source: doc/refrat/refrat.tex @ aa99647

\documentclass[openright,twoside]{report}
\usepackage{fullpage,times}
\usepackage{xspace}
\usepackage{varioref}
\usepackage{listings}
\usepackage{latexsym}                                   % \Box
\usepackage{mathptmx}                                   % better math font with "times"
\usepackage[pagewise]{lineno}
\renewcommand{\linenumberfont}{\scriptsize\sffamily}
\usepackage[dvips,plainpages=false,pdfpagelabels,pdfpagemode=UseNone,colorlinks=true,pagebackref=true,linkcolor=blue,citecolor=blue,urlcolor=blue,pagebackref=true,breaklinks=true]{hyperref}
\usepackage{breakurl}
\urlstyle{sf}

%\input code.sty
\input xref.tex

\newcommand{\define}[1]{\emph{#1\/}\index{#1}}
\newenvironment{rationale}{%
  \begin{quotation}\noindent$\Box$\enspace
}{%
  \hfill\enspace$\Box$\end{quotation}
}%
\newcommand{\rewrite}{\(\Rightarrow\)}
\newcommand{\rewriterules}{\paragraph{Rewrite Rules}\hskip1em\par\noindent}
\newcommand{\examples}{\paragraph{Examples}\hskip1em\par\noindent}
\newcommand{\semantics}{\paragraph{Semantics}\hskip1em\par\noindent}
\newcommand{\constraints}{\paragraph{Constraints}\hskip1em\par\noindent}
\newenvironment{predefined}{%
  \paragraph{Predefined Identifiers}%
%  \begin{code}%
}{%
%  \end{code}
}%

\def\syntax{\paragraph{Syntax}\trivlist\parindent=.5in\item[\hskip.5in]}
\let\endsyntax=\endtrivlist
\newcommand{\lhs}[1]{\par{\it #1:}\index{#1@{\it #1}|italic}}
\newcommand{\rhs}{\hfil\break\hbox{\hskip1in}}
\newcommand{\oldlhs}[1]{{\it #1: \ldots}\index{#1@{\it #1}|italic}}
\newcommand{\nonterm}[1]{{\it #1\/}\index{#1@{\it #1}|italic}}
\newcommand{\opt}{$_{opt}$\ }

\renewcommand{\reftextfaceafter}{\unskip}
\renewcommand{\reftextfacebefore}{\unskip}
\renewcommand{\reftextafter}{\unskip}
\renewcommand{\reftextbefore}{\unskip}
\renewcommand{\reftextfaraway}[1]{\unskip, p.~\pageref{#1}}
\renewcommand{\reftextpagerange}[2]{\unskip, pp.~\pageref{#1}--\pageref{#2}}
\newcommand{\VRef}[2][Section]{\ifx#1\@empty\else{#1}\nobreakspace\fi\vref{#2}}
\newcommand{\VPageref}[2][page]{\ifx#1\@empty\else{#1}\nobreakspace\fi\pageref{#2}}

\newcommand{\CFA}{Cforall\xspace}
\newcommand{\CFAA}{C$\forall$\xspace}
\newcommand{\CC}{C\kern-.1em\hbox{+\kern-.25em+}\xspace}
\def\c11{ISO/IEC C}% cannot have numbers in latex command name

\lstdefinelanguage{CFA}[ANSI]{C}%
  {morekeywords={asm,_Atomic,catch,choose,_Complex,context,dtype,fallthru,forall,ftype,_Imaginary,lvalue,restrict,throw,try,type,},
}

\lstset{
language=CFA,
columns=fullflexible,
basicstyle=\sf\small,
tabsize=4,
xleftmargin=\parindent,
escapechar=@,
%showtabs=true,
%tab=\rightarrowfill,
}

\setcounter{secnumdepth}{3}     % number subsubsections
\makeindex

\begin{document}
\pagestyle{headings}
\linenumbers                                    % comment out to turn off line numbering

\title{\CFA (\CFAA) Reference Manual and Rationale}
\author{Glen Ditchfield}
\date{DRAFT\\\today}

\pagenumbering{roman}
\pagestyle{plain}

\maketitle

\vspace*{\fill}
\thispagestyle{empty}
\noindent
\copyright\,2015 Glen Ditchfield \\ \\
\noindent
This work is licensed under the Creative Commons Attribution 4.0 International License. To view a
copy of this license, visit {\small\url{http://creativecommons.org/licenses/by/4.0}}.
\vspace*{1in}

\clearpage
\pdfbookmark[1]{Contents}{section}
\tableofcontents

\clearpage
\pagenumbering{arabic}


\chapter*{Introduction}\addcontentsline{toc}{chapter}{Introduction}

This document is a reference manual and rationale for \CFA, a polymorphic extension of the C
programming language. It makes frequent reference to the {\c11} standard \cite{ANS:C11}, and
occasionally compares \CFA to {\CC} \cite{c++}.

The manual deliberately imitates the ordering of the {\c11} standard (although the section numbering
differs). Unfortunately, this means that the manual contains more ``forward references'' than
usual, and that it will be hard to follow if the reader does not have a copy of the {\c11} standard
near-by. For a gentle introduction to \CFA, see the companion document ``An Overview of
\CFA'' \cite{Ditchfield96:Overview}.

\begin{rationale}
Commentary (like this) is quoted with quads. Commentary usually deals with subtle points, the
rationale behind a rule, and design decisions.
\end{rationale}

% No ``Scope'' or ``Normative references'' chapters yet.
\setcounter{chapter}{2}
\chapter{Terms, definitions, and symbols}
Terms from the {\c11} standard used in this document have the same meaning as in the {\c11}
standard.

% No ``Conformance'' or ``Environment'' chapters yet.
\setcounter{chapter}{5}
\chapter{Language}
\section{Notation}
The syntax notation used in this document is the same as is used in the {\c11} standard, with one
exception: ellipsis in the definition of a nonterminal, as in ``\emph{declaration:} \ldots'',
indicates that these rules extend a previous definition, which occurs in this document or in the
{\c11} standard.


\section{Concepts}


\subsection{Scopes of identifiers}\index{scopes}

\CFA's scope rules differ from C's in one major respect: a declaration of an identifier may
overload\index{overloading} outer declarations of lexically identical identifiers in the same name
space\index{name spaces}, instead of hiding them. The outer declaration is hidden if the two
declarations have compatible type\index{compatible type}, or if one declares an array type and the
other declares a pointer type and the element type and pointed-at type are compatible, or if one has
function type and the other is a pointer to a compatible function type, or if one declaration is a
\lstinline$type$\use{type} or \lstinline$typedef$\use{typedef} declaration and the other is not.
The outer declaration becomes visible\index{visible} when the scope of the inner declaration
terminates.
\begin{rationale}
Hence, a \CFA program can declare an \lstinline$int v$ and a \lstinline$float v$ in the same
scope; a {\CC} program can not.
\end{rationale}


\subsection{Linkage of identifiers}\index{linkage}

\CFA's linkage rules differ from C's in only one respect: instances of a particular identifier
with external or internal linkage do not necessarily denote the same object or function. Instead,
in the set of translation units and libraries that constitutes an entire program, any two instances
of a particular identifier with external linkage\index{external linkage} denote the same object or
function if they have compatible types\index{compatible type}, or if one declares an array type and
the other declares a pointer type and the element type and pointed-at type are compatible, or if one
has function type and the other is a pointer to a compatible function type. Within one translation
unit, each instance of an identifier with internal linkage\index{internal linkage} denotes the same
object or function in the same circumstances. Identifiers with no linkage\index{no linkage} always
denote unique entities.
\begin{rationale}
A \CFA program can declare an \lstinline$extern int v$ and an \lstinline$extern float v$; a C
program cannot.
\end{rationale}

\section{Conversions}
\CFA defines situations where values of one type are automatically converted to another type.
These conversions are called \define{implicit conversions}. The programmer can request
\define{explicit conversions} using cast expressions.


\subsection{Arithmetic operands}
\setcounter{subsubsection}{7}


\subsubsection{Safe arithmetic conversions}
In C, a pattern of conversions known as the \define{usual arithmetic conversions} is used with most
binary arithmetic operators to convert the operands to a common type and determine the type of the
operator's result. In \CFA, these conversions play a role in overload resolution, and
collectively are called the \define{safe arithmetic conversions}.

Let \(int_r\) and \(unsigned_r\) be the signed and unsigned integer types with integer conversion
rank\index{integer conversion rank} \index{rank|see{integer conversion rank}} $r$. Let
\(unsigned_{mr}\) be the unsigned integer type with maximal rank.

The following conversions are \emph{direct} safe arithmetic conversions.
\begin{itemize}
\item
The integer promotions\index{integer promotions}.

\item
For every rank $r$ greater than or equal to the rank of \lstinline$int$, conversion from \(int_r\)
to \(unsigned_r\).

\item
For every rank $r$ greater than or equal to the rank of \lstinline$int$, where \(int_{r+1}\) exists
and can represent all values of \(unsigned_r\), conversion from \(unsigned_r\) to \(int_{r+1}\).

\item
Conversion from \(unsigned_{mr}\) to \lstinline$float$.

\item
Conversion from an enumerated type to its compatible integer type.

\item
Conversion from \lstinline$float$ to \lstinline$double$, and from \lstinline$double$ to
\lstinline$long double$.

\item
Conversion from \lstinline$float _Complex$ to \lstinline$double _Complex$,
and from \lstinline$double _Complex$ to \lstinline$long double _Complex$.

\begin{sloppypar}
\item
Conversion from \lstinline$float _Imaginary$ to \lstinline$double _Imaginary$, and from
\lstinline$double _Imaginary$ to \lstinline$long double$ \lstinline$_Imaginary$, if the
implementation supports imaginary types.
\end{sloppypar}
\end{itemize}

If type \lstinline$T$ can be converted to type \lstinline$U$ by a safe direct arithmetic conversion
and type \lstinline$U$ can be converted to type \lstinline$V$ by a safe arithmetic conversion, then
the conversion from \lstinline$T$ to type \lstinline$V$ is an \emph{indirect} safe arithmetic
conversion.

\begin{rationale}
Note that {\c11} does not include conversion from real types\index{real type} to complex
types\index{complex type} in the usual arithmetic conversions, and \CFA does not include them as
safe conversions.
\end{rationale}


\subsection{Other operands}
\setcounter{subsubsection}{3}


\subsubsection{Anonymous structures and unions}
\label{anon-conv}

If an expression's type is a pointer to a structure or union type that has a member that is an
anonymous structure\index{anonymous structure} or an anonymous union\index{anonymous union}, it can
be implicitly converted\index{implicit conversions} to a pointer to the anonymous structure's or
anonymous union's type. The result of the conversion is a pointer to the member.

\examples
\begin{lstlisting}
struct point {
        int x, y;
};
void move_by(struct point * p1, struct point * p2) {@\impl{move_by}@
        p1->x += p2.x;
        p1->y += p2.y;
}

struct color_point {
        enum { RED, BLUE, GREEN } color;
        struct point;
} cp1, cp2;
move_to(&cp1, &cp2);
\end{lstlisting}
Thanks to implicit conversion, the two arguments that \lstinline$move_by()$ receives are pointers to
\lstinline$cp1$'s second member and \lstinline$cp2$'s second member.


\subsubsection{Specialization}
A function or value whose type is polymorphic may be implicitly converted to one whose type is less
polymorphic\index{less polymorphic} by binding values to one or more of its inferred
parameters\index{inferred parameter}. Any value that is legal for the inferred parameter may be
used, including other inferred parameters.

If, after the inferred parameter binding, an assertion parameter\index{assertion parameters} has no
inferred parameters in its type, then an object or function must be visible at the point of the
specialization that has the same identifier as the assertion parameter and has a type that is
compatible\index{compatible type} with or can be specialized to the type of the assertion parameter.
The assertion parameter is bound to that object or function.

The type of the specialization is the type of the original with the bound inferred parameters and
the bound assertion parameters replaced by their bound values.

\examples
The type
\begin{lstlisting}
forall( type T, type U ) void (*)( T, U );
\end{lstlisting}
can be specialized to (among other things)
\begin{lstlisting}
forall( type T ) void (*)( T, T );              // U bound to T 
forall( type T ) void (*)( T, real );   // U bound to real 
forall( type U ) void (*)( real, U );   // T bound to real 
void f( real, real );                                   // both bound to real 
\end{lstlisting}

The type
\begin{lstlisting}
forall( type T | T ?+?( T, T )) T (*)( T );
\end{lstlisting}
can be specialized to (among other things)
\begin{lstlisting}
int (*)( int );                                         // T bound to int, and T ?+?(T, T ) bound to int ?+?( int, int ) 
\end{lstlisting}


\subsubsection{Safe conversions}

A \define{direct safe conversion} is one of the following conversions:
\begin{itemize}
\item
a direct safe arithmetic conversion;
\item
from any object type or incomplete type to \lstinline$void$;
\item
from a pointer to any non-\lstinline$void$ type to a pointer to \lstinline$void$;
\item
from a pointer to any type to a pointer to a more qualified version of the type\index{qualified
type};
\item
from a pointer to a structure or union type to a pointer to the type of a member of the structure or
union that is an anonymous structure\index{anonymous structure} or an anonymous
union\index{anonymous union};
\item
within the scope of an initialized type declaration\index{type declaration}, conversions between a
type and its implementation or between a pointer to a type and a pointer to its implementation.
\end{itemize}

Conversions that are not safe conversions are \define{unsafe conversions}.
\begin{rationale}
As in C, there is an implicit conversion from \lstinline$void *$ to any pointer type. This is
clearly dangerous, and {\CC} does not have this implicit conversion.
\CFA\index{deficiencies!void * conversion} keeps it, in the interest of remaining as pure a
superset of C as possible, but discourages it by making it unsafe.
\end{rationale}


\subsection{Conversion cost}

The \define{conversion cost} of a safe\index{safe conversions}
conversion\footnote{Unsafe\index{unsafe conversions} conversions do not have defined conversion
costs.} is a measure of how desirable or undesirable it is. It is defined as follows.
\begin{itemize}
\item
The cost of a conversion from any type to itself is 0.

\item
The cost of a direct safe conversion is 1.

\item
The cost of an indirect safe arithmetic conversion is the smallest number of direct conversions
needed to make up the conversion.
\end{itemize}

\examples
In the following, assume an implementation that does not provide any extended integer types.

\begin{itemize}
\item
The cost of an implicit conversion from \lstinline$int$ to \lstinline$long$ is 1. The cost of an
implicit conversion from \lstinline$long$ to \lstinline$double$ is 3, because it is defined in terms
of conversions from \lstinline$long$ to \lstinline$unsigned long$, then to \lstinline$float$, and
then to \lstinline$double$.

\item
If \lstinline$int$ can represent all the values of \lstinline$unsigned short$, then the cost of an
implicit conversion from \lstinline$unsigned short$ to \lstinline$unsigned$ is 2:
\lstinline$unsigned short$ to \lstinline$int$ to \lstinline$unsigned$. Otherwise,
\lstinline$unsigned short$ is converted directly to \lstinline$unsigned$, and the cost is 1.

\item
If \lstinline$long$ can represent all the values of \lstinline$unsigned$, then the conversion cost
of \lstinline$unsigned$ to \lstinline$long$ is 1. Otherwise, the conversion is an unsafe
conversion, and its conversion cost is undefined.
\end{itemize}

\section{Lexical elements}
\subsection{Keywords}
\begin{syntax}
\oldlhs{keyword}
        \rhs \lstinline$forall$
        \rhs \lstinline$lvalue$
        \rhs \lstinline$context$
        \rhs \lstinline$dtype$
        \rhs \lstinline$ftype$
        \rhs \lstinline$type$
\end{syntax}


\subsection{Identifiers}

\CFA allows operator overloading\index{overloading} by associating operators with special
function identifiers. Furthermore, the constants ``\lstinline$0$'' and ``\lstinline$1$'' have
special status for many of C's data types (and for many programmer-defined data types as well), so
\CFA treats them as overloadable identifiers. Programmers can use these identifiers to declare
functions and objects that implement operators and constants for their own types.


\setcounter{subsubsection}{2}
\subsubsection{Constant identifiers}

\begin{syntax}
\oldlhs{identifier}
\rhs \lstinline$0$
\rhs \lstinline$1$
\end{syntax}

\index{constant identifiers}\index{identifiers!for constants} The tokens ``\lstinline$0$''\impl{0}
and ``\lstinline$1$''\impl{1} are identifiers. No other tokens defined by the rules for integer
constants are considered to be identifiers.
\begin{rationale}
Why ``\lstinline$0$'' and ``\lstinline$1$''? Those integers have special status in C. All scalar
types can be incremented and decremented, which is defined in terms of adding or subtracting 1. The
operations ``\lstinline$&&$'', ``\lstinline$||$'', and ``\lstinline$!$'' can be applied to any
scalar arguments, and are defined in terms of comparison against 0. A \nonterm{constant-expression}
that evaluates to 0 is effectively compatible with every pointer type.

In C, the integer constants 0 and 1 suffice because the integer promotion rules can convert them to
any arithmetic type, and the rules for pointer expressions treat constant expressions evaluating to
0 as a special case. However, user-defined arithmetic types often need the equivalent of a 1 or 0
for their functions or operators, polymorphic functions often need 0 and 1 constants of a type
matching their polymorphic parameters, and user-defined pointer-like types may need a null value.
Defining special constants for a user-defined type is more efficient than defining a conversion to
the type from \lstinline$_Bool$.

Why \emph{just} ``\lstinline$0$'' and ``\lstinline$1$''? Why not other integers? No other integers
have special status in C. A facility that let programmers declare specific
constants---``\lstinline$const Rational 12$'', for instance---would not be much of an improvement.
Some facility for defining the creation of values of programmer-defined types from arbitrary integer
tokens would be needed. The complexity of such a feature doesn't seem worth the gain.
\end{rationale}


\subsubsection{Operator identifiers}

\index{operator identifiers}\index{identifiers!for operators} Table \ref{opids} lists the
programmer-definable operator identifiers and the operations they are associated with. Functions
that are declared with (or pointed at by function pointers that are declared with) these identifiers
can be called by expressions that use the operator tokens and syntax, or the operator identifiers
and ``function call'' syntax. The relationships between operators and function calls are discussed
in descriptions of the operators.

\begin{table}[hbt]
\hfil
\begin{tabular}[t]{ll}
%identifier & operation \\ \hline
\lstinline$?[?]$ & subscripting \impl{?[?]}\\
\lstinline$?()$ & function call \impl{?()}\\
\lstinline$?++$ & postfix increment \impl{?++}\\
\lstinline$?--$ & postfix decrement \impl{?--}\\
\lstinline$++?$ & prefix increment \impl{++?}\\
\lstinline$--?$ & prefix decrement \impl{--?}\\
\lstinline$*?$ & dereference \impl{*?}\\
\lstinline$+?$ & unary plus \impl{+?}\\
\lstinline$-?$ & arithmetic negation \impl{-?}\\
\lstinline$~?$ & bitwise negation \impl{~?}\\
\lstinline$!?$ & logical complement \impl{"!?}\\
\lstinline$?*?$ & multiplication \impl{?*?}\\
\lstinline$?/?$ & division \impl{?/?}\\
\end{tabular}\hfil
\begin{tabular}[t]{ll}
%identifier & operation \\ \hline
\lstinline$?%?$ & remainder \impl{?%?}\\
\lstinline$?+?$ & addition \impl{?+?}\\
\lstinline$?-?$ & subtraction \impl{?-?}\\
\lstinline$?<<?$ & left shift \impl{?<<?}\\
\lstinline$?>>?$ & right shift \impl{?>>?}\\
\lstinline$?<?$ & less than \impl{?<?}\\
\lstinline$?<=?$ & less than or equal \impl{?<=?}\\
\lstinline$?>=?$ & greater than or equal \impl{?>=?}\\
\lstinline$?>?$ & greater than \impl{?>?}\\
\lstinline$?==?$ & equality \impl{?==?}\\
\lstinline$?!=?$ & inequality \impl{?"!=?}\\
\lstinline$?&?$ & bitwise AND \impl{?&?}\\
\end{tabular}\hfil
\begin{tabular}[t]{ll}
%identifier & operation \\ \hline
\lstinline$?^?$ & exclusive OR \impl{?^?}\\
\lstinline$?|?$ & inclusive OR \impl{?"|?}\\
\lstinline$?=?$ & simple assignment \impl{?=?}\\
\lstinline$?*=?$ & multiplication assignment \impl{?*=?}\\
\lstinline$?/=?$ & division assignment \impl{?/=?}\\
\lstinline$?%=?$ & remainder assignment \impl{?%=?}\\
\lstinline$?+=?$ & addition assignment \impl{?+=?}\\
\lstinline$?-=?$ & subtraction assignment \impl{?-=?}\\
\lstinline$?<<=?$ & left-shift assignment \impl{?<<=?}\\
\lstinline$?>>=?$ & right-shift assignment \impl{?>>=?}\\
\lstinline$?&=?$ & bitwise AND assignment \impl{?&=?}\\
\lstinline$?^=?$ & exclusive OR assignment \impl{?^=?}\\
\lstinline$?|=?$ & inclusive OR assignment \impl{?"|=?}\\
\end{tabular}
\hfil
\caption{Operator Identifiers}
\label{opids}
\end{table}

\begin{rationale}
Operator identifiers are made up of the characters of the operator token, with question marks added
to mark the positions of the arguments of operators. The question marks serve as mnemonic devices;
programmers can not create new operators by arbitrarily mixing question marks and other
non-alphabetic characters. Note that prefix and postfix versions of the increment and decrement
operators are distinguished by the position of the question mark.
\end{rationale}

\begin{rationale}
The use of ``\lstinline$?$'' in identifiers means that some C programs are not \CFA programs.
For instance, the sequence of characters ``\lstinline$(i < 0)?--i:i$'' is legal in a C program, but
a \CFA compiler will detect a syntax error because it will treat ``\lstinline$?--$'' as an
identifier, not as the two tokens ``\lstinline$?$'' and ``\lstinline$--$''.
\end{rationale}

\begin{rationale}
Certain operators \emph{cannot} be defined by the programmer:
\begin{itemize}
\item
The logical operators ``\lstinline$&&$'' and ``\lstinline$||$'', and the conditional operator
``\lstinline$?:$''. These operators do not always evaluate their operands, and hence can not be
properly defined by functions unless some mechanism like call-by-name is added to the language.
Note that the definitions of ``\lstinline$&&$'' and ``\lstinline$||$'' say that they work by
checking that their arguments are unequal to 0, so defining ``\lstinline$!=$'' and ``\lstinline$0$''
for user-defined types is enough to allow them to be used in logical expressions.

\item
The comma operator\index{comma expression}. It is a control-flow operator like those above.
Changing its meaning seems pointless and confusing.

\item
The ``address of'' operator. It would seem useful to define a unary ``\lstinline$&$'' operator that
returns values of some programmer-defined pointer-like type. The problem lies with the type of the
operator. Consider the expression ``\lstinline$p = &x$'', where \lstinline$x$ is of type
\lstinline$T$ and \lstinline$p$ has the programmer-defined type \lstinline$T_ptr$. The expression
might be treated as a call to the unary function ``\lstinline$&?$''. Now what is the type of the
function's parameter? It can not be \lstinline$T$, because then \lstinline$x$ would be passed by
value, and there is no way to create a useful pointer-like result from a value. Hence the parameter
must have type \lstinline$T *$. But then the expression must be rewritten as ``\lstinline$p = &?( &x )$''
---which doesn't seem like progress!

