fix haddock submodule pointer

[ghc-hetmet.git] / docs / users_guide / ffi-chap.xml

<?xml version="1.0" encoding="iso-8859-1"?>
<!-- FFI docs as a chapter -->

<chapter id="ffi">
 <title>
Foreign function interface (FFI)
 </title>

  <para>GHC (mostly) conforms to the Haskell Foreign Function Interface,
  whose definition is part of the Haskell Report on <ulink url="http://www.haskell.org/"><literal>http://www.haskell.org/</literal></ulink>.</para>

  <para>FFI support is enabled by default, but can be enabled or disabled explicitly with the <option>-XForeignFunctionInterface</option><indexterm><primary><option>-XForeignFunctionInterface</option></primary>
    </indexterm> flag.</para>

  <para>GHC implements a number of GHC-specific extensions to the FFI
    Addendum.  These extensions are described in <xref linkend="ffi-ghcexts" />, but please note that programs using
    these features are not portable.  Hence, these features should be
    avoided where possible.</para>

  <para>The FFI libraries are documented in the accompanying library
  documentation; see for example the
    <ulink url="&libraryBaseLocation;/Control-Concurrent.html"><literal>Foreign</literal></ulink> module.</para>

  <sect1 id="ffi-ghcexts">
    <title>GHC extensions to the FFI Addendum</title>

    <para>The FFI features that are described in this section are specific to
    GHC.  Your code will not be portable to other compilers if you use them.</para>

    <sect2>
      <title>Unboxed types</title>

      <para>The following unboxed types may be used as basic foreign types
      (see FFI Addendum, Section 3.2): <literal>Int#</literal>,
      <literal>Word#</literal>, <literal>Char#</literal>,
      <literal>Float#</literal>, <literal>Double#</literal>,
      <literal>Addr#</literal>, <literal>StablePtr# a</literal>,
      <literal>MutableByteArray#</literal>, <literal>ForeignObj#</literal>,
      and <literal>ByteArray#</literal>.</para>
    </sect2>

    <sect2 id="ffi-newtype-io">
      <title>Newtype wrapping of the IO monad</title>
      <para>The FFI spec requires the IO monad to appear in various  places,
      but it can sometimes be convenient to wrap the IO monad in a
      <literal>newtype</literal>, thus:
<programlisting>
  newtype MyIO a = MIO (IO a)
</programlisting>
     (A reason for doing so might be to prevent the programmer from
        calling arbitrary IO procedures in some part of the program.)
</para>
<para>The Haskell FFI already specifies that arguments and results of
foreign imports and exports will be automatically unwrapped if they are
newtypes (Section 3.2 of the FFI addendum).  GHC extends the FFI by automatically unwrapping any newtypes that
wrap the IO monad itself.
More precisely, wherever the FFI specification requires an IO type, GHC will
accept any newtype-wrapping of an IO type.  For example, these declarations are
OK:
<programlisting>
   foreign import foo :: Int -> MyIO Int
   foreign import "dynamic" baz :: (Int -> MyIO Int) -> CInt -> MyIO Int
</programlisting>
</para>
      </sect2>

      <sect2 id="ffi-prim">
        <title>Primitive imports</title>
        <para>
          GHC extends the FFI with an additional calling convention
          <literal>prim</literal>, e.g.:
<programlisting>
   foreign import prim "foo" foo :: ByteArray# -> (# Int#, Int# #)
</programlisting>
          This is used to import functions written in Cmm code that follow an
          internal GHC calling convention. This feature is not intended for
          use outside of the core libraries that come with GHC. For more
          details see the GHC developer wiki.
        </para>
      </sect2>

      <sect2 id="ffi-interruptible">
        <title>Interruptible foreign calls</title>
        <para>
          This concerns the interaction of foreign calls
          with <literal>Control.Concurrent.throwTo</literal>.
          Normally when the target of a <literal>throwTo</literal> is
          involved in a foreign call, the exception is not raised
          until the call returns, and in the meantime the caller is
          blocked.  This can result in unresponsiveness, which is
          particularly undesirable in the case of user interrupt
          (e.g. Control-C).  The default behaviour when a Control-C
          signal is received (<literal>SIGINT</literal> on Unix) is to raise
          the <literal>UserInterrupt</literal> exception in the main
          thread; if the main thread is blocked in a foreign call at
          the time, then the program will not respond to the user
          interrupt.
        </para>

