function adopted from UNIX System V.
SEEK_CUR, SEEK_END, and SEEK_SET have been moved to <stdio.h> from a header
specified in /usr/group and not retained in the Standard.
45
FOPEN_MAX and TMP_MAX were added as environmental limits of some interest to programs that
129
WG14/N897 J11/99-032 C9X RATIONALE
FILENAME_MAX is provided so that buffers to hold file names can be conveniently declared. If
the target system supports arbitrarily long filenames, the implementor should provide some
5 reasonable value (80, 255, 509, etc.) rather than something unusable like USHRT_MAX.
The fpos_t wording has been changed in C9X to exclude array type objects. If fpos_t were an
array, then a function would not be able to handle fpos_t parameters in the same manner as other
fpos_t variables.
10
7.19.2 Streams
C inherited its notion of text streams from the UNIX environment in which it was born. Having
each line delimited by a single newline character, regardless of the characteristics of the actual
15 terminal, supported a simple model of text as a sort of arbitrary length scroll or “galley.” Having a
channel that is “transparent” (no file structure or reserved data encodings) eliminated the need for a
distinction between text and binary streams.
The I/O facilities defined by the Standard are both more complex and more restrictive than the
ancestral I/O facilities of UNIX. This is justified on pragmatic grounds: most of the differences,
25 restrictions and omissions exist to permit C I/O implementations in environments which differ from
the UNIX I/O model.
30 The definition of lines. In the UNIX model, division of a file into lines is effected by newline
characters. Different techniques are used by other systems: lines may be separated by
CR-LF (carriage return, line feed) or by unrecorded areas on the recording medium; or each
line may be prefixed by its length. The Standard addresses this diversity by specifying that
newline be used as a line separator at the program level, but then permitting an
35 implementation to transform the data read or written to conform to the conventions of the
environment.
Some environments represent text lines as blank-filled fixed-length records. Thus the
Standard specifies that it is implementation-defined whether trailing blanks are removed
40 from a line on input. (This specification also addresses the problems of environments
which represent text as variable-length records, but do not allow a record length of 0: an
empty line may be written as a one-character record containing a blank, and the blank is
stripped on input.)
45 Transparency. Some programs require access to external data without modification. For instance,
transformation of CR-LF to a newline character is usually not desirable when object code is
130
C9X RATIONALE WG14/N897 J11/99-032
processed. The Standard defines two stream types, text and binary, to allow a program to
define, when a file is opened, whether the preservation of its exact contents or of its line
structure is more important in an environment which cannot accurately reflect both.
5 Random access. The UNIX I/O model features random access to data in a file, indexed by
character number. On systems where a newline character processed by the program
represents an unknown number of physically recorded characters, this simple mechanism
cannot be consistently supported for text streams. The Standard abstracts the significant
properties of random access for text streams: the ability to determine the current file
10 position and then later reposition the file to the same location. ftell returns a file position
indicator, which has no necessary interpretation except that an fseek operation with that
indicator value will position the file to the same place. Thus an implementation may
encode whatever file positioning information is most appropriate for a text file, subject only
to the constraint that the encoding be representable as a long. Use of fgetpos and
15 fsetpos removes even this constraint.
Buffering. UNIX allows the program to control the extent and type of buffering for various
purposes. For example, a program can provide its own large I/O buffer to improve
efficiency, or can request unbuffered terminal I/O to process each input character as it is
20 entered. Other systems do not necessarily support this generality. Some systems provide
only line-at-a-time access to terminal input; some systems support program-allocated
buffers only by copying data to and from system-allocated buffers for processing. Buffering
is addressed in the Standard by specifying UNIX-like setbuf and setvbuf functions,
but permitting great latitude in their implementation. A conforming library need neither
25 attempt the impossible nor respond to a program attempt to improve efficiency by
introducing additional overhead.
Thus, the Standard imposes a clear distinction between text streams, which must be mapped to suit
local custom, and binary streams, for which no mapping takes place. Local custom on UNIX and
30 related systems is of course to treat the two sorts of streams identically, and nothing in the Standard
requires any change to this practice.
Even the specification of binary streams requires some changes to accommodate a wide range of
systems. Because many systems do not keep track of the length of a file to the nearest byte, an
35 arbitrary number of characters may appear on the end of a binary stream directed to a file. The
Standard cannot forbid this implementation, but does require that this padding consist only of null
characters. The alternative would be to restrict C to producing binary files digestible only by other
C programs; this alternative runs counter to the spirit of C.
40 The set of characters required to be preserved in text stream I/O are those needed for writing C
programs; the intent is that the Standard should permit a C translator to be written in a maximally
portable fashion. Control characters such as backspace are not required for this purpose, so their
handling in text streams is not mandated.
45 It was agreed that some minimum maximum line length must be mandated, and 254 was chosen for
C89. C9X increases this limit to 4095.
131
WG14/N897 J11/99-032 C9X RATIONALE
7.19.3 Files
The as if principle is once again invoked to define the nature of input and output in terms of just
5 two functions, fgetc and fputc. The actual primitives in a given system may be quite different.
The distinction between buffered and unbuffered streams suggests the desired interactive behavior;
but an implementation may still be conforming even if delays in a network or terminal controller
prevent output from appearing in time. It is the intent that matters here.
10
No constraints are imposed upon file names except that they must be representable as strings with
no embedded null characters.
/usr/group provides the unlink system call to remove files. The UNIX-specific definition of this
function prompted the C89 Committee to replace it with a portable function.
20
7.19.4.2 The rename function
This function was added to provide a system-independent atomic operation to change the name of
an existing file; /usr/group only provided the link system call, which gives the file a new name
25 without removing the old one, and which is extremely system-dependent.
The C89 Committee considered a proposal that rename should quietly copy a file if simple
renaming couldn’t be performed in some context, but rejected this as potentially too expensive at
execution time.
30
rename is meant to give access to an underlying facility of the execution environment’s operating
system. When the new name is the name of an existing file, some systems allow the renaming and
delete the old file or make it inaccessible by that name, while others prohibit the operation. The
effect of rename is thus implementation-defined.
35
7.19.4.3 The tmpfile function
The tmpfile function is intended to allow users to create binary “scratch” files. The as if
principle implies that the information in such a file need never actually be stored on a file-structured
40 device.
The temporary file is created in binary update mode because it will presumably be first written and
then read as transparently as possible. Trailing null-character padding may cause problems for
some existing programs.
45
132
C9X RATIONALE WG14/N897 J11/99-032
This function allows for more control than tmpfile: a file can be opened in binary mode or text
mode, and files are not erased at completion.
5
There is always some time between the call to tmpnam and the use in fopen of the returned
name. Hence it is conceivable that in some implementations the name, which named no file at the
call to tmpnam, has been used as a filename by the time of the call to fopen. Implementations
should devise name generation strategies which minimize this possibility, but users should allow
10 for it.
A new feature of C9X: the C9X committee recognized that the C89 specification had a serious
flaw: if tmpnam were called fewer than TMP_MAX times but was unable to generate a suitable
string because every potential string named an existing file, there was no way to report failure and
15 no undefined beharior, so there was no option other than to never return. C9X resolved the
problem by allowing tmpnam to return a null pointer when it cannot generate a suitable string and
by specifying that TMP_MAX is the number of potential strings, any or all of which may name
existing files and thus not be suitable return values.
Code that calls tmpnam and doesn’t check for a null return value may produce
undefined behavior.
On some operating systems, it is difficult or impossible to create a file unless something is written
30 to the file. A maximally portable program which relies on a file being created must write
something to the associated stream before closing it.
35 The fflush function ensures that output has been forced out of internal I/O buffers for a specified
stream. Occasionally, however, it is necessary to ensure that all output is forced out, and the
programmer may not conveniently be able to specify all the currently open streams, perhaps
because some streams are manipulated within library packages.9 To provide an implementation-
independent method of flushing all output buffers, the Standard specifies that this is the result of
40 calling fflush with a NULL argument.
9
For instance, on a system (such as UNIX) which supports process forks, it is usually necessary to flush all output buffers just prior to the fork.
133
WG14/N897 J11/99-032 C9X RATIONALE
The b type modifier was added to deal with the text/binary dichotomy (see §7.19.2). Because of
the limited ability to seek within text files (see §7.19.9.1), an implementation is at liberty to treat the
old update + modes as if b were also specified.
5 Table 7.1 tabulates the capabilities and actions associated with the various specified mode string
arguments to fopen.
134
C9X RATIONALE WG14/N897 J11/99-032
r w a r+ w+ a+
file must exist before open ✔ ✔
old file contents discarded on open ✔ ✔
stream can be read ✔ ✔ ✔ ✔
stream can be written ✔ ✔ ✔ ✔ ✔
stream can be written only at end ✔ ✔
5
Other specifications for files, such as record length and block size, are not specified in the Standard
due to their widely varying characteristics in different operating environments. Changes to file
access modes and buffer sizes may be specified using the setvbuf function (see §7.19.5.6). An
implementation may choose to allow additional file specifications as part of the mode string
10 argument. For instance,
might be a reasonable extension on a system that provides record-oriented binary files and allows a
15 programmer to specify record length.
setbuf is subsumed by setvbuf; but it has been retained for compatibility with old code.
135
WG14/N897 J11/99-032 C9X RATIONALE
provided for a stream, so a program must never expect the buffer’s contents to reflect I/O
operations. Furthermore, the Standard does not require that the requested buffering be
implemented; it merely mandates a standard mechanism for requesting whatever buffering services
might be provided.
5
Although three types of buffering are defined, an implementation may choose to make one or more
of them equivalent. For example, a library may choose to implement line buffering for binary files
as equivalent to unbuffered I/O, or it may choose to always implement full buffering as equivalent
to line buffering.
10
The general principle is to provide portable code with a means of requesting the most appropriate
popular buffering style, but not to require an implementation to support these styles.
A new feature of C9X: C90 was not clear about what, if anything, the size argument means when
15 buf is a null pointer. Existing practice is mixed: some implementations ignore it completely,
other implementations use it as guidance for determining the size of the buffer allocated by
setvbuf. C9X gives warning that size might not be ignored in this case, so portable programs
must be sure to supply a reasonable value.
The %hh, %ll and %lf length modifiers were added in C9X (see §7.19.6.2).
25
Use of the L modifier with floating conversions was added in C89 to deal with formatted output of
the long double type.
Note that the %X and %x formats expect a corresponding unsigned int argument, and %lX and
30 %lx must be supplied with an unsigned long argument.
The %i conversion specifier was added in C89 for programmer convenience to provide symmetry
with fscanf’s %i conversion specifier, even though it has exactly the same meaning as the %d
conversion specifier when used with fprintf.
35
The %p conversion specifier was added to C89 for pointer conversion since the size of a pointer is
not necessarily the same as the size of any integer type. Because an implementation may support
more than one size of pointer, the corresponding argument is expected to be a pointer to void.
40 The %n format was added to C89 to permit ascertaining the number of characters converted up to
that point in the current invocation of the formatter.
136
C9X RATIONALE WG14/N897 J11/99-032
Some existing implementations provide %D and %O as synonyms or replacements for %ld and %lo.
The C89 Committee considered the latter notation preferable.
The C89 Committee reserved lower case conversion specifiers for future standardization.
5
The use of leading zero in field widths to specify zero padding is superseded by a precision field.
The older mechanism was retained.