The rule for address-of expressions would have to be something like ``keep applying address-of
functions until you get one that takes a pointer argument, then use the built-in operator and
stop''. It seems simpler to define a conversion function from \lstinline$T *$ to \lstinline$T_ptr$.

\item
The \lstinline$sizeof$ operator. It is already defined for every object type, and intimately tied
into the language's storage allocation model. Redefining it seems pointless.

\item
The ``member of'' operators ``\lstinline$.$'' and ``\lstinline$->$''. These are not really infix
operators, since their right ``operand'' is not a value or object.

\item
Cast operators\index{cast expression}. Anything that can be done with an explicit cast can be done
with a function call. The difference in syntax is small.
\end{itemize}
\end{rationale}


\section{Expressions}
\CFA allows operators and identifiers to be overloaded. Hence, each expression can have a number
of \define{interpretations}, each of which has a different type. The interpretations that are
potentially executable are called \define{valid interpretations}. The set of interpretations
depends on the kind of expression and on the interpretations of the subexpressions that it contains.
The rules for determining the valid interpretations of an expression are discussed below for each
kind of expression. Eventually the context of the outermost expression chooses one interpretation
of that expression.

An \define{ambiguous interpretation} is an interpretation which does not specify the exact object or
function denoted by every identifier in the expression. An expression can have some interpretations
that are ambiguous and others that are unambiguous. An expression that is chosen to be executed
shall not be ambiguous.

The \define{best valid interpretations} are the valid interpretations that use the fewest
unsafe\index{unsafe conversions} conversions. Of these, the best are those where the functions and
objects involved are the least polymorphic\index{less polymorphic}. Of these, the best have the
lowest total conversion cost\index{conversion cost}, including all implicit conversions in the
argument expressions. Of these, the best have the highest total conversion cost for the implicit
conversions (if any) applied to the argument expressions. If there is no single best valid
interpretation, or if the best valid interpretation is ambiguous, then the resulting interpretation
is ambiguous\index{ambiguous interpretation}.

\begin{rationale}
\CFA's rules for selecting the best interpretation are designed to allow overload resolution to
mimic C's operator semantics. In C, the ``usual arithmetic conversions'' are applied to the
operands of binary operators if necessary to convert the operands to types with a common real type.
In \CFA, those conversions are ``safe''. The ``fewest unsafe conversions'' rule ensures that the
usual conversions are done, if possible. The ``lowest total expression cost'' rule chooses the
proper common type. The odd-looking ``highest argument conversion cost'' rule ensures that, when
unary expressions must be converted, conversions of function results are preferred to conversion of
function arguments: \lstinline$(double)-i$ will be preferred to \lstinline$-(double)i$.

The ``least polymorphic'' rule reduces the number of polymorphic function calls, since such
functions are presumably more expensive than monomorphic functions and since the more specific
function is presumably more appropriate. It also gives preference to monomorphic values (such as
the \lstinline$int$ \lstinline$0$) over polymorphic values (such as the null pointer
\lstinline$0$)\use{0}\index{null pointer}. However, interpretations that call polymorphic functions
are preferred to interpretations that perform unsafe conversions, because those conversions
potentially lose accuracy or violate strong typing.

There are two notable differences between \CFA's overload resolution rules and the rules for
{\CC} defined in \cite{c++}. First, the result type of a function plays a role. In {\CC}, a
function call must be completely resolved based on the arguments to the call in most circumstances.
In \CFA, a function call may have several interpretations, each with a different result type, and
the interpretations of the containing context choose among them. Second, safe conversions are used
to choose among interpretations of all sorts of functions; in {\CC}, the ``usual arithmetic
conversions'' are a separate set of rules that apply only to the built-in operators.
\end{rationale}

Expressions involving certain operators\index{operator identifiers} are considered to be equivalent
to function calls. A transformation from ``operator'' syntax to ``function call'' syntax is defined
by \define{rewrite rules}. Each operator has a set of predefined functions that overload its
identifier. Overload resolution determines which member of the set is executed in a given
expression. The functions have internal linkage\index{internal linkage} and are implicitly declared
with file scope\index{file scope}. The predefined functions and rewrite rules are discussed below
for each of these operators.
\begin{rationale}
Predefined functions and constants have internal linkage because that simplifies optimization in
traditional compile-and-link environments. For instance, ``\lstinline$an_int + an_int$'' is
equivalent to ``\lstinline$?+?(an_int, an_int)$''. If integer addition has not been redefined in
the current scope, a compiler can generate code to perform the addition directly. If predefined
functions had external linkage, this optimization would be difficult.
\end{rationale}

\begin{rationale}
Since each subsection describes the interpretations of an expression in terms of the interpretations
of its subexpressions, this chapter can be taken as describing an overload resolution algorithm that
uses one bottom-up pass over an expression tree. Such an algorithm was first described (for Ada) by
Baker \cite{Bak:overload}. It is extended here to handle polymorphic functions and arithmetic
conversions. The overload resolution rules and the predefined functions have been chosen so that,
in programs that do not introduce overloaded declarations, expressions will have the same meaning in
C and in \CFA.
\end{rationale}

\begin{rationale}
Expression syntax is quoted from the {\c11} standard. The syntax itself defines the precedence and
associativity of operators. The sections are arranged in decreasing order of precedence, with all
operators in a section having the same precedence.
\end{rationale}

\subsection{Primary expressions}
\begin{syntax}
\lhs{primary-expression}
\rhs \nonterm{identifier}
\rhs \nonterm{constant}
\rhs \nonterm{string-literal}
\rhs \lstinline$($ \nonterm{expression} \lstinline$)$
\rhs \nonterm{generic-selection}
\end{syntax}

\paragraph{Predefined Identifiers}%
\begin{lstlisting}
const int 1;@\use{1}@
const int 0;@\use{0}@
forall( dtype DT ) DT *const 0;
forall( ftype FT ) FT *const 0;
\end{lstlisting}

\semantics
The valid interpretations\index{valid interpretations} of an \nonterm{identifier} are given by the
visible\index{visible} declarations of the identifier.

A \nonterm{constant} or \nonterm{string-literal} has one valid interpretation, which has the type
and value defined by {\c11}. The predefined integer identifiers ``\lstinline$1$'' and
``\lstinline$0$'' have the integer values 1 and 0, respectively. The other two predefined
``\lstinline$0$'' identifiers are bound to polymorphic pointer values that, when
specialized\index{specialization} with a data type or function type respectively, produce a null
pointer of that type.

A parenthesised expression has the same interpretations as the contained \nonterm{expression}.

\examples
The expression \lstinline$(void *)0}$\use{0} specializes the (polymorphic) null pointer to a null
pointer to \lstinline$void$. \lstinline$(const void *)0$ does the same, and also uses a safe
conversion from \lstinline$void *$ to \lstinline$const void *$. In each case, the null pointer
conversion is better\index{best valid interpretations} than the unsafe conversion of the integer
\lstinline$0$ to a pointer.

\begin{rationale}
Note that the predefined identifiers have addresses.

\CFA does not have C's concept of ``null pointer constants'', which are not typed values but
special strings of tokens. The C token ``\lstinline$0$'' is an expression of type \lstinline$int$
with the value ``zero'', and it \emph{also} is a null pointer constant. Similarly,
``\lstinline$(void *)0$ is an expression of type \lstinline$(void *)$ whose value is a null pointer,
and it also is a null pointer constant. However, in C, ``\lstinline$(void *)(void *)0$'' is
\emph{not} a null pointer constant, even though it is null-valued, a pointer, and constant! The
semantics of C expressions contain many special cases to deal with subexpressions that are null
pointer constants.

\CFA handles these cases through overload resolution. The declaration
\begin{lstlisting}
forall( dtype DT ) DT *const 0;
\end{lstlisting}
means that \lstinline$0$ is a polymorphic object, and contains a value that can have \emph{any}
pointer-to-object type or pointer-to-incomplete type. The only such value is the null pointer.
Therefore the type \emph{alone} is enough to identify a null pointer. Where C defines an operator
with a special case for the null pointer constant, \CFA defines predefined functions with a
polymorphic object parameter.
\end{rationale}

\subsubsection{Generic selection}
\constraints The best interpretation of the controlling expression shall be
unambiguous\index{ambiguous interpretation}, and shall have type compatible with at most one of the
types named in its generic association list. If a generic selection has no \lstinline$default$
generic association, the best interpretation of its controlling expression shall have type
compatible with exactly one of the types named in its generic association list.

\semantics
A generic selection has the same interpretations as its result expression.


\subsection{Postfix operators}

\begin{syntax}
\lhs{postfix-expression}
\rhs \nonterm{primary-expression}
\rhs \nonterm{postfix-expression} \lstinline$[$ \nonterm{expression} \lstinline$]$
\rhs \nonterm{postfix-expression} \lstinline$($ 
         \nonterm{argument-expression-list}\opt \lstinline$)$
\rhs \nonterm{postfix-expression} \lstinline$.$ \nonterm{identifier}
\rhs \nonterm{postfix-expression} \lstinline$->$ \nonterm{identifier}
\rhs \nonterm{postfix-expression} \lstinline$++$
\rhs \nonterm{postfix-expression} \lstinline$--$
\rhs \lstinline$($ \nonterm{type-name} \lstinline$)$ \lstinline${$ \nonterm{initializer-list} \lstinline$}$
\rhs \lstinline$($ \nonterm{type-name} \lstinline$)$ \lstinline${$ \nonterm{initializer-list} \lstinline$,$ \lstinline$}$
\lhs{argument-expression-list}
\rhs \nonterm{assignment-expression}
\rhs \nonterm{argument-expression-list} \lstinline$,$
         \nonterm{assignment-expression}
\end{syntax}

\rewriterules
\begin{lstlisting}
a[b] @\rewrite@ ?[?]( b, a ) // if a has integer type */@\use{?[?]}@
a[b] @\rewrite@ ?[?]( a, b ) // otherwise
a( ${\em arguments }$ ) @\rewrite@ ?()( a, ${\em arguments} )$@\use{?()}@
a++ @\rewrite@ ?++(&( a ))@\use{?++}@
a-- @\rewrite@ ?--(&( a ))@\use{?--}@
\end{lstlisting}

\subsubsection{Array subscripting}
\begin{lstlisting}
forall( type T ) lvalue T ?[?]( T *, ptrdiff_t );@\use{ptrdiff_t}@
forall( type T ) lvalue _Atomic T ?[?]( _Atomic T *, ptrdiff_t );
forall( type T ) lvalue const T ?[?]( const T *, ptrdiff_t );
forall( type T ) lvalue restrict T ?[?]( restrict T *, ptrdiff_t );
forall( type T ) lvalue volatile T ?[?]( volatile T *, ptrdiff_t );
forall( type T ) lvalue _Atomic const T ?[?]( _Atomic const T *, ptrdiff_t );
forall( type T ) lvalue _Atomic restrict T ?[?]( _Atomic restrict T *, ptrdiff_t );
forall( type T ) lvalue _Atomic volatile T ?[?]( _Atomic volatile T *, ptrdiff_t );
forall( type T ) lvalue const restrict T ?[?]( const restrict T *, ptrdiff_t );
forall( type T ) lvalue const volatile T ?[?]( const volatile T *, ptrdiff_t );
forall( type T ) lvalue restrict volatile T ?[?]( restrict volatile T *, ptrdiff_t );
forall( type T ) lvalue _Atomic const restrict T ?[?]( _Atomic const restrict T *, ptrdiff_t );
forall( type T ) lvalue _Atomic const volatile T ?[?]( _Atomic const volatile T *, ptrdiff_t );
forall( type T ) lvalue _Atomic restrict volatile T ?[?]( _Atomic restrict volatile T *, ptrdiff_t );
forall( type T ) lvalue const restrict volatile T ?[?]( const restrict volatile T *, ptrdiff_t );
forall( type T ) lvalue _Atomic const restrict volatile T ?[?]( _Atomic const restrict volatile T *, ptrdiff_t );
\end{lstlisting}
\semantics
The interpretations of subscript expressions are the interpretations of the corresponding function
call expressions.
\begin{rationale}
C defines subscripting as pointer arithmetic in a way that makes \lstinline$a[i]$ and
\lstinline$i[a]$ equivalent. \CFA provides the equivalence through a rewrite rule to reduce the
number of overloadings of \lstinline$?[?]$.

Subscript expressions are rewritten as function calls that pass the first parameter by value. This
is somewhat unfortunate, since array-like types tend to be large. The alternative is to use the
rewrite rule ``\lstinline$a[b]$ \rewrite \lstinline$?[?](&(a), b)$''. However, C semantics forbid
this approach: the \lstinline$a$ in ``\lstinline$a[b]$'' can be an arbitrary pointer value, which
does not have an address.

The repetitive form of the predefined identifiers shows up a deficiency\index{deficiencies!pointers
 to qualified types} of \CFA's type system. Type qualifiers are not included in type values, so
polymorphic functions that take pointers to arbitrary types often come in one flavor for each
possible qualification of the pointed-at type.
\end{rationale}


\subsubsection{Function calls}

\semantics
A \define{function designator} is an interpretation of an expression that has function type. The
\nonterm{postfix-expression} in a function call may have some interpretations that are function
designators and some that are not.

For those interpretations of the \nonterm{postfix-expression} that are not function designators, the
expression is rewritten and becomes a call of a function named ``\lstinline$?()$''. The valid
interpretations of the rewritten expression are determined in the manner described below.

Each combination of function designators and argument interpretations is considered. For those
interpretations of the \nonterm{postfix-expression} that are monomorphic\index{monomorphic function}
function designators, the combination has a valid interpretation\index{valid interpretations} if the
function designator accepts the number of arguments given, and each argument interpretation matches
the corresponding explicit parameter:
\begin{itemize}
\item
if the argument corresponds to a parameter in the function designator's prototype, the argument
interpretation must have the same type as the corresponding parameter, or be implicitly convertible
to the parameter's type
\item
if the function designator's type does not include a prototype or if the argument corresponds to
``\lstinline$...$'' in a prototype, a default argument promotion\index{default argument promotions}
is applied to it.
\end{itemize}
The type of the valid interpretation is the return type of the function designator.

For those combinations where the interpretation of the \nonterm{postfix-expression} is a
polymorphic\index{polymorphic function} function designator and the function designator accepts the
number of arguments given, there shall be at least one set of \define{implicit arguments} for the
implicit parameters such that
\begin{itemize}
\item
If the declaration of the implicit parameter uses type-class\index{type-class}
\lstinline$type$\use{type}, the implicit argument must be an object type; if it uses
\lstinline$dtype$, the implicit argument must be an object type or an incomplete type; and if it
uses \lstinline$ftype$, the implicit argument must be a function type.

\item
if an explicit parameter's type uses any implicit parameters, then the corresponding explicit
argument must have a type that is (or can be safely converted\index{safe conversions} to) the type
produced by substituting the implicit arguments for the implicit parameters in the explicit
parameter type.

\item
the remaining explicit arguments must match the remaining explicit parameters, as described for
monomorphic function designators.

\item
for each assertion parameter\index{assertion parameters} in the function designator's type, there
must be an object or function with the same identifier that is visible at the call site and whose
type is compatible with or can be specialized to the type of the assertion declaration.
\end{itemize}
There is a valid interpretation for each such set of implicit parameters. The type of each valid
interpretation is the return type of the function designator with implicit parameter values
substituted for the implicit arguments.

A valid interpretation is ambiguous\index{ambiguous interpretation} if the function designator or
any of the argument interpretations is ambiguous.

Every valid interpretation whose return type is not compatible with any other valid interpretation's
return type is an interpretation of the function call expression.

Every set of valid interpretations that have mutually compatible\index{compatible type} result types
also produces an interpretation of the function call expression. The type of the interpretation is
the composite\index{composite type} type of the types of the valid interpretations, and the value of
the interpretation is that of the best valid interpretation\index{best valid interpretations}.
\begin{rationale}
One desirable property of a polymorphic programming language is \define{generalizability}: the
ability to replace an abstraction with a more general but equivalent abstraction without requiring
changes in any of the uses of the original\cite{Cormack90}. For instance, it should be possible to
replace a function ``\lstinline$int f( int );$'' with ``\lstinline$forall( type T ) T f( T );$''
without affecting any calls of \lstinline$f$.

\CFA\index{deficiencies!generalizability} does not fully possess this property, because
unsafe\index{unsafe conversions} conversions are not done when arguments are passed to polymorphic
parameters. Consider
\begin{lstlisting}
float g( float, float );
int i;
float f;
double d;
f = g( f, f );  // (1) 
f = g( i, f );  // (2) (safe conversion to float) 
f = g( d, f );  // (3) (unsafe conversion to float) 
\end{lstlisting}
If \lstinline$g$ was replaced by ``\lstinline$forall( type T ) T g( T, T );$'', the first and second
calls would be unaffected, but the third would change: \lstinline$f$ would be converted to
\lstinline$double$, and the result would be a \lstinline$double$.

Another example is the function ``\lstinline$void h( int *);$''. This function can be passed a
\lstinline$void *$ argument, but the generalization ``\lstinline$forall( type T ) void h( T *);$''
can not. In this case, \lstinline$void$ is not a valid value for \lstinline$T$ because it is not an
object type. If unsafe conversions were allowed, \lstinline$T$ could be inferred to be \emph{any}
object type, which is undesirable.
\end{rationale}

\examples
A function called ``\lstinline$?()$'' might be part of a numerical differentiation package.
\begin{lstlisting}
extern type Derivative;
extern double ?()( Derivative, double );
extern Derivative derivative_of( double (*f)( double ) );
extern double sin( double );

Derivative sin_dx = derivative_of( sin );
double d;
d = sin_dx( 12.9 );
\end{lstlisting}
Here, the only interpretation of \lstinline$sin_dx$ is as an object of type \lstinline$Derivative$.
For that interpretation, the function call is treated as ``\lstinline$?()( sin_dx, 12.9 )$''.
\begin{lstlisting}
int f( long );          // (1) 
int f( int, int );      // (2) 
int f( int *);          // (3) 

int i = f( 5 );         // calls (1) 
\end{lstlisting}
Function (1) provides a valid interpretation of ``\lstinline$f( 5 )$'', using an implicit
\lstinline$int$ to \lstinline$long$ conversion. The other functions do not, since the second
requires two arguments, and since there is no implicit conversion from \lstinline$int$ to
\lstinline$int *$ that could be used with the third function.

\begin{lstlisting}
forall( type T ) T h( T );
double d = h( 1.5 );
\end{lstlisting}
``\lstinline$1.5$'' is a \lstinline$double$ constant, so \lstinline$T$ is inferred to be
\lstinline$double$, and the result of the function call is a \lstinline$double$.

\begin{lstlisting}
forall( type T, type U ) void g( T, U );        // (4) 
forall( type T ) void g( T, T );                        // (5) 
forall( type T ) void g( T, long );                     // (6) 
void g( long, long );                                           // (7) 
double d;
int i;
int *p;

g( d, d );                      // calls (5) 
g( d, i );                      // calls (6) 
g( i, i );                      // calls (7) 
g( i, p );                      // calls (4) 
\end{lstlisting}
The first call has valid interpretations for all four versions of \lstinline$g$. (6) and (7) are
discarded because they involve unsafe \lstinline$double$-to-\lstinline$long$ conversions. (5) is
chosen because it is less polymorphic than (4).

For the second call, (7) is again discarded. Of the remaining interpretations for (4), (5), and (6)
(with \lstinline$i$ converted to \lstinline$long$), (6) is chosen because it is the least
polymorphic.

The third call has valid interpretations for all of the functions; (7) is chosen since it is not
polymorphic at all.

The fourth call has no interpretation for (5), because its arguments must have compatible type. (4)
is chosen because it does not involve unsafe conversions.
\begin{lstlisting}
forall( type T ) T min( T, T );
double max( double, double );
context min_max( T ) {@\impl{min_max}@
        T min( T, T );
        T max( T, T );
}
forall( type U | min_max( U ) ) void shuffle( U, U );
shuffle(9, 10);
\end{lstlisting}
The only possibility for \lstinline$U$ is \lstinline$double$, because that is the type used in the
only visible \lstinline$max$ function. 9 and 10 must be converted to \lstinline$double$, and
\lstinline$min$ must be specialized with \lstinline$T$ bound to \lstinline$double$.
\begin{lstlisting}
extern void q( int );           // (8) 
extern void q( void * );        // (9) 
extern void r();
q( 0 );
r( 0 );
\end{lstlisting}
The \lstinline$int 0$ could be passed to (8), or the \lstinline$(void *)$
specialization\index{specialization} of the null pointer\index{null pointer} \lstinline$0$\use{0}
could be passed to (9). The former is chosen because the \lstinline$int$ \lstinline$0$ is less
polymorphic\index{less polymorphic}. For the same reason, \lstinline$int$ \lstinline$0$ is passed
to \lstinline$r()$, even though it has \emph{no} declared parameter types.


\subsubsection{Structure and union members}

\semantics In the member selection expression ``\lstinline$s$.\lstinline$m$'', there shall be at
least one interpretation of \lstinline$s$ whose type is a structure type or union type containing a
member named \lstinline$m$. If two or more interpretations of \lstinline$s$ have members named
\lstinline$m$ with mutually compatible types, then the expression has an ambiguous
interpretation\index{ambiguous interpretation} whose type is the composite type of the types of the
members. If an interpretation of \lstinline$s$ has a member \lstinline$m$ whose type is not
compatible with any other \lstinline$s$'s \lstinline$m$, then the expression has an interpretation
with the member's type. The expression has no other interpretations.

The expression ``\lstinline$p->m$'' has the same interpretations as the expression
``\lstinline$(*p).m$''.