        <para>
          The problem is that it is not possible in general to
          interrupt a foreign call safely.  However, GHC does provide
          a way to interrupt blocking system calls which works for
          most system calls on both Unix and Windows.  A foreign call
          can be annotated with <literal>interruptible</literal> instead
          of <literal>safe</literal> or <literal>unsafe</literal>:

<programlisting>
foreign import ccall interruptible
   "sleep" :: CUint -> IO CUint
</programlisting>

          <literal>interruptible</literal> behaves exactly as
          <literal>safe</literal>, except that when
          a <literal>throwTo</literal> is directed at a thread in an
          interruptible foreign call, an OS-specific mechanism will be
          used to attempt to cause the foreign call to return:

          <variablelist>
            <varlistentry>
              <term>Unix systems</term>
              <listitem>
                <para>
                  The thread making the foreign call is sent
                  a <literal>SIGPIPE</literal> signal
                  using <literal>pthread_kill()</literal>.  This is
                  usually enough to cause a blocking system call to
                  return with <literal>EINTR</literal> (GHC by default
                  installs an empty signal handler
                  for <literal>SIGPIPE</literal>, to override the
                  default behaviour which is to terminate the process
                  immediately).
                </para>
              </listitem>
            </varlistentry>
            <varlistentry>
              <term>Windows systems</term>
              <listitem>
                <para>
                  [Vista and later only] The RTS calls the Win32
                  function <literal>CancelSynchronousIO</literal>,
                  which will cause a blocking I/O operation to return
                  with the
                  error <literal>ERROR_OPERATION_ABORTED</literal>.
                </para>
              </listitem>
            </varlistentry>
          </variablelist>

          If the system call is successfully interrupted, it will
          return to Haskell whereupon the exception can be raised.  Be
          especially careful when
          using <literal>interruptible</literal> that the caller of
          the foreign function is prepared to deal with the
          consequences of the call being interrupted; on Unix it is
          good practice to check for <literal>EINTR</literal> always,
          but on Windows it is not typically necessary to
          handle <literal>ERROR_OPERATION_ABORTED</literal>.
        </para>
      </sect2>
  </sect1>

  <sect1 id="ffi-ghc">
    <title>Using the FFI with GHC</title>

    <para>The following sections also give some hints and tips on the
    use of the foreign function interface in GHC.</para>

    <sect2 id="foreign-export-ghc">
      <title>Using <literal>foreign export</literal> and <literal>foreign import ccall "wrapper"</literal> with GHC</title>

      <indexterm><primary><literal>foreign export
      </literal></primary><secondary>with GHC</secondary>
      </indexterm>

      <para>When GHC compiles a module (say <filename>M.hs</filename>)
      which uses <literal>foreign export</literal> or
      <literal>foreign import "wrapper"</literal>, it generates two
      additional files, <filename>M_stub.c</filename> and
      <filename>M_stub.h</filename>.  GHC will automatically compile
      <filename>M_stub.c</filename> to generate
      <filename>M_stub.o</filename> at the same time.</para>

      <para>For a plain <literal>foreign export</literal>, the file
      <filename>M_stub.h</filename> contains a C prototype for the
      foreign exported function, and <filename>M_stub.c</filename>
      contains its definition.  For example, if we compile the
      following module:</para>

<programlisting>
module Foo where

foreign export ccall foo :: Int -> IO Int

foo :: Int -> IO Int
foo n = return (length (f n))

f :: Int -> [Int]
f 0 = []
f n = n:(f (n-1))</programlisting>

      <para>Then <filename>Foo_stub.h</filename> will contain
      something like this:</para>

<programlisting>
#include "HsFFI.h"
extern HsInt foo(HsInt a0);</programlisting>

      <para>and <filename>Foo_stub.c</filename> contains the
      compiler-generated definition of <literal>foo()</literal>.  To
      invoke <literal>foo()</literal> from C, just <literal>#include
      "Foo_stub.h"</literal> and call <literal>foo()</literal>.</para>

