def-foreign-call

Arguments: name-and-options arglist &key call-direct callback convention returning lisp-return-will-not-move method-index release-heap release-heap-ignorable arg-checking optimize-for-space strings-convert error-value pass-structs-by-value allow-gc release-heap-implies-allow-gc documentation

This macro creates the specification which allows Lisp to correctly call non-Lisp code. Like other defining forms, its macroexpansion clearly shows what will occur and at what eval-when times. The execution of the expanded form always returns the Lisp name being defined. The definition that is installed is a Lisp function that serves as a wrapper and initiates the foreign call.

Macroexpansion of a def-foreign-call form provides useful information about how the call is interpreted. See below.

The following table shows the arguments. The first two entries are the required arguments and the remainder are keyword arguments.

Note 1: Argument Specifications

Argument specifications are available with a rich set of syntax and defaults which allow for a C "feel" while still retaining the Lisp semantics and power.

The elements of an argument list can be:

• A name (symbol). The type defaults to a C int, and the value will be converted in the usualy manner from the Lisp integer subtype passed in.
• A two element list, (name type), where name is a sumbol and type is either a built-in foreign type, or a type defined by def-foreign-type, or one of the keywords :lisp, :foreign-address, or :aligned. We discuss the keywords below.
• A three element list (name type lisp-type), where name and type are as above, and lisp-type can be any valid Lisp type. It specifies the type of the value passed. If the lisp-type is not consistent with the foreign type, a warning may be signaled.
• A four element list (name type lisp-type user-conversion-function), where name, type, and lisp-type are as above, and user-conversion-function is a symbol naming a function, described in section The user-conversion function in a complex-type-spec in foreign-functions.htm. An example is

The allowable keywords for the foreign type (second list element) are:

• :int: the default, the argument will be converted to a C integer.
• :lisp: the argument will be passed unchanged. analogous to the returning type :lisp.
• :aligned: like :lisp (in that the argument is passed through unchanged) but the argument value must be a fixnum. If arguments are checked, a value other than a fixnum will be rejected. See Aligned Pointers and the :aligned type in ftype.htm for more information.
• :foreign-address: the argument value should be an integer which will be interpreted as a foreign address.

Here is an example:

  (a (b :int fixnum) (c :lisp))

A foreign type specification that includes a reference spec such as (& :int) will be interpreted as a pass-by-reference argument.

For boolean values, specify the argument (argname :int boolean). Then any non-nil Lisp value (including 0) will be converted into a C value of 1, and nil will be converted into a C value of 0. For returned values, a C value of 0 is converted into nil and a non-zero C value is converted into t.

The special Lisp type specification :no-proto is provided for use with a :float foreign type (but its use is deprecated); it is equivalent to the :single-float-no-proto foreign type (not listed above because it is deprecated). The preferred specification is (name :double single-float).

Specifying (* :float) or (* :double) as the type of an argument is not recommended. The function will expect a foreign address (not a Lisp address) and it will not pass the address of a Lisp float if a Lisp float is given as a value.

See A note on foreign addresses in foreign-functions.htm for a discussion of foreign addresses and what is expected when :foreign-address is specified.

Note 2: Possible Values for Convention

Other than :c, (the default and suitable for most purposes), the :fortran convention is defined. This convention generally causes a conversion of most atomic arguments to pass-by-reference.

:fastcall convention on Windows does not work

:fastcall is a special convention used by some Windows operating systems to speed up some calls, by passing two arguments in registers. However, it does not work with Allegro CL. In Allegro CL on X86 architectures, the first two arguments are usually passed in registers anyway in Lisp, but they are different registers than used in fastcall. Also, since the calling sequence itself overshadows the speed that would be gained by saving a couple of push instructions, the foreign call to a fastcall function would not in fact be very fast at all.

Note 3: the release-heap keyword argument

The native-threads implementation of Allegro CL changes some basic assumptions of the foreign functions user interface. There is always exactly one native thread per Lisp Process, but there is not necessarily a Lisp process for every thread. Threads are free to run whenever they want; however, only one thread at a time can access the Lisp heap (for read or write); a thread cannot access the Lisp heap unless it has "acquired" the heap, which is only possible after another thread has "released" the heap.