Some implementations have provided the %r format as a means of indirectly passing a variable-
10 length argument list. The functions vfprintf, etc., are considered to be a more controlled
method of effecting this indirection, so %r was not adopted in the Standard (see §7.19.6.8).
15 The C89 translation limit of 509 characters produced from a single conversion specifier was
increased to 4095 in C9X (approximately an eight-fold increase) to reflect the increase in the
minimum amount of memory available in the C9X target machine (see §5.2.4.1).
The printing formats for numbers are not entirely specified. The requirements of the Standard are
20 loose enough to allow implementations to handle such cases as signed zero, NaN, and infinity in an
appropriate fashion. These have been improved in C9X.
Binary implementations can choose the hexadecimal digit to the left of the decimal-point
character so that subsequent digits align to nibble boundaries. For example, the next value
25 greater than one in the common IEC 60559 80-bit extended format could be
0x8.000000000000001p-3
The next value less than one in IEC 60559 double could be
30
0x1.fffffffffffffp-1
Note that if the precision is missing, trailing zeros may be omitted. For example, the value
positive zero might be
35
0x0.p+0
The more suggestive conversion specifiers for hexadecimal formatting, namely x and h, were
unavailable; and since h was taken, H was ruled out in favor of a lower/upper case option.
40 Possibilities other than a included: b j k m q r t v w y z. The optional h to indicate
hexadecimal floating, as in %he, was deemed a less natural fit with the established scheme for
specifiers and options.
137
WG14/N897 J11/99-032 C9X RATIONALE
For binary-to-decimal conversion, the infinitely precise result is just the source value, and the
destination format’s values are the numbers representable with the given format specifier. The
5 number of significant digits is determined by the format specifier, and in the case of fixed-point
conversion by the source value as well.
• As soon as one specified conversion fails, the whole function invocation fails.
Input pointer conversion with %p was added to C89, although it is obviously risky, for symmetry
with fprintf. The %i format was added to permit the scanner to determine the radix of the
number in the input stream; the %n format was added to make available the number of characters
25 scanned thus far in the current invocation of the scanner. C9X adds %a and %A.
An implementation must not use the ungetc function to perform the necessary one-character
30 pushback. In particular, since the unmatched text is left “unread,” the file position indicator as
reported by the ftell function must be the position of the character remaining to be read.
Furthermore, if the unread characters were themselves pushed back via ungetc, the pushback in
fscanf could not affect the pushback stack in ungetc. A scanf call that matches N characters
from a stream must leave the stream in the same state as if N consecutive getc calls had been
35 made.
A new feature of C9X: The hh and ll length modifiers were added in C9X. ll supports the new
long long int type. hh adds the ability to treat character types the same as all other integer
types; this can be useful in implementing macros such as SCNd8 in <inttypes.h> (see 7.18).
40
7.19.6.3 The printf function
138
C9X RATIONALE WG14/N897 J11/99-032
15 In the interests of minimizing redundancy, sprintf has subsumed the older, rather uncommon,
ecvt, fcvt, and gcvt.
20 The behavior of sscanf on encountering end of string has been clarified. See also comments in
§7.19.6.2 above.
25 The functions vfprintf, vprintf, and vsprintf were adopted from UNIX System V to
facilitate writing special purpose formatted output functions.
30 The functions vfscanf, vscanf, and vsscanf were adopted analogously to the vfprintf
functions to facilitate writing special-purpose formatted intput functions.
35 See §7.19.6.8.
See §7.19.6.9.
40
7.19.6.12 The vsnprintf function
See §7.19.6.5.
139
WG14/N897 J11/99-032 C9X RATIONALE
See §7.19.6.8.
Because much existing code assumes that fgetc and fputc are the actual functions equivalent to
the macros getc and putc, the Standard requires that they not be implemented as macros.
This function subsumes gets which has no limit to prevent storage overwrite on arbitrary input
(see §7.19.7.7).
See §7.19.7.1.
Because gets does not check for buffer overrun, it is generally unsafe to use when its input is not
under the programmer’s control. This has caused some to question whether it should appear in the
Standard at all. The Committee decided that gets was useful and convenient in those special
35 circumstances when the programmer does have adequate control over the input, and as
longstanding existing practice, it needed a standard specification. In general, however, the
preferred function is fgets (see §7.19.7.2).
45 puts(s) is not exactly equivalent to fputs(stdout,s); and puts also writes a newline after
140
C9X RATIONALE WG14/N897 J11/99-032
5 /usr/group requires that at least one character be read before ungetc is called in certain
implementation-specific cases. The C89 Committee removed this requirement, thus obliging a
FILE structure to have room to store one character of pushback regardless of the state of the buffer.
It felt that this degree of generality makes clearer the ways in which the function may be used. The
C9X Committee decided to deprecate the use of ungetc on a binary file at the beginning of the
10 file because of the impossibility of distinguishing between successful and error returns from the
ftell function, both of which would be –1L.
It is permissible to push back a different character than that which was read, which accords with
common existing practice. The last-in, first-out nature of ungetc has been clarified.
15
ungetc is typically used to handle algorithms such as tokenization which involve one-character
lookahead in text files. fseek and ftell are used for random access, typically in binary files.
So that these disparate file-handling disciplines are not unnecessarily linked, the value of a text
file’s file position indicator immediately after ungetc has been specified as indeterminate.
20
Existing practice relies on two different models of the effect of ungetc. One model can be
characterized as writing the pushed-back character “on top of” the previous character. This model
implies an implementation in which the pushed-back characters are stored within the file buffer and
bookkeeping is performed by setting the file position indicator to the previous character position.
25 (Care must be taken in this model to recover the overwritten character values when the pushed-back
characters are discarded as a result of other operations on the stream.) The other model can be
characterized as pushing the character “between” the current character and the previous character.
This implies an implementation in which the pushed-back characters are specially buffered (within
the FILE structure, say) and accounted for by a flag or count. In this model it is natural not to
30 move the file position indicator. The indeterminacy of the file position indicator while pushed-back
characters exist accommodates both models.
Mandating either model by specifying the effect of ungetc on a text file’s file position indicator
creates problems with implementations that have assumed the other model. Requiring the file
35 position indicator not to change after ungetc would necessitate changes in programs which
combine random access and tokenization on text files, and rely on the file position indicator
marking the end of a token even after pushback. Requiring the file position indicator to back up
would create severe implementation problems in certain environments, since in some file
organizations it can be impossible to find the previous input character position without having read
40 the file sequentially to the point in question.10
10
Consider, for instance, a sequential file of variable-length records in which a line is represented as a count field followed by the characters in the
line. The file position indicator must encode a character position as the position of the count field plus an offset into the line; from the position of the
count field and the length of the line, the next count field can be found. Insufficient information is available for finding the previous count field, so
backing up from the first character of a line necessitates, in the general case, a sequential read from the start of the file.
141
WG14/N897 J11/99-032 C9X RATIONALE
5 size_t is the appropriate type both for an object size and for an array bound (see §6.5.3.4), so this
is the type of both size and nelem.
10 See §7.19.8.1.
Given these restrictions, the Committee still felt that this function has enough utility, and is used in
sufficient existing code, to warrant its retention in the Standard. fgetpos and fsetpos were
added to deal with files that are too large to handle with fseek and ftell.
30
The fseek function will reset the end-of-file flag for the stream; the error flag is not changed
unless an error occurs, in which case it will be set.
• the stream is associated with a terminal or some other file type for which file
position indicator is meaningless.
40
• the file may be positioned at a location not representable in a long.
Thus a method for ftell to report failure was specified (see also §7.19.9.1).
142
C9X RATIONALE WG14/N897 J11/99-032
Resetting the end-of-file and error indicators was added to the specification of rewind to make the
specification more logically consistent.
At various times, the C89 Committee considered providing a form of perror that delivers up an
10 error string version of errno without performing any output. It ultimately decided to provide this
capability in a separate function, strerror (see §7.21.6.2).
15 The header <stdlib.h> was invented by the C89 Committee to hold an assortment of functions
that were otherwise homeless.
atof, atoi, and atol are subsumed by strtod and strtol, but were retained because they
are used extensively in existing code. They are less reliable, but may be faster if the argument is
known to be in a valid range.
25
This specification does not require float and long double versions of atof, but instead
encourages the use of strtof and strtold which have a more generally useful interface.
35 strtod was adopted for C89 from UNIX System V because it offers more control over the
conversion process, and because it is required not to produce unexpected results on overflow during
conversion. strtol (§7.20.1.4) was adopted for the same reason. C9X adds strtof and
strtold.
143
WG14/N897 J11/99-032 C9X RATIONALE
Requiring a sign for NaN or infinity input was considered as a way of minimizing the chance of
mistakenly accepting nonnumeric input. The need for this was deemed insufficient, partly on the
basis of prior art.
5
QUIET CHANGE IN C9X
For simplicity, the infinity and NaN representations are provided through
straightforward enhancements to C89 rather than through a new locale. Note also that
10 standard C locale categories do not affect the representations of infinities and NaNs.
A previous specification that strtod return a NaN for invalid numeric input as recommended
by IEEE 854 was withdrawn because of the incompatibility with C89, which demands that
strtod return zero for invalid numeric input.
15
7.20.1.4 The strtol, strtoll, strtoul, and strtoull functions
strtol was adopted for C89 as was strtod (§7.20.1.3); C9X adds strtoll and strtoull.
20 strtoul was introduced by the C89 Committee to provide a facility like strtol for
unsigned long values. Simply using strtol in such cases could result in overflow upon
conversion.
The C89 Committee decided that an implementation should be allowed to provide a rand function
which generates the best random sequence possible in that implementation, and therefore mandated
30 no standard algorithm. It recognized the value, however, of being able to generate the same
pseudo-random sequence in different implementations, and so it published as an example in the
Standard an algorithm that generates the same pseudo-random sequence in any conforming
implementation, given the same seed.
35 The rand and srand functions were based on existing practice; indeed the example
implementation was actually used in some versions of UNIX. Pseudo-random numbers have many
uses; and it should be noted that the example generator, while adequate for casual purposes, is
insufficiently random for demanding applications such as Monte-Carlo sampling and cryptography.
Also, only 32,768 distinct values are returned, which may be insufficiently fine resolution for some
40 purposes. Implementations may substitute improved algorithms and wider ranges of values; it is
incumbent on the programmer to ensure that the particular generator has appropriate statistical
properties for the intended application.
144
C9X RATIONALE WG14/N897 J11/99-032
The treatment of null pointers and zero-length allocation requests in the definition of these
5 functions was in part guided by a desire to support this paradigm:
/* initial allocation */
10 p = (OBJ *) calloc(0, sizeof(OBJ));
/* ... */
This coding style, not necessarily endorsed by the Committee, is reported to be in widespread use.
20
Some implementations have returned non-null values for allocation requests of zero bytes.
Although this strategy has the theoretical advantage of distinguishing between “nothing” and “zero”
(an unallocated pointer vs. a pointer to zero-length space), it has the more compelling theoretical
disadvantage of requiring the concept of a zero-length object. Since such objects cannot be
25 declared, the only way they could come into existence would be through such allocation requests.
The C89 Committee decided not to accept the idea of zero-length objects. The allocation functions
may therefore return a null pointer for an allocation request of zero bytes. Note that this treatment
does not preclude the paradigm outlined above.
30
QUIET CHANGE IN C89
Both nelem and elsize must be of type size_t for reasons similar to those for fread (see
§7.19.8.1).