\subsubsection{Postfix increment and decrement operators}

\begin{lstlisting}
_Bool ?++( volatile _Bool * ),
        ?++( _Atomic volatile _Bool * );
char ?++( volatile char * ),
        ?++( _Atomic volatile char * );
signed char ?++( volatile signed char * ),
        ?++( _Atomic volatile signed char * );
unsigned char ?++( volatile signed char * ),
        ?++( _Atomic volatile signed char * );
short int ?++( volatile short int * ),
        ?++( _Atomic volatile short int * );
unsigned short int ?++( volatile unsigned short int * ),
        ?++( _Atomic volatile unsigned short int * );
int ?++( volatile int * ),
        ?++( _Atomic volatile int * );
unsigned int ?++( volatile unsigned int * ),
        ?++( _Atomic volatile unsigned int * );
long int ?++( volatile long int * ),
        ?++( _Atomic volatile long int * );
long unsigned int ?++( volatile long unsigned int * ),
        ?++( _Atomic volatile long unsigned int * );
long long int ?++( volatile long long int * ),
        ?++( _Atomic volatile long long int * );
long long unsigned ?++( volatile long long unsigned int * ),
        ?++( _Atomic volatile long long unsigned int * );
float ?++( volatile float * ),
        ?++( _Atomic volatile float * );
double ?++( volatile double * ),
        ?++( _Atomic volatile double * );
long double ?++( volatile long double * ),
        ?++( _Atomic volatile long double * );

forall( type T ) T * ?++( T * restrict volatile * ),
        * ?++( T * _Atomic restrict volatile * );

forall( type T ) _Atomic T * ?++( _Atomic T * restrict volatile * ),
        * ?++( _Atomic T * _Atomic restrict volatile * );

forall( type T ) const T * ?++( const T * restrict volatile * ),
        * ?++( const T * _Atomic restrict volatile * );

forall( type T ) volatile T * ?++( volatile T * restrict volatile * ),
        * ?++( volatile T * _Atomic restrict volatile * );

forall( type T ) restrict T * ?++( restrict T * restrict volatile * ),
        * ?++( restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic const T * ?++( _Atomic const T * restrict volatile * ),
        * ?++( _Atomic const T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict T * ?++( _Atomic restrict T * restrict volatile * ),
        * ?++( _Atomic restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic volatile T * ?++( _Atomic volatile T * restrict volatile * ),
        * ?++( _Atomic volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict T * ?++( const restrict T * restrict volatile * ),
        * ?++( const restrict T * _Atomic restrict volatile * );

forall( type T ) const volatile T * ?++( const volatile T * restrict volatile * ),
        * ?++( const volatile T * _Atomic restrict volatile * );

forall( type T ) restrict volatile T * ?++( restrict volatile T * restrict volatile * ),
        * ?++( restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict T * ?++( _Atomic const restrict T * restrict volatile * ),
        * ?++( _Atomic const restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic const volatile T * ?++( _Atomic const volatile T * restrict volatile * ),
        * ?++( _Atomic const volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict volatile T * ?++( _Atomic restrict volatile T * restrict volatile * ),
        * ?++( _Atomic restrict volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict volatile T * ?++( const restrict volatile T * restrict volatile * ),
        * ?++( const restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict volatile T * ?++( _Atomic const restrict volatile T * restrict volatile * ),
        * ?++( _Atomic const restrict volatile T * _Atomic restrict volatile * );

_Bool ?--( volatile _Bool * ),
        ?--( _Atomic volatile _Bool * );
char ?--( volatile char * ),
        ?--( _Atomic volatile char * );
signed char ?--( volatile signed char * ),
        ?--( _Atomic volatile signed char * );
unsigned char ?--( volatile signed char * ),
        ?--( _Atomic volatile signed char * );
short int ?--( volatile short int * ),
        ?--( _Atomic volatile short int * );
unsigned short int ?--( volatile unsigned short int * ),
        ?--( _Atomic volatile unsigned short int * );
int ?--( volatile int * ),
        ?--( _Atomic volatile int * );
unsigned int ?--( volatile unsigned int * ),
        ?--( _Atomic volatile unsigned int * );
long int ?--( volatile long int * ),
        ?--( _Atomic volatile long int * );
long unsigned int ?--( volatile long unsigned int * ),
        ?--( _Atomic volatile long unsigned int * );
long long int ?--( volatile long long int * ),
        ?--( _Atomic volatile long long int * );
long long unsigned ?--( volatile long long unsigned int * ),
        ?--( _Atomic volatile long long unsigned int * );
float ?--( volatile float * ),
        ?--( _Atomic volatile float * );
double ?--( volatile double * ),
        ?--( _Atomic volatile double * );
long double ?--( volatile long double * ),
        ?--( _Atomic volatile long double * );

forall( type T ) T * ?--( T * restrict volatile * ),
        * ?--( T * _Atomic restrict volatile * );

forall( type T ) _Atomic T * ?--( _Atomic T * restrict volatile * ),
        * ?--( _Atomic T * _Atomic restrict volatile * );

forall( type T ) const T * ?--( const T * restrict volatile * ),
        * ?--( const T * _Atomic restrict volatile * );

forall( type T ) volatile T * ?--( volatile T * restrict volatile * ),
        * ?--( volatile T * _Atomic restrict volatile * );

forall( type T ) restrict T * ?--( restrict T * restrict volatile * ),
        * ?--( restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic const T * ?--( _Atomic const T * restrict volatile * ),
        * ?--( _Atomic const T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict T * ?--( _Atomic restrict T * restrict volatile * ),
        * ?--( _Atomic restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic volatile T * ?--( _Atomic volatile T * restrict volatile * ),
        * ?--( _Atomic volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict T * ?--( const restrict T * restrict volatile * ),
        * ?--( const restrict T * _Atomic restrict volatile * );

forall( type T ) const volatile T * ?--( const volatile T * restrict volatile * ),
        * ?--( const volatile T * _Atomic restrict volatile * );

forall( type T ) restrict volatile T * ?--( restrict volatile T * restrict volatile * ),
        * ?--( restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict T * ?--( _Atomic const restrict T * restrict volatile * ),
        * ?--( _Atomic const restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic const volatile T * ?--( _Atomic const volatile T * restrict volatile * ),
        * ?--( _Atomic const volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict volatile T * ?--( _Atomic restrict volatile T * restrict volatile * ),
        * ?--( _Atomic restrict volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict volatile T * ?--( const restrict volatile T * restrict volatile * ),
        * ?--( const restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict volatile T * ?--( _Atomic const restrict volatile T * restrict volatile * ),
        * ?--( _Atomic const restrict volatile T * _Atomic restrict volatile * );
\end{lstlisting}
For every extended integer type \lstinline$X$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X ?++( volatile X * ), ?++( _Atomic volatile X * ),
 ?--( volatile X * ), ?--( _Atomic volatile X * );
\end{lstlisting}
For every complete enumerated type \lstinline$E$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
E ?++( volatile E * ), ?++( _Atomic volatile E * ),
 ?--( volatile E * ), ?--( _Atomic volatile E * );
\end{lstlisting}

\begin{rationale}
Note that ``\lstinline$++$'' and ``\lstinline$--$'' are rewritten as function calls that are given a
pointer to that operand. (This is true of all operators that modify an operand.) As Hamish Macdonald
has pointed out, this forces the modified operand of such expressions to be an lvalue. This
partially enforces the C semantic rule that such operands must be \emph{modifiable} lvalues.
\end{rationale}

\begin{rationale}
In C, a semantic rule requires that pointer operands of increment and decrement be pointers to
object types. Hence, \lstinline$void *$ objects cannot be incremented. In \CFA, the restriction
follows from the use of a \lstinline$type$ parameter in the predefined function definitions, as
opposed to \lstinline$dtype$, since only object types can be inferred arguments corresponding to the
type parameter \lstinline$T$.
\end{rationale}

\semantics
First, each interpretation of the operand of an increment or decrement expression is considered
separately. For each interpretation that is a bit-field or is declared with the
\lstinline$register$\index{register@{\lstinline$register$}} storage-class
specifier\index{storage-class specifier}, the expression has one valid interpretation, with the type
of the operand, and the expression is ambiguous if the operand is.

For the remaining interpretations, the expression is rewritten, and the interpretations of the
expression are the interpretations of the corresponding function call. Finally, all interpretations
of the expression produced for the different interpretations of the operand are combined to produce
the interpretations of the expression as a whole; where interpretations have compatible result
types, the best interpretations are selected in the manner described for function call expressions.

\examples
\begin{lstlisting}
volatile short int vs;  vs++; // rewritten as ?++( &(vs) )
short int s;                    s++;
const short int cs;             cs++;
_Atomic short int as;   as++;
\end{lstlisting}
\begin{sloppypar}
Since \lstinline$&(vs)$ has type \lstinline$volatile short int *$, the best valid interpretation of
\lstinline$vs++$ calls the \lstinline$?++$ function with the \lstinline$volatile short *$ parameter.
\lstinline$s++$ does the same, applying the safe conversion from \lstinline$short int *$ to
\lstinline$volatile short int *$. Note that there is no conversion that adds an \lstinline$_Atomic$
qualifier, so the \lstinline$_Atomic volatile short int$ overloading does not provide a valid
interpretation.
\end{sloppypar}

There is no safe conversion from \lstinline$const short int *$ to \lstinline$volatile short int *$,
and no \lstinline$?++$ function that accepts a \lstinline$const *$ parameter, so \lstinline$cs++$
has no valid interpretations.

The best valid interpretation of \lstinline$as++$ calls the \lstinline$short ?++$ function with the
\lstinline$_Atomic volatile short int *$ parameter, applying a safe conversion to add the
\lstinline$volatile$ qualifier.

\begin{lstlisting}
char * const restrict volatile * restrict volatile pqpc; pqpc++
char * * restrict volatile ppc; ppc++; 
\end{lstlisting}
Since \lstinline$&(pqpc)$ has type \lstinline$char * const restrict volatile * restrict volatile *$,
the best valid interpretation of \lstinline$pqpc++$ calls the polymorphic \lstinline$?++$ function
with the \lstinline$const restrict volatile T * restrict volatile *$ parameter, inferring
\lstinline$T$ to be \lstinline$char *$.

\begin{sloppypar}
\lstinline$ppc++$ calls the same function, again inferring \lstinline$T$ to be \lstinline$char *$,
and using the safe conversions from \lstinline$T$ to \lstinline$T const restrict volatile$.
\end{sloppypar}

\begin{rationale}
Increment and decrement expressions show up a deficiency of \CFA's type system. There is no such
thing as a pointer to a register object or bit-field\index{deficiencies!pointers to bit-fields}.
Therefore, there is no way to define a function that alters them, and hence no way to define
increment and decrement functions for them. As a result, the semantics of increment and decrement
expressions must treat them specially. This holds true for all of the operators that may modify
such objects.
\end{rationale}

\begin{rationale}
The polymorphic overloadings for pointer increment and decrement can be understood by considering
increasingly complex types.
\begin{enumerate}
\item
``\lstinline$char * p; p++;$''. The argument to \lstinline$?++$ has type \lstinline$char * *$, and
the result has type \lstinline$char *$. The expression would be valid if \lstinline$?++$ were
declared by
\begin{lstlisting}
forall( type T ) T * ?++( T * * );
\end{lstlisting}
with \lstinline$T$ inferred to be \lstinline$char$.

\item
``\lstinline$char *restrict volatile qp; qp++$''. The result again has type \lstinline$char *$, but
the argument now has type \lstinline$char *restrict volatile *$, so it cannot be passed to the
hypothetical function declared in point 1. Hence the actual predefined function is
\begin{lstlisting}
forall( type T ) T * ?++( T * restrict volatile * );
\end{lstlisting}
which also accepts a \lstinline$char * *$ argument, because of the safe conversions that add
\lstinline$volatile$ and \lstinline$restrict$ qualifiers. (The parameter is not const-qualified, so
constant pointers cannot be incremented.)

\item
``\lstinline$char *_Atomic ap; ap++$''. The result again has type \lstinline$char *$, but no safe
conversion adds an \lstinline$_Atomic$ qualifier, so the function in point 2 is not applicable. A
separate overloading of \lstinline$?++$ is required.

\item
``\lstinline$char const volatile * pq; pq++$''. Here the result has type
\lstinline$char const volatile *$, so a new overloading is needed:
\begin{lstlisting}
forall( type T ) T const volatile * ?++( T const volatile *restrict volatile * );
\end{lstlisting}
One overloading is needed for each combination of qualifiers in the pointed-at
type\index{deficiencies!pointers to qualified types}.
 
\item
``\lstinline$float *restrict * prp; prp++$''. The \lstinline$restrict$ qualifier is handled just
like \lstinline$const$ and \lstinline$volatile$ in the previous case:
\begin{lstlisting}
forall( type T ) T restrict * ?++( T restrict *restrict volatile * );
\end{lstlisting}
with \lstinline$T$ inferred to be \lstinline$float *$. This looks odd, because {\c11} contains a
constraint that requires restrict-qualified types to be pointer-to-object types, and \lstinline$T$
is not syntactically a pointer type. \CFA loosens the constraint.
\end{enumerate}
\end{rationale}


\subsubsection{Compound literals}

\semantics 
A compound literal has one interpretation, with the type given by the \nonterm{type-name} of the
compound literal.


\subsection{Unary operators}

\begin{syntax}
\lhs{unary-expression}
\rhs \nonterm{postfix-expression}
\rhs \lstinline$++$ \nonterm{unary-expression}
\rhs \lstinline$--$ \nonterm{unary-expression}
\rhs \nonterm{unary-operator} \nonterm{cast-expression}
\rhs \lstinline$sizeof$ \nonterm{unary-expression}
\rhs \lstinline$sizeof$ \lstinline$($ \nonterm{type-name} \lstinline$)$
\lhs{unary-operator} one of \rhs \lstinline$&$ \lstinline$*$ \lstinline$+$ \lstinline$-$ \lstinline$~$ \lstinline$!$
\end{syntax}

\rewriterules
\begin{lstlisting}
*a      @\rewrite@ *?(a) @\use{*?}@
+a      @\rewrite@ +?(a) @\use{+?}@
-a      @\rewrite@ -?(a) @\use{-?}@
~a      @\rewrite@ ~?(a) @\use{~?}@
!a      @\rewrite@ !?(a) @\use{"!?}@
++a     @\rewrite@ ++?(&(a)) @\use{++?}@
--a     @\rewrite@ --?(&(a)) @\use{--?}@
\end{lstlisting}


\subsubsection{Prefix increment and decrement operators}

\begin{lstlisting}
_Bool ++?( volatile _Bool * ),
        ++?( _Atomic volatile _Bool * );
char ++?( volatile char * ),
        ++?( _Atomic volatile char * );
signed char ++?( volatile signed char * ),
        ++?( _Atomic volatile signed char * );
unsigned char ++?( volatile signed char * ),
        ++?( _Atomic volatile signed char * );
short int ++?( volatile short int * ),
        ++?( _Atomic volatile short int * );
unsigned short int ++?( volatile unsigned short int * ),
        ++?( _Atomic volatile unsigned short int * );
int ++?( volatile int * ),
        ++?( _Atomic volatile int * );
unsigned int ++?( volatile unsigned int * ),
        ++?( _Atomic volatile unsigned int * );
long int ++?( volatile long int * ),
        ++?( _Atomic volatile long int * );
long unsigned int ++?( volatile long unsigned int * ),
        ++?( _Atomic volatile long unsigned int * );
long long int ++?( volatile long long int * ),
        ++?( _Atomic volatile long long int * );
long long unsigned ++?( volatile long long unsigned int * ),
        ++?( _Atomic volatile long long unsigned int * );
float ++?( volatile float * ),
        ++?( _Atomic volatile float * );
double ++?( volatile double * ),
        ++?( _Atomic volatile double * );
long double ++?( volatile long double * ),
        ++?( _Atomic volatile long double * );

forall( type T ) T * ++?( T * restrict volatile * ),
        * ++?( T * _Atomic restrict volatile * );

forall( type T ) _Atomic T * ++?( _Atomic T * restrict volatile * ),
        * ++?( _Atomic T * _Atomic restrict volatile * );

forall( type T ) const T * ++?( const T * restrict volatile * ),
        * ++?( const T * _Atomic restrict volatile * );

forall( type T ) volatile T * ++?( volatile T * restrict volatile * ),
        * ++?( volatile T * _Atomic restrict volatile * );

forall( type T ) restrict T * ++?( restrict T * restrict volatile * ),
        * ++?( restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic const T * ++?( _Atomic const T * restrict volatile * ),
        * ++?( _Atomic const T * _Atomic restrict volatile * );

forall( type T ) _Atomic volatile T * ++?( _Atomic volatile T * restrict volatile * ),
        * ++?( _Atomic volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict T * ++?( _Atomic restrict T * restrict volatile * ),
        * ++?( _Atomic restrict T * _Atomic restrict volatile * );

forall( type T ) const volatile T * ++?( const volatile T * restrict volatile * ),
        * ++?( const volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict T * ++?( const restrict T * restrict volatile * ),
        * ++?( const restrict T * _Atomic restrict volatile * );

forall( type T ) restrict volatile T * ++?( restrict volatile T * restrict volatile * ),
        * ++?( restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const volatile T * ++?( _Atomic const volatile T * restrict volatile * ),
        * ++?( _Atomic const volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict T * ++?( _Atomic const restrict T * restrict volatile * ),
        * ++?( _Atomic const restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict volatile T * ++?( _Atomic restrict volatile T * restrict volatile * ),
        * ++?( _Atomic restrict volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict volatile T * ++?( const restrict volatile T * restrict volatile * ),
        * ++?( const restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict volatile T * ++?( _Atomic const restrict volatile T * restrict volatile * ),
        * ++?( _Atomic const restrict volatile T * _Atomic restrict volatile * );

_Bool --?( volatile _Bool * ),
        --?( _Atomic volatile _Bool * );
char --?( volatile char * ),
        --?( _Atomic volatile char * );
signed char --?( volatile signed char * ),
        --?( _Atomic volatile signed char * );
unsigned char --?( volatile signed char * ),
        --?( _Atomic volatile signed char * );
short int --?( volatile short int * ),
        --?( _Atomic volatile short int * );
unsigned short int --?( volatile unsigned short int * ),
        --?( _Atomic volatile unsigned short int * );
int --?( volatile int * ),
        --?( _Atomic volatile int * );
unsigned int --?( volatile unsigned int * ),
        --?( _Atomic volatile unsigned int * );
long int --?( volatile long int * ),
        --?( _Atomic volatile long int * );
long unsigned int --?( volatile long unsigned int * ),
        --?( _Atomic volatile long unsigned int * );
long long int --?( volatile long long int * ),
        --?( _Atomic volatile long long int * );
long long unsigned --?( volatile long long unsigned int * ),
        --?( _Atomic volatile long long unsigned int * );
float --?( volatile float * ),
        --?( _Atomic volatile float * );
double --?( volatile double * ),
        --?( _Atomic volatile double * );
long double --?( volatile long double * ),
        --?( _Atomic volatile long double * );

forall( type T ) T * --?( T * restrict volatile * ),
        * --?( T * _Atomic restrict volatile * );

forall( type T ) _Atomic T * --?( _Atomic T * restrict volatile * ),
        * --?( _Atomic T * _Atomic restrict volatile * );

forall( type T ) const T * --?( const T * restrict volatile * ),
        * --?( const T * _Atomic restrict volatile * );

forall( type T ) volatile T * --?( volatile T * restrict volatile * ),
        * --?( volatile T * _Atomic restrict volatile * );

forall( type T ) restrict T * --?( restrict T * restrict volatile * ),
        * --?( restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic const T * --?( _Atomic const T * restrict volatile * ),
        * --?( _Atomic const T * _Atomic restrict volatile * );

forall( type T ) _Atomic volatile T * --?( _Atomic volatile T * restrict volatile * ),
        * --?( _Atomic volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict T * --?( _Atomic restrict T * restrict volatile * ),
        * --?( _Atomic restrict T * _Atomic restrict volatile * );

forall( type T ) const volatile T * --?( const volatile T * restrict volatile * ),
        * --?( const volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict T * --?( const restrict T * restrict volatile * ),
        * --?( const restrict T * _Atomic restrict volatile * );

forall( type T ) restrict volatile T * --?( restrict volatile T * restrict volatile * ),
        * --?( restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const volatile T * --?( _Atomic const volatile T * restrict volatile * ),
        * --?( _Atomic const volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict T * --?( _Atomic const restrict T * restrict volatile * ),
        * --?( _Atomic const restrict T * _Atomic restrict volatile * );

forall( type T ) _Atomic restrict volatile T * --?( _Atomic restrict volatile T * restrict volatile * ),
        * --?( _Atomic restrict volatile T * _Atomic restrict volatile * );

forall( type T ) const restrict volatile T * --?( const restrict volatile T * restrict volatile * ),
        * --?( const restrict volatile T * _Atomic restrict volatile * );

forall( type T ) _Atomic const restrict volatile T * --?( _Atomic const restrict volatile T * restrict volatile * ),
        * --?( _Atomic const restrict volatile T * _Atomic restrict volatile * );
\end{lstlisting}
For every extended integer type \lstinline$X$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X       ++?( volatile X * ),
        ++?( _Atomic volatile X * ),
        --?( volatile X * ),
        --?( _Atomic volatile X * );
\end{lstlisting}
For every complete enumerated type \lstinline$E$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
E ++?( volatile E * ),
        ++?( _Atomic volatile E * ),
        ?--( volatile E * ),
        ?--( _Atomic volatile E * );
\end{lstlisting}

\semantics
The interpretations of prefix increment and decrement expressions are
determined in the same way as the interpretations of postfix increment and
decrement expressions.


\subsubsection{Address and indirection operators}

\begin{lstlisting}
forall( type T ) lvalue T *?( T * );
forall( type T ) _Atomic lvalue T *?( _Atomic T * );
forall( type T ) const lvalue T *?( const T * );
forall( type T ) volatile lvalue T *?( volatile T * );
forall( type T ) restrict lvalue T *?( restrict T * );
forall( type T ) _Atomic const lvalue T *?( _Atomic const T * );
forall( type T ) _Atomic volatile lvalue T *?( _Atomic volatile T * );
forall( type T ) _Atomic restrict lvalue T *?( _Atomic restrict T * );
forall( type T ) const volatile lvalue T *?( const volatile T * );
forall( type T ) const restrict lvalue T *?( const restrict T * );
forall( type T ) restrict volatile lvalue T *?( restrict volatile T * );
forall( type T ) _Atomic const volatile lvalue T *?( _Atomic const volatile T * );
forall( type T ) _Atomic const restrict lvalue T *?( _Atomic const restrict T * );
forall( type T ) _Atomic restrict volatile lvalue T *?( _Atomic restrict volatile T * );
forall( type T ) const restrict volatile lvalue T *?( const restrict volatile T * );
forall( type T ) _Atomic const restrict volatile lvalue T *?( _Atomic const restrict volatile T * );

forall( ftype FT ) FT *?( FT * );
\end{lstlisting}

\constraints
The operand of the unary ``\lstinline$&$'' operator shall have exactly one
interpretation\index{ambiguous interpretation}\index{interpretations}, which shall be unambiguous.

\semantics
The ``\lstinline$&$'' expression has one interpretation which is of type \lstinline$T *$, where
\lstinline$T$ is the type of the operand.

The interpretations of an indirection expression are the interpretations of the corresponding
function call.


\subsubsection{Unary arithmetic operators}

\begin{lstlisting}
int
        +?( int ),
        -?( int ),
        ~?( int );
unsigned int
        +?( unsigned int ),
        -?( unsigned int ),
         ~?( unsigned int );
long int
        +?( long int ),
        -?( long int ),
        ~?( long int );
long unsigned int
        +?( long unsigned int ),
        -?( long unsigned int ),
        ~?( long unsigned int );
long long int
        +?( long long int ),
        -?( long long int ),
        ~?( long long int );
long long unsigned int
        +?( long long unsigned int ),
        -?( long long unsigned int ),
        ~?( long long unsigned int );
float
        +?( float ),
        -?( float );
double
        +?( double ),
        -?( double );
long double
        +?( long double ),
        -?( long double );
_Complex float
        +?( _Complex float ),
        -?( _Complex float );
_Complex double
        +?( _Complex double ),
        -?( _Complex double );
_Complex long double
        +?( _Complex long double ),
        -?( _Complex long double );

int !?( int ),
        !?( unsigned int ),
        !?( long ),
        !?( long unsigned int ),
        !?( long long int ),
        !?( long long unsigned int ),
        !?( float ),
        !?( double ),
        !?( long double ),
        !?( _Complex float ),
        !?( _Complex double ),
        !?( _Complex long double );

forall( dtype DT ) int !?( const restrict volatile DT * );
forall( dtype DT ) int !?( _Atomic const restrict volatile DT * );
forall( ftype FT ) int !?( FT * );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X +?( X ), -?( X ), ~?( X );
int !?( X );
\end{lstlisting}

\semantics
The interpretations of a unary arithmetic expression are the interpretations of the corresponding
function call.

