Skip to content

Latest commit

 

History

History
924 lines (716 loc) · 28.5 KB

README.md

File metadata and controls

924 lines (716 loc) · 28.5 KB

Now with libffi!

Using its own facilities, autowrap now includes autowrap/libffi. This allows functions that pass and return structs to be called using autowrap. To use this, just load or :depends-on cl-autowrap/libffi instead of cl-autowrap:

  :depends-on (... :cl-autowrap/libffi ...)
  ...

Of course, this requires libffi be compiled and available to your lisp.

Usage mostly identical; functions called via libffi look and act the same as any others. The one exception is functions that return a struct by value:

(c-with ((return-value some-struct))
  (some-call-returning-some-struct return-value ...))

Calls returning a struct take the return as their first parameter. This should be evident in SLIME/Sly. This ultimately allows easier management of freeing as well as more control over where they're stored. (They also still return the value, if you're trying to chain calls.)

Issues?

If you have issues, do not hesitate to file an issue! See the FAQ for some quick tips.

cl-autowrap

This is a new c2ffi-based wrapper generator for Common Lisp with a focus, performance, convenience, and completeness. It works like this:

(c-include "file.h")

That's it. This calls c2ffi and generates architecture-specific .spec files you can distribute with your project. Neither c2ffi, nor any compiler (or even .h files!) are necessary for your users!

  • Types: structs (including bitfields!), unions, enums, typedefs, etc.
  • Wrappers: Thin wrappers for structs, unions, and aliases of structs and unions: pointers and validation only, no expensive translation
  • Accessors: Complete recursive accessors for structs and unions, including anonymous struct members, arrays, and pointers for each field.
  • Functions: Macros which expand into foreign calls via CFFI-SYS
  • Metadata: Full access to all the information about all types, functions, etc.

For instance:

/* test.h - abbreviated from example */
typedef struct foo {
  int a, b;
  char c[3];

  struct {
    unsigned int b0 : 2, b1 : 3;

    struct {
      char x, y;
    } s;
  } x[2];
} foo_t;

foo_t* get_foo();
void free_foo(foo_t *foo);
int* get_int();

Out of this, we can do the following. (Note: dots are just part of the function names for disambiguation, this doesn't alter the reader):

(c-include "test.h")

(let ((foo (get-foo)))
  (setf (foo-t.a foo) 5)             ;; foo.a = 5;
  (setf (foo-t.x[].b0 foo 0) #b10)   ;; foo.x[0].b0 = 2;
  (print (foo-t.x[].s.x foo 1))      ;; anonymous struct
  (foo-t.x[].s foo 0)                ;; => child wrapper
  (foo-t.x[].s& foo 0)               ;; &(foo.x[0].s) => pointer
  (free-foo foo))

Alternatively, there is now cl-plus-c, which can optionally be loaded for a different access mechanism and much quicker compile times:

(asdf:load-system :cl-plus-c)
(use-package :plus-c)

;;; This allocates a FOO-T and frees it at the end:
(c-let ((foo foo-t :free t))
  (print foo)                        ;; => wrapper
  (setf (foo :a) 5)                  ;; foo.a = 5;
  (setf (foo :x 0 :b0) #b10)         ;; foo.x[0].b0 = 2;
  (print (foo :x 1 :s :x))           ;; anonymous struct: foo.x[1].s.x
  (foo :x 0 :s)                      ;; => child wrapper
  (foo :x 0 :s &))                   ;; &(foo.x[0].s) => pointer

See cl-plus-c.md for more information.

Overview

Using cl-autowrap is meant to get you to the "lispifying" stage of your wrapper as quickly and conveniently as possible:

  • Make sure you have c2ffi
  • Load your library as you normally would with CFFI
  • Make a file for your c-include
  • Examine wrappers and tweak if necessary
  • Back to lisp!

c2ffi

You will need to build c2ffi if you have not already done so. This requires a repository version of LLVM and Clang, but the build process is straightforward.

Again, note that your users do not need this, assuming you distribute the .spec files appropriate to their architecture. cl-autowrap should generate everything for you, though.