45
If a scalar with all bits zero is not interpreted as a zero value by an implementation, then calloc
145
WG14/N897 J11/99-032 C9X RATIONALE
5 The Standard makes clear that a program may only free that which has been allocated, that an
allocation may only be freed once, and that a region may not be accessed once it is freed. Some
implementations allow more dangerous license. The null pointer is specified as a valid argument to
this function to reduce the need for special-case coding.
A null first argument is permissible. If the first argument is not null, and the second argument is 0,
then the call frees the memory pointed to by the first argument, and a null argument may be
returned; this specification is consistent with the policy of not allowing zero-sized objects.
15
A new feature of C9X: the realloc function was changed to make it clear that the pointed-to
object is deallocated, a new object is allocated, and the content of the new object is the same as that
of the old object up to the lesser of the two sizes. C89 attempted to specify that the new object was
the same object as the old object but might have a different address. This conflicts with other parts
20 of the Standard that assume that the address of an object is constant during its lifetime. Also,
implementations that support an actual allocation when the size is zero do not necessarily return a
null pointer for this case. C89 appeared to require a null return value, and the Committee felt that
this was too restrictive.
The C89 Committee vacillated over whether a call to abort should return if the SIGABRT signal
30 is caught or ignored. To minimize astonishment, the final decision was that abort never returns.
atexit provides a program with a convenient way to clean up the environment before it exits. It
35 was adapted from the Whitesmiths C run-time library function onexit.
A suggested alternative was to use the SIGTERM facility of the signal/raise machinery, but
that would not give the last-in-first-out stacking of multiple functions so useful with atexit.
40 It is the responsibility of the library to maintain the chain of registered functions so that they are
invoked in the correct sequence upon program exit.
45 The argument to exit is a status indication returned to the invoking environment. In the UNIX
operating system, a value of zero is the successful return code from a program. As usage of C has
146
C9X RATIONALE WG14/N897 J11/99-032
spread beyond UNIX, exit(0) has often been retained as an idiom indicating successful
termination, even on operating systems with different systems of return codes. This usage is thus
recognized as standard. There has never been a portable way of indicating a non-successful
termination, since the arguments to exit are implementation-defined. The EXIT_FAILURE macro
5 was added to C89 to provide such a capability. EXIT_SUCCESS was added as well.
Aside from calls explicitly coded by a programmer, exit is invoked on return from main. Thus
in at least this case, the body of exit cannot assume the existence of any objects with automatic
storage duration except those declared in exit.
10
The Committee considered the addition of _exit, but rejected it based on concerns of
incompatible with the POSIX specification upon which it is based. For example, one concern
expressed is that _exit was specified as a way to get out of a signal handler without triggering
another signal, but that is not actually the way _exit behaves in POSIX environments. The
15 Committee did not wish to give programmers this kind of false hope. (But see §7.20.4.4 for C9X.)
A new feature of C9X: the C9X Committee considered it desirable to have an _exit-like function
20 that would result in immediate program termination without triggering signals or atexit-
registered functions, but chose the name, _Exit, rather than _exit, because of the potential
conflict with existing practice mentioned above.
A corresponding putenv function was omitted from the Standard, since its utility outside a multi-
process environment is questionable, and since its definition is properly the domain of an operating
system standard.
35
7.20.4.6 The system function
The system function allows a program to suspend its execution temporarily in order to run
another program to completion.
40
Information may be passed to the called program in three ways: through command-line argument
strings, through the environment, and (most portably) through data files. Before calling the
system function, the calling program should close all such data files.
45 Information may be returned from the called program in two ways: through the implementation-
147
WG14/N897 J11/99-032 C9X RATIONALE
defined return value (In many implementations, the termination status code which is the argument
to the exit function is returned by the implementation to the caller as the value returned by the
system function.), and (most portably) through data files.
5 If the environment is interactive, information may also be exchanged with users of interactive
devices. Some implementations offer built-in programs called “commands” (for example, “date”)
which may provide useful information to an application program via the system function. The
Standard does not attempt to characterize such commands, and their use is not portable.
10 On the other hand, the use of the system function is portable, provided the implementation
supports the capability. The Standard permits the application to ascertain this by calling the
system function with a null pointer argument. Whether more levels of nesting are supported can
also be ascertained this way; but assuming more than one such level is obviously dangerous.
abs was moved from <math.h> as it was the only function in that library which did not involve
double arithmetic. Some programs have included <math.h> solely to gain access to abs, but
25 in some implementations this results in unused floating-point run-time routines becoming part of
the translated program.
The C89 Committee rejected proposals to add an absolute value operator to the language. An
implementation can provide a built-in function for efficiency.
30
7.20.6.2 The div, ldiv, and lldiv functions
Because C89 had implementation-defined semantics for division of signed integers when negative
operands were involved, div and ldiv, and lldiv in C9X, were invented to provide well-
35 specified semantics for signed integer division and remainder operations. The semantics were
adopted to be the same as in Fortran. Since these functions return both the quotient and the
remainder, they also serve as a convenient way of efficiently modeling underlying hardware that
computes both results as part of the same operation. Table 7.2 summarizes the semantics of these
functions.
40
Table 7.2: Results of div, ldiv and lldiv
numer denom quot rem
7 3 2 1
148
C9X RATIONALE WG14/N897 J11/99-032
-7 3 -2 -1
7 -3 -2 1
-7 -3 2 -1
Division by zero is described as undefined behavior rather than as setting errno to EDOM. The
program can just as easily check for a zero divisor before a division as for an error code afterwards,
5 and the adopted scheme reduces the burden on the function.
Now that C9X requires similar semantics for the division operator, the main reason for new
programs to use div, ldiv or lldiv is to simultaneously obtain quotient and remainder.
See §5.2.1.2, §MSE.8 and §MSE.9.1 for an overall discussion of multibyte character
representations and wide characters.
15 Implementation note: although these functions are described as having a conversion state for
consistency with the restartable equivalents in §7.24.6.3 and §7.24.6.4, the only piece of the
conversion state that is required for these functions is the current shift state, if any.
25 The C89 Committee felt that the functions in this subclause were all excellent candidates for
replacement by high-performance built-in operations. Hence many simple functions have been
retained, and several added, just to leave the door open for better implementations of these common
operations.
30 The Standard reserves function names beginning with str or mem for possible future use.
memcpy, memset, memcmp, and memchr were adopted in C89 from several existing
35 implementations. The general goal was to provide equivalent capabilities for three types of byte
sequences:
149
WG14/N897 J11/99-032 C9X RATIONALE
These arguments are contradictory but each is compelling. Therefore the Standard mandates two
block copy functions: memmove is required to work correctly even if the source and destination
overlap, while memcpy can assume non-overlapping operands and be optimized accordingly.
20
7.21.2.4 The strncpy function
strncpy was initially introduced into the C library to deal with fixed-length name fields in
structures such as directory entries. Such fields are not used in the same way as strings: the trailing
25 null is unnecessary for a maximum-length field, and setting trailing bytes for shorter names to null
assures efficient field-wise comparisons. strncpy is not by origin a “bounded strcpy,” and the
Committee preferred to recognize existing practice rather than alter the function to better suit it to
such use.
Note that this function may add n+1 characters to the string.
35
7.21.4 Comparison functions
40 See §7.21.1.
strcoll and strxfrm provide for locale-specific string sorting. strcoll is intended for
45 applications in which the number of comparisons is small; strxfrm is more appropriate when
150
C9X RATIONALE WG14/N897 J11/99-032
items are to be compared a number of times and the cost of transformation is paid only once.
5 See §7.21.4.3.
15 The strstr function is an invention of the C89 Committee. It is included as a hook for efficient
substring algorithms, or for built-in substring instructions.
20 This function was included in C89 to provide a convenient solution to many simple problems of
lexical analysis, such as scanning command line arguments.
The strsep function was proposed as an enhanced replacement for the strtok function. While
this is a common extension, it is easy enough for a user to provide this functionality, and it is
25 unclear that an implementor can do a substantially better job; so, there was not sufficient support
for adding this feature.
151
WG14/N897 J11/99-032 C9X RATIONALE
Type-generic macros allow calling a function whose type is determined by the argument type, as
is the case for C operators such as + and *. For example, with a type-generic cos macro, the
expression cos((float)x) will have type float. This feature enables writing more
5 portably efficient code and alleviates need for awkward casting and suffixing in the process of
porting or adjusting precision. Generic math functions are a widely appreciated feature of
Fortran.
The only arguments that affect the type resolution are the arguments corresponding to the
10 parameters that have type double in the synopsis. Hence the type of a type-generic call to
nexttoward, whose second parameter is long double in the synopsis, is determined solely
by the type of the first argument.
The term type-generic was chosen over the proposed alternatives of intrinsic and overloading.
15 The term is more specific than intrinsic, which already is widely used with a more general
meaning, and reflects a closer match to Fortran’s generic functions than to C++ overloading.
The macros are placed in their own header in order not to silently break old programs that
include <math.h>, for example with printf("%e", sin(x)).
20
modf(double, double *) is excluded because no way was seen to make it safe without
complicating the type resolution.
This specification differs from an earlier proposal in that the type is determined solely by the
25 argument, and may be narrower than the type for expression evaluation. This change was made
because the performance costs for computing functions with narrow arguments to wide range and
precision might be too high, even if the implementation efficiently evaluates basic operations to
wider format.
30 Also, this differs from earlier proposals in that integer-type arguments are converted to double
instead of float. Although converting to float would have been more consistent with the
usual arithmetic conversions, converting to double has the advantages of preserving the value
more often on many systems, and of being more compatible with C89 where unsuffixed calls to
math functions with integer arguments were calls to double functions.
35
Having a g suffix for the generic macros was considered but thought unnecessary.
The implementation might, as an extension, endow appropriate ones of the macros that this
standard specifies only for real arguments with the ability to invoke the complex functions.
40
This specification does not prescribe any particular implementation mechanism for generic
macros. It could be implemented simply with built-in macros. The generic macro for sqrt, for
example, could be implemented with
45 #undef sqrt
#define sqrt(x) _ _BUILTIN_GENERIC_sqrt(x)
152
C9X RATIONALE WG14/N897 J11/99-032
Generic macros are designed for a useful level of consistency with C++ overloaded math
functions.
5 The great majority of existing C programs are expected to be unaffected when <tgmath.h> is
included instead of <math.h> or <complex.h>. Generic macros are similar to the C89
library masking macros, though the semantic types of return values differ.
The ability to overload on integer as well as floating types would have been useful for some
10 functions, for example copysign. Overloading with different numbers of arguments would
have allowed reusing names, for example remainder for remquo. However, these facilities
would have complicated the specification; and their natural consistent use, such as for a floating
abs or a two-argument atan, would have introduced further inconsistencies with C89 for
insufficient benefit.
15
This specification in no way limits the implementation’s options for efficiency, including inlining
library functions.
The types clock_t and time_t are arithmetic because values of these types must, in accordance
25 with existing practice, on occasion be compared with –1 (a “don’t-know” indication), suitably cast.
No arithmetic properties of these types are defined by the Standard, however, in order to allow
implementations the maximum flexibility in choosing ranges, precisions, and representations most
appropriate to their intended application. The representation need not be a count of some basic
unit; an implementation might conceivably represent different components of a temporal value as
30 subfields of an integer type.