      <para>The <filename>foo_stub.c</filename> and
        <filename>foo_stub.h</filename> files can be redirected using the
        <option>-stubdir</option> option; see <xref linkend="options-output"
          />.</para>

      <para>When linking the program, remember to include
        <filename>M_stub.o</filename> in the final link command line, or
        you'll get link errors for the missing function(s) (this isn't
        necessary when building your program with <literal>ghc
        &ndash;&ndash;make</literal>, as GHC will automatically link in the
        correct bits).</para>

      <sect3 id="using-own-main">
        <title>Using your own <literal>main()</literal></title>

        <para>Normally, GHC's runtime system provides a
        <literal>main()</literal>, which arranges to invoke
        <literal>Main.main</literal> in the Haskell program.  However,
        you might want to link some Haskell code into a program which
        has a main function written in another language, say C.  In
        order to do this, you have to initialize the Haskell runtime
        system explicitly.</para>

        <para>Let's take the example from above, and invoke it from a
        standalone C program.  Here's the C code:</para>

<programlisting>
#include &lt;stdio.h&gt;
#include "HsFFI.h"

#ifdef __GLASGOW_HASKELL__
#include "foo_stub.h"
#endif

int main(int argc, char *argv[])
{
  int i;

  hs_init(&amp;argc, &amp;argv);

  for (i = 0; i &lt; 5; i++) {
    printf("%d\n", foo(2500));
  }

  hs_exit();
  return 0;
}</programlisting>

        <para>We've surrounded the GHC-specific bits with
        <literal>#ifdef __GLASGOW_HASKELL__</literal>; the rest of the
        code should be portable across Haskell implementations that
        support the FFI standard.</para>

        <para>The call to <literal>hs_init()</literal>
        initializes GHC's runtime system.  Do NOT try to invoke any
        Haskell functions before calling
        <literal>hs_init()</literal>: bad things will
        undoubtedly happen.</para>

        <para>We pass references to <literal>argc</literal> and
        <literal>argv</literal> to <literal>hs_init()</literal>
        so that it can separate out any arguments for the RTS
        (i.e. those arguments between
        <literal>+RTS...-RTS</literal>).</para>

      <informaltable>
        <tgroup cols="2" align="left" colsep="1" rowsep="1">
          <thead>
            <row>
              <entry>Character</entry>
              <entry>Replacement</entry>
            </row>
          </thead>
          <tbody>
            <row>
              <entry><literal>.</literal></entry>
              <entry><literal>zd</literal></entry>
            </row>
            <row>
              <entry><literal>_</literal></entry>
              <entry><literal>zu</literal></entry>
            </row>
            <row>
              <entry><literal>`</literal></entry>
              <entry><literal>zq</literal></entry>
            </row>
            <row>
              <entry><literal>Z</literal></entry>
              <entry><literal>ZZ</literal></entry>
            </row>
            <row>
              <entry><literal>z</literal></entry>
              <entry><literal>zz</literal></entry>
            </row>
          </tbody>
        </tgroup>
      </informaltable>

        <para>After we've finished invoking our Haskell functions, we
        can call <literal>hs_exit()</literal>, which terminates the
        RTS.</para>

        <para>There can be multiple calls to
        <literal>hs_init()</literal>, but each one should be matched
        by one (and only one) call to
        <literal>hs_exit()</literal><footnote><para>The outermost
        <literal>hs_exit()</literal> will actually de-initialise the
        system.  NOTE that currently GHC's runtime cannot reliably
        re-initialise after this has happened,
        see <xref linkend="ffi-divergence" />.</para>
        </footnote>.</para>

        <para>NOTE: when linking the final program, it is normally
        easiest to do the link using GHC, although this isn't
        essential.  If you do use GHC, then don't forget the flag
        <option>-no-hs-main</option><indexterm><primary><option>-no-hs-main</option></primary>
          </indexterm>, otherwise GHC will try to link
        to the <literal>Main</literal> Haskell module.</para>
      </sect3>

      <sect3 id="ffi-library">
        <title>Making a Haskell library that can be called from foreign
          code</title>

        <para>The scenario here is much like in <xref linkend="using-own-main"
            />, except that the aim is not to link a complete program, but to
          make a library from Haskell code that can be deployed in the same
          way that you would deploy a library of C code.</para>

        <para>The main requirement here is that the runtime needs to be
          initialized before any Haskell code can be called, so your library
          should provide initialisation and deinitialisation entry points,
          implemented in C or C++.  For example:</para>

<programlisting>
 HsBool mylib_init(void){
   int argc = ...
   char *argv[] = ...