\examples
\begin{lstlisting}
long int li;
void eat_double( double );@\use{eat_double}@

eat_double(-li ); // @\rewrite@ eat_double( -?( li ) ); 
\end{lstlisting}
The valid interpretations of ``\lstinline$-li$'' (assuming no extended integer types exist) are
\begin{center}
\begin{tabular}{llc}
interpretation & result type & expression conversion cost \\
\hline
\lstinline$-?( (int)li )$                                       & \lstinline$int$                                       & (unsafe) \\
\lstinline$-?( (unsigned)li)$                           & \lstinline$unsigned int$                      & (unsafe) \\
\lstinline$-?( (long)li)$                                       & \lstinline$long$                                      & 0 \\
\lstinline$-?( (long unsigned int)li)$          & \lstinline$long unsigned int$         & 1 \\
\lstinline$-?( (long long int)li)$                      & \lstinline$long long int$                     & 2 \\
\lstinline$-?( (long long unsigned int)li)$     & \lstinline$long long unsigned int$& 3 \\
\lstinline$-?( (float)li)$                                      & \lstinline$float$                                     & 4 \\
\lstinline$-?( (double)li)$                                     & \lstinline$double$                            & 5 \\
\lstinline$-?( (long double)li)$                        & \lstinline$long double$                       & 6 \\
\lstinline$-?( (_Complex float)li)$                     & \lstinline$float$                                     & (unsafe) \\
\lstinline$-?( (_Complex double)li)$            & \lstinline$double$                            & (unsafe) \\
\lstinline$-?( (_Complex long double)li)$       & \lstinline$long double$                       & (unsafe) \\
\end{tabular}
\end{center}
The valid interpretations of the \lstinline$eat_double$ call, with the cost of the argument
conversion and the cost of the entire expression, are
\begin{center}
\begin{tabular}{lcc}
interpretation & argument cost & expression cost \\
\hline
\lstinline$eat_double( (double)-?( (int)li) )$                                  & 7                     & (unsafe) \\
\lstinline$eat_double( (double)-?( (unsigned)li) )$                             & 6                     & (unsafe) \\
\lstinline$eat_double( (double)-?(li) )$                                                & 5                     & \(0+5=5\) \\
\lstinline$eat_double( (double)-?( (long unsigned int)li) )$    & 4                     & \(1+4=5\) \\
\lstinline$eat_double( (double)-?( (long long int)li) )$                & 3                     & \(2+3=5\) \\
\lstinline$eat_double( (double)-?( (long long unsigned int)li) )$& 2            & \(3+2=5\) \\
\lstinline$eat_double( (double)-?( (float)li) )$                                & 1                     & \(4+1=5\) \\
\lstinline$eat_double( (double)-?( (double)li) )$                               & 0                     & \(5+0=5\) \\
\lstinline$eat_double( (double)-?( (long double)li) )$                  & (unsafe)      & (unsafe) \\
\lstinline$eat_double( (double)-?( (_Complex float)li) )$               & (unsafe)      & (unsafe) \\
\lstinline$eat_double( (double)-?( (_Complex double)li) )$              & (unsafe)      & (unsafe) \\
\lstinline$eat_double( (double)-?( (_Complex long double)li) )$ & (unsafe)      & (unsafe) \\
\end{tabular}
\end{center}
Each has result type \lstinline$void$, so the best must be selected. The interpretations involving
unsafe conversions are discarded. The remainder have equal expression conversion costs, so the
``highest argument conversion cost'' rule is invoked, and the chosen interpretation is
\lstinline$eat_double( (double)-?(li) )$.


\subsubsection{The {\tt sizeof} and {\tt \_Alignof} operators}

\constraints
The operand of \lstinline$sizeof$ or \lstinline$_Alignof$ shall not be \lstinline$type$,
\lstinline$dtype$, or \lstinline$ftype$.

When the \lstinline$sizeof$\use{sizeof} operator is applied to an expression, the expression shall
have exactly one interpretation\index{ambiguous interpretation}\index{interpretations}, which shall
be unambiguous. \semantics A \lstinline$sizeof$ or \lstinline$_Alignof$ expression has one
interpretation, of type \lstinline$size_t$.

When \lstinline$sizeof$ is applied to an identifier declared by a \nonterm{type-declaration} or a
\nonterm{type-parameter}, it yields the size in bytes of the type that implements the operand. When
the operand is an opaque type or an inferred type parameter\index{inferred parameter}, the
expression is not a constant expression.

When \lstinline$_Alignof$ is applied to an identifier declared by a \nonterm{type-declaration} or a
\nonterm{type-parameter}, it yields the alignment requirement of the type that implements the
operand. When the operand is an opaque type or an inferred type parameter\index{inferred
parameter}, the expression is not a constant expression.
\begin{rationale}
\begin{lstlisting}
type Pair = struct { int first, second; };
size_t p_size = sizeof(Pair);           // constant expression 

extern type Rational;@\use{Rational}@
size_t c_size = sizeof(Rational);       // non-constant expression 

forall(type T) T f(T p1, T p2) {
        size_t t_size = sizeof(T);              // non-constant expression 
        ...
}
\end{lstlisting}
``\lstinline$sizeof Rational$'', although not statically known, is fixed. Within \lstinline$f()$,
``\lstinline$sizeof(T)$'' is fixed for each call of \lstinline$f()$, but may vary from call to call.
\end{rationale}

\subsection{Cast operators}
\begin{syntax}
\lhs{cast-expression}
\rhs \nonterm{unary-expression}
\rhs \lstinline$($ \nonterm{type-name} \lstinline$)$ \nonterm{cast-expression}
\end{syntax}

\constraints
The \nonterm{type-name} in a \nonterm{cast-expression} shall not be \lstinline$type$,
\lstinline$dtype$, or \lstinline$ftype$.

\semantics

In a cast expression\index{cast expression} ``\lstinline$($\nonterm{type-name}\lstinline$)e$'', if
\nonterm{type-name} is the type of an interpretation of \lstinline$e$, then that interpretation is the
only interpretation of the cast expression; otherwise, \lstinline$e$ shall have some interpretation that
can be converted to \nonterm{type-name}, and the interpretation of the cast expression is the cast
of the interpretation that can be converted at the lowest cost. The cast expression's interpretation
is ambiguous\index{ambiguous interpretation} if more than one interpretation can be converted at the
lowest cost or if the selected interpretation is ambiguous.

\begin{rationale}
Casts can be used to eliminate ambiguity in expressions by selecting interpretations of
subexpressions, and to specialize polymorphic functions and values.
\end{rationale}

\subsection{Multiplicative operators}
\begin{syntax}
\lhs{multiplicative-expression}
\rhs \nonterm{cast-expression}
\rhs \nonterm{multiplicative-expression} \lstinline$*$ \nonterm{cast-expression}
\rhs \nonterm{multiplicative-expression} \lstinline$/$ \nonterm{cast-expression}
\rhs \nonterm{multiplicative-expression} \lstinline$%$ \nonterm{cast-expression}
\end{syntax}

\rewriterules
\begin{lstlisting}
a * b @\rewrite@ ?*?( a, b )@\use{?*?}@
a / b @\rewrite@ ?/?( a, b )@\use{?/?}@
a % b @\rewrite@ ?%?( a, b )@\use{?%?}@
\end{lstlisting}

\begin{lstlisting}
int?*?( int, int ),
        ?/?( int, int ),
        ?%?( int, int );
unsigned int?*?( unsigned int, unsigned int ),
        ?/?( unsigned int, unsigned int ),
        ?%?( unsigned int, unsigned int );
long int?*?( long int, long int ),
        ?/?( long, long ),
        ?%?( long, long );
long unsigned int?*?( long unsigned int, long unsigned int ),
        ?/?( long unsigned int, long unsigned int ),
        ?%?( long unsigned int, long unsigned int );
long long int?*?( long long int, long long int ),
        ?/?( long long int, long long int ),
        ?%?( long long int, long long int );
long long unsigned int ?*?( long long unsigned int, long long unsigned int ),
        ?/?( long long unsigned int, long long unsigned int ),
        ?%?( long long unsigned int, long long unsigned int );
float?*?( float, float ),
        ?/?( float, float );
double?*?( double, double ),
        ?/?( double, double );
long double?*?( long double, long double ),
        ?/?( long double, long double );
_Complex float?*?( float, _Complex float ),
        ?/?( float, _Complex float ),
        ?*?( _Complex float, float ),
        ?/?( _Complex float, float ),
        ?*?( _Complex float, _Complex float ),
        ?/?( _Complex float, _Complex float );
_Complex double?*?( double, _Complex double ),
        ?/?( double, _Complex double ),
        ?*?( _Complex double, double ),
        ?/?( _Complex double, double ),
        ?*?( _Complex double, _Complex double ),
        ?/?( _Complex double, _Complex double );
_Complex long double?*?( long double, _Complex long double ),
        ?/?( long double, _Complex long double ),
        ?*?( _Complex long double, long double ),
        ?/?( _Complex long double, long double ),
        ?*?( _Complex long double, _Complex long double ),
        ?/?( _Complex long double, _Complex long double );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X ?*?( X ), ?/?( X ), ?%?( X );
\end{lstlisting}

\begin{rationale}
{\c11} does not include conversions from the real types\index{real type} to complex
types\index{complex type} in the usual arithmetic conversions\index{usual arithmetic conversions}.
Instead it specifies conversion of the result of binary operations on arguments from mixed type
domains. \CFA's predefined operators match that pattern.
\end{rationale}

\semantics
The interpretations of multiplicative expressions are the interpretations of the corresponding
function call.

\examples
\begin{lstlisting}
int i;
long li;
void eat_double( double );@\use{eat_double}@
eat_double( li % i );
\end{lstlisting}
``\lstinline$li % i$'' is rewritten as ``\lstinline$?%?(li, i )$''. The valid interpretations
of \lstinline$?%?(li, i )$, the cost\index{conversion cost} of converting their arguments, and
the cost of converting the result to \lstinline$double$ (assuming no extended integer types are
present ) are
\begin{center}
\begin{tabular}{lcc}
interpretation & argument cost & result cost \\
\hline 
\lstinline$ ?%?( (int)li, i )$                                                                          & (unsafe)      & 6     \\
\lstinline$ ?%?( (unsigned)li,(unsigned)i )$                                            & (unsafe)      & 5     \\
\lstinline$ ?%?(li,(long)i )$                                                                           & 1                     & 4     \\
\lstinline$ ?%?( (long unsigned)li,(long unsigned)i )$                          & 3                     & 3     \\
\lstinline$ ?%?( (long long)li,(long long)i )$                                          & 5                     & 2     \\
\lstinline$ ?%?( (long long unsigned)li, (long long unsigned)i )$       & 7                     & 1     \\
\end{tabular}
\end{center}
The best interpretation of \lstinline$eat_double( li, i )$ is
\lstinline$eat_double( (double)?%?(li, (long)i ))$, which has no unsafe conversions and the
lowest total cost.

\begin{rationale}
{\c11} defines most arithmetic operations to apply an integer promotion\index{integer promotions} to
any argument that belongs to a type that has an integer conversion rank\index{integer conversion
 rank} less than that of \lstinline$int$.If \lstinline$s$ is a \lstinline$short int$,
``\lstinline$s *s$'' does not have type \lstinline$short int$; it is treated as
``\lstinline$( (int)s ) * ( (int)s )$'', and has type \lstinline$int$. \CFA matches that pattern;
it does not predefine ``\lstinline$short ?*?( short, short )$''.

These ``missing'' operators limit polymorphism. Consider
\begin{lstlisting}
forall( type T | T ?*?( T, T ) ) T square( T );
short s;
square( s );
\end{lstlisting}
Since \CFA does not define a multiplication operator for \lstinline$short int$,
\lstinline$square( s )$ is treated as \lstinline$square( (int)s )$, and the result has type
\lstinline$int$. This is mildly surprising, but it follows the {\c11} operator pattern.

A more troubling example is
\begin{lstlisting}
forall( type T | ?*?( T, T ) ) T product( T[], int n );
short sa[5];
product( sa, 5);
\end{lstlisting}
This has no valid interpretations, because \CFA has no conversion from ``array of
\lstinline$short int$'' to ``array of \lstinline$int$''. The alternatives in such situations
include
\begin{itemize}
\item
Defining monomorphic overloadings of \lstinline$product$ for \lstinline$short$ and the other
``small'' types.
\item
Defining ``\lstinline$short ?*?( short, short )$'' within the scope containing the call to
\lstinline$product$.
\item
Defining \lstinline$product$ to take as an argument a conversion function from the ``small'' type to
the operator's argument type.
\end{itemize}
\end{rationale}


\subsection{Additive operators}

\begin{syntax}
\lhs{additive-expression}
\rhs \nonterm{multiplicative-expression}
\rhs \nonterm{additive-expression} \lstinline$+$ \nonterm{multiplicative-expression}
\rhs \nonterm{additive-expression} \lstinline$-$ \nonterm{multiplicative-expression}
\end{syntax}

\rewriterules
\begin{lstlisting}
a + b @\rewrite@ ?+?( a, b )@\use{?+?}@
a - b @\rewrite@ ?-?( a, b )@\use{?-?}@
\end{lstlisting}

\begin{lstlisting}
int?+?( int, int ),
        ?-?( int, int );
unsigned int?+?( unsigned int, unsigned int ),
        ?-?( unsigned int, unsigned int );
long int?+?( long int, long int ),
        ?-?( long int, long int );
long unsigned int?+?( long unsigned int, long unsigned int ),
        ?-?( long unsigned int, long unsigned int );
long long int?+?( long long int, long long int ),
        ?-?( long long int, long long int );
long long unsigned int ?+?( long long unsigned int, long long unsigned int ),
        ?-?( long long unsigned int, long long unsigned int );
float?+?( float, float ),
        ?-?( float, float );
double?+?( double, double ),
        ?-?( double, double );
long double?+?( long double, long double ),
        ?-?( long double, long double );
_Complex float?+?( _Complex float, float ),
        ?-?( _Complex float, float ),
        ?+?( float, _Complex float ),
        ?-?( float, _Complex float ),
        ?+?( _Complex float, _Complex float ),
        ?-?( _Complex float, _Complex float );
_Complex double?+?( _Complex double, double ),
        ?-?( _Complex double, double ),
        ?+?( double, _Complex double ),
        ?-?( double, _Complex double ),
        ?+?( _Complex double, _Complex double ),
        ?-?( _Complex double, _Complex double );
_Complex long double?+?( _Complex long double, long double ),
        ?-?( _Complex long double, long double ),
        ?+?( long double, _Complex long double ),
        ?-?( long double, _Complex long double ),
        ?+?( _Complex long double, _Complex long double ),
        ?-?( _Complex long double, _Complex long double );

forall( type T ) T
        * ?+?( T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, T * ),
        * ?-?( T *, ptrdiff_t );

forall( type T ) _Atomic T
        * ?+?( _Atomic T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, _Atomic T * ),
        * ?-?( _Atomic T *, ptrdiff_t );

forall( type T ) const T
        * ?+?( const T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, const T * ),
        * ?-?( const T *, ptrdiff_t );

forall( type T ) restrict T
        * ?+?( restrict T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, restrict T * ),
        * ?-?( restrict T *, ptrdiff_t );

forall( type T ) volatile T
        * ?+?( volatile T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, volatile T * ),
        * ?-?( volatile T *, ptrdiff_t );

forall( type T ) _Atomic const T
        * ?+?( _Atomic const T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, _Atomic const T * ),
        * ?-?( _Atomic const T *, ptrdiff_t );

forall( type T ) _Atomic restrict T
        * ?+?( _Atomic restrict T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, _Atomic restrict T * ),
        * ?-?( _Atomic restrict T *, ptrdiff_t );

forall( type T ) _Atomic volatile T
        * ?+?( _Atomic volatile T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, _Atomic volatile T * ),
        * ?-?( _Atomic volatile T *, ptrdiff_t );

forall( type T ) const restrict T
        * ?+?( const restrict T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, const restrict T * ),
        * ?-?( const restrict T *, ptrdiff_t );

forall( type T ) const volatile T
        * ?+?( const volatile T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, const volatile T * ),
        * ?-?( const volatile T *, ptrdiff_t );

forall( type T ) restrict volatile T
        * ?+?( restrict volatile T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, restrict volatile T * ),
        * ?-?( restrict volatile T *, ptrdiff_t );

forall( type T ) _Atomic const restrict T
        * ?+?( _Atomic const restrict T *, ptrdiff_t ),
        * ?+?( ptrdiff_t, _Atomic const restrict T * ),
        * ?-?( _Atomic const restrict T *, ptrdiff_t );

forall( type T ) ptrdiff_t
        * ?-?( const restrict volatile T *, const restrict volatile T * ),
        * ?-?( _Atomic const restrict volatile T *, _Atomic const restrict volatile T * );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank
\index{integer conversion rank}greater than the rank of \lstinline$int$ there
exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X ?+?( X ), ?-?( X );
\end{lstlisting}

\semantics
The interpretations of additive expressions are the interpretations of the corresponding function
calls.

\begin{rationale}
\lstinline$ptrdiff_t$ is an implementation-defined identifier defined in \lstinline$<stddef.h>$ that
is synonymous with a signed integral type that is large enough to hold the difference between two
pointers. It seems reasonable to use it for pointer addition as well. (This is technically a
difference between \CFA and C, which only specifies that pointer addition uses an \emph{integral}
argument.) Hence it is also used for subscripting, which is defined in terms of pointer addition.
The {\c11} standard uses \lstinline$size_t$ in several cases where a library function takes an
argument that is used as a subscript, but \lstinline$size_t$ is unsuitable here because it is an
unsigned type.
\end{rationale}


\subsection{Bitwise shift operators}

\begin{syntax}
\lhs{shift-expression}
\rhs \nonterm{additive-expression}
\rhs \nonterm{shift-expression} \lstinline$<<$ \nonterm{additive-expression}
\rhs \nonterm{shift-expression} \lstinline$>>$ \nonterm{additive-expression}
\end{syntax}

\rewriterules \use{?>>?}%use{?<<?}
\begin{lstlisting}
a << b @\rewrite@ ?<<?( a, b )
a >> b @\rewrite@ ?>>?( a, b )
\end{lstlisting}

\begin{lstlisting}
int ?<<?( int, int ),
         ?>>?( int, int );
unsigned int ?<<?( unsigned int, int ),
         ?>>?( unsigned int, int );
long int ?<<?( long int, int ),
         ?>>?( long int, int );
long unsigned int ?<<?( long unsigned int, int ),
         ?>>?( long unsigned int, int );
long long int ?<<?( long long int, int ),
         ?>>?( long long int, int );
long long unsigned int ?<<?( long long unsigned int, int ),
         ?>>?( long long unsigned int, int);
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
 rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X ?<<?( X, int ), ?>>?( X, int );
\end{lstlisting}

\begin{rationale}
The bitwise shift operators break the usual pattern: they do not convert both operands to a common
type. The right operand only undergoes integer promotion\index{integer promotion}.
\end{rationale}

\semantics
The interpretations of a bitwise shift expression are the interpretations of the corresponding
function calls.


\subsection{Relational operators}

\begin{syntax}
\lhs{relational-expression}
\rhs \nonterm{shift-expression}
\rhs \nonterm{relational-expression} \lstinline$< $ \nonterm{shift-expression}
\rhs \nonterm{relational-expression} \lstinline$> $ \nonterm{shift-expression}
\rhs \nonterm{relational-expression} \lstinline$<=$ \nonterm{shift-expression}
\rhs \nonterm{relational-expression} \lstinline$>=$ \nonterm{shift-expression}
\end{syntax}

\rewriterules\use{?>?}\use{?>=?}%use{?<?}%use{?<=?}
\begin{lstlisting}
a < b @\rewrite@ ?<?( a, b )
a > b @\rewrite@ ?>?( a, b )
a <= b @\rewrite@ ?<=?( a, b )
a >= b @\rewrite@ ?>=?( a, b )
\end{lstlisting}

\begin{lstlisting}
int ?<?( int, int ),
        ?<=?( int, int ),
        ?>?( int, int ),
        ?>=?( int, int );
int ?<?( unsigned int, unsigned int ),
        ?<=?( unsigned int, unsigned int ),
        ?>?( unsigned int, unsigned int ),
        ?>=?( unsigned int, unsigned int );
int ?<?( long int, long int ),
        ?<=?( long int, long int ),
        ?>?( long int, long int ),
        ?>=?( long int, long int );
int ?<?( long unsigned int, long unsigned ),
        ?<=?( long unsigned int, long unsigned ),
        ?>?( long unsigned int, long unsigned ),
        ?>=?( long unsigned int, long unsigned );
int ?<?( long long int, long long int ),
        ?<=?( long long int, long long int ),
        ?>?( long long int, long long int ),
        ?>=?( long long int, long long int );
int ?<?( long long unsigned int, long long unsigned ),
        ?<=?( long long unsigned int, long long unsigned ),
        ?>?( long long unsigned int, long long unsigned ),
        ?>=?( long long unsigned int, long long unsigned );
int ?<?( float, float ),
        ?<=?( float, float ),
        ?>?( float, float ),
        ?>=?( float, float );
int ?<?( double, double ),
        ?<=?( double, double ),
        ?>?( double, double ),
        ?>=?( double, double );
int ?<?( long double, long double ),
        ?<=?( long double, long double ),
        ?>?( long double, long double ),
        ?>=?( long double, long double );

forall( dtype DT ) int
        ?<?( const restrict volatile DT *, const restrict volatile DT * ),
        ?<?( _Atomic const restrict volatile DT *, _Atomic const restrict volatile DT * ),
        ?<=?( const restrict volatile DT *, const restrict volatile DT * ),
        ?<=?( _Atomic const restrict volatile DT *, _Atomic const restrict volatile DT * ),
        ?>?( const restrict volatile DT *, const restrict volatile DT * ),
        ?>?( _Atomic const restrict volatile DT *, _Atomic const restrict volatile DT * ),
        ?>=?( const restrict volatile DT *, const restrict volatile DT * ),
        ?>=?( _Atomic const restrict volatile DT *, _Atomic const restrict volatile DT * );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
 rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
int ?<?( X, X ),
        ?<=?( X, X ),
        ?<?( X, X ),
        ?>=?( X, X );
\end{lstlisting}

\semantics
The interpretations of a relational expression are the interpretations of the corresponding function
call.