Many C environments do not support /usr/group library concepts of daylight saving time (DST, also
called summer time) or time zones. Both notions are defined geographically and politically, and
thus may require more knowledge about the real world than an implementation can support. Hence
35 the Standard specifies the date and time functions such that information about DST and time zones
is not required. /usr/group’s tzset function, which would require dealing with time zones, was
excluded altogether. An implementation reports that information about DST is not available by
setting the tm_isdst field in a broken-down time to a negative value. An implementation may
return a null pointer from a call to gmtime if information about the offset between Coordinated
40 Universal Time (UTC, née GMT) and local time is not available.
153
WG14/N897 J11/99-032 C9X RATIONALE
This function is intended for measuring intervals of execution time in whatever units an
implementation desires. The conflicting goals of high resolution, long interval capacity, and low
timer overhead must be balanced carefully in the light of this intended use.
mktime was invented by the C89 Committee to complete the set of time functions. With this
function it becomes possible to perform portable calculations involving clock times and broken-
15 down times.
The rules on the ranges of the fields within the *timeptr record are crafted to permit useful
arithmetic to be done. For instance, here is a paradigm for continuing some loop for an hour:
20 #include <time.h>
struct tm when;
time_t now;
time_t deadline;
25 /* ... */
now = time(0);
when = *localtime(&now);
when.tm_hour += 1; // result is in the range [1,24]
deadline = mktime(&when);
30
printf("Loop will finish: %s\n", asctime(&when));
while (difftime(deadline, time(0)) > 0) whatever();
The specification of mktime guarantees that the addition to the tm_hour field produces the
35 correct result even when the new value of tm_hour is 24, that is, a value outside the range ever
returned by a library function in a struct tm object.
One of the reasons for adding this function is to replace the capability to do such arithmetic which
is lost when a programmer cannot depend on time_t being an integral multiple of some known
40 time unit.
Several readers of earlier versions of this Rationale have pointed out apparent problems in this
example if now is just before a transition into or out of daylight saving time. However,
when.tm_isdst indicates what sort of time was the basis of the calculation. Implementors, take
45 heed. If this field is set to –1 on input, one truly ambiguous case involves the transition out of
daylight saving time. As DST is currently legislated in the United States, the hour from 0100 to
154
C9X RATIONALE WG14/N897 J11/99-032
0159 occurs twice, first as DST and then as standard time. Hence an unlabeled 0130 on this date is
problematic. An implementation may choose to take this as DST or standard time, marking its
decision in the tm_isdst field. It may also legitimately take this as invalid input and return
(time_t)(-1).
5
7.23.2.4 The time function
Since no measure is given for how precise an implementation’s best approximation to the current
time must be, an implementation could always return the same date instead of a more honest –1.
10 This is, of course, not the intent.
Despite objections that GMT, that is, Coordinated Universal Time (UTC), is not available in some
implementations, this function was retained because UTC is a useful and widespread standard
representation of time. If UTC is not available, a null pointer may be returned.
30
7.23.3.5 The strftime function
strftime provides a way of formatting the date and time in the appropriate locale-specific
fashion using the %c, %x, and %X format specifiers. More generally, it allows the programmer to
35 tailor whatever date and time format is appropriate for a given application. The facility is based on
the UNIX system date command. See §7.5 for further discussion of locale specification. For the
field controlled by %P, an implementation may wish to provide special symbols to mark noon and
midnight.
40 A new feature of C9X: C9X extends the strftime specifiers, introducing %C, %D, %e, %F, %g,
%G, %h, %n, %r, %R, %t, %T, %u and %V, as well as the E and O modifiers. These specifiers were
chosen according to existing practice to cover long-standing POSIX practice and to allow all the
formatting available in ISO 8601.
45
155
WG14/N897 J11/99-032 C9X RATIONALE
This subclause includes specific mention of the future direction in which the Committee intends to
extend and/or restrict the library. The contents of this subclause should be considered as quite
5 likely to become a part of the next version of the Standard. Implementors are advised that failure to
take heed of the points mentioned herein is considered undesirable for a conforming
implementation. Users are advised that failure to take heed of the points mentioned herein is
considered undesirable for a conforming program.
10 8. Annexes
Most of the material in the annexes is not new. It is simply a summary of information in the
Standard, collated for the convenience of users of the Standard.
15 New (advisory) information is found in Annex J (Common warnings) and in Annex K.5 (Common
extensions). The subclause on common extensions is provided in part to give programmers even
further information which may be useful in avoiding features of local dialects of C.
Vagaries of floating-point arithmetic have plagued programmers and users since its inception;
and they still do, even though hardware floating-point is now largely standardized. When IEEE
25 binary floating-point standard 754 became an official standard in July 1985, 26 months before
the radix-independent standard 854, several IEEE implementations were already shipping. In
1993, IEEE 754 was published as international standard IEC 559, now IEC 60559. Now
virtually all new floating-point implementations conform to IEC 60559, at least in format if not
to the last detail. Although these standards have been enormously successful in influencing
30 hardware implementation, many of their features, including predictability, remain impractical or
unavailable for use by programmers. IEC 60559 does not include language bindings, a cost of
delivering the basic standard in a timely fashion. The C89 Committee attempted to remove
conflicts with IEEE arithmetic, but did not specify IEEE support. Expediencies of programming
language implementation and optimization can deny the features offered by modern hardware. In
35 the meantime, particular companies have defined their own IEEE language extensions and
libraries; and not surprisingly, lack of portability has impeded programming for these interfaces.
The Numerical C Extensions Group, NCEG, at its initial meeting in May 1989, identified support
for IEEE floating-point arithmetic as one of its focus areas and organized a subgroup to produce
40 a technical report. The subgroup benefited from the considerable C language and IEEE floating-
point expertise associated with NCEG. It included individuals with substantial experience with
language extensions (albeit proprietary) for IEEE floating-point. And, following after the
standardization of C, it had a stable, well defined base for its extensions. Thus NCEG had a
unique opportunity to solve this problem. The floating-point part of NCEG’s technical report
156
C9X RATIONALE WG14/N897 J11/99-032
published in 1995 was the basis for the C9X floating-point specification.
F.2 Types
5 Minimal conformance to the IEC 60559 floating-point standards does not require a format wider
than single. The narrowest C double type allowed by standard C is wider than IEC 60559
single, and wider than the minimum IEC 60559 single-extended format. (IEC 60559 single-
extended is an optional format intended only for those implementations that don’t support
double; it has at least 32 bits of precision.) Both standard C and the IEC 60559 standards would
10 be satisfied if float were IEC 60559 single and double were an IEC 60559 single-extended
format with at least 35 bits of precision. However, this specification goes slightly further by
requiring double to be IEC 60559 double rather than just a wide IEC 60559 single-extended.
The primary objective of the IEC 60559 part of this specification is to facilitate writing portable
15 code that exploits the floating-point standard, including its standardized single and double data
formats. Bringing the C data types and the IEC 60559 standard formats into line advances this
objective.
This specification accommodates what are expected to be the most important IEC 60559
20 floating-point architectures for general C implementations.
Because of standard C’s bias toward double, extended-based architectures might appear to be
better served by associating the C double type with IEC 60559 extended. However, such an
approach would not allow standard C types for both IEC 60559 double and single and would go
25 against current industry naming, in addition to undermining this specification’s portability goal.
Other features in the Standard, for example the type definitions float_t and double_t
(defined in <math.h>), are intended to allow effective use of architectures with more efficient,
wider formats.
30 The long double type is not required to be IEC 60559 extended because
1. some of the major IEC 60559 floating-point architectures for C implementations do not
support extended.
3. extended is less standard than single or double in that only bounds for its range and precision
are specified in IEC 60559.
40 For implementations without extended in hardware, non-IEC 60559 extended arithmetic written
in software, exploiting double in hardware, provides some of the advantages of IEC 60559
extended but with significantly better performance than true IEC 60559 extended in software.
Specification for a variable-length extended type, one whose width could be changed by the user,
45 was deemed premature. However, not unduly encumbering experimentation and future
extensions, for example for variable length extended, is a goal of this specification.
157
WG14/N897 J11/99-032 C9X RATIONALE
Narrow-format implementations
Some C implementations, namely ones for digital signal processing, provide only the IEC 60559
5 single format, possibly augmented by single-extended, which may be narrower than IEC 60559
double or standard C double, and possibly further augmented by double in software. These
non-conforming implementations might generally adopt this specification, though not matching
its requirements for types.
10 One approach would be to match standard C float with single, match standard C double
with single-extended or single; and match standard C long double with double, single-
extended, or single. Then most of this specification could be applied straightforwardly. Users
should be clearly warned that the types may not meet expectations.
15 Another approach would be to refer to a single-extended format as long float and then not
recognize any C types not truly supported. This would provide ample warning for programs
requiring double. The translation part of porting programs could be accomplished easily with the
help of type definitions. In the absence of a double type, most of this specification for double
could be adopted for the long float type. Having distinct types for long float and
20 double, previously synonyms, requires more imagination.
The IEC 60559 floating-point standard requires perfect rounding for a large though incomplete
25 subset of decimal conversions. This specification goes beyond the IEC 60559 floating-point
standard by requiring perfect rounding for all decimal conversions involving DECIMAL_DIG or
fewer decimal digits and a supported IEC 60559 format, because practical methods are now
available. Although not requiring correct rounding for arbitrarily wide decimal numbers, this
specification is sufficient in the sense that it ensures that every internal numeric value in an IEC
30 60559 format can be determined as a decimal constant.
F.7 Environment
C89 did not specify when aggregate and union initialization is done. Otherwise, this section is
merely a clarification. Note that, under the effect of an enabling FENV_ACCESS pragma, any
45 exception resulting from execution-time initialization must be raised at execution time.
158
C9X RATIONALE WG14/N897 J11/99-032
The specification for constant expressions and initialization does not suit C++, whose static and
aggregate initializers need not be constant. Specifying all floating-point constant arithmetic and
initialization to be as if at execution time would be suitable for C++, and given the
5 FENV_ACCESS mechanism, still would allow the bulk of constant arithmetic to be done, in
actuality, at translation time.
whose use may raise the divide-by-zero exception. Similarly, INFINITY and NAN cannot be
15 implemented as ((float)(1.0/0.0)) and ((float)(0.0/0.0).
Special cases
2. assure predictable special-case behavior the programmer can exploit for simpler, more
efficient code.
25
3. allow implementations enough flexibility to provide needed performance.
Compatibility with IEC 60559 is a foremost strategy. The C9X annexes adopt the IEC 60559
specification for the functions covered by that standard, such as sqrt and rint, and follows
30 the spirit of IEC 60559 for other functions. This means the special values (infinities, NaNs, and
signed zeros) and the floating-point exceptions have a consistent meaning throughout the basic
arithmetic and the libraries. At a higher level, C9X shares the IEC 60559 goal to enhance
robustness through predictable behavior. For special cases, this behavior is chosen to be useful
for most applications wherever possible, recognizing that it is in the nature of exceptional cases
35 that one behavior is not best in all situations.
