\section{Overview}

Camlidl generates stub code for interfacing Caml with C
(as described in chapter ``Interfacing with C'' of the
\footahref{http://caml.inria.fr/ocaml/htmlman/index.html}{Objective Caml
reference manual}) from an IDL description of the C functions to be
made available in Caml.
Thus, Camlidl automates the most tedious task in interfacing C
libraries with Caml programs.  It can also be used to interface Caml
programs with other languages, as long as those languages have a
well-defined C interface.

In addition, Camlidl provides basic support for COM interfaces and
components.  It supports both using COM components (usually written in
C++ or C) from Caml programs, and packaging Caml objects as
COM components that can then be used from C++ or C.

\subsection{What is IDL?}

IDL stands for Interface Description Language.  This is a generic term
for a family of small languages that have been developed to provide
type specifications for libraries written in C and C++.  Those
languages resembles C declarations (as found in C header files), with
extra annotations to provide more precise types for the arguments and
results of the functions.

The particular IDL used by Camlidl is inspired by Microsoft's IDL,
which itself is an extension of the IDL found in DCE (The Open Group's
Distributed Common Environment).  The initial motivation for those IDLs
was to automate the generation of stub code for remote procedure calls
and network objects, where the arguments to the function are marshaled
at the calling site, then sent across the network or through
interprocess communications to a server process, which unmarshals the
arguments, compute the function application, marshal the results,
sends them back to the calling site, where they are unmarshaled and
returned to the caller.  IDLs were also found to be very useful for
inter-language communications, since the same type information that
guides the generation of marshaling stubs can be used to generate
stubs to convert between the data representations of several
languages.

\subsection{What is COM?}

COM is Microsoft's Common Object Model.  It provides a set of
programming conventions as well as system support for packaging
C++ objects as executable components that can be used in other
programs, either by dynamic linking of the component inside the
program, or through interprocess or internetwork communications
between the program and a remote server.  COM components implement one
or several interfaces, (similar to Caml object types or Java
interfaces) identified by unique 128-bit interface identifiers (IIDs).
COM specifies a standard protocol for reference counting of
components, and for asking a component which interfaces it
implements.

While the full range of COM services and third-party components is
available only on Microsoft's Windows operating systems, the basic COM
conventions can also be used on Unix and other operating systems
to exchange objects between Caml and C or C++.  Of particular
interest is the encapsulation of Caml objects as COM components, which
can then be used inside larger C or C++ applications; those
applications do not need to know anything about Caml: they just call
the component methods as if they were C++ methods or C
functions, without knowing that they are actually implemented in Caml.

For more information about COM, see for instance {\em Inside COM}
by Dale Rogerson (Microsoft Press), or the 
\footahref{http://msdn.microsoft.com}{Microsoft developer Web site}.

\section{IDL syntax}

This section describes the syntax of IDL files.  IDL syntax is very
close to that of C declarations, with extra attributes between
brackets adding information to the C types.  The following example
should give the flavor of the syntax:
\begin{verbatim}
        int f([in,string] char * msg);
\end{verbatim}
This reads: ``"f" is a function taking a character string as input and
returning an "int"''.

\subsection{Lexical conventions}

\paragraph{Blanks.} Space, newline, horizontal tabulation, carriage
return, line feed and form feed are considered as blanks.  Blanks are
ignored, but they separate adjacent tokens.

\paragraph{Comments.} Both C-style comments "/* ... */" and
Java-style comments "// ..." are supported.  C-style comments are
introduced by "/*" and terminated by "*/".  Java-style comments are
introduced by "//" and extend to the end of the line.  Comments are
treated as blank characters.  Comments do not occur inside string or
character literals. Nested C-style comments are not supported.

\paragraph{Identifiers.} Identifiers have the same syntax as in C.
\begin{syntax}
ident: ("A" \ldots "Z" || "a" \ldots "z" || "_")
       {"A" \ldots "Z" || "a" \ldots "z" || "0" \ldots "9" || "_"};
\end{syntax}

\paragraph{Literals.}  Integer literals, character literals and string
literals have the same syntax as in C.
\begin{syntax}
integer: ['-'] {'0' \ldots '9'}
         ['-'] '0x' {'0' \ldots '9' || 'a' \ldots 'f' || 'A' \ldots 'F'}
         ['-'] '0' {'0' \ldots '7'};
character: "'" (regular-char || escape-char) "'";
string: '"' {regular-char || escape-char} '"';
escape-char: '\' ('b'||'n'||'r'||'t') 
           | '\' ('0'\ldots'7') ['0'\ldots'7'] ['0'\ldots'7'];
\end{syntax}

\paragraph{UUID.}  Unique identifiers are composed of 16 hexadecimal
digits, in groups of 8, 4, 4, 4 and 12, separated by dashes.
\begin{syntax}
uuid: hex^8 '-' hex^4 '-' hex^4 '-' hex^4 '-' hex^4 hex^4 hex^4;
hex: '0' \ldots '9' || 'a' \ldots 'f' || 'A' \ldots 'F';
\end{syntax}

\subsection{Limited expressions}

Limited expressions are similar to C expressions, with the omission of
assignment operators ("=", "+=", etc), and the addition of the
unsigned (logical) right shift operator ">>>".  Operators have the same
precedences and associativities as in C.  They are listed below in
decreasing priority order.
\begin{syntax} 
lexpr:
  ident
| integer
| character
| 'true'
| 'false'
| string
| sizeof '(' type-expr ')'
| '(' lexpr ')'
| lexpr ('.' || '->') ident
| '(' type-expr ')' lexpr
| ('&'||'*'||'!'||'~'||'-'||'+') lexpr
| lexpr ('*'||'/'||'%') lexpr
| lexpr ('+'||'-') lexpr
| lexpr ('<<'||'>>'|'>>>') lexpr
| lexpr ('=='||'!='||'>='||'<='||'>'||'<')
| lexpr ('&'||'^'||'|') lexpr
| lexpr ('&&'||'||') lexpr  
| lexpr '?' lexpr ':' lexpr
;
\end{syntax}
Constant limited expressions, written @const-lexpr@ below, can only
reference identifiers that are bound by the IDL "const" declaration.