See Releasing the heap when calling foreign functions in foreign-functions.htm for more information on this point.

def-foreign-call allows for the specification of whether to release the heap or not during a call. The possibilities for the release-heap keyword argument are:

• :never - This is the default and is compatible to the original defforeign interface; the caller necessarily has the lisp heap, this call will not release it. Note that if the foreign code being called spawns any new threads, or if it allows another thread to run, and the other thread attempts to call back into lisp, it will have to wait for the lisp heap. The danger is that the original thread may be waiting for results from its partner thread, but it has not yet given up the heap (this constitutes a deadlock situation). If this situation holds (foreign code does spawn new thread which call back into Lisp), :when-ok is the appropriate value for the release-heap argument.
• :always - The foreign call always gives up the heap before making the transition into non-lisp execution. If the form is being called within the dynamic context of a without-interrupts form, an error is signaled.
• :when-ok - The current dynamic context is examined, and if without-interrupts is in effect, the heap is not released; otherwise the heap is released.

Garbage collections in an SMP Lisp

There is a big difference in dealing with other-thread and garbage collection (gc) action intermingled with foreign calls in all of virtual-threads (processes are managed by Lisp), os-threads (processes are managed by the OS, but only one hardware processor is used for Lisp code), and SMP (multiple hardware processors can be used for Lisp and foreign code) Lisps. All three are different. This makes designing functionality which will run correctly on all three complicated.

The release-heap argument was added for os-threads, and has options :never, :always, and :when-ok. These are still the options. They pertain to whether the thread calling a foreign-function wants to allow another thread to access the Lisp heap during the call (note that amonst other things, this could lead to a garbage-collection).

The virtual-threads Lisps are simpler. Because they are more constrained, :release-heap :never is the only option that makes sense, since there is no concept of releasing the heap by a thread separate from pausing the thread.

If the user is coding for both virtual and os-threads, he can code for os-threads, and then add the :release-heap-ignorable t option to allow virtual-threads calls to compile without warning (and with the implied descriptive value of :never for the :release-heap option).

SMP Lisps are more complicated As with virtual threads, there is again no concept of heap ownership, in this case not because the heap is automatically assigned to the running thread, but because there are no longer any heap management constraints except one (explained later) which means that any thread can access the heap at any time. So the concept of releasing the heap makes even less sense in an SMP Lisp than it does in a virtual-threads Lisp.

In a pure SMP Lisp, there would be no restrictions at all on foreign calls. But, as mentioned above, although the heap itself is no longer a one-at-a-time resource needing gating, the garbage-collector, which is a stop-the-world collector still, must have complete and exclusive access to the heap when it runs, and so threads must specify whether to allow the gc to run.

def-foreign-call has two new keyword arguments: allow-gc and release-heap-implies-allow-gc to deals with these issues.

allow-gc,defaults to :never, to keep it in line with the release-heap argument. It can have the same values as release-heap, and they roughly mean the same thing (with the difference being that the garbage-collector is allowed to run either never (with :never), when interrupts are enabled (with :when-ok) or always (with :always). Note that as with os-threads, in an SMP Lisp, :allow-gc :never does not lock out the gc if the foreign code performs a callback to Lisp. So appropriae measures must still be taken, if a callback is possible, to reacquire Lisp objects after a possible gc, because they may have moved.

Note 4: String conversion when no arguments are specified

(ff:def-foreign-call foo ()) has no arguments specified, meaning that any number of arguments of any type will be passed. If any of these arguments are strings, it may be that string conversion should be performed.

String conversion is done by default. The above def-foreign-call form will generate a warning message (to indicate that string arguments will be converted). To suppress the warning but still convert the strings, specify the value t for the strings-convert keyword argument. To suppress the warning and to suppress all automatic strings conversion, specify the value nil for the strings-convert keyword argument.

Note 5: More on the error-value argument

Getting an error value from a foreign call is somewhat complicated, so users should not casually use this feature, but instead consider whether the information is necessary and be aware of various ways in which the information may be incorrect.