Typically, the tradeoff is between a numeric result that is useful in only some applications and a
more pessimistic NaN result. As in IEC 60559, choosing utility over conservatism exacts a cost
in specification complexity. For example, regarding NaNs only as error indicators and rules like
40 “NaN in, NaN out” are simple but not always most useful. A NaN argument is often better
interpreted as an indeterminate value. This supports the programming practice of initializing
with NaNs those variables whose true values are yet to be determined, and permits returning the
obvious numeric value for functions that are independent of one of their arguments. Thus
hypot(∞, NAN) is infinity, as this would be the result regardless of the numeric value of the
45 second argument. The C9X fmax and fmin functions return the maximum or minimum of their
numerical arguments, hence fmax(NAN, 1.2) is 1.2, which is the desired behavior for
159
WG14/N897 J11/99-032 C9X RATIONALE
Although a definition of fmax implying a NaN result for fmax(NAN, 1.2) might be equally
useful, choosing one of the viable specifications instead of leaving the choice to the
5 implementation has the inherent value of suiting some portable code instead of none. In other
special cases, however, choices are left to the implementation because of existing practice (for
example, the return value of ilogb) or performance issues (for example, whether certain
rounding functions raise the inexact exception).
10 Generally, C9X eschews a NaN result where a numerical value is useful. IEC 60559 follows the
same approach, as in defining overflow results to be infinite (given default rounding), which is
neither mathematically correct nor useful in all cases. The results of pow(∞,0) and
pow(0,0) are both 1, because there are applications that can exploit this definition. For
example, if x(p) and y(p) are any analytic functions that become zero at p = a, then pow(x,y),
15 which equals exp(y*log(x)), approaches 1 as p approaches a. The result of pow(–2,∞) is
+∞, because all large positive floating-point values are even integers. The result of
atan2(+0,+0), which is equivalent to carg(+0+i0), is defined to be +0. A significant
benefit is supporting a clog that is equivalent to log on the nonnegative real axis.
20 The choice for special-case behavior, which typically is arbitrary to some degree, was made in
favor of preserving identities (involving numeric, not NaN, values), specification consistency
among functions, and efficiency in implementation. The functions hypot(x,y) and
cabs(x+yi) are equivalent, as are atan2(y,x) and carg(x+yi), and these behave so as to be
useful building blocks for other complex functions.
25
In certain respects, C9X is less demanding than might be expected, in order to give some
flexibility in implementation, especially where the loss of utility is believed to be negligible or
the cost is not justifiable. For example, C9X leaves to the implementation to decide whether
functions (like sin) that are essentially always inexact raise the inexact flag, as there doesn’t
30 seem to be significant utility in testing an expression involving such a function for exactness.
Functions in <math.h> are allowed to raise undeserved inexact and underflow exceptions,
because determination may be difficult. C9X allows complex multiply and divide to raise
spurious exceptions because of the performance cost of avoiding them.
35 The cost for exception behavior is intended to be modest enough for most purposes. And, as the
exceptions are accessible only in code under the effect of an enabling FENV_ACCESS pragma,
an implementation could invoke, perhaps even by default, routines that didn’t have the specified
exception behavior. (The pragma does not exempt the implementation from having to return
specified result values.)
40
Underflow
The IEC 60559 floating-point standard offers the implementation multiple definitions of
underflow. All resulting in the same values, the options differ only in that the thresholds when
45 the exception is raised may differ by a rounding error. It is not intended that library functions
necessarily use the same definition of underflow as the arithmetic, because the difference so
160
C9X RATIONALE WG14/N897 J11/99-032
rarely matters.
Exactness
5 For some functions, pow for example, determining exactness in all cases may be too costly.
Functions have certain restrictions against raising spurious exceptions detectable by the user. For
example, the implementation must hide an underflow generated by an intermediate computation
of a non-tiny result.
10
F.9.1 Trigonometric functions
15 The more contentious cases are y and x both infinite or both zero. These deliver numeric results
instead of NaNs in order to preserve more identities and for better utility. The specification of
atan2(0,0) to be 0 facilitates the definition of carg(x+yi) as atan2(x,y) and clog(z)
as log(|x|) + I*carg(z) so that clog(z) agrees with log(x) on the real axis.
20 The specification of atan2(∞,∞) as π/4 indicates the reasonable quadrant, preserving some
information in preference to none.
pow(x,0) is specified to return 1 for any x, because there are significant applications where 1 is
more useful than NaN. pow(f(t),g(t)) approaches 1 in all cases where f and g are analytic
functions and g(t) approaches zero. The result 1 better supports applications where the second
35 argument is integral. pow(NAN,0) returns 1.0 on the general principle that if a result is
independent of the numerical value of an argument, then that result is appropriate if that
argument is a NaN.
Some applications may be better served by a max function that would return a NaN if one of its
arguments were a NaN:
161
WG14/N897 J11/99-032 C9X RATIONALE
Note that two branches still are required for symmetry in NaN cases.
5
10 Although the specification in Annex G is fundamental for IEC 60559 style complex arithmetic,
the annex is designated informative because of insufficient prior art for normative status.
G.2 Types
15 Although not present in older complex arithmetic facilities such as Fortran’s, the imaginary types
naturally model the imaginary axis of complex analysis, promote computational and storage
efficiency, and capture the completeness and consistency of IEC 60559 arithmetic for the
complex domain. See also rationale for §G.5.
20 The representation and alignment requirements of imaginary types are intended to allow
imaginary arguments to fprintf and fscanf; however, technically speaking, this invokes
undefined behavior because corresponding real and imaginary types are not compatible types.
The recommended practice is that implementations promote arguments with float
imaginary types to double imaginary, and treat the arguments as if they had the
25 corresponding real type. When no prototype is in scope, function calls involving arguments with
imaginary types should behave in a similar manner.
Text book formulas for complex arithmetic tend to turn infinite inputs into NaNs, often losing
useful information unnecessarily. For example,
and
162
C9X RATIONALE WG14/N897 J11/99-032
basic properties of infinities, and provides the cproj function to map all infinities to a canonical
one. For example, a finite non-zero value times an infinity must be an infinity, hence
(1+i0)*(∞+i∞) must be an infinity. In the same spirit, cexp(∞+iNaN) is an infinity and
cexp(-∞+iNaN) is a complex zero, which preserve cabs(cexp(x+iy)) = exp(x).
5
C9X treats multiple infinities so as to preserve directional information where possible, despite
the inherent limitations of the ordered-pair (Cartesian) representation. The product of the
imaginary unit and a real infinity is a correctly signed imaginary infinity, i × ∞ = i∞. And
10 i × (∞ – i∞) = ∞ + i∞
C9X allows complex multiply and divide to raise spurious exceptions because of the
15 performance cost of avoiding them.
IEC 60559 compatibility is a primary rationale for the imaginary types. Without them the
traditional complex arithmetic programming facilities prove fundamentally incompatible with
IEC 60559 in the treatment of special values; with them compatibility comes surprisingly
30 naturally. Very little special-case specification is required for imaginary types.
The imaginary types, together with the usual arithmetic conversion rules and operator
specifications (see §G.4), allow substantially more efficient code. For example, multiplication of
an imaginary by a complex can be implemented straightforwardly with two multiplications,
35 instead of four multiplications and two additions.
In the absence of imaginary types, macros would be required in order to create certain special
values. For example, 0+∞i could be created by CMPLX(0.0,INFINITY). With the
imaginary types, imaginary infinity is simply the value of I*INFINITY. (If imaginary types are
40 not supported and I is _Complex_I, then INFINITY*I would result in a real part of NaN
and an invalid exception.) With imaginary types, values of I*y and x + I*y, where x and y are
real floating values, cover all values of the imaginary and complex types, hence eliminating this
need for the complex macros.
45 Some programs are expected to use the imaginary types implicitly in constructions with the
163
WG14/N897 J11/99-032 C9X RATIONALE
imaginary unit I, such as x + I*y, and not explicitly in declarations. This suggests making the
imaginary types private to the implementation and not available for explicit program
declarations. However, such an approach was rejected as being less in the open spirit of C, and
not much simpler. For the same reasons, the approach of treating imaginariness as an attribute of
5 certain complex expressions, rather than as additional types, was rejected.
Another proposal was to regard the special values (infinities, NaNs, and signed zeros) as outside
the model. This would allow any behavior when special values occur, including much that is
prescribed by this specification. However, this approach would not serve the growing majority
10 of implementations, including all IEC 60559 ones, that support the special values. These
implementations would require additional specification in order to provide a consistent extension
of their treatment of special cases in the real domain. On the other hand, implementations not
supporting special values should have little additional trouble implementing imaginary types as
proposed here.
15
The efficiency benefits of the imaginary types goes beyond what the implementation provides. In
many cases programmers have foregone a programming language’s complex arithmetic facilities,
which, lacking an imaginary type, required contiguous storage of both real and imaginary parts;
programmers could store and manipulate complex values more efficiently using real arithmetic
20 directly. The imaginary types enable programmers to exploit the efficiency of the real formats
without having to give up support for complex arithmetic semantics.
Care is taken throughout so that the sign of zero is available for distinguishing the sides of a
branch cut along the axes, even at infinities. Therefore
25 _____
√–∞+i0 = 0+i∞
and by conjugation
_____
30 √–∞–i0 = 0–i∞
Exploiting the fact that some functions map the imaginary axis onto the real or imaginary axis
35 gains more efficient calculation involving imaginaries, and better meets user expectations in
some cases. However, dropping out of the complex domain may lead to surprises as subsequent
operations may be done with real functions, which generally are more restrictive than their
complex counterparts. For example, sqrt(-cos(I)) invokes the real sqrt function, which
is invalid for the negative real value -cos(I), whereas the complex sqrt is valid everywhere.
40
45 LIA-1 was not made a normative part of C9X for three reasons: implementation vendors saw no
164
C9X RATIONALE WG14/N897 J11/99-032
need to add LIA-1 support because customers are not asking for it, LIA-1 may change now that
work on LIA-2 is finishing and work on LIA-3 is starting, and the Committee did not wish to rush a
possibly incomplete specification into C9X at the last moment. A proposed binding between C and
LIA-1 was produced a few months before C9X was frozen. That binding was a compromise
5 between those who believe LIA-1 should be forgotten and those who wanted full LIA-1 and more
(for example, C signal handlers that could patch up exceptions on the fly and restart the exceptional
floating-point instruction). It took several years for the NCEG floating-point specification to settle
down, so it was assumed that it would take a similar timeframe to get the LIA-1 binding stable.
The Committee did not wish to delay C9X for this one item. An informative LIA-1 annex was
10 added, however, because all programming languages covered by ISO/IEC JTC1 SC22 standards are
expected to review LIA-1 and incorporate and further define the binding between LIA-1 and each
programming language.
C9X’s binding between C and LIA-1 differs from LIA-1’s C binding in three cases in which the
15 Committee believes that LIA-1 is incorrect. First, LIA-2 and LIA-1 have different behaviors for
mathematical pole exceptions (similar to finite non-zero divided by zero and log(0)). The
Committee believes that LIA-2 is better and that LIA-1 will be changed to match LIA-2. Second,
the existing hardware that many computers use for conversion from floating-point type to integer
type raises the undefined exception, instead of the required integer overflow, for values that are out
20 of bounds. Third, requiring support for signaling NaNs on IEC 60559 implementations should be
optional because existing hardware support for signaling NaNs is inconsistent.
H.3.1.2 Traps
25 The math library is required by both C89 and C9X to be atomic in that no exceptions (raise of a
signal and invocation of a user’s signal handler) may be visible in the user’s program. Because of
that requirement, C9X cannot meet LIA’s requirement of either trap-and-terminate or trap-and-
resume for math library errors. On the other hand, both kinds of traps are allowed for the arithmetic
operations.