\subsection{Attributes}

\begin{syntax}
attributes:
  '[' attribute { ',' attribute } ']' ;
attribute:
  ident
| ident '(' [lexpr] { ',' [lexpr] } ')'
| ident '(' uuid ')'
| attribute '*'
| '*' attribute
;
\end{syntax}
Attribute lists are written in brackets "[...]", and are always
optional.  Each attribute is identified by a name, and may carry
optional arguments.  Starred attributes apply to the element type of a
pointer or array type, rather than to the pointer or array type itself.
The following table summarizes the recognized attributes and their
arguments.

\begin{tableau}{|l|l|}{Attribute}{Context where it can appear}
\entree{@"abstract"@      }
       {"typedef" }
\entree{@"bigarray"@      }
       {array type}
\entree{@"blocking"@      }
       {function declaration }
\entree{@"camlint"@       }
       {"int" or "long" integer type}
\entree{@"compare"'('fun-name')'@ }
       {"typedef" }
\entree{@"c2ml"'('fun-name')'@ }
       {"typedef" }
\entree{@"errorcheck"'('fun-name')'@ }
       {"typedef" }
\entree{@"errorcode"@     }
       {"typedef" }
\entree{@"finalize"'('fun-name')'@ }
       {"typedef" }
\entree{@"fortran"@       }
       {array type with "bigarray" attribute}
\entree{@"hash"'('fun-name')'@ }
       {"typedef" }
\entree{@"ignore"@        }
       {any pointer type  }
\entree{@"in"@            }
       {function parameter }
\entree{@"int_default"'(' 'camlint'||'nativeint'||'int32'||'int64'')'@ }
       {interface }
\entree{@"int32"@         }
       {"int" or "long" integer type}
\entree{@"int64"@         }
       {"int" or "long" integer type}
\entree{@"length_is"'('le_1','le_2','\ldots')'@  }
       {array type  }
\entree{@"long_default"'(' 'camlint'||'nativeint'||'int32'||'int64'')'@ }
       {interface }
\entree{@"managed"@       }
       {array type with "bigarray" attribute}
\entree{@"ml2c"'('fun-name')'@ }
       {"typedef" }
\entree{@'mlname(' fun-or-field-name ')' @ }
       {function declaration, "struct" field }
\entree{@'mltype("' caml-type-expr '")' @ }
       {"typedef" }
\entree{@"nativeint"@       }
       {"int" or "long" integer type}
\entree{@"null_terminated"@ }
       {array of pointers  }
\entree{@"object"@}
       {interface}
\entree{@"out"@           }
       {function parameter }
\entree{@"pointer_default"'(' 'ref'||'unique'||'ptr'')'@ }
       {interface }
\entree{@"propget"@}{function declaration}
\entree{@"propput"@}{function declaration}
\entree{@"propputref"@}{function declaration}
\entree{@"ptr"@           }
       {any pointer type  }
\entree{@"ref"@           }
       {any pointer type  }
\entree{@"set"@           }
       {enum type  }
\entree{@"size_is"'('le_1','le_2','\ldots')'@  }
       {array type  }
\entree{@"string"@        }
       {character array or pointer  }
\entree{@"switch_is"'('le')'@ }
       {union type or pointer to union  }
\entree{@"switch_type"'('ty')'@ }
       {union or pointer to union  }
\entree{@"unique"@        }
       {any pointer, array, or bigarray type  }
\entree{@"uuid"'(' uuid ')'@ }
       {interface }
\end{tableau}

\subsection{Types and declarators}

The declaration of an identifier along with its type is as in C:
a type specification comes first, followed by the identifier possibly
decorated with "*" and "[...]" to denote pointers and array types.
For instance, "int x" declares an identifier "x" of type "int",
while "int (*x)[]" declares an identifier "x" that is a pointer to an
array of integers.
\begin{syntax}
type-spec:
  ['unsigned'||'signed']
  ('int'||'short'||'long'||'char'||'hyper'||'long' 'long'||'__int64')
| 'byte'
| 'float'
| 'double'
| 'boolean'
| 'void'
| ident
| 'wchar_t'
| 'handle_t'
| 'struct' ident
| 'union' ident
| 'enum' ident
| struct-decl
| union-decl
| enum-decl
;
declarator:
  {'*'} direct-declarator
;
direct-declarator:
  ident
| '(' declarator ')'
| direct-declarator '[' [const-lexpr] ']'
;
\end{syntax}

\subsection{Structures, unions and enumerations}

\begin{syntax}
struct-decl:
  'struct' [ident] '{' {field-decl} '}'
;
field-decl:
  attributes type-spec declarator { ',' declarator } ';'
;
union-decl:
  'union' [ident] '{' {union-case} '}'
| 'union' [ident] 'switch' '(' type-spec ident ')' '{' {union-case} '}'
;
union-case:
  {{'case' ident ':'}} [field-decl] ';'
  'default' ':' [field-decl] ';'
;
enum-decl:
  'enum' [ident] '{' enum-case {',' enum-case} [','] '}'
;
enum-case:
  ident ['=' const-lexpr]
;
\end{syntax}

IDL "struct" declarations are like those of C, with the addition of
optional attributes on each field.  "union" declarations are also as
in C, except that each case of an union must be labeled by one or
several @'case' ident ':'@.  The first form of union declaration
assumes that the discriminant of the union is provided separately
via a "switch_is" annotation on the union type, while the second form
encapsulates the discriminant along with the union itself
(like in Pascal's "record case of" construct).

\subsection{Function declarations}

\begin{syntax}
function-decl:
  attributes type-spec {'*'} ident '(' params ')' {'quote''('ident','string')'}
;
params:
  epsilon
| 'void'
| param { ',' param }
;
param:
  attributes type-spec declarator
;
\end{syntax}
Function declarations are like in ANSI C, with the addition of
attributes on each parameter and on the function itself.
Parameters must be named.  The optional @quote@ statements following the
declaration are user-provided calling sequences and deallocation
sequences that replaces the default sequences in the
"camlidl"-generated stub code for the function.

\subsection{Constant definitions}

\begin{syntax}
constant-decl: 'const' attributes type-spec {'*'} ident '=' const-lexpr ';'
\end{syntax}
A constant declaration associates a name to a limited expression.
The limited expression can refer to constant names declared earlier,
but cannot refer to other kinds of identifiers.  The optional
attributes influence the interpretation of the type specification,
e.g. "const int x = 3" defines "x" with Caml type "int", but
"const [int64] long x = 5" defines "x" with Caml type "int64".

\subsection{IDL files}

\begin{syntax}
file: {decl} ;
decl:
  function-decl ';'
| constant-decl ';'
| struct-decl ';'
| union-decl ';'
| enum-decl ';'
| 'typedef' attributes type-spec declarator { ',' declarator } ';'
| attributes 'interface' ident [ ':' ident ] '{' {decl} '}'
| 'struct' ident ';'
| 'union' ident ';'
| 'union' 'switch' '(' type-spec ident ')' ';'
| attributes 'interface' ident ';'
| 'import' string ';'
| 'quote' '(' [ident ','] string ')'
| 'cpp_quote' '(' string ')'
\end{syntax}
An IDL file is a sequence of IDL declarations.  Declarations include
function declarations, constant declarations, type declarations
(structs, unions, enums, as well as a C-style "typedef" declaration to
name a type expression), and interfaces.