   // Initialize Haskell runtime
   hs_init(&amp;argc, &amp;argv);

   // do any other initialization here and
   // return false if there was a problem
   return HS_BOOL_TRUE;
 }

 void mylib_end(void){
   hs_exit();
 }
</programlisting>

        <para>The initialisation routine, <literal>mylib_init</literal>, calls
          <literal>hs_init()</literal> as
          normal to initialise the Haskell runtime, and the corresponding
          deinitialisation function <literal>mylib_end()</literal> calls
          <literal>hs_exit()</literal> to shut down the runtime.</para>
      </sect3>

    </sect2>

    <sect2 id="glasgow-foreign-headers">
      <title>Using header files</title>

      <indexterm><primary>C calls, function headers</primary></indexterm>

      <para>C functions are normally declared using prototypes in a C
        header file.  Earlier versions of GHC (6.8.3 and
        earlier) <literal>&num;include</literal>d the header file in
        the C source file generated from the Haskell code, and the C
        compiler could therefore check that the C function being
        called via the FFI was being called at the right type.</para>

      <para>GHC no longer includes external header files when
        compiling via C, so this checking is not performed.  The
        change was made for compatibility with the native code backend
        (<literal>-fasm</literal>) and to comply strictly with the FFI
        specification, which requires that FFI calls are not subject
        to macro expansion and other CPP conversions that may be
        applied when using C header files.  This approach also
        simplifies the inlining of foreign calls across module and
        package boundaries: there's no need for the header file to be
        available when compiling an inlined version of a foreign call,
        so the compiler is free to inline foreign calls in any
        context.</para>

      <para>The <literal>-&num;include</literal> option is now
        deprecated, and the <literal>include-files</literal> field
        in a Cabal package specification is ignored.</para>

    </sect2>

    <sect2>
      <title>Memory Allocation</title>

      <para>The FFI libraries provide several ways to allocate memory
      for use with the FFI, and it isn't always clear which way is the
      best.  This decision may be affected by how efficient a
      particular kind of allocation is on a given compiler/platform,
      so this section aims to shed some light on how the different
      kinds of allocation perform with GHC.</para>

      <variablelist>
        <varlistentry>
          <term><literal>alloca</literal> and friends</term>
          <listitem>
            <para>Useful for short-term allocation when the allocation
            is intended to scope over a given <literal>IO</literal>
            computation.  This kind of allocation is commonly used
            when marshalling data to and from FFI functions.</para>

            <para>In GHC, <literal>alloca</literal> is implemented
            using <literal>MutableByteArray#</literal>, so allocation
            and deallocation are fast: much faster than C's
            <literal>malloc/free</literal>, but not quite as fast as
            stack allocation in C.  Use <literal>alloca</literal>
            whenever you can.</para>
          </listitem>
        </varlistentry>

        <varlistentry>
          <term><literal>mallocForeignPtr</literal></term>
          <listitem>
            <para>Useful for longer-term allocation which requires
            garbage collection.  If you intend to store the pointer to
            the memory in a foreign data structure, then
            <literal>mallocForeignPtr</literal> is
            <emphasis>not</emphasis> a good choice, however.</para>

            <para>In GHC, <literal>mallocForeignPtr</literal> is also
            implemented using <literal>MutableByteArray#</literal>.
            Although the memory is pointed to by a
            <literal>ForeignPtr</literal>, there are no actual
            finalizers involved (unless you add one with
            <literal>addForeignPtrFinalizer</literal>), and the
            deallocation is done using GC, so
            <literal>mallocForeignPtr</literal> is normally very
            cheap.</para>
          </listitem>
        </varlistentry>

        <varlistentry>
          <term><literal>malloc/free</literal></term>
          <listitem>
            <para>If all else fails, then you need to resort to
            <literal>Foreign.malloc</literal> and
            <literal>Foreign.free</literal>.  These are just wrappers
            around the C functions of the same name, and their
            efficiency will depend ultimately on the implementations
            of these functions in your platform's C library.  We
            usually find <literal>malloc</literal> and
            <literal>free</literal> to be significantly slower than
            the other forms of allocation above.</para>
          </listitem>
        </varlistentry>