30
165
WG14/N897 J11/99-032 C9X RATIONALE
Most traditional computer systems and computer languages, including traditional C, have an
10 assumption, sometimes undocumented, that a “character” can be handled as an atomic quantity
associated with a single memory storage unit — a “byte” or something similar. This is not true in
general. For example, a Japanese, Chinese or Korean character usually requires a representation of
two or three bytes; this is a multibyte character as defined by §3.72 and §5.2.1.2. Even in the Latin
world, a multibyte coded character set appears. This conflict is called the byte and character
15 problem.
A related concern in this area is how to address having at least two different meanings for string
length: number of bytes and number of characters.
20 To cope with these problems, many technical experts, particularly in Japan, have developed their
own sets of additional multibyte character functions, sometimes independently and sometimes
cooperatively. Fortunately, the developed extensions are actually quite similar. It can be said that
in the process they have found common features for multibyte character support. Moreover, the
industry currently has many good implementations of such support.
25
The above in no way denigrates the important groundwork in multibyte- and wide-character
programming provided by C90:
• Both the source and execution character sets can contain multibyte characters with possibly
30 different encodings, even in the “C” locale.
• Multibyte characters are permitted in comments, string literals, character constants, and
header names.
• The library has five basic functions that convert between multibyte and wide characters.
However, the five functions are often too restrictive and too primitive to develop portable
40 international programs that manage characters. Consider a simple program that wants to count the
number of characters, not bytes, in its input.
45 #include <stdio.h>
166
C9X RATIONALE WG14/N897 J11/99-032
int main(void)
{
int c, n = 0;
while ((c = getchar()) != EOF)
5 n++;
printf("Count = %d\n", n);
return 0;
}
10 does not work as expected if the input contains multibyte characters; it always counts the number of
bytes. It is certainly possible to rewrite this program using just some of the five basic conversion
functions, but the simplicity and elegance of the above are lost.
C90 deliberately chose not to invent a more complete multibyte- and wide-character library,
15 choosing instead to await their natural development as the C community acquired more experience
with wide characters. The task of the Committee was to study the various existing implementations
and, with care, develop the first amendment to C90. The set of developed library functions is
commonly called the MSE (Multibyte Support Extension).
20 Similarly, C90 deliberately chose not to address in detail the problem of writing C source code with
character sets such as the national variants of ISO/IEC 646. These variants often redefine several of
the punctuation characters used to write a number of C tokens. The partial solution adopted was to
add trigraphs to the language. Thus, for example, ??< can appear anywhere in a C program that {
can appear, even within a character constant or a string literal.
25
AMD1 responds to an international sentiment that more readable alternatives should also be
provided wherever possible. Thus, it adds to the language alternate spellings of several tokens. It
also adds a library header, <iso646.h>, that defines a number of macros that expand to still other
tokens which are less readable when spelled with trigraphs. Note, however, that trigraphs are still
30 the only alternative to writing certain characters within a character constant or a string literal.
• digraphs — The tokens %: and %:%: are just sequences of preprocessing tokens in C90 but
40 become single preprocessing tokens with specific meanings in AMD1. An existing
program that uses either of these tokens in a macro argument can behave differently as a
result of AMD1.
• new function names — Several names with external linkage, such as btowc, not reserved
45 to the implementation in C90, are now so reserved. An existing program that uses any of
these names can behave differently as a result of AMD1.
167
WG14/N897 J11/99-032 C9X RATIONALE
Using the MSE functions, a multibyte-character-handling program can be written as easily and in
the same style as a traditional single-byte-based program. A programming model based on MSE
5 function is as follows. First, a multibyte character or a multibyte string is read from an external file
into a wchar_t object or a wchar_t array object by the fgetwc function, or other input
functions based on the fgetwc function such as getwchar, getwc, or fgetws. During this
read operation, a code conversion occurs: the input function converts the multibyte character to the
corresponding wide character as if by a call to the mbrtowc function.
10
After all necessary multibyte characters are read and converted, the wchar_t objects are processed
in memory by the MSE functions such as iswxxx, wcstod, wcscpy, wmemcmp, and so on.
Finally, the resulting wchar_t objects are written to an external file as a sequence of multibyte
characters by the fputwc function or other output functions based on the fputwc function such
15 as putwchar, putwc, or fputws. During this write operation, a code conversion occurs: the
output function converts the wide character to the corresponding multibyte character as if by a call
to the wcrtomb function.
In the case of the formatted input/output functions, a similar programming style can be applied,
20 except that the character code conversion may also be done through extended conversion specifiers
such as %ls and %lc. For example, the wide-character-based program corresponding to that
shown in §MSE.1 can be written as follows:
#include <stdio.h>
25 #include <wchar.h>
int main(void)
{
wint_t wc;
30 int n = 0;
40 When defining the MSE library functions, the Committee could have chosen a design policy based
either on parallelism or on improvement. “Parallelism” means that a function interface defined in
AMD1 is similar to the corresponding single-byte function in C90. The number of parameters in
corresponding functions are exactly same, and the types of parameters and the types of return values
have a simple correspondence:
45
char ↔ wchar_t int ↔ wint_t
168
C9X RATIONALE WG14/N897 J11/99-032
On the other hand, “improvement” means that a function interface in AMD1 is changed from the
5 corresponding single-byte functions in C90 in order to resolve problems potentially contained in the
existing functions. Or, a corresponding function is not introduced in AMD1 when the functionality
can be better attained through other functions.
In an attempt to achieve improvement, there were numerous collisions of viewpoints on how to get
10 the most appropriate interface. Moreover, much careful consideration and discussion among
various experts in this area was necessary to decide which policy should be taken for each function.
AMD1 is the result of this process.
isalnum iswalnum
isalpha iswalpha
20 iscntrl iswcntrl
isdigit iswdigit
isgraph iswgraph
islower iswlower
isprint iswprint
25 ispunct iswpunct
isspace iswspace
isupper iswupper
isxdigit iswxdigit
tolower towlower
30 toupper towupper
fgetc fgetwc
fgets fgetws
fputc fputwc
35 fputs fputws
fprintf fwprintf
fscanf fwscanf
getc getwc
getchar getwchar
40 printf wprintf
putc putwc
putchar putwchar
scanf wscanf
sprintf swprintf
45 sscanf swscanf
ungetc ungetwc
vfprintf vwfprintf
169
WG14/N897 J11/99-032 C9X RATIONALE
vprintf vwprintf
vsprintf vswprintf
memchr wmemchr
5 memcmp wmemcmp
memcpy wmemcpy
memmove wmemmove
memset wmemset
strcat wcscat
10 strcmp wcscmp
strcoll wcscoll
strcpy wcscpy
strchr wcschr
strcspn wcscspn
15 strftime wcsftime
strlen wcslen
strncat wcsncat
strncmp wcsncmp
strncpy wcsncpy
20 strpbrk wcspbrk
strrchr wcsrchr
strspn wcsspn
strstr wcsstr
strtod wcstod
25 strtok wcstok
strtol wcstol
strtoul wcstoul
strxfrm wcsxfrm
isblank iswblank
snprintf (swprintf)
35 vfscanf vfwscanf
vscanf vwscanf
vsnprintf (vswprintf)
vsscanf vswscanf
40 strtof wcstof
strtold wcstold
strtoll wcstoll
strtoull wcstoull
45
Note that there may still be subtle differences (see for example §MSE.6.2).
The following functions have different interfaces between single-byte and wide-character versions:
170
C9X RATIONALE WG14/N897 J11/99-032
- Members of the sprintf family based on wide characters all have an extra size_t parameter
in order to avoid overflowing the buffer. Compare:
- wcstok, the wide-character version of strtok, has an extra wchar_t ** parameter in order
to eliminate the internal memory that the strtok function has to maintain. Compare:
The following is a list of the functions in C90, with atoll added in C9X, that do not have parallel
20 functions in AMD1 for any of several reasons such as redundancy, dangerous behavior, or a lack of
need in a wide-character-based program. Most of these can be rather directly replaced by other
functions:
atof
25 atoi
atol
atoll
gets
perror
30 puts
strerror
Finally, the following is a list of the functions in AMD1 that do not have parallel functions in C90.
They were introduced either to achieve better control over the conversion between multibyte
35 characters and wide characters, or to give character handling programs greater flexibility and
simplicity:
btowc
fwide
40 iswctype
mbrlen
mbrtowc
mbsinit
mbsrtowcs
45 towctrans
wcrtomb
wcsrtombs
171
WG14/N897 J11/99-032 C9X RATIONALE
wctob
wctrans
wctype
With its rich set of operators and punctuators, the C language makes heavy demands on the ASCII
character set. Even before the language was standardized, it presented problems to those who
would move C to EBCDIC machines. More than one vendor provided alternate spellings for some
10 of the tokens that used characters with no EBCDIC equivalent. With the spread of C throughout
the world, such representation problems have only grown worse.
ISO/IEC 646, the international standard corresponding to ASCII, permits national variants of a
number of the characters used by C. Strictly speaking, this is not a problem in representing C
15 programs, since the necessary characters exist in all such variants: they just print oddly. Displaying
C programs for human edification suffers, however, since the operators and punctuators can be hard
to recognize in their various altered forms.
C90 addresses the problem in a different way. It provides replacements at the level of individual
20 characters using three-character sequences called trigraphs (see §5.2.1.1). For example, ??< is
entirely equivalent to {, even within a character constant or string literal. While this approach
provides a solution for the known limitations of EBCDIC (except for the exclamation mark) and
ISO/IEC 646, the result is arguably not highly readable.
25 Thus, AMD1 provides a set of more readable digraphs (see §6.4.6). These are two-character
alternate spellings for several of the operators and punctuators that can be hard to read with
ISO/IEC 646 national variants. Trigraphs are still required within character constants and string
literals, but at least the more common operators and punctuators can have more suggestive spellings
using digraphs.
30
The added digraphs were intentionally kept to a minimum. Wherever possible, the Committee
instead provided alternate spellings for operators in the form of macros defined in the new header
<iso646.h>. Alternate spellings are provided for the preprocessing operators # and ## because
they cannot be replaced by macro names. Digraphs are also provided for the punctuators [, ], {,
35 and } because macro names proved to be a less readable alternative. The Committee recognizes
that the solution offered in this header is incomplete and involves a mixture of approaches, but
nevertheless believes that it can help make Standard C programs more readable.
MSE.5 Headers
40
MSE.5.1 <wchar.h>
45 Function prototypes for the MSE library functions had to be included in some header. The
172
C9X RATIONALE WG14/N897 J11/99-032
2. Declare all the MSE function prototypes in <stdlib.h> where wchar_t is already defined.
3. Introduce a new header and declare all the MSE function prototypes in the new header.
10
4. Declare the MSE function prototypes in the existing headers specified in C90 which are most
closely related to these functions.
The drawback to idea 1 is that the relationship between new headers and existing ones becomes
15 complicated. For example, there may be dependencies between the old and the new headers, so one
or more headers may have to be included prior to including <wstdio.h>, as in:
#include <stdlib.h>
#include <stdio.h>
20 #include <wstdio.h>
The drawback to idea 2 is that the program has to include many prototype declarations even if the
program does not need declarations in <stdlib.h> other than existing ones. At the time, the
Committee strongly opposed adding several identifiers to existing headers for this purpose.
25
The drawback to idea 3 is that it introduces an asymmetry between existing headers and the new
headers.