An interface declaration gives a name and attributes to a collection
of declarations.  For interfaces with the "object" attribute,
an optional super-interface can be provided, as in
@'interface' intf ':' super-intf@.  The name of the interface can be
used as a type name in the remainder of the file.

Forward declarations of structs, unions and interfaces are supported
in the usual C manner, by just giving the name of the struct, union or
interface, but not its actual contents.

The "import" statement reads another IDL file and makes available its
type and constant declarations in the remainder of the file.
No code is generated for the functions and interfaces declared in the
imported file.  The same file can be imported several times, but is
read in only the first time.

The @"quote" "(" ident ',' str ')'@ diversion copies the string @str@
verbatim to one of the files generated by the "camlidl" compiler.
The @ident@ determines the file where @str@ is copied:
it can be "ml" for the Caml implementation file (".ml"),
"mli" for the Caml interface file (".mli"),
"mlmli" for both Caml files,
"h" for the C header file (".h"),
and "c" for the C source file containing the generated stub code (".c"
file).  For backward compatibility, @"cpp_quote" "(" str ")"@ is
recognized as synonymous for @"quote" "(" "h" "," str ")"@.

\section{The Caml-IDL mapping}

This section describes how IDL types, function declarations, and
interfaces are mapped to Caml types, functions and classes.

\subsection{Base types}

\begin{tableau}{|l|l|}{IDL type \var{ty}}{Caml type \transl{\var{ty}}}
\entree{"byte", "short"}{"int"}
\entree{"int", "long" with "[camlint]" attribute}{"int"}
\entree{"int", "long" with "[nativeint]" attribute}{"nativeint"}
\entree{"int", "long" with "[int32]" attribute}{"int32"}
\entree{"int", "long" with "[int64]" attribute}{"int64"}
\entree{"hyper", "long long", "__int64"}{"int64"}
\entree{"char"}{"char"}
\entree{"float", "double"}{"float"}
\entree{"boolean"}{"bool"}
\end{tableau}
(For integer types, "signed" and "unsigned" variants of the same IDL
integer type translate to the same Caml type.)

Depending on the attributes, the "int" and "long" integer types are
converted to one of the Caml integer types "int", "nativeint",
"int32", or "int64".  Values of Caml type "int32" are exactly 32-bit wide
and values of type "int64" are exactly 64-bit wide on all platforms.
Values of type "nativeint" have the natural word size of the platform,
and are large enough to accommodate any C "int" or "long int" without
loss of precision.  Values of Caml type "int" have the natural word
size of the platform minus one bit of tag, hence the conversion from IDL
types "int" and "long" loses the most significant bit on 32-bit
platforms.  On 64-bit platforms, the conversion from "int" is exact,
but the conversion from "long" loses the most significant bit.

If no explicit integer attribute is given for an "int" or "long" type,
the "int_default" or "long_default" attribute of the enclosing
interface, if any, determines the kind of the integer.
If no "int_default" or "long_default" attribute is in scope, the kind
"camlint" is assumed, which maps IDL "int" and "long" types to the
Caml "int" type.

\subsection{Pointers}

The mapping of IDL pointer types depends on their kinds.  Writing
\transl{\var{ty}} for the Caml type corresponding to the IDL type
$ty$, we have:
\begin{alltt}
       [ref] \var{ty} *  \(\Rightarrow\)  \transl{\var{ty}}
    [unique] \var{ty} *  \(\Rightarrow\)  \transl{\var{ty}} option
       [ptr] \var{ty} *  \(\Rightarrow\)  \transl{\var{ty}} Com.opaque
\end{alltt}
In other terms, IDL pointers of kind "ref" are ignored during the mapping:
"[ref] "\var{ty}" *" is mapped to the same Caml type as \var{ty}.
A pointer \var{p} to a C value \var{c}" = *"\var{p} is translated to
the Caml value corresponding to \var{c}.

IDL pointers of kind "unique" are mapped to an "option" type.  The
option value is "None" for a null pointer, and "Some("\var{v}")"
for a non-null pointer to a C value \var{c} that translates to the ML
value \var{v}.

IDL pointers of kind "ptr" are mapped to a "Com.opaque" type.
This is an abstract type that encapsulates the C pointer without
attempting to convert it to an ML data structure.

IDL pointers of kind "ignore" denote struct fields and function
parameters that need not be exposed in the Caml code.  Those pointers
are simply set to null when converting from Caml to C, and ignored
when converting from C to Caml.  They cannot occur elsewhere.

If no explicit pointer kind is given, the "pointer_default" attribute
of the enclosing interface, if any, determines the kind of the pointer.
If no "pointer_default" attribute is in scope, the kind "unique" is
assumed.

\subsection{Arrays}

IDL arrays of characters that carry the "[string]" attribute are mapped
to the Caml "string" type:
\begin{tableau}{|l|l|}{IDL type \var{ty}}{Caml type \transl{\var{ty}}}
\entree{"[string] char []"}{"string"}
\entree{"[string] unsigned char []"}{"string"}
\entree{"[string] signed char []"}{"string"}
\entree{"[string] byte []"}{"string"}
\end{tableau}
Caml string values are translated to standard null-terminated C strings.
Be careful about embedded null characters in the Caml string, which
will be recognized as end of string by C functions.

IDL arrays of characters that carry the "[byte]" attribute are mapped
to the Caml "bytes" type of byte arrays.

IDL arrays carrying the "[bigarray]" attribute are translated to Caml
``big arrays'', as described in the next section.

All other IDL arrays are translated to ML arrays:
\begin{alltt}
        \var{ty} []  \(\Rightarrow\)  \transl{\var{ty}} array
\end{alltt}
For instance, "double []" becomes "float array".
Consequently, multi-dimensional arrays are translated to Caml arrays
of arrays.  For instance, "int [][]" becomes "int array array".

If the "unique" attribute is given, the IDL array is translated to an
ML option type:
\begin{alltt}
        [string,unique] char []  \(\Rightarrow\)  string option
        [unique] \var{ty} []     \(\Rightarrow\)  \transl{\var{ty}} array option
\end{alltt}
As in the case of pointers of kind "unique", the option value is
"None" for a null C pointer, and "Some("\var{v}")" for a non-null
C pointer to a C array that translates to the ML string or array \var{v}.

Conversion between a C array and an ML array proceed element by
element.  For the conversion from C to ML, the number of elements of
the ML array is determined as follows (in the order presented):
\begin{itemize}
\item By the "length_is" attribute, if present.
\item By the "size_is" attribute, if present.
\item By the bound written in the array type, if any.
\item By searching the first null element of the C array, if the
"null_terminated" attribute is present.
\end{itemize}
For instance, C values of IDL type "[length_is(n)] double[]" are
mapped to Caml "float array" of "n" elements.  C values of IDL type
"double[10]" are mapped to Caml "float array" of 10 elements.