First note the following:

• The cost of obtaining the error value and returning it as the second value is not insignificant, so the value should only be obtained (by specifying a non-nil value for the error-value argument) if there is reason to believe it may be needed.
• The only way to get the correct value is to specify the correct non-nil value for for the error-value argument. Foreign function call are complicated and typically additional functionality is wrapped around the actual function call by the system (that is, by Allegro CL). This additional functionality may affect the value of the errno system variable in the current system thread or the value returned by GetLastError() so simply looking at errno or calling GetLastError() after the forreign call has returned will not work: the value seen can be different from the correct value. When error-value has an appropriate non-nil value (appropriate values are discussed below), code is added by the system to get and store the value of errno or the return value of GetLastError() immediately after the actual foreign function returns (and before the additional wrapper code is run). Using error-value is the only way to get the correct value, and even that may not always work, as the next bullet discusses.
• The returned as the error value assumes the foreign function being called behaves correctly with regard to errors. Thus, the value of errno has to be set to 0 at the beginning of the call or the error value seen by GetLastError() has to be cleared. Then the value seen when the call returns will be correct. If this is not done, then the value seen need not be correct. So, for example, if the foreign call does nothing to set errno, and its value happens to be 2 when the call is made, and the function runs without error so errno is not affected by any error handler, then the return error value will be 2 even though it should be zero. Note that this has nothing to do with Allegro CL and the code it generates. Allegro CL is returning what it promises: the value of errno or that retruned by GetLastError() immediately after the foreign function returns. It is just that value in such a case is bogus. The Allegro CL shared library does have a function clear_errno_value(). While this can be called by the actual foreign function, if you have such a function where you cannot modify the source, you could call this function, either in a wrapper around the foreign function of interest or even in a separate foreign call prior to the call of interest and, while not guaranteed for reasons just given, will make the second return value more likely to be correct.

The two acceptable non-nil values for error-value are :errno and :os-specific.

On UNIX and UNIX-like platforms

The two non-nil values have the same effect: the value of errno is stored immediately after the actual foreign function returns (and before additional wrapper code is run), and then returned as the second return value of the foreign call.

On Windows

The values have different effects.

If you specify the value :os-specific, then GetLastError() is called immediately after the actual foreign function returns (and before additional wrapper code is run), and then returned as the second return value of the foreign call. This is the standard Windows procedure, and works for all standard Windows routines, for example LoadLibrary() (see the description of that function in the MSDN where it says that you get error information by calling GetLastError()).

On Windows, you must specify the appropriate way to get the error value for the routine you are calling. Specifying :errno will give a wrong value for LoadLibrary(). Specifying :os-specific will give a wrong value for (many, perhaps all implementations of) open().

Note 6: More on the method-index argument

This argument is supported on Windows only. If a non-nil value is specified for the method-index keyword argument, then the value must be a vector whose first value is the integer index into the virtual table of the C++ class. And then, when the foreign function is called, the first argument of the call must be the vector whose first element is an integer which is a pointer to the table of method addresses. The arguments specified to def-foreign-call follow that first argument. Note that the function name specified in the def-foreign-call form is used only for the Lisp function name, and does not refer to any function in a shared library. Instead, the function's address (as a table address and an index into the table) is passed when the function is called.

Note 7: Potential problems with foreign functions returning :int in 64-bit Lisps

def-foreign-call is defined to default its :returning keyword argument (specifying the expected type of value returned by the foreign call) to :int. :int corresponds to C's int type. However, if the foreign function does not actually return an int, subtle bugs could be introduced in programs, particularly if the C function returns a long, an unsigned long, or a pointer of some sort. In 32-bit Lisps, returning those values is not a problem (when :returning :int is specified or defaulted to) because int is always 32 bits on every architecture we support. But on 64-bit Lisps, if a 64 bit value is returned, the upper 32 bits are lost. If the value was not correctly sign-extended by the foreign code, a negative value in the foreign code could be seen by Lisp as a large positive value. Also, the :long or :unsigned-long type is an inadequate specfication because on Windows long types are always 32 bits. So on either 32 or 64 bit lisps, for portability, use :returning :unsigned-nat when the return value is some kind of pointer. (See The :nat and :unsigned-nat types in implementation.htm for a description of the :nat and :unsigned-nat types.) When the returned value is an integer value, be sure to use the correct type and be sure that the foreign code actually produces that type.