The drawback to idea 4, as with idea 2, is that the Committee strongly opposed adding many
30 identifiers to existing headers.
So the Committee decided to introduce a new header, <wchar.h>, as the least objectionable way
to declare all MSE function prototypes. Later, the Committee split off the functions analogous to
those in <ctype.h> and placed their declarations in the header, <wctype.h>, as described in
35 §MSE.5.2.
The Committee was concerned that the definitions of types and macros in <wchar.h> be
40 specified efficiently. One goal was to require that only the header <wchar.h> need be included
to use the MSE library functions; but there were strong objections to declaring existing types such
as FILE in the new header.
The definitions in <wchar.h> are thus limited to those types and macros that are largely
45 independent of the existing library. The existing header <stdio.h> must also be included along
with <wchar.h> when the program needs explicit definitions of either of the types FILE and
173
WG14/N897 J11/99-032 C9X RATIONALE
fpos_t.
MSE.5.2 <wctype.h>
5 The Committee originally intended to place all MSE functionality in a single header, <wchar.h>,
as explained in §MSE.5.1.1. It found, however, that this header was excessively large, even
compared to the existing large headers, <stdio.h> and <stdlib.h>. The Committee also
observed that the wide-character classification and mapping functions, typically have names of the
form iswxxx or towxxx, seemed to form a separate group. A translation unit could well make use
10 of most of the functionality of the MSE without using this separate group. Equally, a translation
unit might need the wide-character classification and mapping functions without needing the other
MSE functions.
The Committee therefore decided to form a separate header, <wctype.h>, closely analogous to
15 the existing <ctype.h>. That division also reduced the size of <wchar.h> to more
manageable proportions.
20 Eleven iswxxx functions were introduced to correspond to the character-testing functions defined
in C90. Each wide-character testing function is specified in parallel with the matching single-byte
character handling function, however the following changes were also introduced.
30 This existing “C” locale restriction for character testing functions in C90 has been replaced with a
superseding constraint for wide-character-testing functions. There is no special description of “C”
locale behavior for the iswxxx functions. Instead, the following rule is applied to any locale: when
a character c causes isxxx(c) to return true, the corresponding wide character wc shall cause the
corresponding iswxxx(wc) to return true.
35
isxxx(c) != 0 ⇒ iswxxx(wc) != 0
where c == wctob(wc). Note that the converse relationship does not necessarily hold.
The space character, ' ', is treated specially in isprint, isgraph, and ispunct. Handling
of the space character in the corresponding wide-character functions differs from that specified in
C90. The corresponding wide-character functions return true for all wide characters for which
174
C9X RATIONALE WG14/N897 J11/99-032
iswspace returns true, instead of just the single space character; therefore the behaviors of the
iswgraph and iswpunct functions may differ from their matching functions in C90 in this
regard (see the footnote concerning iswgraph in §7.25.2.1.6).
There are eleven standard character-testing functions defined in C90. As the number of supported
locales increases, the requirements for additional character classifications grows, and varies from
locale to locale. To satisfy this requirement, many existing implementations, especially for non-
10 English-speaking countries, have been defining new isxxx functions, such as iskanji,
ishanja, and so forth.
This approach, however, adds to the global namespace clutter (although the names have been
reserved) and is not flexible at all in supporting additional classification requirements. Therefore,
15 in AMD1, a pair of extensible wide character classification functions, wctype and iswctype,
are introduced to satisfy the open-ended requirements for character classification. Since the name
of a character classification is passed as an argument to the wctype function, it does not add to
problem of global namespace pollution; and these generic interfaces allow a program to test if the
classification is available in the current locale, and to test for locale-specific character
20 classifications, such as Kanji or Hiragana in Japanese.
In the same way, a pair of wide character mapping functions, wctrans and towctrans, are
introduced to support locale-specific character mappings. One of the example of applying this
functionality is the mappings between Hiragana and Katakana in a Japanese character set.
25
MSE.8 Generalized multibyte characters
C90 intentionally restricted the class of acceptable encodings for multibyte characters. One goal
was to ensure that, at least in the initial shift state, the characters in the basic C character set have
30 multibyte representations that are single characters with the same code as the single-byte
representation. The other was to ensure that the null byte should always be available as an end-of-
string indicator. Hence, it should never appear as the second or subsequent byte of any multibyte
code. For example, the one-byte sequence 'a' should always represent L'a', at least initially,
and '\0' should always represent L'\0'.
35
While these may be reasonable restrictions within a C program, they hamper the ability of the MSE
functions to read arbitrary wide-oriented files. For example, a system may wish to represent files as
sequences of ISO/IEC 10646 characters. Reading or writing such a file as a wide-oriented stream
should be an easy matter. At most, the library may have to map between native and some canonical
40 byte order in the file. In fact, it is easy to think of an ISO/IEC 10646 file as being some form of
multibyte file except that it violates both restrictions described above: the code for 'a' can look
like the four-byte sequence \0\0\0a for example.)
Thus, the MSE introduces the notion of a generalized multibyte encoding. It subsumes all the ways
45 the Committee can currently imagine that operating systems will represent files containing
175
WG14/N897 J11/99-032 C9X RATIONALE
characters from a large character set. (Such encodings are valid only in files; they are still not
permitted as internal multibyte encodings.)
It is necessary to convert between multibyte characters and wide characters within wide character
input/output functions. The conversion state introduced in §7.24.6 is used to help perform this
10 conversion. Every wide character input/output function makes use of, and updates, the conversion
state held in the FILE object controlling the wide-oriented stream.
The conversion state in the FILE object augments the file position within the corresponding
multibyte character stream with the parse state for the next multibyte character to be obtained from
15 that stream. For state-dependent encodings, the remembered shift state is a part of this parse state,
and therefore a part of the conversion state. (Note that a multibyte encoding that has any characters
requiring two or more bytes needs a nontrivial conversion state even if it is not a state-dependent
encoding.)
• the wide character input/output functions use this hidden object as the state argument to the
25 mbrtowc or wcrtomb functions that perform the conversion between multibyte
characters in the file and wide characters inside the program.
MSE.9.2 Implementation
30 The Committee assumed that only wide character input/output functions can maintain consistency
between the conversion state information and the stream. The byte input/output functions do
nothing with the conversion state information in the FILE object. The wide character input/output
functions are designed on the premise that they always begin executing with the stream positioned
at the boundary between two multibyte characters.
35
The Committee felt that it would be intolerable to require implementors to implement these
functions without such a guarantee. Since executing a byte input/output function on a
wide-oriented stream may well leave the file position indicator at other than the boundary between
two multibyte characters, the Committee decided to prohibit such use of the byte input/output
40 functions.
An fpos_t object for a stream in a state-dependent encoding includes the shift state information
45 for the corresponding stream. In order to ensure the behavior of subsequent wide character
176
C9X RATIONALE WG14/N897 J11/99-032
input/output functions in a state-dependent encoding environment, a seek operation should reset the
conversion state corresponding to the file position as well as restoring the file position.
The traditional seek functions, fseek and ftell, may not be adequate in such an environment
5 because even an object of type long int may be too small to hold both the conversion state
information and the file position indicator. Thus, the newer fsetpos and fgetpos are
preferred, since they can store as much information as necessary in an fpos_t object.
A multiple encoding environment has two or more different encoding schemes for files. In such an
environment, some programmers will want to handle two or more multibyte character encodings on
20 a single platform, possibly within a single program. There is, for example, an environment in Japan
that has two or more encoding rules for a single character set. Most implementations for Japanese
environments should provide for such multiple encodings.
During program execution, the wide character input/output functions get information about the
25 current encodings from the LC_CTYPE category of the current locale when the conversion state is
bound, as described immediately below. When writing a program for a multiple encoding
environment, the programmer should be aware of the proper LC_CTYPE category when opening a
file and establishing its orientation. During subsequent accesses to the file, the LC_CTYPE
category need not be restored by the program.
30
The encoding rule information is effectively a part of the conversion state. Thus, the information
about the encoding rule should be stored with the hidden mbstate_t object within the FILE
object. Some implementations may even choose to store the encoding rule as part of the value of an
fpos_t object.
35
The conversion state just created when a file is opened is said to have unbound state because it has
no relations to any of the encoding rules. Just after the first wide character input/output operation,
the conversion state is bound to the encoding rule which corresponds to the LC_CTYPE category of
the current locale. The following is a summary of the relations between various objects, the shift
40 state, and the encoding rules:
fpos_t FILE
177
WG14/N897 J11/99-032 C9X RATIONALE
5 Both the wide character input/output functions and the byte input/output functions refer the same
type of object, a FILE object. As described in §MSE.9.2, however, there is a constraint on mixed
usage of the two types of input/output functions. That is, if a wide character input/output function
is executed for a FILE object, its stream becomes wide-oriented and no byte input/output function
may be applied later, and conversely.
10
The reason for this constraint is to ensure consistency between the current file position and the
current conversion state in the FILE object. Executing one of the byte input/output functions for a
wide-oriented stream breaks this consistency because the byte input/output functions may, and
should, ignore the conversion state information in the FILE object.
15
The diagram A1 shows the state transitions of a stream in response to various input/output
functions.
178
C9X RATIONALE WG14/N897 J11/99-032
Diagram A1
179
WG14/N897 J11/99-032 C9X RATIONALE
In some implementations such as UNIX, there are streams which look the same whether read or
written as text or binary. For example, arbitrary file positioning operations are supported even in
5 text mode. In such an implementation, the Committee specified a file opened as a binary stream
should obey the usage constraints placed upon text streams when accessed as a wide-oriented
stream (for example, the restrictions on file positioning operations should be obeyed).
So an implementation of the wide character input/output functions can rely on the premise that
10 programmers use the wide character input/output functions with a binary stream under the same
constraints as for a text stream. An implementation may also provide wide character input/output
functions that behave correctly on an unconstrained binary stream, however the behavior of the
wide character input/output functions on such an unconstrained binary stream cannot be ensured by
all implementations.
15
MSE.10 Formatted input/output functions
20 The simplest extension for wide character input/output is to use existing formatted input/output
functions with existing byte-oriented streams. In this case, data such as strings that consist of
characters only are treated as sequences of wide characters, and other data such as numerical values
are treated as sequences of single-byte characters. Though this is not a complete model for wide
character processing, it is a common extension among some existing implementations in Japan, and
25 so the Committee decided to include a similar extension.
The original intent was to add the new conversion specifiers %S and %C to the existing formatted
input and output functions to handle a wide character string and a wide character respectively.
After long discussions about the actual implementation and future library directions, these
30 specifiers were withdrawn. They were replaced with the qualified conversion specifiers, %ls and
%lc, with the addition of %l[...] in the formatted input functions. Note that even though the new
qualifier is introduced as an extension for wide character processing, the field width and the
precision still specify the number of bytes in the multibyte representation in the stream.
35 To implement these new conversion specifiers efficiently, a new set of functions is required to parse
or generate multibyte sequences “restartably.” Thus, the functions described in §7.24.6.4 were
introduced.
Because these new conversions are pure extensions to C90, they have several essential restrictions
40 on them, and so it is expected that they will be most useful in implementations that are not
state-dependent. The restrictions are:
180
C9X RATIONALE WG14/N897 J11/99-032
function treats these shift sequences in exactly the same way as for single byte characters, an
unexpected match may occur or an expected match might not occur (see §4.6.2.3.2 of
AMD1 for some examples).