        <varlistentry>
          <term><literal>Foreign.Marshal.Pool</literal></term>
          <listitem>
            <para>Pools are currently implemented using
            <literal>malloc/free</literal>, so while they might be a
            more convenient way to structure your memory allocation
            than using one of the other forms of allocation, they
            won't be any more efficient.  We do plan to provide an
            improved-performance implementation of Pools in the
            future, however.</para>
          </listitem>
        </varlistentry>
      </variablelist>
    </sect2>

    <sect2 id="ffi-threads">
      <title>Multi-threading and the FFI</title>

      <para>In order to use the FFI in a multi-threaded setting, you must
        use the <option>-threaded</option> option
        (see <xref linkend="options-linker" />).</para>

      <sect3>
        <title>Foreign imports and multi-threading</title>

        <para>When you call a <literal>foreign import</literal>ed
          function that is annotated as <literal>safe</literal> (the
          default), and the program was linked
          using <option>-threaded</option>, then the call will run
          concurrently with other running Haskell threads.  If the
          program was linked without <option>-threaded</option>,
          then the other Haskell threads will be blocked until the
          call returns.</para>

        <para>This means that if you need to make a foreign call to
          a function that takes a long time or blocks indefinitely,
          then you should mark it <literal>safe</literal> and
          use <option>-threaded</option>.  Some library functions
          make such calls internally; their documentation should
          indicate when this is the case.</para>

        <para>If you are making foreign calls from multiple Haskell
          threads and using <option>-threaded</option>, make sure that
          the foreign code you are calling is thread-safe.  In
          particularly, some GUI libraries are not thread-safe and
          require that the caller only invokes GUI methods from a
          single thread.  If this is the case, you may need to
          restrict your GUI operations to a single Haskell thread,
          and possibly also use a bound thread (see
          <xref linkend="haskell-threads-and-os-threads" />).</para>

        <para>Note that foreign calls made by different Haskell
          threads may execute in <emphasis>parallel</emphasis>, even
          when the <literal>+RTS -N</literal> flag is not being used
          (<xref linkend="parallel-options" />).  The <literal>+RTS
          -N</literal> flag controls parallel execution of Haskell
          threads, but there may be an arbitrary number of foreign
          calls in progress at any one time, regardless of
          the <literal>+RTS -N</literal> value.</para>

        <para>If a call is annotated as <literal>interruptible</literal>
          and the program was multithreaded, the call may be
          interrupted in the event that the Haskell thread receives an
          exception.  The mechanism by which the interrupt occurs
          is platform dependent, but is intended to cause blocking
          system calls to return immediately with an interrupted error
          code.  The underlying operating system thread is not to be
          destroyed.  See <xref linkend="ffi-interruptible"/> for more details.</para>
      </sect3>

      <sect3 id="haskell-threads-and-os-threads">
        <title>The relationship between Haskell threads and OS
          threads</title>

        <para>Normally there is no fixed relationship between Haskell
          threads and OS threads.  This means that when you make a
          foreign call, that call may take place in an unspecified OS
          thread.  Furthermore, there is no guarantee that multiple
          calls made by one Haskell thread will be made by the same OS
          thread.</para>

        <para>This usually isn't a problem, and it allows the GHC
          runtime system to make efficient use of OS thread resources.
          However, there are cases where it is useful to have more
          control over which OS thread is used, for example when
          calling foreign code that makes use of thread-local state.
          For cases like this, we provide <emphasis>bound
          threads</emphasis>, which are Haskell threads tied to a
          particular OS thread.  For information on bound threads, see
          the documentation
          for the <ulink url="&libraryBaseLocation;/Control-Concurrent.html"><literal>Control.Concurrent</literal></ulink>
          module.</para>
      </sect3>

      <sect3>
        <title>Foreign exports and multi-threading</title>

        <para>When the program is linked
          with <option>-threaded</option>, then you may
          invoke <literal>foreign export</literal>ed functions from
          multiple OS threads concurrently.  The runtime system must
          be initialised as usual by
          calling <literal>hs_init()</literal>, and this call must
          complete before invoking any <literal>foreign
          export</literal>ed functions.</para>
      </sect3>