\subsection{Equality operators}

\begin{syntax}
\lhs{equality-expression}
\rhs \nonterm{relational-expression}
\rhs \nonterm{equality-expression} \lstinline$==$ \nonterm{relational-expression}
\rhs \nonterm{equality-expression} \lstinline$!=$ \nonterm{relational-expression}
\end{syntax}

\rewriterules
\begin{lstlisting}
a == b @\rewrite@ ?==?( a, b )@\use{?==?}@
a != b @\rewrite@ ?!=?( a, b )@\use{?"!=?}@
\end{lstlisting}

\begin{lstlisting}
int ?==?( int, int ),
        ?!=?( int, int ),
        ?==?( unsigned int, unsigned int ),
        ?!=?( unsigned int, unsigned int ),
        ?==?( long int, long int ),
        ?!=?( long int, long int ),
        ?==?( long unsigned int, long unsigned int ),
        ?!=?( long unsigned int, long unsigned int ),
        ?==?( long long int, long long int ),
        ?!=?( long long int, long long int ),
        ?==?( long long unsigned int, long long unsigned int ),
        ?!=?( long long unsigned int, long long unsigned int ),
        ?==?( float, float ),
        ?!=?( float, float ),
        ?==?( _Complex float, float ),
        ?!=?( _Complex float, float ),
        ?==?( float, _Complex float ),
        ?!=?( float, _Complex float ),
        ?==?( _Complex float, _Complex float ),
        ?!=?( _Complex float, _Complex float ),
        ?==?( double, double ),
        ?!=?( double, double ),
        ?==?( _Complex double, double ),
        ?!=?( _Complex double, double ),
        ?==?( double, _Complex double ),
        ?!=?( double, _Complex double ),
        ?==?( _Complex double, _Complex double ),
        ?!=?( _Complex double, _Complex double ),
        ?==?( long double, long double ),
        ?!=?( long double, long double ),
        ?==?( _Complex long double, long double ),
        ?!=?( _Complex long double, long double ),
        ?==?( long double, _Complex long double ),
        ?!=?( long double, _Complex long double ),
        ?==?( _Complex long double, _Complex long double ),
        ?!=?( _Complex long double, _Complex long double );

forall( dtype DT ) int
        ?==?( const restrict volatile DT *, const restrict volatile DT * ),
        ?!=?( const restrict volatile DT *, const restrict volatile DT * ),
        ?==?( const restrict volatile DT *, const restrict volatile void * ),
        ?!=?( const restrict volatile DT *, const restrict volatile void * ),
        ?==?( const restrict volatile void *, const restrict volatile DT * ),
        ?!=?( const restrict volatile void *, const restrict volatile DT * ),
        ?==?( const restrict volatile DT *, forall( dtype DT2) const DT2 * ),
        ?!=?( const restrict volatile DT *, forall( dtype DT2) const DT2 * ),
        ?==?( forall( dtype DT2) const DT2*, const restrict volatile DT * ),
        ?!=?( forall( dtype DT2) const DT2*, const restrict volatile DT * ),
        ?==?( forall( dtype DT2) const DT2*, forall( dtype DT3) const DT3 * ),
        ?!=?( forall( dtype DT2) const DT2*, forall( dtype DT3) const DT3 * ),

        ?==?( _Atomic const restrict volatile DT *, _Atomic const restrict volatile DT * ),
        ?!=?( _Atomic const restrict volatile DT *, _Atomic const restrict volatile DT * ),
        ?==?( _Atomic const restrict volatile DT *, const restrict volatile void * ),
        ?!=?( _Atomic const restrict volatile DT *, const restrict volatile void * ),
        ?==?( const restrict volatile void *, _Atomic const restrict volatile DT * ),
        ?!=?( const restrict volatile void *, _Atomic const restrict volatile DT * ),
        ?==?( _Atomic const restrict volatile DT *, forall( dtype DT2) const DT2 * ),
        ?!=?( _Atomic const restrict volatile DT *, forall( dtype DT2) const DT2 * ),
        ?==?( forall( dtype DT2) const DT2*, _Atomic const restrict volatile DT * ),
        ?!=?( forall( dtype DT2) const DT2*, _Atomic const restrict volatile DT * );

forall( ftype FT ) int
        ?==?( FT *, FT * ),
        ?!=?( FT *, FT * ),
        ?==?( FT *, forall( ftype FT2) FT2 * ),
        ?!=?( FT *, forall( ftype FT2) FT2 * ),
        ?==?( forall( ftype FT2) FT2*, FT * ),
        ?!=?( forall( ftype FT2) FT2*, FT * ),
        ?==?( forall( ftype FT2) FT2*, forall( ftype FT3) FT3 * ),
        ?!=?( forall( ftype FT2) FT2*, forall( ftype FT3) FT3 * );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
 rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
int ?==?( X, X ),
        ?!=?( X, X );
\end{lstlisting}

\begin{rationale}
The polymorphic equality operations come in three styles: comparisons between pointers of compatible
types, between pointers to \lstinline$void$ and pointers to object types or incomplete types, and
between the null pointer constant\index{null pointer} and pointers to any type. In the last case, a
special constraint rule for null pointer constant operands has been replaced by a consequence of the
\CFA type system.
\end{rationale}

\semantics
The interpretations of an equality expression are the interpretations of the corresponding function
call.

\begin{sloppypar}
The result of an equality comparison between two pointers to predefined functions or predefined
values is implementation-defined.
\end{sloppypar}
\begin{rationale}
The implementation-defined status of equality comparisons allows implementations to use one library
routine to implement many predefined functions. These optimization are particularly important when
the predefined functions are polymorphic, as is the case for most pointer operations
\end{rationale}


\subsection{Bitwise AND operator}

\begin{syntax}
\lhs{AND-expression}
\rhs \nonterm{equality-expression}
\rhs \nonterm{AND-expression} \lstinline$&$ \nonterm{equality-expression}
\end{syntax}

\rewriterules
\begin{lstlisting}
a & b @\rewrite@ ?&?( a, b )@\use{?&?}@
\end{lstlisting}

\begin{lstlisting}
int ?&?( int, int );
unsigned int ?&?( unsigned int, unsigned int );
long int ?&?( long int, long int );
long unsigned int ?&?( long unsigned int, long unsigned int );
long long int ?&?( long long int, long long int );
long long unsigned int ?&?( long long unsigned int, long long unsigned int );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
int ?&?( X, X );
\end{lstlisting}

\semantics
The interpretations of a bitwise AND expression are the interpretations of the corresponding
function call.


\subsection{Bitwise exclusive OR operator}

\begin{syntax}
\lhs{exclusive-OR-expression}
\rhs \nonterm{AND-expression}
\rhs \nonterm{exclusive-OR-expression} \lstinline$^$ \nonterm{AND-expression}
\end{syntax}

\rewriterules
\begin{lstlisting}
a ^ b @\rewrite@ ?^?( a, b )@\use{?^?}@
\end{lstlisting}

\begin{lstlisting}
int ?^?( int, int );
unsigned int ?^?( unsigned int, unsigned int );
long int ?^?( long int, long int );
long unsigned int ?^?( long unsigned int, long unsigned int );
long long int ?^?( long long int, long long int );
long long unsigned int ?^?( long long unsigned int, long long unsigned int );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
 rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
int ?^?( X, X );
\end{lstlisting}

\semantics
The interpretations of a bitwise exclusive OR expression are the interpretations of the
corresponding function call.


\subsection{Bitwise inclusive OR operator}

\begin{syntax}
\lhs{inclusive-OR-expression}
\rhs \nonterm{exclusive-OR-expression}
\rhs \nonterm{inclusive-OR-expression} \lstinline$|$ \nonterm{exclusive-OR-expression}
\end{syntax}

\rewriterules\use{?"|?}
\begin{lstlisting}
a | b @\rewrite@ ?|?( a, b )
\end{lstlisting}

\begin{lstlisting}
int ?|?( int, int );
unsigned int ?|?( unsigned int, unsigned int );
long int ?|?( long int, long int );
long unsigned int ?|?( long unsigned int, long unsigned int );
long long int ?|?( long long int, long long int );
long long unsigned int ?|?( long long unsigned int, long long unsigned int );
\end{lstlisting}
For every extended integer type \lstinline$X$ with integer conversion rank \index{integer conversion
rank}greater than the rank of \lstinline$int$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
int ?|?( X, X );
\end{lstlisting}

\semantics 
The interpretations of a bitwise inclusive OR expression are the interpretations of the
corresponding function call.


\subsection{Logical AND operator}

\begin{syntax}
\lhs{logical-AND-expression}
\rhs \nonterm{inclusive-OR-expression}
\rhs \nonterm{logical-AND-expression} \lstinline$&&$ \nonterm{inclusive-OR-expression}
\end{syntax}

\semantics The operands of the expression ``\lstinline$a && b$'' are treated as
``\lstinline$(int)((a)!=0)$'' and ``\lstinline$(int)((b)!=0)$'', which shall both be
unambiguous. The expression has only one interpretation, which is of type \lstinline$int$.
\begin{rationale}
When the operands of a logical expression are values of built-in types, and ``\lstinline$!=$'' has
not been redefined for those types, the compiler can optimize away the function calls.

A common C idiom omits comparisons to \lstinline$0$ in the controlling expressions of loops and
\lstinline$if$ statements. For instance, the loop below iterates as long as \lstinline$rp$ points
at a \lstinline$Rational$ value that is non-zero.

\begin{lstlisting}
extern type Rational;@\use{Rational}@
extern const Rational 0;@\use{0}@
extern int ?!=?( Rational, Rational );
Rational *rp;

while ( rp && *rp ) { ... }
\end{lstlisting}
The logical expression calls the \lstinline$Rational$ inequality operator, passing
it \lstinline$*rp$ and the \lstinline$Rational 0$, and getting a 1 or 0 as a result. In
contrast, {\CC} would apply a programmer-defined \lstinline$Rational$-to-\lstinline$int$
conversion to \lstinline$*rp$ in the equivalent situation. The conversion to \lstinline$int$ would
produce a general integer value, which is unfortunate, and possibly dangerous if the conversion was
not written with this situation in mind.
\end{rationale}


\subsection{Logical OR operator}

\begin{syntax}
\lhs{logical-OR-expression}
\rhs \nonterm{logical-AND-expression}
\rhs \nonterm{logical-OR-expression} \lstinline$||$ \nonterm{logical-AND-expression}
\end{syntax}

\semantics

The operands of the expression ``\lstinline$a || b$'' are treated as ``\lstinline$(int)((a)!=0)$''
and ``\lstinline$(int)((b})!=0)$'', which shall both be unambiguous. The expression has only one
interpretation, which is of type \lstinline$int$.


\subsection{Conditional operator}

\begin{syntax}
\lhs{conditional-expression}
\rhs \nonterm{logical-OR-expression}
\rhs \nonterm{logical-OR-expression} \lstinline$?$ \nonterm{expression}
         \lstinline$:$ \nonterm{conditional-expression}
\end{syntax}

\semantics
In the conditional expression\use{?:} ``\lstinline$a?b:c$'', if the second and
third operands both have an interpretation with \lstinline$void$ type, then the expression has an
interpretation with type \lstinline$void$, equivalent to
\begin{lstlisting}
( int)(( a)!=0) ? ( void)( b) : ( void)( c)
\end{lstlisting}

If the second and third operands both have interpretations with non-\lstinline$void$ types, the
expression is treated as if it were the call ``\lstinline$cond((a)!=0, b, c)$'',
with \lstinline$cond$ declared as
\begin{lstlisting}
forall( type T ) T cond( int, T, T );
 
forall( dtype D ) void
        * cond( int, D *, void * ),
        * cond( int, void *, D * );
        
forall( dtype D ) _atomic void
        * cond( int, _Atomic D *, _Atomic void * ),
        * cond( int, _Atomic void *, _Atomic D * );

forall( dtype D ) const void
        * cond( int, const D *, const void * ),
        * cond( int, const void *, const D * );

forall( dtype D ) restrict void
        * cond( int, restrict D *, restrict void * ),
        * cond( int, restrict void *, restrict D * );

forall( dtype D ) volatile void
        * cond( int, volatile D *, volatile void * ),
        * cond( int, volatile void *, volatile D * );

forall( dtype D ) _Atomic const void
        * cond( int, _Atomic const D *, _Atomic const void * ),
        * cond( int, _Atomic const void *, _Atomic const D * );

forall( dtype D ) _Atomic restrict void
        * cond( int, _Atomic restrict D *, _Atomic restrict void * ),
        * cond( int, _Atomic restrict void *, _Atomic restrict D * );

forall( dtype D ) _Atomic volatile void
        * cond( int, _Atomic volatile D *, _Atomic volatile void * ),
        * cond( int, _Atomic volatile void *, _Atomic volatile D * );

forall( dtype D ) const restrict void
        * cond( int, const restrict D *, const restrict void * ),
        * cond( int, const restrict void *, const restrict D * );

forall( dtype D ) const volatile void
        * cond( int, const volatile D *, const volatile void * ),
        * cond( int, const volatile void *, const volatile D * );

forall( dtype D ) restrict volatile void
        * cond( int, restrict volatile D *, restrict volatile void * ),
        * cond( int, restrict volatile void *, restrict volatile D * );

forall( dtype D ) _Atomic const restrict void
        * cond( int, _Atomic const restrict D *, _Atomic const restrict void * ),
        * cond( int, _Atomic const restrict void *, _Atomic const restrict D * );

forall( dtype D ) _Atomic const volatile void
        * cond( int, _Atomic const volatile D *, _Atomic const volatile void * ),
        * cond( int, _Atomic const volatile void *, _Atomic const volatile D * );

forall( dtype D ) _Atomic restrict volatile void
        * cond( int, _Atomic restrict volatile D *,
         _Atomic restrict volatile void * ),
        * cond( int, _Atomic restrict volatile void *,
         _Atomic restrict volatile D * );

forall( dtype D ) const restrict volatile void
        * cond( int, const restrict volatile D *,
         const restrict volatile void * ),
        * cond( int, const restrict volatile void *,
         const restrict volatile D * );

forall( dtype D ) _Atomic const restrict volatile void
        * cond( int, _Atomic const restrict volatile D *,
         _Atomic const restrict volatile void * ),
        * cond( int, _Atomic const restrict volatile void *,
         _Atomic const restrict volatile D * );
\end{lstlisting}

\begin{rationale}
The object of the above is to apply the usual arithmetic conversions\index{usual arithmetic
conversions} when the second and third operands have arithmetic type, and to combine the
qualifiers of the second and third operands if they are pointers.
\end{rationale}

\examples
\begin{lstlisting}
#include <stdlib.h>
int i;
long l;
rand() ? i : l;
\end{lstlisting}
The best interpretation infers the expression's type to be \lstinline$long$ and applies the safe
\lstinline$int$-to-\lstinline$long$ conversion to \lstinline$i$.

\begin{lstlisting}
const int *cip;
volatile int *vip;
rand() ? cip : vip;
\end{lstlisting}
The expression has type \lstinline$const volatile int *$, with safe conversions applied to the second
and third operands to add \lstinline$volatile$ and \lstinline$const$ qualifiers, respectively.

\begin{lstlisting}
rand() ? cip : 0;
\end{lstlisting}
The expression has type \lstinline$const int *$, with a specialization conversion applied to
\lstinline$0$.


\subsection{Assignment operators}

\begin{syntax}
\lhs{assignment-expression}
\rhs \nonterm{conditional-expression}
\rhs \nonterm{unary-expression} \nonterm{assignment-operator}
         \nonterm{assignment-expression}
\lhs{assignment-operator} one of
\rhs \lstinline$=$\ \ \lstinline$*=$\ \ \lstinline$/=$\ \ \lstinline$%=$\ \ \lstinline$+=$\ \ \lstinline$-=$\ \  
         \lstinline$<<=$\ \ \lstinline$>>=$\ \ \lstinline$&=$\ \ \lstinline$^=$\ \ \lstinline$|=$
\end{syntax}

\rewriterules
Let ``\(\leftarrow\)'' be any of the assignment operators. Then
\use{?=?}\use{?*=?}\use{?/=?}\use{?%=?}\use{?+=?}\use{?-=?}
\use{?>>=?}\use{?&=?}\use{?^=?}\use{?"|=?}%use{?<<=?}
\begin{lstlisting}
a @$\leftarrow$@ b @\rewrite@ ?@$\leftarrow$@?( &( a ), b )
\end{lstlisting}

\semantics
Each interpretation of the left operand of an assignment expression is considered separately. For
each interpretation that is a bit-field or is declared with the \lstinline$register$ storage class
specifier, the expression has one valid interpretation, with the type of the left operand. The
right operand is cast to that type, and the assignment expression is ambiguous if either operand is.
For the remaining interpretations, the expression is rewritten, and the interpretations of the
assignment expression are the interpretations of the corresponding function call. Finally, all
interpretations of the expression produced for the different interpretations of the left operand are
combined to produce the interpretations of the expression as a whole; where interpretations have
compatible result types, the best interpretations are selected in the manner described for function
call expressions.


\subsubsection{Simple assignment}

\begin{lstlisting}
_Bool
        ?=?( volatile _Bool *, _Bool ),
        ?=?( volatile _Bool *, forall( dtype D ) D * ),
        ?=?( volatile _Bool *, forall( ftype F ) F * ),
        ?=?( _Atomic volatile _Bool *, _Bool ),
        ?=?( _Atomic volatile _Bool *, forall( dtype D ) D * ),
        ?=?( _Atomic volatile _Bool *, forall( ftype F ) F * );
char
        ?=?( volatile char *, char ),
        ?=?( _Atomic volatile char *, char );
unsigned char
        ?=?( volatile unsigned char *, unsigned char ),
        ?=?( _Atomic volatile unsigned char *, unsigned char );
signed char
        ?=?( volatile signed char *, signed char ),
        ?=?( _Atomic volatile signed char *, signed char );
short int
        ?=?( volatile short int *, short int ),
        ?=?( _Atomic volatile short int *, short int );
unsigned short
        ?=?( volatile unsigned int *, unsigned int ),
        ?=?( _Atomic volatile unsigned int *, unsigned int );
int
        ?=?( volatile int *, int ),
        ?=?( _Atomic volatile int *, int );
unsigned int
        ?=?( volatile unsigned int *, unsigned int ),
        ?=?( _Atomic volatile unsigned int *, unsigned int );
long int
        ?=?( volatile long int *, long int ),
        ?=?( _Atomic volatile long int *, long int );
unsigned long int
        ?=?( volatile unsigned long int *, unsigned long int ),
        ?=?( _Atomic volatile unsigned long int *, unsigned long int );
long long int
        ?=?( volatile long long int *, long long int ),
        ?=?( _Atomic volatile long long int *, long long int );
unsigned long long int
        ?=?( volatile unsigned long long int *, unsigned long long int ),
        ?=?( _Atomic volatile unsigned long long int *, unsigned long long int );
float
        ?=?( volatile float *, float ),
        ?=?( _Atomic volatile float *, float );
double
        ?=?( volatile double *, double ),
        ?=?( _Atomic volatile double *, double );
long double
        ?=?( volatile long double *, long double ),
        ?=?( _Atomic volatile long double *, long double );
_Complex float
        ?=?( volatile float *, float ),
        ?=?( _Atomic volatile float *, float );
_Complex double
        ?=?( volatile double *, double ),
        ?=?( _Atomic volatile double *, double );
_Complex long double
        ?=?( volatile _Complex long double *, _Complex long double ),
        ?=?( _Atomic volatile _Complex long double *, _Atomic _Complex long double );

forall( ftype FT ) FT
        * ?=?( FT * volatile *, FT * ),
        * ?=?( FT * volatile *, forall( ftype F ) F * );

forall( ftype FT ) FT const
        * ?=?( FT const * volatile *, FT const * ),
        * ?=?( FT const * volatile *, forall( ftype F ) F * );

forall( ftype FT ) FT volatile
        * ?=?( FT volatile * volatile *, FT * ),
        * ?=?( FT volatile * volatile *, forall( ftype F ) F * );

forall( ftype FT ) FT const
        * ?=?( FT const volatile * volatile *, FT const * ),
        * ?=?( FT const volatile * volatile *, forall( ftype F ) F * );

forall( dtype DT ) DT
        * ?=?( DT * restrict volatile *, DT * ),
        * ?=?( DT * restrict volatile *, void * ),
        * ?=?( DT * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT * _Atomic restrict volatile *, DT * ),
        * ?=?( DT * _Atomic restrict volatile *, void * ),
        * ?=?( DT * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic
        * ?=?( _Atomic DT * restrict volatile *, DT _Atomic * ),
        * ?=?( _Atomic DT * restrict volatile *, void * ),
        * ?=?( _Atomic DT * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( _Atomic DT * _Atomic restrict volatile *, DT _Atomic * ),
        * ?=?( _Atomic DT * _Atomic restrict volatile *, void * ),
        * ?=?( _Atomic DT * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT const
        * ?=?( DT const * restrict volatile *, DT const * ),
        * ?=?( DT const * restrict volatile *, void const * ),
        * ?=?( DT const * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT const * _Atomic restrict volatile *, DT const * ),
        * ?=?( DT const * _Atomic restrict volatile *, void const * ),
        * ?=?( DT const * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT restrict
        * ?=?( restrict DT * restrict volatile *, DT restrict * ),
        * ?=?( restrict DT * restrict volatile *, void * ),
        * ?=?( restrict DT * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( restrict DT * _Atomic restrict volatile *, DT restrict * ),
        * ?=?( restrict DT * _Atomic restrict volatile *, void * ),
        * ?=?( restrict DT * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT volatile
        * ?=?( DT volatile * restrict volatile *, DT volatile * ),
        * ?=?( DT volatile * restrict volatile *, void volatile * ),
        * ?=?( DT volatile * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT volatile * _Atomic restrict volatile *, DT volatile * ),
        * ?=?( DT volatile * _Atomic restrict volatile *, void volatile * ),
        * ?=?( DT volatile * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic const
        * ?=?( DT _Atomic const * restrict volatile *, DT _Atomic const * ),
        * ?=?( DT _Atomic const * restrict volatile *, void const * ),
        * ?=?( DT _Atomic const * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT _Atomic const * _Atomic restrict volatile *, DT _Atomic const * ),
        * ?=?( DT _Atomic const * _Atomic restrict volatile *, void const * ),
        * ?=?( DT _Atomic const * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic restrict
        * ?=?( _Atomic restrict DT * restrict volatile *, DT _Atomic restrict * ),
        * ?=?( _Atomic restrict DT * restrict volatile *, void * ),
        * ?=?( _Atomic restrict DT * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( _Atomic restrict DT * _Atomic restrict volatile *, DT _Atomic restrict * ),
        * ?=?( _Atomic restrict DT * _Atomic restrict volatile *, void * ),
        * ?=?( _Atomic restrict DT * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic volatile
        * ?=?( DT _Atomic volatile * restrict volatile *, DT _Atomic volatile * ),
        * ?=?( DT _Atomic volatile * restrict volatile *, void volatile * ),
        * ?=?( DT _Atomic volatile * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT _Atomic volatile * _Atomic restrict volatile *, DT _Atomic volatile * ),
        * ?=?( DT _Atomic volatile * _Atomic restrict volatile *, void volatile * ),
        * ?=?( DT _Atomic volatile * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT const restrict
        * ?=?( DT const restrict * restrict volatile *, DT const restrict * ),
        * ?=?( DT const restrict * restrict volatile *, void const * ),
        * ?=?( DT const restrict * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT const restrict * _Atomic restrict volatile *, DT const restrict * ),
        * ?=?( DT const restrict * _Atomic restrict volatile *, void const * ),
        * ?=?( DT const restrict * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT const volatile
        * ?=?( DT const volatile * restrict volatile *, DT const volatile * ),
        * ?=?( DT const volatile * restrict volatile *, void const volatile * ),
        * ?=?( DT const volatile * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT const volatile * _Atomic restrict volatile *, DT const volatile * ),
        * ?=?( DT const volatile * _Atomic restrict volatile *, void const volatile * ),
        * ?=?( DT const volatile * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT restrict volatile
        * ?=?( DT restrict volatile * restrict volatile *, DT restrict volatile * ),
        * ?=?( DT restrict volatile * restrict volatile *, void volatile * ),
        * ?=?( DT restrict volatile * restrict volatile *, forall( dtype D ) D * ),
        * ?=?( DT restrict volatile * _Atomic restrict volatile *, DT restrict volatile * ),
        * ?=?( DT restrict volatile * _Atomic restrict volatile *, void volatile * ),
        * ?=?( DT restrict volatile * _Atomic restrict volatile *, forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic const restrict
        * ?=?( DT _Atomic const restrict * restrict volatile *,
         DT _Atomic const restrict * ),
        * ?=?( DT _Atomic const restrict * restrict volatile *,
         void const * ),
        * ?=?( DT _Atomic const restrict * restrict volatile *,
         forall( dtype D ) D * ),
        * ?=?( DT _Atomic const restrict * _Atomic restrict volatile *,
         DT _Atomic const restrict * ),
        * ?=?( DT _Atomic const restrict * _Atomic restrict volatile *,
         void const * ),
        * ?=?( DT _Atomic const restrict * _Atomic restrict volatile *,
         forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic const volatile
        * ?=?( DT _Atomic const volatile * restrict volatile *,
         DT _Atomic const volatile * ),
        * ?=?( DT _Atomic const volatile * restrict volatile *,
         void const volatile * ),
        * ?=?( DT _Atomic const volatile * restrict volatile *,
         forall( dtype D ) D * ),
        * ?=?( DT _Atomic const volatile * _Atomic restrict volatile *,
         DT _Atomic const volatile * ),
        * ?=?( DT _Atomic const volatile * _Atomic restrict volatile *,
         void const volatile * ),
        * ?=?( DT _Atomic const volatile * _Atomic restrict volatile *,
         forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic restrict volatile
        * ?=?( DT _Atomic restrict volatile * restrict volatile *,
         DT _Atomic restrict volatile * ),
        * ?=?( DT _Atomic restrict volatile * restrict volatile *,
         void volatile * ),
        * ?=?( DT _Atomic restrict volatile * restrict volatile *,
         forall( dtype D ) D * ),
        * ?=?( DT _Atomic restrict volatile * _Atomic restrict volatile *,
         DT _Atomic restrict volatile * ),
        * ?=?( DT _Atomic restrict volatile * _Atomic restrict volatile *,
         void volatile * ),
        * ?=?( DT _Atomic restrict volatile * _Atomic restrict volatile *,
         forall( dtype D ) D * );