10 In the early MSE, formatted wide character input/output functions were not introduced because an
extension to existing formatted input/output functions seemed to be sufficient. After considering
the complete model for wide character handling, the need for formatted wide character input/output
functions was recognized.
15 Formatted wide character input/output functions have much the same conversion specifiers and
qualifiers as existing formatted input/output functions, even including the qualified conversion
specifiers, %lc, %ls, and %l[...], but because the format string consists of wide characters and the
field width and precision specify the number of wide characters, some of the restrictions on existing
functions are removed in the new functions. This means that wide characters are read and written
20 under tighter control of the format string.
While the Committee believes that the MSE provides reasonably complete functionality for
25 manipulating wide-oriented files, it noticed that no reliable mechanism existed for testing or setting
the orientation of a stream. The program can try certain operations to see if they fail, but that is
risky and still not a complete strategy. The Committee therefore introduced the fwide function as
a means of forcing a newly opened stream into the desired orientation without attempting any
input/output on the stream. The function also serves as a passive means of testing the orientation of
30 a stream, either before or after the orientation has been fixed; and it serves as a way to bind an
encoding rule to a wide-oriented stream under more controlled circumstances (see §MSE.9.2.3).
35 Two single-byte wide character conversion functions, btowc and wctob, were introduced in
AMD1. These functions simplify mappings between a single-byte character and its corresponding
wide character, if any.
C90 specifies the rule that L'x' == 'x' for any member x of the basic character set. The
40 Committee discussed whether to relax or tighten this rule. In AMD1, this rule is preserved without
any changes. Applying the rule to all single-byte characters, however, imposes an unnecessary
constraint on implementations with regard to wide-character encodings. It prohibits an
implementation from having a common wide-character encoding for multiple multibyte encodings.
45 On the other hand, relaxing or removing the rule was considered to be inappropriate in terms of
181
WG14/N897 J11/99-032 C9X RATIONALE
practical implementations. The new wctob function can be used to test safely and quickly whether
a wide character corresponds to some single-byte character. For example, when the format string
passed to scanf is parsed and searched for a white space character, the wctob function can be
used in conjunction with the isspace function.
5
Similarly, there are frequent occasions in wide-character processing code, especially in the wide
character handling library functions, where it is necessary to test quickly and efficiently whether a
single-byte character is the first and only character of a valid multibyte character. This is the reason
for introducing the btowc function. Note that, for some encodings, btowc can be reduced to a
10 simple inline expression.
15 Although C90 allows multibyte characters to have state-dependent encoding (§5.2.1.2), the original
functions are not always sufficient to efficiently support such encodings due to the following
limitations of the multibyte character conversion functions (§7.20.7):
1. Since the functions maintain shift state information internally, they cannot handle multiple
20 strings at the same time.
2. The formatted output functions may write redundant shift sequences, and the formatted input
functions cannot reliably parse input with arbitrary or redundant shift sequences.
For all these reasons, the Committee felt it necessary to augment the set of conversion functions in
30 AMD1.
To handle multiple strings with a state-dependent encoding, the Committee introduced the concept
35 of conversion state. The conversion state determines the behavior of a conversion between
multibyte and wide-character encodings. For conversion from multibyte characters to wide
characters, the conversion state stores information such as the position within the current multibyte
character (as a sequence of characters or a wide character accumulator). For conversions in either
direction, the conversion state stores the current shift state, if any, and possibly the encoding rule.
40
The non-array object type mbstate_t is defined to encode the conversion state. A zero-valued
mbstate_t object is assumed to describe the initial conversion state. (This is not necessarily the
only way to encode the initial conversion state, however.) Before any operations are performed on
it, such a zero-valued mbstate_t object is unbound. Once a multibyte or wide-character
45 conversion function executes with the mbstate_t object as an argument, however, the object
182
C9X RATIONALE WG14/N897 J11/99-032
The conversion functions maintain the conversion state in an mbstate_t object according to the
encoding specified in the LC_CTYPE category of the current locale. Once the conversion starts, the
5 functions will work as if the encoding scheme were not changed provided all three of the following
conditions obtain:
• the function is applied to the same string as when the mbstate_t object was first bound.
10 • the LC_CTYPE category setting is the same as when the mbstate_t object was first
bound.
• the conversion direction (multibyte to wide character, or wide character to multibyte) is the
same as when the mbstate_t object was first bound.
15
MSE.13.2 Conversion utilities
Once the mbstate_t object was introduced, the Committee discussed the need for additional
functions to manipulate such objects.
20
MSE.13.2.1 Initializing conversion states
Though a method to initialize the object is needed, the Committee decided that it would be better
not to define too many functions in AMD1. Thus the Committee decided to specify only one way
25 to make an mbstate_t object represent the initial conversion state, by initializing it with zero.
No initializing function is supplied.
30 The Committee reached the conclusion that it may be impossible to define the equality between two
conversion states. If two mbstate_t objects have the same values for all attributes, they might
be the same. However, they might also have different values and still represent the same
conversion state. No comparison function is supplied.
The mbsinit function was added to test whether an mbstate_t object describes the initial
conversion state or not, because this state does not always correspond to a certain set of component
values (and the components cannot be portably compared anyway). The function is necessary
40 because many functions in AMD1 treat the initial shift state as a special condition.
Regarding problems 2 and 3 described at the beginning of §MSE.13, the Committee introduced a
45 method to distinguish between an invalid sequence of bytes and a valid prefix to a still incomplete
multibyte character. When encountering such an incomplete multibyte sequence, the mbrlen and
183
WG14/N897 J11/99-032 C9X RATIONALE
5 The new multibyte/wide-string conversion utilities are thus made restartable by using the character
accumulator and shift state information stored in an mbstate_t object. As part of this
enhancement, the functions also have a parameter that is a pointer to a pointer to the source of the
position where the conversion stopped.
The number of characters to be read or written can be specified in existing formatted input/output
functions. On a traditional display device that displays characters with fixed pitch, the number of
characters is directly proportional to the width occupied by these characters; so the display format
15 can be specified through the field width and/or the precision.
In formatted wide character input/output functions, the field width and the precision specify the
number of wide characters to be read or written. The number of wide characters is not always
directly proportional to the width of their display. For example, with Japanese traditional display
20 devices, a single-byte character such as an ASCII character has half the width of a Kanji character,
even though each of them is treated as one wide character. To control the display format for wide
characters, a set of formatted wide character input/output functions were proposed whose metric
was the column width instead of the character count.
25 This proposal was supported only by Japan. Critics observed that the proposal was based on such
traditional display devices with fixed-width characters, while many modern display devices support
a broad assortment of proportional pitch type faces. Hence, it was questioned whether the extra
input/output functions in this proposal were really needed or were sufficiently general. Also
considered was another set of functions that return the column width for any kind of display device
30 for a given wide character or wide-character string; but these seemed to be beyond the scope of C.
Thus all proposals regarding column width were withdrawn.
If an implementor needs this kind of functionality, there are a few ways to extend wide character
output functions and still remain conforming to AMD1. For example, the new conversion specifier
35 can be used to specify the column width as shown below:
%#N — set the counting mode to “printing positions” and reset the %n counter.
%N — set the counting mode back to “wide characters” and reset the %n counter.
40
184
Index
Index
II
Index
mantissa, 19 rewind, 130, 137
matherr, 111 Ritchie, Dennis M., 24
mbrlen, 177 safe evaluation, 100
mbrtomb, 170 same type, 30
mbrtowc, 170 scanf, 100
mbstate_t, 171, 176 scope, 22
memchr, 144 sequence points, 11, 44
memcmp, 144 sequenced expression, 55
memcpy, 144 sequencing, 11
memmove, 144 setbuf, 126, 130
memset, 144, 146 setjmp, 118
minimum maxima, 3 setlocale, 102, 110
mixed code and declarations, 83 setvbuf, 124, 126, 130
mktime, 148 side effect, 28, 55, 71
modf, 113 sig_atomic_t, 17
modifiable lvalue, 34 SIGABRT, 141
MSE, 5, 161 SIGILL, 120
multibyte character, 6 signal, 12, 17, 27, 119, 122, 141
multibyte characters, 16, 143 signal.h, 17
Multibyte Support Extension, 161 signed, 36, 58
Multibyte Support Extensions, 5 significand, 19
Multiple encoding environment, 171 sign-magnitude, 19
multi-processing, 120 SIGTERM, 141
name space, 22 size_t, 50, 122, 136, 140, 146
NaN, 21, 113, 132 sizeof, 6, 57
new-line, 17 sizeof operator, 50
NULL, 54, 122 snprintf, 133
null pointer constant, 122 source file, 9
object, 6 spirit of C, 3
obsolescent, 95 sprintf, 111
old-style declaration, 75 srand, 139
ones-complement, 19 sscanf, 134
onexit, 141 standard pragmas, 94
optimization, 57 statements, 78
order of evaluation, 44 static initializers, 57
overlapping objects, 96 storage duration, 22
Pascal, 29, 81 strcoll, 145
perror, 137, 146 streams, 125
phases of translation, 9, 10 strerror, 146
POSIX, 120, 124 strftime, 149
pragma operator, 94 strictly conforming program, 2, 7, 8, 10, 14, 17, 37, 40, 97,
precedence, 44 99
preprocessing, 9, 10, 35, 41, 42, 43, 100 not, 3, 11, 14, 24, 110
primary expression, 47 stringizing, 91
printing character, 102 strlen, 146
program strncat, 145
erroneous, 10 strncpy, 144
program startup, 11, 57 strstr, 145
prototype, 84 strtod, 138
prototypes, 95 strtok, 145
ptrdiff_t, 53, 122 strtol, 138, 139
pure function, 55 strtoul, 139
putc, 100, 135
struct hack, 62, 63
putenv, 142 structure constant, 37
puts, 135 structures, 62
quality of implementation, 6, 10 strxfrm, 145
quiet change, 2, 36 system, 142
rand, 139 tags, 57
range error, 113 text streams, 126
register, 57 time, 149
remove, 127
time_t, 147
rename, 127
tm_isdst, 148
repertoire, 15
TMP_MAX, 125
restrict, 66, 96
tmpfile, 127
III
Index
tmpnam, 127
token pasting, 92
translation limits, 3
translation phases, 9
trigraph, 15, 166
Trigraphs, 161
twos-complement, 28
type modifier, 70
type qualifiers, 64
typedef, 70, 76, 84
UCN, 37, 110
undefined behavior, 6, 7, 10, 12, 14, 21, 23, 27, 28, 35, 39,
41, 45, 48, 49, 52, 63, 90, 96, 120, 142, 143
ungetc, 135
universal character, 6
universal character name, 110
Universal Character Name, 37
UNIX, 32, 96, 111, 123, 127
unlink, 127
unsequenced expression, 55
unsigned preserving, 30, 31
unspecified behavior, 6, 7, 69, 93
va_arg, 120, 121
va_end, 121
va_list, 121, 122
va_start, 120, 121
value preserving, 30
variable length array, 26, 30, 51, 53, 57, 71, 75, 83
variably modified type, 71
VAX/VMS, 111
vfprintf, 132, 134
void, 36, 58
void *, 34
void*, 28, 52, 54, 55, 131
volatile, 36
vprintf, 134
vsnprintf, 134
vsprintf, 134
wchar_t, 122
wctob, 175
wctype, 169
WG14, 1
white space, 35
wide character, 6, 40
wide string, 42
widened types, 100
IV