      <sect3 id="hs-exit">
        <title>On the use of <literal>hs_exit()</literal></title>

        <para><literal>hs_exit()</literal> normally causes the termination of
          any running Haskell threads in the system, and when
          <literal>hs_exit()</literal> returns, there will be no more Haskell
          threads running.  The runtime will then shut down the system in an
          orderly way, generating profiling
          output and statistics if necessary, and freeing all the memory it
          owns.</para>

        <para>It isn't always possible to terminate a Haskell thread forcibly:
          for example, the thread might be currently executing a foreign call,
          and we have no way to force the foreign call to complete.  What's
          more, the runtime must
          assume that in the worst case the Haskell code and runtime are about
          to be removed from memory (e.g. if this is a <link linkend="win32-dlls">Windows DLL</link>,
          <literal>hs_exit()</literal> is normally called before unloading the
          DLL).  So <literal>hs_exit()</literal> <emphasis>must</emphasis> wait
          until all outstanding foreign calls return before it can return
          itself.</para>

        <para>The upshot of this is that if you have Haskell threads that are
          blocked in foreign calls, then <literal>hs_exit()</literal> may hang
          (or possibly busy-wait) until the calls return.  Therefore it's a
          good idea to make sure you don't have any such threads in the system
          when calling <literal>hs_exit()</literal>.  This includes any threads
          doing I/O, because I/O may (or may not, depending on the
          type of I/O and the platform) be implemented using blocking foreign
          calls.</para>

        <para>The GHC runtime treats program exit as a special case, to avoid
          the need to wait for blocked threads when a standalone
          executable exits.  Since the program and all its threads are about to
          terminate at the same time that the code is removed from memory, it
          isn't necessary to ensure that the threads have exited first.
          (Unofficially, if you want to use this fast and loose version of
          <literal>hs_exit()</literal>, then call
          <literal>shutdownHaskellAndExit()</literal> instead).</para>
      </sect3>
    </sect2>

    <sect2 id="ffi-floating-point">
      <title>Floating point and the FFI</title>

      <para>
        The standard C99 <literal>fenv.h</literal> header
        provides operations for inspecting and modifying the state of
        the floating point unit.  In particular, the rounding mode
        used by floating point operations can be changed, and the
        exception flags can be tested.
      </para>

      <para>
        In Haskell, floating-point operations have pure types, and the
        evaluation order is unspecified.  So strictly speaking, since
        the <literal>fenv.h</literal> functions let you change the
        results of, or observe the effects of floating point
        operations, use of <literal>fenv.h</literal> renders the
        behaviour of floating-point operations anywhere in the program
        undefined.
      </para>

      <para>
        Having said that, we <emphasis>can</emphasis> document exactly
        what GHC does with respect to the floating point state, so
        that if you really need to use <literal>fenv.h</literal> then
        you can do so with full knowledge of the pitfalls:
        <itemizedlist>
          <listitem>
            <para>
              GHC completely ignores the floating-point
              environment, the runtime neither modifies nor reads it.
            </para>
          </listitem>
          <listitem>
            <para>
              The floating-point environment is not saved over a
              normal thread context-switch.  So if you modify the
              floating-point state in one thread, those changes may be
              visible in other threads.  Furthermore, testing the
              exception state is not reliable, because a context
              switch may change it.  If you need to modify or test the
              floating point state and use threads, then you must use
              bound threads
              (<literal>Control.Concurrent.forkOS</literal>), because
              a bound thread has its own OS thread, and OS threads do
              save and restore the floating-point state.
            </para>
          </listitem>
          <listitem>
            <para>
              It is safe to modify the floating-point unit state
              temporarily during a foreign call, because foreign calls
              are never pre-empted by GHC.
            </para>
          </listitem>
        </itemizedlist>
      </para>
    </sect2>
  </sect1>
</chapter>

<!-- Emacs stuff:
     ;;; Local Variables: ***
     ;;; sgml-parent-document: ("users_guide.xml" "book" "chapter") ***
     ;;; End: ***
 -->