The "length_is" and "size_is" attributes take as argument one or
several limited expressions.  Each expression applies to one dimension
of the array.  For instance, "[size_is(*dimx, *dimy)] double d[][]"
specifies a matrix of "double" whose first dimension has size
"*dimx" and the second has size "*dimy".

\subsection{Big arrays}

IDL arrays of integers or floats that carry the "[bigarray]" attribute
are mapped to one of the Caml "Bigarray" types: "Array1.t" for
one-dimensional arrays, "Array2.t" for 2-dimensional arrays,
"Array3.t" for 3-dimensional arrays, and "Genarray.t" for arrays of 4
dimensions or more.

If the "[fortran]" attribute is given, the big array is accessed
from Caml using the Fortran conventions (array indices start at 1;
column-major memory layout).  By default, the big array is accessed
from Caml using the C conventions (array indices start at 0; row-major
memory layout).

If the "[managed]" attribute is given on a big array type that is
result type or out parameter type of a function, Caml assumes that the
corresponding C array was allocated using "malloc()", and is not
referenced anywhere else; then, the Caml garbage collector will free
the C array when the corresponding Caml big array becomes unreachable.
By default, Caml assumes that result or out C arrays are statically or
permanently allocated, and keeps a pointer to them during conversion
to Caml big arrays, and does not free them when the Caml bigarrays
become unreachable.

Finally, the "[unique]" attribute applies to bigarrays as to arrays,
that is, it maps a null C pointer to "None", and a non-null C pointer
\var{p} to "Some("\var{v}")" where \var{v} is the ML bigarray
resulting from the translation of \var{p}.

\subsection{Structs}

IDL structs are mapped to Caml record types.  The names and types of
the IDL struct fields determine the names and types of the Caml record
type:
\begin{alltt}
struct \var{s} \{ ... ; \nth{ty}{i} \nth{id}{i} ; ... \} {\rm becomes}  type \var{s} = \{ ... ; \nth{id}{i} : \transl{\nth{ty}{i}} ; ... \}
\end{alltt}
Example: "struct s { int n; double d[4]; }" becomes
"type s = {n: int; d: float array}".

Exceptions to this rule are as follows:
\begin{itemize}
\item Fields of the IDL struct that are pointers with the "[ignore]"
attribute do not appear in the Caml record type.
Example: "struct s { double x,y; [ignore] void * data; }"
becomes "type struct_s = {x : float; y: float}".
Those ignored pointer fields are set to "NULL" when converting from a
Caml record to a C struct.

\item Integer fields of the IDL struct that appear in a "length_is",
"size_is" or "switch_is" attribute of another field also do not appear
in the Caml record type.  (We call those fields {\em dependent} fields.)
Example: "struct s { int idx; int len; [size_is(len)] double d[]; }"
is translated to the Caml record type 
"type struct_s = {idx: int; d: float array}".
The value of "len" is recovered from the size of the Caml array "d",
and thus doesn't need to be represented explicitly in the Caml record.

\item If, after elimination of ignored pointer fields and dependent
fields as described above, the IDL struct has only one field
$ty~id$, we avoid creating a one-field Caml record type
and translate the IDL struct type directly to the Caml type
\transl{\var{ty}}.
Example: "struct s { int len; [size_is(len)] double d[]; }"
is translated to the Caml type abbreviation "type struct_s = double array".

\item The names of labels in the Caml record type can be changed by
using the @"mlname"@ attribute on struct field declarations.  For instance,
\begin{alltt}
struct s \{ int n; [mlname(p)] int q; \}
        {\rm becomes} type s = \{ n : int; p : int \}
\end{alltt}

\item The Caml type system makes it difficult to use two record types
defined in the same module and having some label names in common.
Thus, if CamlIDL encounters two or more structs having
identically-named fields, it prefixes the Caml label names by the names
of the structs in order to distinguish them.  For instance:
\begin{alltt}
struct s1 \{ int x; int y; \}
struct s2 \{ double x; double t; \}
struct s3 \{ int z; \}
        {\rm becomes} type s1 = \{ s1_x: int; s1_y: int \}
                 and s2 = \{ s2_x: float; s2_t: float \}
                 and s3 = \{ z: int \}
\end{alltt}
The labels for "s1" and "s2" have been prefixed by "s1_" and
"s2_" respectively, to avoid ambiguity on the "x" label.  However, the
label "z" for "s3" is not prefixed, since it is not used elsewhere.

The prefix added in front of multiply-defined labels is taken from the
struct name, if any, and otherwise from the name of the nearest
enclosing struct, union or typedef.  For instance:
\begin{alltt}
typedef struct \{ int x; \} t;
struct s4 \{ struct \{ int x; \} z; \};
        {\rm becomes} type t = \{ t_x: int \}
                 and s4 = \{ z: struct_1 \}
                 and struct_1 = \{ s4_x: int \}
\end{alltt}

The ``minimal prefixing'' strategy described above is the default
behavior of "camlidl".  If the "-prefix-all-labels" option is given,
all record labels are prefixed, whether they occur several times or
not.  If the "-keep-labels" option is given, no automatic prefixing
takes place; the naming of record labels is left entirely under the
user's control, via @"mlname"@ annotations.

\end{itemize}

\subsection{Unions}

IDL discriminated unions are translated to Caml sum types.  Each case
of the union corresponds to a constructor of the sum type.  The
constructor is constant if the union case has no associated field,
otherwise has one argument corresponding to the union case field.  If
the union has a "default" case, an extra constructor
"Default_"\var{unionname} is added to the Caml sum type, carrying an
"int" argument (the value of the discriminating field),
and possibly another argument corresponding to the default field.
Examples:
\begin{alltt}
union u1 \{ case A: int x; case B: case C: double d; case D: ; \}
        {\rm becomes} type u1 = A of int | B of float | C of float | D
union u2 \{ case A: int x; case B: double d; default: ; \}
        {\rm becomes} type u2 = A of int | B of float | Default_u of int
union u3 \{ case A: int x; default: double d; \}
        {\rm becomes} type u3 = A of int | Default_v of int * double
\end{alltt}
All IDL unions must be discriminated, either via the special syntax
"union "\var{name}" switch(int "\var{discr}")"\ldots, or via the
attribute "switch_is("\var{discr}")", where \var{discr} is a C l-value
built from other parameters of the current function, or other fields
of the current "struct".  Both the discriminant and the
case labels must be of an integer type.  Unless a "default" case is
given, the value of the discriminant must be one of the cases of the
union.

\subsection{Enums}

IDL enums are translated to Caml enumerated types (sum types with only
constant constructors).  The names of the constructors are determined
by the names of the enum labels.  The values attached to the enum
labels are ignored.
Example:
"enum e { A, B = 2, C = 4 }" becomes "type enum_e = A | B | C".