forall( dtype DT ) DT const restrict volatile
        * ?=?( DT const restrict volatile * restrict volatile *,
         DT const restrict volatile * ),
        * ?=?( DT const restrict volatile * restrict volatile *,
         void const volatile * ),
        * ?=?( DT const restrict volatile * restrict volatile *,
         forall( dtype D ) D * ),
        * ?=?( DT const restrict volatile * _Atomic restrict volatile *,
         DT const restrict volatile * ),
        * ?=?( DT const restrict volatile * _Atomic restrict volatile *,
         void const volatile * ),
        * ?=?( DT const restrict volatile * _Atomic restrict volatile *,
         forall( dtype D ) D * );

forall( dtype DT ) DT _Atomic const restrict volatile
        * ?=?( DT _Atomic const restrict volatile * restrict volatile *,
         DT _Atomic const restrict volatile * ),
        * ?=?( DT _Atomic const restrict volatile * restrict volatile *,
         void const volatile * ),
        * ?=?( DT _Atomic const restrict volatile * restrict volatile *,
         forall( dtype D ) D * ),
        * ?=?( DT _Atomic const restrict volatile * _Atomic restrict volatile *,
         DT _Atomic const restrict volatile * ),
        * ?=?( DT _Atomic const restrict volatile * _Atomic restrict volatile *,
         void const volatile * ),
        * ?=?( DT _Atomic const restrict volatile * _Atomic restrict volatile *,
         forall( dtype D ) D * );

forall( dtype DT ) void
        * ?=?( void * restrict volatile *, DT * );

forall( dtype DT ) void const
        * ?=?( void const * restrict volatile *, DT const * );

forall( dtype DT ) void volatile
        * ?=?( void volatile * restrict volatile *, DT volatile * );

forall( dtype DT ) void const volatile
        * ?=?( void const volatile * restrict volatile *, DT const volatile * );
\end{lstlisting}
\begin{rationale}
The pattern of overloadings for simple assignment resembles that of pointer increment and decrement,
except that the polymorphic pointer assignment functions declare a \lstinline$dtype$ parameter,
instead of a \lstinline$type$ parameter, because the left operand may be a pointer to an incomplete
type.
\end{rationale}

For every complete structure or union type \lstinline$S$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
S ?=?( S volatile *, S ), ?=?( S _Atomic volatile *, S );
\end{lstlisting}

For every extended integer type \lstinline$X$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
X ?=?( X volatile *, X ), ?=?( X _Atomic volatile *, X );
\end{lstlisting}

For every complete enumerated type \lstinline$E$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
E ?=?( E volatile *, int ), ?=?( E _Atomic volatile *, int );
\end{lstlisting}
\begin{rationale}
The right-hand argument is \lstinline$int$ because enumeration constants have type \lstinline$int$.
\end{rationale}

\semantics
The structure assignment functions provide member-wise assignment; each non-array member and each
element of each array member of the right argument is assigned to the corresponding member or
element of the left argument using the assignment function defined for its type. All other
assignment functions have the same effect as the corresponding C assignment expression.
\begin{rationale}
Note that, by default, union assignment\index{deficiencies!union assignment} uses C semantics---that
is, bitwise copy---even if some of the union members have programmer-defined assignment functions.
\end{rationale}


\subsubsection{Compound assignment}

\begin{lstlisting}
forall( type T ) T
        * ?+=?( T * restrict volatile *, ptrdiff_t ),
        * ?-=?( T * restrict volatile *, ptrdiff_t ),
        * ?+=?( T * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic
        * ?+=?( T _Atomic * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T const
        * ?+=?( T const * restrict volatile *, ptrdiff_t ),
        * ?-=?( T const * restrict volatile *, ptrdiff_t ),
        * ?+=?( T const * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T const * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T restrict
        * ?+=?( T restrict * restrict volatile *, ptrdiff_t ),
        * ?-=?( T restrict * restrict volatile *, ptrdiff_t ),
        * ?+=?( T restrict * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T restrict * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T volatile
        * ?+=?( T volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic const
        * ?+=?( T _Atomic const restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic const _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic restrict
        * ?+=?( T _Atomic restrict * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic restrict * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic restrict * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic restrict * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic volatile
        * ?+=?( T _Atomic volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T const restrict
        * ?+=?( T const restrict * restrict volatile *, ptrdiff_t ),
        * ?-=?( T const restrict * restrict volatile *, ptrdiff_t ),
        * ?+=?( T const restrict * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T const restrict * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T const volatile
        * ?+=?( T const volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T const volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T const volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T const volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T restrict volatile
        * ?+=?( T restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T restrict volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T restrict volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic const restrict
        * ?+=?( T _Atomic const restrict * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const restrict * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic const restrict * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const restrict * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic const volatile
        * ?+=?( T _Atomic const volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic const volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic restrict volatile
        * ?+=?( T _Atomic restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic restrict volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic restrict volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T const restrict volatile
        * ?+=?( T const restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T const restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T const restrict volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T const restrict volatile * _Atomic restrict volatile *, ptrdiff_t );

forall( type T ) T _Atomic const restrict volatile
        * ?+=?( T _Atomic const restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const restrict volatile * restrict volatile *, ptrdiff_t ),
        * ?+=?( T _Atomic const restrict volatile * _Atomic restrict volatile *, ptrdiff_t ),
        * ?-=?( T _Atomic const restrict volatile * _Atomic restrict volatile *, ptrdiff_t );

_Bool
        ?*=?( _Bool volatile *, _Bool ),
        ?/=?( _Bool volatile *, _Bool ),
        ?+=?( _Bool volatile *, _Bool ),
        ?-=?( _Bool volatile *, _Bool ),
        ?%=?( _Bool volatile *, _Bool ),
        ?<<=?( _Bool volatile *, int ),
        ?>>=?( _Bool volatile *, int ),
        ?&=?( _Bool volatile *, _Bool ),
        ?^=?( _Bool volatile *, _Bool ),
        ?|=?( _Bool volatile *, _Bool );
char
        ?*=?( char volatile *, char ),
        ?/=?( char volatile *, char ),
        ?+=?( char volatile *, char ),
        ?-=?( char volatile *, char ),
        ?%=?( char volatile *, char ),
        ?<<=?( char volatile *, int ),
        ?>>=?( char volatile *, int ),
        ?&=?( char volatile *, char ),
        ?^=?( char volatile *, char ),
        ?|=?( char volatile *, char );
unsigned char
        ?*=?( unsigned char volatile *, unsigned char ),
        ?/=?( unsigned char volatile *, unsigned char ),
        ?+=?( unsigned char volatile *, unsigned char ),
        ?-=?( unsigned char volatile *, unsigned char ),
        ?%=?( unsigned char volatile *, unsigned char ),
        ?<<=?( unsigned char volatile *, int ),
        ?>>=?( unsigned char volatile *, int ),
        ?&=?( unsigned char volatile *, unsigned char ),
        ?^=?( unsigned char volatile *, unsigned char ),
        ?|=?( unsigned char volatile *, unsigned char );
signed char
        ?*=?( signed char volatile *, signed char ),
        ?/=?( signed char volatile *, signed char ),
        ?+=?( signed char volatile *, signed char ),
        ?-=?( signed char volatile *, signed char ),
        ?%=?( signed char volatile *, signed char ),
        ?<<=?( signed char volatile *, int ),
        ?>>=?( signed char volatile *, int ),
        ?&=?( signed char volatile *, signed char ),
        ?^=?( signed char volatile *, signed char ),
        ?|=?( signed char volatile *, signed char );
short int
        ?*=?( short int volatile *, short int ),
        ?/=?( short int volatile *, short int ),
        ?+=?( short int volatile *, short int ),
        ?-=?( short int volatile *, short int ),
        ?%=?( short int volatile *, short int ),
        ?<<=?( short int volatile *, int ),
        ?>>=?( short int volatile *, int ),
        ?&=?( short int volatile *, short int ),
        ?^=?( short int volatile *, short int ),
        ?|=?( short int volatile *, short int );
unsigned short int
        ?*=?( unsigned short int volatile *, unsigned short int ),
        ?/=?( unsigned short int volatile *, unsigned short int ),
        ?+=?( unsigned short int volatile *, unsigned short int ),
        ?-=?( unsigned short int volatile *, unsigned short int ),
        ?%=?( unsigned short int volatile *, unsigned short int ),
        ?<<=?( unsigned short int volatile *, int ),
        ?>>=?( unsigned short int volatile *, int ),
        ?&=?( unsigned short int volatile *, unsigned short int ),
        ?^=?( unsigned short int volatile *, unsigned short int ),
        ?|=?( unsigned short int volatile *, unsigned short int );
int
        ?*=?( int volatile *, int ),
        ?/=?( int volatile *, int ),
        ?+=?( int volatile *, int ),
        ?-=?( int volatile *, int ),
        ?%=?( int volatile *, int ),
        ?<<=?( int volatile *, int ),
        ?>>=?( int volatile *, int ),
        ?&=?( int volatile *, int ),
        ?^=?( int volatile *, int ),
        ?|=?( int volatile *, int );
unsigned int
        ?*=?( unsigned int volatile *, unsigned int ),
        ?/=?( unsigned int volatile *, unsigned int ),
        ?+=?( unsigned int volatile *, unsigned int ),
        ?-=?( unsigned int volatile *, unsigned int ),
        ?%=?( unsigned int volatile *, unsigned int ),
        ?<<=?( unsigned int volatile *, int ),
        ?>>=?( unsigned int volatile *, int ),
        ?&=?( unsigned int volatile *, unsigned int ),
        ?^=?( unsigned int volatile *, unsigned int ),
        ?|=?( unsigned int volatile *, unsigned int );
long int
        ?*=?( long int volatile *, long int ),
        ?/=?( long int volatile *, long int ),
        ?+=?( long int volatile *, long int ),
        ?-=?( long int volatile *, long int ),
        ?%=?( long int volatile *, long int ),
        ?<<=?( long int volatile *, int ),
        ?>>=?( long int volatile *, int ),
        ?&=?( long int volatile *, long int ),
        ?^=?( long int volatile *, long int ),
        ?|=?( long int volatile *, long int );
unsigned long int
        ?*=?( unsigned long int volatile *, unsigned long int ),
        ?/=?( unsigned long int volatile *, unsigned long int ),
        ?+=?( unsigned long int volatile *, unsigned long int ),
        ?-=?( unsigned long int volatile *, unsigned long int ),
        ?%=?( unsigned long int volatile *, unsigned long int ),
        ?<<=?( unsigned long int volatile *, int ),
        ?>>=?( unsigned long int volatile *, int ),
        ?&=?( unsigned long int volatile *, unsigned long int ),
        ?^=?( unsigned long int volatile *, unsigned long int ),
        ?|=?( unsigned long int volatile *, unsigned long int );
long long int
        ?*=?( long long int volatile *, long long int ),
        ?/=?( long long int volatile *, long long int ),
        ?+=?( long long int volatile *, long long int ),
        ?-=?( long long int volatile *, long long int ),
        ?%=?( long long int volatile *, long long int ),
        ?<<=?( long long int volatile *, int ),
        ?>>=?( long long int volatile *, int ),
        ?&=?( long long int volatile *, long long int ),
        ?^=?( long long int volatile *, long long int ),
        ?|=?( long long int volatile *, long long int );
unsigned long long int
        ?*=?( unsigned long long int volatile *, unsigned long long int ),
        ?/=?( unsigned long long int volatile *, unsigned long long int ),
        ?+=?( unsigned long long int volatile *, unsigned long long int ),
        ?-=?( unsigned long long int volatile *, unsigned long long int ),
        ?%=?( unsigned long long int volatile *, unsigned long long int ),
        ?<<=?( unsigned long long int volatile *, int ),
        ?>>=?( unsigned long long int volatile *, int ),
        ?&=?( unsigned long long int volatile *, unsigned long long int ),
        ?^=?( unsigned long long int volatile *, unsigned long long int ),
        ?|=?( unsigned long long int volatile *, unsigned long long int );
float
        ?*=?( float volatile *, float ),
        ?/=?( float volatile *, float ),
        ?+=?( float volatile *, float ),
        ?-=?( float volatile *, float );
double
        ?*=?( double volatile *, double ),
        ?/=?( double volatile *, double ),
        ?+=?( double volatile *, double ),
        ?-=?( double volatile *, double );
long double
        ?*=?( long double volatile *, long double ),
        ?/=?( long double volatile *, long double ),
        ?+=?( long double volatile *, long double ),
        ?-=?( long double volatile *, long double );
_Complex float
        ?*=?( _Complex float volatile *, _Complex float ),
        ?/=?( _Complex float volatile *, _Complex float ),
        ?+=?( _Complex float volatile *, _Complex float ),
        ?-=?( _Complex float volatile *, _Complex float );
_Complex double
        ?*=?( _Complex double volatile *, _Complex double ),
        ?/=?( _Complex double volatile *, _Complex double ),
        ?+=?( _Complex double volatile *, _Complex double ),
        ?-=?( _Complex double volatile *, _Complex double );
_Complex long double
        ?*=?( _Complex long double volatile *, _Complex long double ),
        ?/=?( _Complex long double volatile *, _Complex long double ),
        ?+=?( _Complex long double volatile *, _Complex long double ),
        ?-=?( _Complex long double volatile *, _Complex long double );
\end{lstlisting}

For every extended integer type \lstinline$X$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
?*=?( X volatile *, X ),
?/=?( X volatile *, X ),
?+=?( X volatile *, X ),
?-=?( X volatile *, X ),
?%=?( X volatile *, X ),
?<<=?( X volatile *, int ),
?>>=?( X volatile *, int ),
?&=?( X volatile *, X ),
?^=?( X volatile *, X ),
?|=?( X volatile *, X );
\end{lstlisting}

For every complete enumerated type \lstinline$E$ there exist
% Don't use predefined: keep this out of prelude.cf.
\begin{lstlisting}
?*=?( E volatile *, E ),
?/=?( E volatile *, E ),
?+=?( E volatile *, E ),
?-=?( E volatile *, E ),
?%=?( E volatile *, E ),
?<<=?( E volatile *, int ),
?>>=?( E volatile *, int ),
?&=?( E volatile *, E ),
?^=?( E volatile *, E ),
?|=?( E volatile *, E );
\end{lstlisting}


\subsection{Comma operator}

\begin{syntax}
\lhs{expression}
\rhs \nonterm{assignment-expression}
\rhs \nonterm{expression} \lstinline$,$ \nonterm{assignment-expression}
\end{syntax}

\semantics
In the comma expression ``\lstinline$a, b$'', the first operand is interpreted as
``\lstinline$( void )(a)$'', which shall be unambiguous\index{ambiguous interpretation}. The
interpretations of the expression are the interpretations of the second operand.


\section{Constant expressions}


\section{Declarations}

\begin{syntax}
\oldlhs{declaration}
\rhs \nonterm{type-declaration}
\rhs \nonterm{spec-definition}
\end{syntax}

\constraints
If an identifier has no linkage\index{no linkage}, there shall be no more than one declaration of
the identifier ( in a declarator or type specifier ) with compatible types in the same scope and in
the same name space, except that:
\begin{itemize}
\item
a typedef name may be redefined to denote the same type as it currently does, provided that type is
not a variably modified type;
\item
tags may be redeclared as specified in section 6.7.2.3 of the {\c11} standard.
\end{itemize}
\begin{rationale}
This constraint adds the phrase ``with compatible types'' to the {\c11} constraint, to allow
overloading.
\end{rationale}

An identifier declared by a type declaration shall not be redeclared as a parameter in a function
definition whose declarator includes an identifier list.
\begin{rationale}
This restriction echos {\c11}'s ban on the redeclaration of typedef names as parameters. This
avoids an ambiguity between old-style function declarations and new-style function prototypes:
\begin{lstlisting}
void f( Complex,        // ... 3000 characters ...
void g( Complex,        // ... 3000 characters ...
int Complex; { ... }
\end{lstlisting}
Without the rule, \lstinline$Complex$ would be a type in the first case, and a parameter name in the
second.
\end{rationale}


\setcounter{subsection}{1}
\subsection{Type specifiers}

\begin{syntax}
\oldlhs{type-specifier}
\rhs \nonterm{forall-specifier}
\end{syntax}

\semantics
Forall specifiers are discussed in \VRef{forall}.


\subsubsection{Structure and union specifiers}

\semantics 
\CFA extends the {\c11} definition of \define{anonymous structure} to include structure
specifiers with tags, and extends the {\c11} definition of \define{anonymous union} to include union
specifiers with tags.
\begin{rationale}
This extension imitates an extension in the Plan 9 C compiler \cite{Thompson90new}.
\end{rationale}

\examples
\begin{lstlisting}
struct point {@\impl{point}@
        int x, y;
};
struct color_point {@\impl{color_point}@
        enum { RED, BLUE, GREEN } color;
        struct point;
};
struct color_point cp;
cp.x = 0;
cp.color = RED;

struct literal {@\impl{literal}@
        enum { NUMBER, STRING } tag;
        union {
         double n;
         char *s;
        };
};
struct literal *next;
int length;
extern int strlen( const char * );
...
if ( next->tag == STRING ) length = strlen( next->s );
\end{lstlisting}


\setcounter{subsubsection}{4}
\subsubsection{Forall specifiers}\label{forall}

\begin{syntax}
\lhs{forall-specifier}
\rhs \lstinline$forall$ \lstinline$($ \nonterm{type-parameter-list} \lstinline$)$
\end{syntax}

\constraints
If the \nonterm{declaration-specifiers} of a declaration that contains a \nonterm{forall-specifier}
declares a structure or union tag, the types of the members of the structure or union shall not use
any of the type identifiers declared by the \nonterm{type-parameter-list}.
\begin{rationale}
This sort of declaration is illegal because the scope of the type identifiers ends at the end of the
declaration, but the scope of the structure tag does not.
\begin{lstlisting}
forall( type T ) struct Pair { T a, b; } mkPair( T, T ); // illegal 
\end{lstlisting}
If an instance of \lstinline$struct Pair$ was declared later in the current scope, what would the
members' type be?
\end{rationale}

\semantics
The \nonterm{type-parameter-list}s and assertions of the \nonterm{forall-specifier}s declare type
identifiers, function and object identifiers with no linkage\index{no linkage}.

If, in the declaration ``\lstinline$T D1$'', \lstinline$T$ contains \nonterm{forall-specifier}s and
\lstinline$D1$ has the form
\begin{lstlisting}
D( @\normalsize\nonterm{parameter-type-list}@ )
\end{lstlisting}
then a type identifier declared by one of the \nonterm{forall-specifier}s is an \define{inferred
 parameter} of the function declarator if and only if it is not an inferred parameter of a function
declarator in \lstinline$D$, and it is used in the type of a parameter in the following
\nonterm{type-parameter-list} or it and an inferred parameter are used as arguments of a
specification\index{specification} in one of the \nonterm{forall-specifier}s. The identifiers
declared by assertions that use an inferred parameter of a function declarator are assertion
parameters\index{assertion parameters} of that function declarator.
\begin{rationale}
Since every inferred parameter is used by some parameter, inference can be understood as a single
bottom-up pass over the expression tree, that only needs to apply local reasoning at each node.

If this restriction were lifted, it would be possible to write
\begin{lstlisting}
forall( type T ) T * alloc( void );@\use{alloc}@
int *p = alloc();
\end{lstlisting}
Here \lstinline$alloc()$ would receive \lstinline$int$ as an inferred argument, and return an
\lstinline$int *$. In general, if a call to \lstinline$alloc()$ is a subexpression of an expression
involving polymorphic functions and overloaded identifiers, there could be considerable distance
between the call and the subexpression that causes \lstinline$T$ to be bound.

With the current restriction, \lstinline$alloc()$ must be given an argument that determines
\lstinline$T$:
\begin{lstlisting}
forall( type T ) T * alloc( T initial_value );@\use{alloc}@
\end{lstlisting}
\end{rationale}

If a function declarator is part of a function definition, its inferred parameters and assertion
parameters have block scope\index{block scope}; otherwise, identifiers declared by assertions have a
\define{declaration scope}, which terminates at the end of the \nonterm{declaration}.

A function type that has at least one inferred parameter is a \define{polymorphic function} type.
Function types with no inferred parameters are \define{monomorphic function} types. One function
type is \define{less polymorphic} than another if it has fewer inferred parameters, or if it has the
same number of inferred parameters and fewer of its explicit parameters have types that depend on an
inferred parameter.

The names of inferred parameters and the order of identifiers in forall specifiers are not relevant
to polymorphic function type compatibility. Let $f$ and $g$ be two polymorphic function types with
the same number of inferred parameters, and let $f_i$ and $g_i$ be the inferred parameters of $f$
and $g$ in their order of occurance in the function types' \nonterm{parameter-type-list}s. Let $f'$
be $f$ with every occurrence of $f_i$ replaced by $g_i$, for all $i$. Then $f$ and $g$ are
compatible types\index{compatible type} if $f'$'s and $g$'s return types and parameter lists are
compatible, and if for every assertion parameter of $f'$ there is an assertion parameter in $g$ with
the same identifier and compatible type, and vice versa.