Note 8: Returning booleans

def-foreign-call can accept :returning :boolean as a return type, and it will be automatically translated to a canonical form which is :returning (:int boolean). This works with some foreign languages (where 0 is considered false and anything else is considered true - it translates 0 into nil on the lisp side and anything else to the value t), but in some cases (C++, for instance) the boolean type is equivalent to signed-char type, and so it must be specified in the def-foreign-call as :returning (:char boolean).

Note 9 on the :lisp return type

The :lisp return type means that the foreign code will return a Lisp object. For most Lisp objects (other than characters, fixnums, and a few others), this means returning a pointer to the object. If for whatever reason the object has been moved before the pointer is in a place where it can be seen by the garbage collector and properly forwarded, the pointer will be invalid and the result may be Lisp failure because of gc failure. For this reason, a warning is signaled when :lisp is specified as a value for :returning.

The warning can be avoided by specifying :aligned instead of :lisp as the :returning value. Although aligned pointers could also represent objects that move during gc, it is much less likely so, because they tend to only be used when the objects they represent are static, like stack-based objects or allocated in aclmalloc space. Note associating the actual Lisp value with an :aligned value is complicated. See Aligned Pointers and the :aligned type in ftype.htm for more information.

If you are sure the item will not move (because it is located in statis space, for example), you can give t as the value of the lisp-return-will-not-move keyword argument. That supresses any warning about :returning :lisp.

Examples:

(def-foreign-call add2 (x y))

The symbol add2 will have a function definition calling the foreign function probably named "add2" in C, whose first arg is named "x" and is an integer in Lisp and which is converted to an int for passing to C. If the integer is larger than can be held in a C int, it is truncated. As with the first arg, the second arg named "y" is an integer converted to a C int. The return value is interpreted as a C int type, and is converted to a Lisp integer (which may either be a fixnum or consed as a bignum).

We say the foreign function is "probably" named "add2" because since no specific name or conversion function is specified, the default system conversion function is used. It depends on the platform and platform-specific rules but typically downcases the symbol name.

(def-foreign-call t_double ((x :double)
                            (y :double single-float)
                            (z :int fixnum))
  :returning :double)

Call a function, probably named "t_double" in C (again "probably" because the actual name depends on platform-dependent defaults), whose first arg is a double-float both in Lisp and in C, and whose second arg is a single-float in Lisp but is converted to double float for passing into C (this is the calling convention used by some non ANSI C compilers and by others when the arguments are not prototyped), and the third argument is a fixnum Lisp passed as an integer to C. The function returns and boxes a double-float value to Lisp.

(def-foreign-call (t-float dash-to-underscore) ((x :double)
                                                (y (:float :no-proto))
                                                (z :int fixnum)
                                                (w (* :char) string))
  :returning #-(or (and sun4 (not svr4)) sun3q) :float
             #+(or (and sun4 (not svr4)) sun3q) (:double single-float)

(def-foreign-call (t-float "t_float") ((x :double)
                                                (y (:float :no-proto))
                                                (z :int fixnum)
                                                (w (* :char) string))
  :returning #-(or (and sun4 (not svr4)) sun3q) :float
             #+(or (and sun4 (not svr4)) sun3q) (:double single-float)

These two examples do the same thing: call a function, named "t_float" in C (assuming in the first case proper conversion by dash-to-underscore, which must already be defined and should downcase the symbol name and replace dashes with underscores), whose first arg is a double-float both in Lisp and in C. Like the previous example, the second arg is a float in Lisp, and is converted to double float for passing into C. The third arg named "z" is a fixnum passed as an int, and "w" is a (null-terminated) Lisp string, whose first-character-address is passed to C (beware, the string may move if a gc is allowed). Depending on the architecture, the C function will return either a double (from older C compilers) or a float, each interpreted and boxed as a Lisp single-float value.

We give both examples to show how a lisp name (the symbol t-float) is converted to a foreign name ("t_float"). You can either specify a function that takes a symbol as an argument and returns the correct string (so (dash-to-underscore 't-float) returns "t_float") or you can simply specify the correct string. Note again that dash-to-underscore must be already defined when the def-foreign-call form is evaluated.