The @"set"@ attribute can be applied to a named enum to denote a
bitfield obtained by logical ``or'' of zero, one or several labels of
the enum.  The corresponding ML value is a list of zero, one or
several constructors of the Caml enumerated type.  Consider for
instance:
\begin{verbatim}
enum e { A = 1, B = 2, C = 4 };
typedef [set] enum e eset;
\end{verbatim}
The Caml type "eset" is equal to "enum_e list".
The C integer 6 (= "B | C") is translated to the ML list "[B; C]".
The ML list "[A; C]" is translated to the C integer "A | C", that is "5".

\subsection{Type definitions}

An IDL "typedef" statement is normally translated
to a Caml type abbreviation.  For instance,
"typedef [string] char * str" becomes "type str = string".

If the @"abstract"@ attribute is given, a Caml abstract type is
generated instead of a type abbreviation, thus hinding from Caml the
representation of the type in question.  For instance,
"typedef [abstract] void * handle" becomes "type handle".
In this case, the IDL type in the "typedef" is ignored.

If the @"mltype" "(" '"'  caml-type-expr '"' ")"@ attribute is given,
the Caml type is made equal to @caml-type-expr@.  This is often used
in conjunction with the @"ml2c"@ and @"c2ml"@ attributes to implement
custom translation of data structures between C and ML.  For instance,
"typedef [mltype(\"int list\")] struct mylist_struct * mylist"
becomes "type mylist = int list".

If the @"c2ml("funct-name")" and @"ml2c("funct-name")" attributes are
given, the user-provided C functions given as attributes will be
called to perform Caml to C and C to Caml conversions for values of
the typedef-ed type, instead of using the "camlidl"-generated
conversion functions.  This allows user-controlled translation of data
structures.  The prototypes of the conversion functions must be
\begin{alltt}
        value c2ml(\var{ty} * input);
        void ml2c(value input, \var{ty} * output);
\end{alltt}
where \var{ty} is the name of the type defined by "typedef".  In other
terms, the "c2ml" function is passed a reference to a \var{ty} and
returns the corresponding Caml value, while the "ml2c" function is
passed a Caml value as first argument and stores the corresponding C
value in the \var{ty} reference passed as second argument.

If the @"finalize("final-fn")"@ attribute is given in combination with the
@"abstract"@ attribute, the function @final-fn@ is called when
the Caml block representing a value of this typedef becomes
unreachable from Caml and is reclaimed by the Caml garbage collector.  
Similarly, @"compare("compare-fn")"@ and @"hash("hash-fn")"@ attach a
comparison function and a hashing function (respectively) to Caml
values for this typedef.  The comparison function is called when two
Caml values of this typedef are compared using the generic comparisons
"compare", "=", "<", etc.  The hashing function is called when
"Hashtbl.hash" is applied to a Caml value of this typedef.
The prototype of the finalization, comparison and hashing functions are:
\begin{alltt}
        value \var{final-fn}(\var{ty} * x);
        int \var{compare-fn}(\var{ty} * x, \var{ty} * y);
        long \var{hash-fn}(\var{ty} * x);
\end{alltt}
That is, their arguments are passed by reference.  The comparison
function must return an integer that is negative, zero, or positive
depending on whether its first argument is smaller, equal or greater
than its second argument.  The hashing function returns a suitable
hash value for its argument.

If the @"errorcheck("fn")"@ attribute is provided for the "typedef" \var{ty},
the error checking function @fn@ is called each time a function result
of type \var{ty} is converted from C to Caml.  The function can then check
the \var{ty} value for values indicating an error condition, and raise the
appropriate exception.  If in addition the @"errorcode"@ attribute is
provided, the conversion from C to Caml is suppressed: values of type
\var{ty} are only passed to @fn@ for error checking, then discarded.

\subsection{Functions}

IDL function declarations are translated to Caml functions.
The parameters and results of the Caml function are determined from
those of the IDL function according to the following rules:
\begin{itemize}
\item First, dependent parameters (parameters that are "size_is",
"length_is" or "switch_is" of other parameters) as well as parameters
that are ignored pointers are removed.

\item The remaining parameters are split into Caml function inputs and
Caml function outputs.  Parameters with the "[in]" attribute are added
to the inputs of the function.  Parameters with the "[out]"
attribute are added to the outputs of the function.  Parameters with
the "[in,out]" attribute are added both to the inputs and to the
outputs of the function, unless they are of type "bytes" or big array,
in which case they are added to the inputs of the function only.
(The reason for this exception is that byte arrays and big arrays are
shared between Caml and C, thus allowing true "in,out" behavior on the
Caml function parameter, while other data types are copied during
Caml/C conversion, thus turning a C "in,out" parameter into a Caml
"copy in, copy out" parameter, that is, one parameter and one result.)

\item The return value of the IDL function is added to the outputs of
the Caml function (in first position), unless it is of type "void" or
of a type name that carries the "errorcode" attribute.  In the latter
two cases, the return value of the IDL function is not transmitted to
Caml.
\item The Caml function is then given type
@in_1 "->" \ldots "->" in_p "->" out_1 "*" \ldots "*" out_q@
where @in_1 \ldots in_p@ are the types of its inputs
and @out_1 \ldots out_q@ are the types of its outputs.
If there are no inputs, a "unit" parameter is added.
If there are no outputs, a "unit" result is added.
\end{itemize}
Examples:
\begin{alltt}
int f([in] double x, [in] double y)             f : float -> float -> int
\end{alltt}
\begin{quote} Two "double" input, one "int" output \end{quote}
\begin{alltt}
void g([in] int x)                              g : int -> unit
\end{alltt}
\begin{quote} One "int" input, no output \end{quote}
\begin{alltt}
int h()                                         h : unit -> int
\end{alltt}
\begin{quote} No input, one "int" result \end{quote}
\begin{alltt}
void i([in] int x, [out] double * y)            i : int -> double
\end{alltt}
\begin{quote} One "int" input, one "double" output (as an "out"
parameter) \end{quote}
\begin{alltt}
int j([in] int x, [out] double * y)             j : int -> int * double
\end{alltt}
\begin{quote} One "int" input, one "int" output (in the result), one
"double" output (as an "out" parameter) \end{quote}
\begin{alltt}
void k([in,out,ref] int * x)                    k : int -> int
\end{alltt}
\begin{quote} The "in,out" parameter is both one "int" input and one
"int" output. \end{quote}
\begin{alltt}
HRESULT l([in] int x, [out] int * res1, [out] int * res2)
                                                l : int -> int * int
\end{alltt}
\begin{quote} "HRESULT" is a predefined type with the "errorcode"
attribute, hence it is ignored.  It remains one "int" input and
two "int" outputs ("out" parameters) \end{quote}
\begin{alltt}
void m([in] int len, [in,size_is(len)] double d[])
                                                m : float array -> int
\end{alltt}
\begin{quote} "len" is a dependent parameter, hence is ignored.  The
only input is the "double" array \end{quote}
\begin{alltt}
void n([in] int inputlen, [out] int * outputlen, 
       [in,out,size_is(inputlen),length_is(*outputlen)] double d[])
                                                n : float array -> float array
\end{alltt}
\begin{quote} The two parameters "inputlen" and "outputlen" are
dependent, hence ignored.  The "double" array is both an input
and an output. \end{quote}
\begin{alltt}
void p([in] int dimx, [in] int dimy,
       [in,out,bigarray,size_is(dimx,dimy)] double d[][])
p : (float, Bigarray.float64_elt, Bigarray.c_layout) Bigarray.Array2.t -> unit
\end{alltt}
\begin{quote} The two parameters "dimx" and "dimy" are dependent
(determined from the dimensions of the big array argument),
hence ignored.  The two-dimensional array "d", although marked "[in,out]",
is a big array, hence passed as an input that will be modified in
place by the C function "p".  The Caml function has no outputs.
\end{quote}