\examples
Consider these analogous monomorphic and polymorphic declarations.
\begin{lstlisting}
int fi( int );
forall( type T ) T fT( T );
\end{lstlisting}
\lstinline$fi()$ takes an \lstinline$int$ and returns an \lstinline$int$. \lstinline$fT()$ takes a
\lstinline$T$ and returns a \lstinline$T$, for any type \lstinline$T$.
\begin{lstlisting}
int (*pfi )( int ) = fi;
forall( type T ) T (*pfT )( T ) = fT;
\end{lstlisting}
\lstinline$pfi$ and \lstinline$pfT$ are pointers to functions. \lstinline$pfT$ is not
polymorphic, but the function it points at is.
\begin{lstlisting}
int (*fvpfi( void ))( int ) {
        return pfi;
}
forall( type T ) T (*fvpfT( void ))( T ) {
        return pfT;
}
\end{lstlisting}
\lstinline$fvpfi()$ and \lstinline$fvpfT()$ are functions taking no arguments and returning pointers
to functions. \lstinline$fvpfT()$ is monomorphic, but the function that its return value points
at is polymorphic.
\begin{lstlisting}
forall( type T ) int ( *fTpfi( T ) )( int );
forall( type T ) T ( *fTpfT( T ) )( T );
forall( type T, type U ) U ( *fTpfU( T ) )( U );
\end{lstlisting}
\lstinline$fTpfi()$ is a polymorphic function that returns a pointer to a monomorphic function
taking an integer and returning an integer. It could return \lstinline$pfi$. \lstinline$fTpfT()$
is subtle: it is a polymorphic function returning a \emph{monomorphic} function taking and returning
\lstinline$T$, where \lstinline$T$ is an inferred parameter of \lstinline$fTpfT()$. For instance,
in the expression ``\lstinline$fTpfT(17)$'', \lstinline$T$ is inferred to be \lstinline$int$, and
the returned value would have type \lstinline$int ( * )( int )$. ``\lstinline$fTpfT(17)(13)$'' and
``\lstinline$fTpfT("yes")("no")$'' are legal, but ``\lstinline$fTpfT(17)("no")$'' is illegal.
\lstinline$fTpfU()$ is polymorphic ( in type \lstinline$T$), and returns a pointer to a function that
is polymorphic ( in type \lstinline$U$). ``\lstinline$f5(17)("no")$'' is a legal expression of type
\lstinline$char *$.
\begin{lstlisting}
forall( type T, type U, type V ) U * f( T *, U, V * const );
forall( type U, type V, type W ) U * g( V *, U, W * const );
\end{lstlisting}
The functions \lstinline$f()$ and \lstinline$g()$ have compatible types. Let \(f\) and \(g\) be
their types; then \(f_1\) = \lstinline$T$, \(f_2\) = \lstinline$U$, \(f_3\) = \lstinline$V$, \(g_1\)
= \lstinline$V$, \(g_2\) = \lstinline$U$, and \(g_3\) = \lstinline$W$. Replacing every \(f_i\)
by \(g_i\) in \(f\) gives
\begin{lstlisting}
forall( type V, type U, type W ) U * f( V *, U, W * const );
\end{lstlisting}
which has a return type and parameter list that is compatible with \(g\).
\begin{rationale}
The word ``\lstinline$type$'' in a forall specifier is redundant at the moment, but I want to leave
room for inferred parameters of ordinary types in case parameterized types get added one day.

Even without parameterized types, I might try to allow
\begin{lstlisting}
forall( int n ) int sum( int vector[n] );
\end{lstlisting}
but C currently rewrites array parameters as pointer parameters, so the effects of such a change
require more thought.
\end{rationale}

\begin{rationale}
A polymorphic declaration must do two things: it must introduce type parameters, and it must apply
assertions to those types. Adding this to existing C declaration syntax and semantics was delicate,
and not entirely successful.

C depends on declaration-before-use, so a forall specifier must introduce type names before they can
be used in the declaration specifiers. This could be done by making the forall specifier part of
the declaration specifiers, or by making it a new introductory clause of declarations.

Assertions are also part of polymorphic function types, because it must be clear which functions
have access to the assertion parameters declared by the assertions. All attempts to put assertions
inside an introductory clause produced complex semantics and confusing code. Building them into the
declaration specifiers could be done by placing them in the function's parameter list, or in a
forall specifier that is a declaration specifier. Assertions are also used with type parameters of
specifications, and by type declarations. For consistency's sake it seems best to attach assertions
to the type declarations in forall specifiers, which means that forall specifiers must be
declaration specifiers.
\end{rationale}
%HERE


\subsection{Type qualifiers}

\CFA defines a new type qualifier \lstinline$lvalue$\impl{lvalue}\index{lvalue}.
\begin{syntax}
\oldlhs{type-qualifier}
\rhs \lstinline$lvalue$
\end{syntax}

\constraints
\lstinline$restrict$\index{register@{\lstinline$restrict$}} Types other than type parameters and
pointer types whose referenced type is an object type shall not be restrict-qualified.

\semantics
An object's type may be a restrict-qualified type parameter. \lstinline$restrict$ does not
establish any special semantics in that case.

\begin{rationale}
\CFA loosens the constraint on the restrict qualifier so that restrict-qualified pointers may be
passed to polymorphic functions.
\end{rationale}

\lstinline$lvalue$ may be used to qualify the return type of a function type. Let \lstinline$T$ be
an unqualified version of a type; then the result of calling a function with return type
\lstinline$lvalue T$ is a modifiable lvalue\index{modifiable lvalue} of type \lstinline$T$.
\lstinline$const$\use{const} and \lstinline$volatile$\use{volatile} qualifiers may also be added to
indicate that the function result is a constant or volatile lvalue.
\begin{rationale}
The \lstinline$const$ and \lstinline$volatile$ qualifiers can only be sensibly used to qualify the
return type of a function if the \lstinline$lvalue$ qualifier is also used.
\end{rationale}

An {lvalue}-qualified type may be used in a cast expression\index{cast expression} if the operand is
an lvalue; the result of the expression is an lvalue.

\begin{rationale}
\lstinline$lvalue$ provides some of the functionality of {\CC}'s ``\lstinline$T&$'' ( reference to
object of type \lstinline$T$) type. Reference types have four uses in {\CC}.
\begin{itemize}
\item
They are necessary for user-defined operators that return lvalues, such as ``subscript'' and
``dereference''.

\item
A reference can be used to define an alias for a complicated lvalue expression, as a way of getting
some of the functionality of the Pascal \lstinline$with$ statement. The following {\CC} code gives
an example.
\begin{lstlisting}
{
        char &code = long_name.some_field[i].data->code;
        code = toupper( code );
}
\end{lstlisting}
This is not very useful.

\item
A reference parameter can be used to allow a function to modify an argument without forcing the
caller to pass the address of the argument. This is most useful for user-defined assignment
operators. In {\CC}, plain assignment is done by a function called ``\lstinline$operator=$'', and
the two expressions
\begin{lstlisting}
a = b;
operator=( a, b );
\end{lstlisting}
are equivalent. If \lstinline$a$ and \lstinline$b$ are of type \lstinline$T$, then the first
parameter of \lstinline$operator=$ must have type ``\lstinline$T&$''. It cannot have type
\lstinline$T$, because then assignment couldn't alter the variable, and it can't have type
``\lstinline$T *$'', because the assignment would have to be written ``\lstinline$&a = b;$''.

In the case of user-defined operators, this could just as well be handled by using pointer types and
by changing the rewrite rules so that ``\lstinline$a = b;$'' is equivalent to
``\lstinline$operator=(&( a), b )$''. Reference parameters of ``normal'' functions are Bad Things,
because they remove a useful property of C function calls: an argument can only be modified by a
function if it is preceded by ``\lstinline$&$''.

\item
References to const-qualified\index{const-qualified} types can be used instead of value parameters.
Given the {\CC} function call ``\lstinline$fiddle( a_thing )$'', where the type of
\lstinline$a_thing$ is \lstinline$Thing$, the type of \lstinline$fiddle$ could be either of
\begin{lstlisting}
void fiddle( Thing );
void fiddle( const Thing & );
\end{lstlisting}
If the second form is used, then constructors and destructors are not invoked to create a temporary
variable at the call site ( and it is bad style for the caller to make any assumptions about such
things), and within \lstinline$fiddle$ the parameter is subject to the usual problems caused by
aliases. The reference form might be chosen for efficiency's sake if \lstinline$Thing$s are too
large or their constructors or destructors are too expensive. An implementation may switch between
them without causing trouble for well-behaved clients. This leaves the implementor to define ``too
large'' and ``too expensive''.

I propose to push this job onto the compiler by allowing it to implement
\begin{lstlisting}
void fiddle( const volatile Thing );
\end{lstlisting}
with call-by-reference. Since it knows all about the size of \lstinline$Thing$s and the parameter
passing mechanism, it should be able to come up with a better definition of ``too large'', and may
be able to make a good guess at ``too expensive''.
\end{itemize}

In summary, since references are only really necessary for returning lvalues, I'll only provide
lvalue functions.
\end{rationale}


\setcounter{subsection}{8}
\subsection{Initialization}

An expression that is used as an \nonterm{initializer} is treated as being cast to the type of the
object being initialized. An expression used in an \nonterm{initializer-list} is treated as being
cast to the type of the aggregate member that it initializes. In either case the cast must have a
single unambiguous interpretation\index{interpretations}.


\setcounter{subsection}{10}
\subsection{Specification definitions}

\begin{syntax}
\lhs{spec-definition}
\rhs \lstinline$spec$ \nonterm{identifier} 
        \lstinline$($ \nonterm{type-parameter-list} \lstinline$)$
        \lstinline${$ \nonterm{spec-declaration-list}\opt \lstinline$}$
\lhs{spec-declaration-list}
\rhs \nonterm{spec-declaration} \lstinline$;$
\rhs \nonterm{spec-declaration-list} \nonterm{spec-declaration} \lstinline$;$
\lhs{spec-declaration}
\rhs \nonterm{specifier-qualifier-list} \nonterm{declarator-list}
\lhs{declarator-list}
\rhs \nonterm{declarator}
\rhs \nonterm{declarator-list} \lstinline$,$ \nonterm{declarator}
\end{syntax}
\begin{rationale}
The declarations allowed in a specification are much the same as those allowed in a structure,
except that bit fields are not allowed, and incomplete types\index{incomplete types} and function
types are allowed.
\end{rationale}

\semantics
A \define{specification definition} defines a name for a \define{specification}: a parameterized
collection of object and function declarations.

The declarations in a specification consist of the declarations in the
\nonterm{spec-declaration-list} and declarations produced by any assertions in the
\nonterm{spec-parameter-list}. If the collection contains two declarations that declare the same
identifier and have compatible types, they are combined into one declaration with the composite type
constructed from the two types.


\subsubsection{Assertions}
\begin{syntax}
\lhs{assertion-list}
\rhs \nonterm{assertion}
\rhs \nonterm{assertion-list} \nonterm{assertion}
\lhs{assertion}
\rhs \lstinline$|$ \nonterm{identifier} \lstinline$($ \nonterm{type-name-list} \lstinline$)$
\rhs \lstinline$|$ \nonterm{spec-declaration}
\lhs{type-name-list}
\rhs \nonterm{type-name}
\rhs \nonterm{type-name-list} \lstinline$,$ \nonterm{type-name}
\end{syntax}

\constraints
The \nonterm{identifier} in an assertion that is not a \nonterm{spec-declaration} shall be the name
of a specification. The \nonterm{type-name-list} shall contain one \nonterm{type-name} argument for
each \nonterm{type-parameter} in that specification's \nonterm{spec-parameter-list}. If the
\nonterm{type-parameter} uses type-class \lstinline$type$\use{type}, the argument shall be the type
name of an object type\index{object types}; if it uses \lstinline$dtype$, the argument shall be the
type name of an object type or an incomplete type\index{incomplete types}; and if it uses
\lstinline$ftype$, the argument shall be the type name of a function type\index{function types}.

\semantics
An \define{assertion} is a declaration of a collection of objects and functions, called
\define{assertion parameters}.

The assertion parameters produced by an assertion that applies the name of a specification to type
arguments are found by taking the declarations specified in the specification and treating each of
the specification's parameters as a synonym for the corresponding \nonterm{type-name} argument.

The collection of assertion parameters produced by the \nonterm{assertion-list} are found by
combining the declarations produced by each assertion. If the collection contains two declarations
that declare the same identifier and have compatible types, they are combined into one declaration
with the composite type\index{composite type} constructed from the two types.

\examples
\begin{lstlisting}
forall( type T | T ?*?( T, T ))@\use{?*?}@
T square( T val ) {@\impl{square}@
        return val + val;
}

context summable( type T ) {@\impl{summable}@
        T ?+=?( T *, T );@\use{?+=?}@
        const T 0;@\use{0}@
};
context list_of( type List, type Element ) {@\impl{list_of}@
        Element car( List );
        List cdr( List );
        List cons( Element, List );
        List nil;
        int is_nil( List );
};
context sum_list( type List, type Element | summable( Element ) | list_of( List, Element ) ) {};
\end{lstlisting}
\lstinline$sum_list$ contains seven declarations, which describe a list whose elements can be added
up. The assertion ``\lstinline$|sum_list( i_list, int )$''\use{sum_list} produces the assertion
parameters
\begin{lstlisting}
int ?+=?( int *, int );
const int 0;
int car( i_list );
i_list cdr( i_list );
i_list cons( int, i_list );
i_list nil;
int is_nil;
\end{lstlisting}

\subsection{Type declarations}
\begin{syntax}
\lhs{type-parameter-list}
\rhs \nonterm{type-parameter}
\rhs \nonterm{type-parameter-list} \lstinline$,$ \nonterm{type-parameter}
\lhs{type-parameter}
\rhs \nonterm{type-class} \nonterm{identifier} \nonterm{assertion-list}\opt
\lhs{type-class}
\rhs \lstinline$type$
\rhs \lstinline$dtype$
\rhs \lstinline$ftype$
\lhs{type-declaration}
\rhs \nonterm{storage-class-specifier}\opt \lstinline$type$ \nonterm{type-declarator-list} \verb|;|
\lhs{type-declarator-list}
\rhs \nonterm{type-declarator}
\rhs \nonterm{type-declarator-list} \lstinline$,$ \nonterm{type-declarator}
\lhs{type-declarator}
\rhs \nonterm{identifier} \nonterm{assertion-list}\opt \lstinline$=$ \nonterm{type-name}
\rhs \nonterm{identifier} \nonterm{assertion-list}\opt
\end{syntax}

\constraints
If a type declaration has block scope, and the declared identifier has external or internal linkage,
the declaration shall have no initializer for the identifier.

\semantics
A \nonterm{type-parameter} or a \nonterm{type-declarator} declares an identifier to be a type
name\index{type names} for a type incompatible with all other types.

An identifier declared by a \nonterm{type-parameter} has no linkage\index{no linkage}. Identifiers
declared with type-class \lstinline$type$\use{type} are object types\index{object types}; those
declared with type-class \lstinline$dtype$\use{dtype} are incomplete types\index{incomplete types};
and those declared with type-class \lstinline$ftype$\use{ftype} are function types\index{function
 types}. The identifier has block scope\index{block scope} that terminates at the end of the
\nonterm{spec-declaration-list} or polymorphic function that contains the \nonterm{type-parameter}.

A \nonterm{type-declarator} with an initializer\index{initializer} is a \define{type definition}.
The declared identifier is an incomplete type\index{incomplete types} within the initializer, and an
object type\index{object types} after the end of the initializer. The type in the initializer is
called the \define{implementation type}. Within the scope of the declaration, implicit
conversions\index{implicit conversions} can be performed between the defined type and the
implementation type, and between pointers to the defined type and pointers to the implementation
type.

A type declaration without an initializer\index{initializer} and without a storage-class
specifier\index{storage-class specifiers} or with storage-class specifier
\lstinline$static$\use{static} defines an incomplete type\index{incomplete types}. If a translation
unit\index{translation unit} or block \index{block} contains one or more such declarations for an
identifier, it must contain exactly one definition of the identifier ( but not in an enclosed block,
which would define a new type known only within that block).
\begin{rationale}
Incomplete type declarations allow compact mutually-recursive types.
\begin{lstlisting}
type t1; // Incomplete type declaration. 
type t2 = struct { t1 * p; ... };
type t1 = struct { t2 * p; ... };
\end{lstlisting}
Without them, mutual recursion could be handled by declaring mutually recursive structures, then
initializing the types to those structures.
\begin{lstlisting}
struct s1;
type t2 = struct s2 { struct s1 * p; ... };
type t1 = struct s1 { struct s2 * p; ... };
\end{lstlisting}
This introduces extra names, and may force the programmer to cast between the types and their
implementations.
\end{rationale}

A type declaration without an initializer and with storage-class specifier \index{storage-class
 specifiers} \lstinline$extern$\use{extern} is an \define{opaque type declaration}. Opaque types
are object types\index{object types}. An opaque type is not a \nonterm{constant-expression};
neither is a structure or union that has a member whose type is not a \nonterm{constant-expression}.
Every other object type\index{object types} is a \nonterm{constant-expression}. Objects with static
storage duration shall be declared with a type that is a \nonterm{constant-expression}.
\begin{rationale}
Type declarations can declare identifiers with external linkage, whereas typedef declarations
declare identifiers that only exist within a translation unit. These opaque types can be used in
declarations, but the implementation of the type is not visible.

Static objects can not have opaque types because space for them would have to be allocated at
program start-up. This is a deficiency\index{deficiencies!static opaque objects}, but I don't want
to deal with ``module initialization'' code just now.
\end{rationale}

An incomplete type\index{incomplete types} which is not a qualified version\index{qualified type} of
a type is a value of type-class\index{type-class} \lstinline$dtype$. An object type\index{object
 types} which is not a qualified version of a type is a value of type-classes \lstinline$type$ and
\lstinline$dtype$. A function type\index{function types} is a value of type-class
\lstinline$ftype$.
\begin{rationale}
Syntactically, a type value is a \nonterm{type-name}, which is a declaration for an object which
omits the identifier being declared.

Object types are precisely the types that can be instantiated. Type qualifiers are not included in
type values because the compiler needs the information they provide at compile time to detect
illegal statements or to produce efficient machine instructions. For instance, the code that a
compiler must generate to manipulate an object that has volatile-qualified type may be different
from the code to manipulate an ordinary object.

Type qualifiers are a weak point of C's type system. Consider the standard library function
\lstinline$strchr()$ which, given a string and a character, returns a pointer to the first
occurrence of the character in the string.
\begin{lstlisting}
char *strchr( const char *s, int c ) {@\impl{strchr}@
        char real_c = c; // done because c was declared as int. 
        for ( ; *s != real_c; s++ )
         if ( *s == '\0' ) return NULL;
        return ( char * )s;
}
\end{lstlisting}
The parameter \lstinline$s$ must be \lstinline$const char *$, because \lstinline$strchr()$ might be
used to search a constant string, but the return type must be \lstinline$char *$, because the result
might be used to modify a non-constant string. Hence the body must perform a cast, and ( even worse)
\lstinline$strchr()$ provides a type-safe way to attempt to modify constant strings. What is needed
is some way to say that \lstinline$s$'s type might contain qualifiers, and the result type has
exactly the same qualifiers. Polymorphic functions do not provide a fix for this
deficiency\index{deficiencies!pointers to qualified types}, because type qualifiers are not part of
type values. Instead, overloading can be used to define \lstinline$strchr()$ for each combination
of qualifiers.
\end{rationale}

\begin{rationale}
Since incomplete types\index{incomplete types} are not type values, they can not be used as the
initializer in a type declaration, or as the type of a structure or union member. This prevents the
declaration of types that contain each other.
\begin{lstlisting}
type t1;
type t2 = t1; // illegal: incomplete type t1. 
type t1 = t2;
\end{lstlisting}

The initializer in a file-scope declaration must be a constant expression. This means type
declarations can not build on opaque types, which is a deficiency\index{deficiencies!nesting opaque
 types}.
\begin{lstlisting}
extern type Huge; // extended-precision integer type. 
type Rational = struct {
        Huge numerator, denominator;    // illegal 
};
struct Pair {
        Huge first, second;                             // legal 
};
\end{lstlisting}
Without this restriction, \CFA might require ``module initialization'' code ( since
\lstinline$Rational$ has external linkage, it must be created before any other translation unit
instantiates it), and would force an ordering on the initialization of the translation unit that
defines \lstinline$Huge$ and the translation that declares \lstinline$Rational$.

A benefit of the restriction is that it prevents the declaration in separate translation units of
types that contain each other, which would be hard to prevent otherwise.
\begin{lstlisting}
//  File a.c:
        extern type t1;
        type t2 = struct { t1 f1; ... } // illegal 
//  File b.c:
        extern type t2;
        type t1 = struct { t2 f2; ... } // illegal 
\end{lstlisting}
\end{rationale}

\begin{rationale}
Since a \nonterm{type-declaration} is a \nonterm{declaration} and not a
\nonterm{struct-declaration}, type declarations can not be structure members. The form of
\nonterm{type-declaration} forbids arrays of, pointers to, and functions returning \lstinline$type$.
Hence the syntax of \nonterm{type-specifier} does not have to be extended to allow type-valued
expressions. It also side-steps the problem of type-valued expressions producing different values
in different declarations.

Since a type declaration is not a \nonterm{parameter-declaration}, functions can not have explicit
type parameters. This may be too restrictive, but it attempts to make compilation simpler. Recall
that when traditional C scanners read in an identifier, they look it up in the symbol table to
determine whether or not it is a typedef name, and return a ``type'' or ``identifier'' token
depending on what they find. A type parameter would add a type name to the current scope. The
scope manipulations involved in parsing the declaration of a function that takes function pointer
parameters and returns a function pointer may just be too complicated.

Explicit type parameters don't seem to be very useful, anyway, because their scope would not include
the return type of the function. Consider the following attempt to define a type-safe memory
allocation function.
\begin{lstlisting}
#include <stdlib.h>
T * new( type T ) { return ( T * )malloc( sizeof( T) ); };
@\ldots@
int * ip = new( int );
\end{lstlisting}
This looks sensible, but \CFA's declaration-before-use rules mean that ``\lstinline$T$'' in the
function body refers to the parameter, but the ``\lstinline$T$'' in the return type refers to the
meaning of \lstinline$T$ in the scope that contains \lstinline$new$; it could be undefined, or a
type name, or a function or variable name. Nothing good can result from such a situation.
\end{rationale}

\examples
Since type declarations create new types, instances of types are always passed by value.
\begin{lstlisting}
type A1 = int[2];
void f1( A1 a ) { a[0] = 0; };
typedef int A2[2];
void f2( A2 a ) { a[0] = 0; };
A1 v1;
A2 v2;
f1( v1 );
f2( v2 );
\end{lstlisting}
\lstinline$V1$ is passed by value, so \lstinline$f1()$'s assignment to \lstinline$a[0]$ does not
modify v1.  \lstinline$V2$ is converted to a pointer, so \lstinline$f2()$ modifies
\lstinline$v2[0]$.