Macroexpansion of a def-foreign-call-form now provides more useful information

def-foreign-call has enhanced the macroexpansion to give useful information for users. Here is an example:

cl-user(11): (pprint 
               (macroexpand '(ff:def-foreign-call foo (x (y (* :char))))))

Warning: A runtime with-native-string call is being generated for argument `y'
         to the foreign-function `foo'.  The with-native-string macro can be
         used for explicit string conversions around the foreign calls.  This
         warning is suppressed when :strings-convert is specified in the
         def-foreign-call.

(progn (eval-when (:compile-toplevel)
         (excl::check-lock-definitions-compile-time 'foo 'function
           'foreign-functions:def-foreign-call (fboundp 'foo))
         (push 'foo excl::.functions-defined.))
       (eval-when (compile load eval) (remprop 'foo 'system::direct-ff-call))
       (setf (fdefinition 'foo)
             (let ((excl::f
                    (named-function foo
                      (lambda (x y)
                        (excl::check-args '((:int (integer * *))
                                            ((* :char) (array character (*))))
                                          'foo x y)
                        (with-native-string (#:g34798
                                             (if*
                                              (stringp y)
                                                then y
                                                else ""))
                                            (unless
                                             (stringp y)
                                             (setq #:g34798 y))
                                            (symbol-macrolet
                                             ((y #:g34798))
                                             (system::ff-funcall
                                              (load-time-value
                                               (excl::determine-foreign-address
                                                '("foo" :language :c)
                                                2
                                                nil))
                                              '(:int (integer * *))
                                              x
                                              '((* :char)
                                                (array character (*)))
                                              y
                                              '(:int (integer * *)))))))))
               (excl::set-func_name excl::f 'foo)
               excl::f))
       (record-source-file 'foo) 'foo)
cl-user(12): 

The effects of the new, longer array implementation on def-foreign-call

The new array implementation is discussed in the Arrays and short arrays section in implementation.htm. In brief, standard Common Lisp arrays now can be significantly larger than in earlier releases, while the new short arrays implement the old arrays (the same size limitations but also the same type codes and structure).

In this discussion, `array' refers to the newly-implemented arrays, while `short array' refers to the old implementation, preserved as short arrays in 7.0.

Foreign calls are made with arrays as arguments by passing the address of the first value. In the new implementation, simple-arrays always have exactly the same first element offset (but some short arrays are aligned to the next higher word boundary so that the elements within are naturally aligned). This sometimes-difference between arrays and short-arrays poses an extra burden on the ff interface, in that the arrays must be distinguished between themselves at runtime.

It is now possible to declare an argument a (short-simple-array ... (*)) and the interface will generate code as it did before for normal arrays, passing the address of the first argument.

For 7.0, a declaration of (simple-array ... (*)) actually generates code that tests at runtime whether the argument is a short array or a normal array. So in effect, a short simple-array passed in as if it were a normal simple-array will be properly passed.

Note that with this setup, if argument checking is specified and a short-array is passed in, the check will fail, because a short-array is not a subtype of simple-array. But for 7.0, if you suppress this argument checking, the interface will pass either array correctly.

However, programmers are urged to provide correct declarations and pass the correct type of array even though 7.0 allows sloppiness. In a future release, we anticipate adding a (dual-simple-array ... (*)) declaration to the direct-call foreign interface, and to move that functionality from its current place in the (simple-array ... (*)) declaration. This would then mean that the (simple-array ... (*)) declaration would only pass simple-arrays, and not short simple-arrays. This change will also make more consistent the arg-checking feature of the direct-call interface, since the dual-simple-array type is defined as an or of simple-array and short-simple-array, so a short-simple-array being passed will pass the test of being a dual-simple-array.

See ftype.htm for information on foreign types in Allegro CL and foreign-functions.htm for general information on foreign functions in Allegro CL. See particularly def-foreign-call in foreign-functions.htm.

Copyright (c) 1998-2022, Franz Inc. Lafayette, CA., USA. All rights reserved.
This page has had moderate revisions compared to the 10.0 page.
Created 2019.8.20.