\paragraph{Error checking:}
For every output that is of a named type with the @"errorcheck("fn")"@
attribute, the error checking function @fn@ is called after the C
function returns.  That function is assumed to raise a Caml exception
if it finds an output denoting an error.

\paragraph{Custom calling and deallocation sequences:}
The IDL declaration for a function can optionally specify a custom
calling sequence and/or a custom deallocation sequence, via @quote@
clauses following the function declaration:
\begin{syntax}
function-decl:
  attributes type-spec {'*'} ident '(' params ')' 
  { 'quote''(' ident ',' string ')' } ;
\end{syntax}
The general shape of a "camlidl"-generated stub function is as
follows:
\begin{alltt}
value caml_wrapper(value camlparam1, ..., value camlparamK)
{
  /* Convert the function parameters from Caml to C */
  param1 = ...;
  ...
  paramN = ...;
  /* Call the C function 'ident' */
  _res = ident(param1, ..., paramN);
  /* Convert the function result and out parameters to Caml values */
  camlres = ...;
  /* Return result to Caml */
  return camlres;
}
\end{alltt}
A @'quote(call,' string ')'@ clause causes the C statements in
@string@ to be inserted in the generated stub code
instead of the default calling sequence "_res = ident(param1, ..., paramN)".
Thus, the statements in @string@ find the converted parameters in
local variables that have the same names as the parameters in the IDL
declaration, and should leave the result of the function, if any, in
the local variable named "_res".

A @'quote(dealloc,' string ')'@ clause causes the C statements in
@string@ to be inserted in the generated stub code just before the
stub function returns, hence after the conversion of the C function
results to Caml values.  Again, the statements in @string@ have access
to the function result in the local variable named "_res", and to out
parameters in local variables having the same names as the
parameters.  Since the function results and out parameters have
already been converted to Caml values, the code in @string@ can safely
deallocate the data structures they point to.

Custom calling sequences are typically used to rearrange or combine
function parameters, and to perform extra error checks on the
arguments and results.  For instance, the Unix "write" system call can
be specified in IDL as follows:
\begin{verbatim}
        int write([in] int fd,
                  [in,string,length_is(len)] char * data,
                  [in] int len,
                  [in] int ofs,
                  [in] int towrite)
          quote(call,
            " /* Validate the arguments */
              if (ofs < 0 || ofs + towrite >= len) failwith(\"write\");
              /* Perform the write */
              _res = write(fd, data + ofs, towrite);
              /* Validate the result */
              if (_res == -1) failwith(\"write\"); ");
\end{verbatim}
%
Custom deallocation sequences are useful to free data structures
dynamically allocated and returned by the C function.  For instance,
a C function "f" that returns a "malloc"-ed string can be specified in
IDL as follows:
\begin{verbatim}
        [string] char * f([in] int x)
          quote(dealloc, "free(_res); ");
\end{verbatim}
If the string is returned as an "out" parameter instead, we would write:
\begin{verbatim}
        void f ([in] int x, [out, string*] char ** str)
          quote(dealloc, "free(*str); ");
\end{verbatim}

\paragraph{Blocking functions:}
A function can be given the attribute "blocking" to indicate that it
may block on an input/output operation.  The generated code will then
allow other Caml threads to execute concurrently with the operation.

\subsection{Interfaces}

IDL interfaces that do not have the @"object"@ attribute are
essentially ignored.  That is, the declarations contained in the
interface are processed as if they occurred at the top-level of the
IDL file.  The @"pointer_default"@, @"int_default"@ and
@"long_default"@ attributes to the interface can be
used to specify the default pointer kind and integer mappings
for the declarations contained in the interface.  Other attributes, as
well as the name of the super-interface if any, are ignored.

IDL interfaces having the @"object"@ attribute specify COM-style object
interfaces.  The function declarations contained in the interface
specify the methods of the COM interface.  Other kinds of declarations
(type declarations, @"import"@ statements, etc) are treated as if they
occurred at the top-level of the IDL file.  An optional
super-interface can be given, in which case the COM interface
implements the methods of the super-interface in addition to those
specified in the IDL interface.
Example:
\begin{verbatim}
[object, uuid(...)] interface IA { typedef int t; int f(int x); }
[object] interface IB : IA { import "foo.idl"; void g([string] char * s); }
\end{verbatim}
This defines a type "t" and imports the file "foo.idl" as usual.  In
addition, two interfaces are declared: "IA", containing one
method "f" from "int" to "int", and "IB", containing
two methods, "f" from "int" to "int" and "g" from "string" to "unit".