A translation unit containing the declarations
\begin{lstlisting}
extern type Complex;@\use{Complex}@ // opaque type declaration. 
extern float abs( Complex );@\use{abs}@
\end{lstlisting}
can contain declarations of complex numbers, which can be passed to \lstinline$abs$. Some other
translation unit must implement \lstinline$Complex$ and \lstinline$abs$. That unit might contain
the declarations
\begin{lstlisting}
type Complex = struct { float re, im; };@\impl{Complex}@
Complex cplx_i = { 0.0, 1.0 };@\impl{cplx_i}@
float abs( Complex c ) {@\impl{abs( Complex )}@
        return sqrt( c.re * c.re + c.im * c.im );
}
\end{lstlisting}
Note that \lstinline$c$ is implicitly converted to a \lstinline$struct$ so that its components can
be retrieved.

\begin{lstlisting}
type Time_of_day = int;@\impl{Time_of_day}@ // seconds since midnight. 
Time_of_day ?+?( Time_of_day t1, int seconds ) {@\impl{?+?}@
        return (( int)t1 + seconds ) % 86400;
}
\end{lstlisting}
\lstinline$t1$ must be cast to its implementation type to prevent infinite recursion.

\begin{rationale}
Within the scope of a type definition, an instance of the type can be viewed as having that type or
as having the implementation type. In the \lstinline$Time_of_day$ example, the difference is
important. Different languages have treated the distinction between the abstraction and the
implementation in different ways.
\begin{itemize}
\item
Inside a Clu cluster \cite{clu}, the declaration of an instance states which view applies. Two
primitives called \lstinline$up$ and \lstinline$down$ can be used to convert between the views.
\item
The Simula class \cite{Simula87} is essentially a record type. Since the only operations on a
record are member selection and assignment, which can not be overloaded, there is never any
ambiguity as to whether the abstraction or the implementation view is being used. In {\CC}
\cite{c++}, operations on class instances include assignment and ``\lstinline$&$'', which can be
overloaded. A ``scope resolution'' operator can be used inside the class to specify whether the
abstract or implementation version of the operation should be used.
\item
An Ada derived type definition \cite{ada} creates a new type from an old type, and also implicitly
declares derived subprograms that correspond to the existing subprograms that use the old type as a
parameter type or result type. The derived subprograms are clones of the existing subprograms with
the old type replaced by the derived type. Literals and aggregates of the old type are also cloned.
In other words, the abstract view provides exactly the same operations as the implementation view.
This allows the abstract view to be used in all cases.

The derived subprograms can be replaced by programmer-specified subprograms. This is an exception
to the normal scope rules, which forbid duplicate definitions of a subprogram in a scope. In this
case, explicit conversions between the derived type and the old type can be used.
\end{itemize}
\CFA's rules are like Clu's, except that implicit conversions and
conversion costs allow it to do away with most uses of \lstinline$up$ and \lstinline$down$. 
\end{rationale}


\subsubsection{Default functions and objects}

A declaration\index{type declaration} of a type identifier \lstinline$T$ with type-class
\lstinline$type$ implicitly declares a \define{default assignment} function
\lstinline$T ?=?( T *, T )$\use{?=?}, with the same scope\index{scopes} and linkage\index{linkage} as
the identifier \lstinline$T$.
\begin{rationale}
Assignment is central to C's imperative programming style, and every existing C object type has
assignment defined for it ( except for array types, which are treated as pointer types for purposes
of assignment). Without this rule, nearly every inferred type parameter would need an accompanying
assignment assertion parameter. If a type parameter should not have an assignment operation,
\lstinline$dtype$ should be used. If a type should not have assignment defined, the user can define
an assignment function that causes a run-time error, or provide an external declaration but no
definition and thus cause a link-time error.
\end{rationale}

A definition\index{type definition} of a type identifier \lstinline$T$ with implementation
type\index{implementation type} \lstinline$I$ and type-class \lstinline$type$ implicitly defines a
default assignment function. A definition\index{type definition} of a type identifier \lstinline$T$
with implementation type \lstinline$I$ and an assertion list implicitly defines \define{default
 functions} and \define{default objects} as declared by the assertion declarations. The default
objects and functions have the same scope\index{scopes} and linkage\index{linkage} as the identifier
\lstinline$T$. Their values are determined as follows:
\begin{itemize}
\item
If at the definition of \lstinline$T$ there is visible a declaration of an object with the same name
as the default object, and if the type of that object with all occurrence of \lstinline$I$ replaced
by \lstinline$T$ is compatible with the type of the default object, then the default object is
initialized with that object. Otherwise the scope of the declaration of \lstinline$T$ must contain
a definition of the default object.

\item 
If at the definition of \lstinline$T$ there is visible a declaration of a function with the same
name as the default function, and if the type of that function with all occurrence of \lstinline$I$
replaced by \lstinline$T$ is compatible with the type of the default function, then the default
function calls that function after converting its arguments and returns the converted result.

Otherwise, if \lstinline$I$ contains exactly one anonymous member\index{anonymous member} such that
at the definition of \lstinline$T$ there is visible a declaration of a function with the same name
as the default function, and the type of that function with all occurrences of the anonymous
member's type in its parameter list replaced by \lstinline$T$ is compatible with the type of the
default function, then the default function calls that function after converting its arguments and
returns the result.

Otherwise the scope of the declaration of \lstinline$T$ must contain a definition of the default
function.
\end{itemize}
\begin{rationale}
Note that a pointer to a default function will not compare as equal to a pointer to the inherited
function.
\end{rationale}

A function or object with the same type and name as a default function or object that is declared
within the scope of the definition of \lstinline$T$ replaces the default function or object.

\examples
\begin{lstlisting}
context s( type T ) {
        T a, b;
}
struct impl { int left, right; } a = { 0, 0 };
type Pair | s( Pair ) = struct impl;
Pair b = { 1, 1 };
\end{lstlisting}
The definition of \lstinline$Pair$ implicitly defines two objects \lstinline$a$ and \lstinline$b$.
\lstinline$Pair a$ inherits its value from the \lstinline$struct impl a$. The definition of
\lstinline$Pair b$ is compulsory because there is no \lstinline$struct impl b$ to construct a value
from.
\begin{lstlisting}
context ss( type T ) {
        T clone( T );
        void munge( T * );
}
type Whatsit | ss( Whatsit );@\use{Whatsit}@
type Doodad | ss( Doodad ) = struct doodad {@\use{Doodad}@
        Whatsit; // anonymous member 
        int extra;
};
Doodad clone( Doodad ) { ... }
\end{lstlisting}
The definition of \lstinline$Doodad$ implicitly defines three functions:
\begin{lstlisting}
Doodad ?=?( Doodad *, Doodad );
Doodad clone( Doodad );
void munge( Doodad * );
\end{lstlisting}
The assignment function inherits \lstinline$struct doodad$'s assignment function because the types
match when \lstinline$struct doodad$ is replaced by \lstinline$Doodad$ throughout.
\lstinline$munge()$ inherits \lstinline$Whatsit$'s \lstinline$munge()$ because the types match when
\lstinline$Whatsit$ is replaced by \lstinline$Doodad$ in the parameter list. \lstinline$clone()$
does \emph{not} inherit \lstinline$Whatsit$'s \lstinline$clone()$: replacement in the parameter
list yields ``\lstinline$Whatsit clone( Doodad )$'', which is not compatible with
\lstinline$Doodad$'s \lstinline$clone()$'s type. Hence the definition of
``\lstinline$Doodad clone( Doodad )$'' is necessary.

Default functions and objects are subject to the normal scope rules.
\begin{lstlisting}
type T = @\ldots@;
T a_T = @\ldots@;               // Default assignment used. 
T ?=?( T *, T );
T a_T = @\ldots@;               // Programmer-defined assignment called. 
\end{lstlisting}
\begin{rationale}
A compiler warning would be helpful in this situation.
\end{rationale}

\begin{rationale}
The \emph{class} construct of object-oriented programming languages performs three independent
functions. It \emph{encapsulates} a data structure; it defines a \emph{subtype} relationship, whereby
instances of one class may be used in contexts that require instances of another; and it allows one
class to \emph{inherit} the implementation of another.

In \CFA, encapsulation is provided by opaque types and the scope rules, and subtyping is provided
by specifications and assertions. Inheritance is provided by default functions and objects.
\end{rationale}


\section{Statements and blocks}
Many statements contain expressions, which may have more than one interpretation. The following
sections describe how the \CFA translator selects an interpretation. In all cases the result of
the selection shall be a single unambiguous interpretation\index{interpretations}.


\setcounter{subsection}{2}
\subsection{Expression and null statements}

The expression in an expression statement is treated as being cast to \lstinline$void$.


\subsection{Selection statements}

The controlling expression \lstinline$E$ in the switch statement
\begin{lstlisting}
switch ( E ) ...
\end{lstlisting}
may have more than one interpretation, but it shall have only one interpretation with an integral
type. An integer promotion\index{integer promotion} is performed on the expression if necessary.
The constant expressions in \lstinline$case$ statements with the switch are converted to the
promoted type.


\subsection{Iteration statements}

The controlling expression \lstinline$E$ in the loops
\begin{lstlisting}
if ( E ) ...
while ( E ) ...
do ... while ( E );
\end{lstlisting}
is treated as ``\lstinline$( int )((E)!=0)$''.

The statement 
\begin{lstlisting}
for ( a; b; c ) @\ldots@
\end{lstlisting}
is treated as
\begin{lstlisting}
for ( ( void )( a ); ( int )(( b )!=0); ( void )( c ) ) @\ldots@
\end{lstlisting}


\subsection{Jump statements}

An expression in a \lstinline$return$ statement is treated as being 
cast to the result type of the function.


\setcounter{section}{9}
\section{Preprocessing directives}


\setcounter{subsection}{7}
\subsection{Predefined macro names}

The implementation shall define the macro names \lstinline$__LINE__$, \lstinline$__FILE__$,
\lstinline$__DATE__$, and \lstinline$__TIME__$, as in the {\c11} standard. It shall not define the
macro name \lstinline$__STDC__$.

In addition, the implementation shall define the macro name \lstinline$__CFORALL__$ to be the
decimal constant 1.


\appendix

\chapter{Examples}

\section{C types}
This section gives example specifications for some groups of types that are important in the C
language, in terms of the predefined operations that can be applied to those types.


\subsection{Scalar, arithmetic, and integral types}

The pointer, integral, and floating-point types are all \define{scalar types}. All of these types
can be logically negated and compared. The assertion ``\lstinline$scalar( Complex )$'' should be read
as ``type \lstinline$Complex$ is scalar''.
\begin{lstlisting}
context scalar( type T ) {@\impl{scalar}@
        int !?( T );
        int ?<?( T, T ), ?<=?( T, T ), ?==?( T, T ), ?>=?( T, T ), ?>?( T, T ), ?!=?( T, T );
};
\end{lstlisting}

The integral and floating-point types are \define{arithmetic types}, which support the basic
arithmetic operators. The use of an assertion in the \nonterm{spec-parameter-list} declares that,
in order to be arithmetic, a type must also be scalar ( and hence that scalar operations are
available ). This is equivalent to inheritance of specifications.
\begin{lstlisting}
context arithmetic( type T | scalar( T ) ) {@\impl{arithmetic}@@\use{scalar}@
        T +?( T ), -?( T );
        T ?*?( T, T ), ?/?( T, T ), ?+?( T, T ), ?-?( T, T );
};
\end{lstlisting}

The various flavors of \lstinline$char$ and \lstinline$int$ and the enumerated types make up the
\define{integral types}.
\begin{lstlisting}
context integral( type T | arithmetic( T ) ) {@\impl{integral}@@\use{arithmetic}@
        T ~?( T );
        T ?&?( T, T ), ?|?( T, T ), ?^?( T, T );
        T ?%?( T, T );
        T ?<<?( T, T ), ?>>?( T, T );
};
\end{lstlisting}


\subsection{Modifiable types}
\index{modifiable lvalue}

The only operation that can be applied to all modifiable lvalues is simple assignment.
\begin{lstlisting}
context m_lvalue( type T ) {@\impl{m_lvalue}@
        T ?=?( T *, T );
};
\end{lstlisting}

Modifiable scalar lvalues are scalars and are modifiable lvalues, and assertions in the
\nonterm{spec-parameter-list} reflect those relationships. This is equivalent to multiple
inheritance of specifications. Scalars can also be incremented and decremented.
\begin{lstlisting}
context m_l_scalar( type T | scalar( T ) | m_lvalue( T ) ) {@\impl{m_l_scalar}@
        T ?++( T * ), ?--( T * );@\use{scalar}@@\use{m_lvalue}@
        T ++?( T * ), --?( T * );
};
\end{lstlisting}

Modifiable arithmetic lvalues are both modifiable scalar lvalues and arithmetic. Note that this
results in the ``inheritance'' of \lstinline$scalar$ along both paths.
\begin{lstlisting}
context m_l_arithmetic( type T | m_l_scalar( T ) | arithmetic( T ) ) {@\impl{m_l_arithmetic}@
        T ?/=?( T *, T ), ?*=?( T *, T );@\use{m_l_scalar}@@\use{arithmetic}@
        T ?+=?( T *, T ), ?-=?( T *, T );
};

context m_l_integral( type T | m_l_arithmetic( T ) | integral( T ) ) {@\impl{m_l_integral}@
        T ?&=?( T *, T ), ?|=?( T *, T ), ?^=?( T *, T );@\use{m_l_arithmetic}@
        T ?%=?( T *, T ), ?<<=?( T *, T ), ?>>=?( T *, T );@\use{integral}@
};
\end{lstlisting}


\subsection{Pointer and array types}

Array types can barely be said to exist in {\c11}, since in most cases an array name is treated as a
constant pointer to the first element of the array, and the subscript expression
``\lstinline$a[i]$'' is equivalent to the dereferencing expression ``\lstinline$(*( a+( i )))$''.
Technically, pointer arithmetic and pointer comparisons other than ``\lstinline$==$'' and
``\lstinline$!=$'' are only defined for pointers to array elements, but the type system does not
enforce those restrictions. Consequently, there is no need for a separate ``array type''
specification.

Pointer types are scalar types. Like other scalar types, they have ``\lstinline$+$'' and
``\lstinline$-$'' operators, but the types do not match the types of the operations in
\lstinline$arithmetic$, so these operators cannot be consolidated in \lstinline$scalar$.
\begin{lstlisting}
context pointer( type P | scalar( P ) ) {@\impl{pointer}@@\use{scalar}@
        P ?+?( P, long int ), ?+?( long int, P ), ?-?( P, long int );
        ptrdiff_t ?-?( P, P );
};

context m_l_pointer( type P | pointer( P ) | m_l_scalar( P ) ) {@\impl{m_l_pointer}@
        P ?+=?( P *, long int ), ?-=?( P *, long int );
        P ?=?( P *, void * );
        void * ?=?( void **, P );
};
\end{lstlisting}

Specifications that define the dereference operator ( or subscript operator ) require two parameters,
one for the pointer type and one for the pointed-at ( or element ) type. Different specifications are
needed for each set of type qualifiers\index{type qualifiers}, because qualifiers are not included
in types. The assertion ``\lstinline$|ptr_to( Safe_pointer, int )$'' should be read as
``\lstinline$Safe_pointer$ acts like a pointer to \lstinline$int$''.
\begin{lstlisting}
context ptr_to( type P | pointer( P ), type T ) {@\impl{ptr_to}@@\use{pointer}@
        lvalue T *?( P ); lvalue T ?[?]( P, long int );
};

context ptr_to_const( type P | pointer( P ), type T ) {@\impl{ptr_to_const}@
        const lvalue T *?( P ); const lvalue T ?[?]( P, long int );@\use{pointer}@
};

context ptr_to_volatile( type P | pointer( P ), type T ) }@\impl{ptr_to_volatile}@
        volatile lvalue T *?( P ); volatile lvalue T ?[?]( P, long int );@\use{pointer}@
};
\end{lstlisting}
\begin{lstlisting}
context ptr_to_const_volatile( type P | pointer( P ), type T ) }@\impl{ptr_to_const_volatile}@
        const volatile lvalue T *?( P );@\use{pointer}@
        const volatile lvalue T ?[?]( P, long int );
};
\end{lstlisting}

Assignment to pointers is more complicated than is the case with other types, because the target's
type can have extra type qualifiers in the pointed-at type: a ``\lstinline$T *$'' can be assigned to
a ``\lstinline$const T *$'', a ``\lstinline$volatile T *$'', and a ``\lstinline$const volatile T *$''.
Again, the pointed-at type is passed in, so that assertions can connect these specifications to the
``\lstinline$ptr_to$'' specifications.
\begin{lstlisting}
context m_l_ptr_to( type P | m_l_pointer( P ),@\use{m_l_pointer}@@\impl{m_l_ptr_to}@ type T | ptr_to( P, T )@\use{ptr_to}@ {
        P ?=?( P *, T * );
        T * ?=?( T **, P );
};

context m_l_ptr_to_const( type P | m_l_pointer( P ),@\use{m_l_pointer}@@\impl{m_l_ptr_to_const}@ type T | ptr_to_const( P, T )@\use{ptr_to_const}@) {
        P ?=?( P *, const T * );
        const T * ?=?( const T **, P );
};

context m_l_ptr_to_volatile( type P | m_l_pointer( P ),@\use{m_l_pointer}@@\impl{m_l_ptr_to_volatile}@ type T | ptr_to_volatile( P, T )) {@\use{ptr_to_volatile}@
        P ?=?( P *, volatile T * );
        volatile T * ?=?( volatile T **, P );
};

context m_l_ptr_to_const_volatile( type P | ptr_to_const_volatile( P ),@\use{ptr_to_const_volatile}@@\impl{m_l_ptr_to_const_volatile}@
                type T | m_l_ptr_to_volatile( P, T ) | m_l_ptr_to_const( P )) {@\use{m_l_ptr_to_const}@@\use{m_l_ptr_to_volatile}@
        P ?=?( P *, const volatile T * );
        const volatile T * ?=?( const volatile T **, P );
};
\end{lstlisting}

Note the regular manner in which type qualifiers appear in those specifications. An alternative
specification can make use of the fact that qualification of the pointed-at type is part of a
pointer type to capture that regularity.
\begin{lstlisting}
context m_l_ptr_like( type MyP | m_l_pointer( MyP ),@\use{m_l_pointer}@@\impl{m_l_ptr_like}@ type CP | m_l_pointer( CP ) ) {
        MyP ?=?( MyP *, CP );
        CP ?=?( CP *, MyP );
};
\end{lstlisting}
The assertion ``\lstinline$| m_l_ptr_like( Safe_ptr, const int * )$'' should be read as
``\lstinline$Safe_ptr$ is a pointer type like \lstinline$const int *$''. This specification has two
defects, compared to the original four: there is no automatic assertion that dereferencing a
\lstinline$MyP$ produces an lvalue of the type that \lstinline$CP$ points at, and the
``\lstinline$|m_l_pointer( CP )$'' assertion provides only a weak assurance that the argument passed
to \lstinline$CP$ really is a pointer type.


\section{Relationships between operations}

Different operators often have related meanings; for instance, in C, ``\lstinline$+$'',
``\lstinline$+=$'', and the two versions of ``\lstinline$++$'' perform variations of addition.
Languages like {\CC} and Ada allow programmers to define operators for new types, but do not
require that these relationships be preserved, or even that all of the operators be implemented.
Completeness and consistency is left to the good taste and discretion of the programmer. It is
possible to encourage these attributes by providing generic operator functions, or member functions
of abstract classes, that are defined in terms of other, related operators.

In \CFA, polymorphic functions provide the equivalent of these generic operators, and
specifications explicitly define the minimal implementation that a programmer should provide. This
section shows a few examples.


\subsection{Relational and equality operators}

The different comparison operators have obvious relationships, but there is no obvious subset of the
operations to use in the implementation of the others. However, it is usually convenient to
implement a single comparison function that returns a negative integer, 0, or a positive integer if
its first argument is respectively less than, equal to, or greater than its second argument; the
library function \lstinline$strcmp$ is an example.

C and \CFA have an extra, non-obvious comparison operator: ``\lstinline$!$'', logical negation,
returns 1 if its operand compares equal to 0, and 0 otherwise.
\begin{lstlisting}
context comparable( type T ) {
        const T 0;
        int compare( T, T );
}

forall( type T | comparable( T ) ) int ?<?( T l, T r ) {
        return compare( l, r ) < 0;
}
// ... similarly for <=, ==, >=, >, and !=. 

forall( type T | comparable( T ) ) int !?( T operand ) {
        return !compare( operand, 0 );
}
\end{lstlisting}


\subsection{Arithmetic and integer operations}

A complete arithmetic type would provide the arithmetic operators and the corresponding assignment
operators. Of these, the assignment operators are more likely to be implemented directly, because
it is usually more efficient to alter the contents of an existing object than to create and return a
new one. Similarly, a complete integral type would provide integral operations based on integral
assignment operations.
\begin{lstlisting}
context arith_base( type T ) {
        const T 1;
        T ?+=?( T *, T ), ?-=?( T *, T ), ?*=?( T *, T ), ?/=?( T *, T );
}

forall( type T | arith_base( T ) ) T ?+?( T l, T r ) {
        return l += r;
}

forall( type T | arith_base( T ) ) T ?++( T * operand ) {
        T temporary = *operand;
        *operand += 1;
        return temporary;
}

forall( type T | arith_base( T ) ) T ++?( T * operand ) {
        return *operand += 1;
}
// ... similarly for -, --, *, and /. 

context int_base( type T ) {
        T ?&=?( T *, T ), ?|=?( T *, T ), ?^=?( T *, T );
        T ?%=?( T *, T ), ?<<=?( T *, T ), ?>>=?( T *, T );
}

forall( type T | int_base( T ) ) T ?&?( T l, T r ) {
        return l &= r;
}
// ... similarly for |, ^, %, <<, and >>.
\end{lstlisting}

Note that, although an arithmetic type would certainly provide comparison functions, and an integral
type would provide arithmetic operations, there does not have to be any relationship among
\lstinline$int_base$, \lstinline$arith_base$ and \lstinline$comparable$. Note also that these
declarations provide guidance and assistance, but they do not define an absolutely minimal set of
requirements. A truly minimal implementation of an arithmetic type might only provide
\lstinline$0$, \lstinline$1$, and \lstinline$?-=?$, which would be used by polymorphic
\lstinline$?+=?$, \lstinline$?*=?$, and \lstinline$?/=?$ functions.

Note also that \lstinline$short$ is an integer type in C11 terms, but has no operations!


\chapter{TODO}
Review index entries.

Restrict allowed to qualify anything, or type/dtype parameters, but only affects pointers. This gets
into \lstinline$noalias$ territory. Qualifying anything (``\lstinline$short restrict rs$'') means
pointer parameters of \lstinline$?++$, etc, would need restrict qualifiers.

Enumerated types. Constants are not ints. Overloading. Definition should be ``representable as an
integer type'', not ``as an int''. C11 usual conversions freely convert to and from ordinary
integer types via assignment, which works between any integer types. Does enum Color ?*?( enum
Color, enum Color ) really make sense? ?++ does, but it adds (int)1.

Operators on {,signed,unsigned} char and other small types. ?<? harmless; ?*? questionable for
chars. Generic selections make these choices visible. Safe conversion operators? Predefined
``promotion'' function?

\lstinline$register$ assignment might be handled as assignment to a temporary with copying back and
forth, but copying must not be done by assignment.

Don't use ptrdiff\_t by name in the predefineds.

Polymorphic objects. Polymorphic typedefs and type declarations.


\bibliographystyle{plain}
\bibliography{refrat}


\addcontentsline{toc}{chapter}{\indexname} % add index name to table of contents
\begin{theindex}
Italic page numbers give the location of the main entry for the referenced term. Plain page numbers
denote uses of the indexed term. Entries for grammar non-terminals are italicized. A typewriter
font is used for grammar terminals and program identifiers.
\indexspace
\input{refrat.ind}
\end{theindex}

\end{document}

% Local Variables: %
% tab-width: 4 %
% fill-column: 100 %
% compile-command: "make" %
% End: %