The definition of an object interface \var{i} generates the following
Caml definitions:
\begin{itemize}
\item An abstract type \var{i} identifying the interface.
COM interfaces of type \var{i} are represented in Caml
with type \var{i} " Com.interface".
\item If a super-interface $s$ is given, a conversion function
$s$"_of_"\var{i} of type
\var{i} " Com.interface -> " $s$ " Com.interface".
\item If the "uuid("$iid$")" attribute is given, a value
"iid_"\var{i} of type \var{i}" Com.iid" holding the given interface
identifier.
\item A Caml class \var{i}"_class", with the same methods as the COM
interface.
\item A function "use_"\var{i} of type \var{i} " Com.interface ->
"\var{i}"_class", to transform a COM object into a Caml object. This
allows the methods of the COM object to be invoked from Caml.
\item A function "make_"\var{i} of type "#"\var{i}"_class -> "\var{i}
" Com.interface", to transform a Caml object into a COM object with
interface \var{i}.  This allows the methods of the Caml object to be
invoked from any COM client.
\end{itemize}

Example: in the "IA" and "IB" example above, the following Caml
definitions are generated for "IA":
\begin{verbatim}
type iA
val iid_iA : iA Com.iid
class iA_class : iA Com.interface -> object method f : int -> int end
val use_iA : iA Com.interface -> iA_class
val make_iA : #iA_class -> iA Com.interface
\end{verbatim}
For "IB", we get:
\begin{verbatim}
type iB
val iA_of_iB : iB Com.interface -> iA Com.interface
class iB_class :
  iB Com.interface -> object inherit iA_class method g : string -> unit end
val use_iB : iB Com.interface -> iB_class
val make_iB : #iB_class -> iB Com.interface
\end{verbatim}

\paragraph{Error handling in interfaces:}  Conventionally, methods of
COM interfaces always return a result of type "HRESULT" that says
whether the method succeeded or failed, and in the latter case returns
an error code to its caller.

When calling an interface method from Caml, if the method returns an
"HRESULT" denoting failure, the exception "Com.Error" is raised with a
message describing the error.  Successful "HRESULT" return values are
ignored.  To make them available to Caml, "camlidl" defines the types
"HRESULT_bool" and "HRESULT_int".  If those types are used as return
types instead of "HRESULT", failure results are mapped to
"Com.Error" exceptions as before, but successful results are mapped to
the Caml types "bool" and "int" respectively.  (For "HRESULT_bool",
the "S_OK" result is mapped to "true" and other successful results are
mapped to "false".  For "HRESULT_int", the low 16 bits of the result
code are returned as a Caml "int".)

When calling a Caml method from a COM client, any exception that
escapes the Caml method is mapped back to a failure "HRESULT".  A
textual description of the uncaught exception is saved using
"SetLastError", and can be consulted by the COM client using
"GetLastError" (this is the standard convention for passing extended
error information in COM).

If the IDL return type of the method is not one of the "HRESULT"
types, any exception escaping the Caml method aborts the whole program
after printing a description of the exception.  Hence, programmers of
Caml components should either use "HRESULT" as result type, or make
very sure that all exceptions are properly caught by the method.

\section{Using "camlidl"}

\subsection{Overview}

The "camlidl" stub generator is invoked as follows:
\begin{alltt}
        camlidl \var{options} \var{file1}.idl \var{file2}.idl ...
\end{alltt}
For each file \var{f}".idl" given on the command line, "camlidl"
generates the following files:
\begin{itemize}
\item A Caml interface file \var{f}".mli" that defines the Caml view
of the IDL file.  It contains Caml definitions for the types declared
in the IDL file, as well as declarations for the functions and the
interfaces.
\item A Caml implementation file \var{f}".ml" that implements the
\var{f}".mli" file.
\item A C source file \var{f}"_stubs.c" that contains the stub functions for
converting between C and Caml data representations.
\item If the "-header" option is given, a C header file \var{f}".h"
containing C declarations for the types declared in the IDL file.
\end{itemize}
The generated ".ml" and ".c" files must be compiled and linked with
the remainder of the Caml program.

\subsection{Options}

The following command-line options are recognized by "camlidl".

\begin{options}

\item["-cpp"]
Pre-process the source IDL files with the C preprocessor.  This option
is set by default.

\item["-D " \var{symbol}"="\var{value}]
Define a preprocessor symbol.  The option "-D"\var{symbol}"="\var{value}
is passed to the C preprocessor.  The \var{value} can be omitted,
as in "-D" \var{symbol}, and defaults to "1".

\item["-header"]
Generate a C header file \var{f}".h" containing C declarations for the
types and functions declared in the IDL file \var{f}".c".

\item["-I " \var{dir}]
Add the directory \var{dir} to the list of directories searched for
".idl" files, as given on the command line or recursively loaded
by @"import"@ statements.

\item["-keep-labels"]
Keep the Caml names of record labels as specified in the IDL file.
Do not prefix them with the name of the enclosing struct, even if they
appear in several struct definitions.

\item["-nocpp"]
Suppresses the pre-processing of source IDL files.

\item["-no-include"]
By default, "camlidl" emits a "#include \""\var{f}".h\"" statement in
the file \var{f}".c" containing the generated C code.
The \var{f}".h" header file being included is
either the one generated by "camlidl -header", or generated by another
tool (such as Microsoft's "midl" compiler) from the IDL file, or
hand-written.  The \var{f}".h" file is assumed to provide all C type
declarations needed for compiling the stub code.

The "-no-include" option suppresses the automatic inclusion of the
\var{f}".h" file.  The IDL file should then include the right header
files and provide the right type declarations via @"quote"@ statements.

\item["-prefix-all-labels"]
Prefix all Caml names of record labels with the name of the enclosing
struct.  The default is to prefix only those labels that could cause
ambiguity because they appear in several struct definitions.

\item["-prepro" \var{preprocessing-command}]
Set the command that is executed to pre-process the source IDL files.
The default is the C preprocessor.

\end{options}

\subsection{The "camlidldll" script}

Under Windows, a "bash" script called "camlidldll" is provided to
automate the construction of a DLL containing a COM component written
in Caml.

The script "camlidldll" accepts essentially the same command-line
arguments and options as the "ocamlc" compiler.  (It also accepts
".tlb" type library files on the command-line; see
section~\ref{s-dispatch}, ``Dispatch interfaces'', for more
information on type libraries.)
It produces a DLL file that encapsulates the Caml and C object files
given on the command line.

Use "regsvr32 /s "\var{file}".dll" to record the components in the
system registry once it is compiled to a DLL.

\section{Hints on writing IDL files}

\subsection{Writing an IDL file for a C library}

When writing an IDL file for a C library that doesn't have an IDL interface
already, the include files for that library are a good starting point:
just copy the relevant type and functin declarations to the IDL file,
then annotate them with IDL attributes to describe more precisely
their actual behavior.  The documentation of the library must be read
carefully to determine the mode of function parameters ("in", "out",
"inout"), the actual sizes of arrays, etc.

The type definitions in the IDL file need not correspond exactly with
those in the include files.  Often, a cleaner Caml interface can be
obtained by omitting irrelevant struct fields, or changing their types.
For instance, the Unix library functions for reading library entries
may use the following structure:
\begin{verbatim}
        struct dirent {
            long int d_ino;
            __off_t d_off;
            unsigned short int d_reclen;
            unsigned char d_type;
            char d_name[256];
        };
\end{verbatim}
Of those fields, only "d_name" and "d_ino" are of interest to the
user; the other fields are internal information for the library
functions, are not specified in the POSIX specs, and therefore must
not be used.  Thus, in the IDL file, you should declare:
\begin{verbatim}
        struct dirent {
            long int d_ino;
            char d_name[256];
        };
\end{verbatim}
Thus, the Caml code will have 
"type struct_dirent = {d_ino: int; d_name: string}"
as desired.  However, the generated stub code, being
compiled against the ``true'' definition of "struct dirent", will find
those two fields at the correct offsets in the actual struct.

Special attention must be paid to integer fields or variables.  
By default, integer IDL types are mapped to the Caml type "int",
which is convenient to use in Caml code, but loses one bit
when converting from a C "long" integer, and may lose one bit (on
32-bit platforms) when converting from a C "int" integer.  When the
range of values represented by the C integer is small enough, this
loss is acceptable.  Otherwise, you should use the attributes
"nativeint", "int32" or "int64" so that integer IDL types are mapped
to one of the Caml boxed integer types.  (We recommend that you use
"int32" or "int64" for integers that are specified as being exactly 32
bit wide or 64 bit wide, and "nativeint" for unspecified "int" or
"long" integers.)

Yet another possibility is to declare certain integer fields or variables
as "double" in the IDL file, so that they are represented by "float"
in Caml, and all 32 bits of the integer are preserved in Caml.  For
instance, the Unix function to get the current type is declared as
\begin{verbatim}
        time_t time(time_t * t);
\end{verbatim}
where "time_t" is usually defined as "long".  We can nonetheless
pretend (in the IDL file) that "time" returns a double:
\begin{verbatim}
        double time() quote(" _res = time(NULL); ");
\end{verbatim}
This way, "time" will have the Caml type "unit -> float".
Again, the stub code ``knows'' that "time" actually returns an integer,
and therefore will insert the right integer-float coercions.

\subsection{Sharing IDL files between MIDL and CamlIDL}

The Microsoft software development kit provides a number of IDL files
describing various libraries and components.  In its current state,
"camlidl" cannot exploit those files directly: they use many
(often poorly documented) Microsoft IDL features that are not
implemented yet in "camlidl"; symmetrically, "camlidl" introduces
several new annotations that are not recognized by Microsoft's "midl"
compiler.  So, significant editing work on the IDL files is required.

The C preprocessor can be used to alleviate the "camlidl"-"midl"
incompatibilities: "camlidl" defines the preprocessor symbol "CAMLIDL"
when preprocessing its input files, while "midl" does not.  Hence,
one can bracket incompatible definitions in 
"#ifdef CAMLIDL ... #else ... #endif".  Along these lines, a C
preprocessor header file, "camlidlcompat.h", is provided: it uses
"#define" to remove "camlidl"-specific attributes when compiling with
"midl", and to remove "midl"-specific attributes when compiling with
"camlidl".  Thus, an IDL file compatible with both "midl" and
"camlidl" would look like this:
\begin{verbatim}
        #include <camlidlcompat.h>

        #ifndef CAMLIDL
        import "unknwn.idl";    // imports specific to MIDL
        import "oaidl.idl";
        #endif
        import "mymodule.idl";  // imports common to MIDL and CamlIDL

        typedef [abstract,marshal_as(int)] void * ptr;

        ...

        #ifndef CAMLIDL
        [...] library MyTypeLib {
          importlib("stdole32.tlb");
          [...] coclass MyComponent { [default] interface IX; }
        }
        #endif
\end{verbatim}
Notice that since "camlidl" doesn't handle type libraries, the type
library part of an "midl" file must be enclosed in "#ifndef CAMLIDL".

\subsection{Dispatch interfaces and type libraries} \label{s-dispatch}

A dispatch interface, in COM lingo, is an interface that supports
dynamic, interpreted dispatch of method interfaces.  This form of
interpreted dispatch is used by Visual Basic and other scripting
languages to perform calls to methods of COM components.

CamlIDL provides minimal support for dispatch interfaces.  To equip a
Caml component with a dispatch interface (thus making it callable from
Visual Basic), you need to do the following:
\begin{enumerate}
\item Use "IDispatch" instead of "IUnknown" as the super-interface of
the component's interfaces.
\item Write a type library for your component and compile it using
"midl".  A type library is a run-time representation of the interfaces
supported by an object.  The "midl" compiler can generate a type
library from the IDL description of the component, enriched with some
special-purpose declarations (the "library" and "coclass"
statements).  Refer to the documentation of "midl" for more
information.
\item Pass the type library files (".tlb" files) generated by "midl"
as extra arguments to "camlidldll" when generating the DLL for your
Caml component.
\end{enumerate}

\section{Release notes}

Here are some caveats and open issues that apply to the current
release.

\paragraph{Deallocation of function results and "out" parameters:}
If a C function dynamically allocates some of its outputs (either
returned or stored in "out" parameters), its IDL declaration must
contain a @'quote(dealloc,' string ')'@ clause to properly free the
space occupied by those outputs after they have been converted to
Caml.  Otherwise, memory leaks will occur.  (The only exception is
results and output parameters of type "[bigarray,managed] "\var{ty}"[]",
where the Caml garbage collector takes care of deallocation.)

This does not conform to the MIDL and COM specifications, which say
that space for "out" data structures must be allocated
with "CoTaskMemAlloc" by the callee, and automatically freed
using "CoTaskMemFree" by the generated stub code.  (The specs don't
say what happens with the return value of the function.)
However, there are many functions in Win32 (not to mention the
Unix world) that do not follow this convention, and
return data structures (e.g. strings) that are statically
allocated, or require special deallocation functions.  Hence,
"camlidl" leaves deallocation of outputs entirely under user control.

\paragraph{Allocation and deallocation of "in,out" parameters:}
For "in,out" parameters, the MIDL/COM rules are that the caller (the
stub code) should allocate the inputs, the callee should free them
and allocate again its outputs, and the caller should free the outputs.
As explained above, "camlidl"-generated stubs don't automatically free
the outputs.  Worse, the inputs passed to the functions are allocated
partially on the stack and partially in the heap
(using "CoTaskMemAlloc"), so the callee may perform an incorrect
free on a stack-allocated argument.  The best thing to do is avoid
"in,out" parameters entirely, and split them into one "in" and one
"out" parameter.

\paragraph{Reference-counting of COM interfaces:}
Caml finalized objects are used to call "Release" automatically on COM
interfaces that become unreachable.  The reference counting of
interfaces passed as "in" and "out" parameters is correctly
implemented.  However, "in,out" parameters that are interfaces are not
correctly handled.  Again, avoid "in,out" parameters.

\paragraph{COM support:}
The support for COM is currently quite small.  COM components
registered in the system registry can be imported via
"Com.create_instance". Components written in Caml can be exported as
DLLs, but not yet as standalone servers.  Preliminary support for
dispatch interfaces is available, however many of the data types used
in the Automation framework are not supported yet (e.g. "SAFEARRAY").