If you decide not to install c2ffi, you can specify its path directly by setting autowrap:*c2ffi-program*, e.g.:

(setf autowrap:*c2ffi-program* "/path/to/my/c2ffi")

This should be part of your local configuration; do not set this in code you distribute. This includes LET forms around C-INCLUDE.

Loading Libraries

This should be done normally with CFFI. Either the high-level interface with CFFI:DEFINE-FOREIGN-LIBRARY and CFFI:USE-FOREIGN-LIBRARY or the low-level interface with CFFI-SYS:%LOAD-FOREIGN-LIBRARY work.

Writing the c-include

It's highly recommended that you use a separate package and file for cl-autowrap. The reasons are simple:

  • A lot of symbols will be generated without regard.
  • Many symbols will also be exported.
  • A rather large number of functions and structs will be generated, resulting in a hefty compile time. With a separate file, this only needs to happen once.

(In fact, you can now specify individual packages for each set of symbols that are generated. See below.)

Once you have this, you can write a simple c-include. This must be a top-level statement:

(c-include "somefile.h")

This will look for somefile.h and generate .spec files in *default-pathname-defaults*, which is probably not very helpful! To fix this, use the following:

(c-include (function-that-finds "somefile.h")
           :spec-path #P"/path/to/spec")

(Note that while these parameters are eval'd, this happens at compile time, so if you use a *special-variable*, its definition needs surrounded by an EVAL-WHEN.)

Hardcoded paths and reinventing functionality aren't very nice though; In both cases you can specify a complete "ASDF path" (starting with the system name), and it'll query the path from ASDF. For example, if we have an ASDF system called my-wrapper, we can do the following:

(c-include '(my-wrapper some-module "localfile.h")
           :spec-path '(my-wrapper spec-module))

Assuming you had defined "localfile.h" as a :static-file of some-module in my-wrapper, as well as spec-module, everything would work as intended.

This is especially useful because you can have a single local header that includes all the files you wish to wrap, and those will be found by c2ffi in the standard include paths.

Tweaking

While c2ffi and cl-autowrap do quite a lot, there are a few times where you may want to or be required to intervene. You can look at any errors that occur, or the symbols that are exported, or even simply macroexpand the c-include and examine the output.

By default, c2ffi outputs everything and likewise cl-autowrap imports everything. Thus you get a rather large sampling of libc where you probably don't need it. Thus you may want to exclude some definitions. You can do this in two ways:

(c-include "file.h"
           :exclude-sources ("/path/to/source1"
                             "/path/.*source2" ...)
           :exclude-definitions ("SomeFunc1"
                                 "_suffix$"))

The first, :exclude-sources, looks at the source information generated by c2ffi for each definition. This is an easy way to exclude the majority of irrelevant definitions. You can make exceptions to this list via :include-sources:

(c-include "file.h"
           :exclude-sources ("/path/to/source1"
                             "/path/.*source2" ...)
           :include-sources ("/path/to/source1/but-include-this"))

While everything else matching "/path/to/source1" will be excluded, in this example, definitions in "/path/to/source1/but-include-this" will still be included (if they exist).

The next specifier, :exclude-definitions, excludes specific definitions by name. These may be conflicting or unnecessary. For instance, SDL2 includes a number of functions ending in _inline and some functions which use stdargs, all of which are unnecessary (or unusable).

Both of these use cl-ppcre regular expressions to match, thus you have a great deal of flexibility with a few strings.

You may also wish to simply rename some symbols. The default routine generally translates symbols like you want, but you may occasionally find C functions named in a way that breaks this. The default rules are as follows:

  • XYZFooBar => XYZ-FOO-BAR
  • foo_barBaz => FOO-BAR-BAZ
  • _x_y => _X_Y (because I think -X-Y looks worse)

However if you encounter something like "FOObar", it is likely you want "FOO-BAR", not "FO-OBAR", which is what you would get. Thus you can specify an exception:

(c-include "file.h"
           :symbol-exceptions (("FOObar" . "FOO-BAR") ...))

These are simple, case-sensitive string matches and replacements. The replacement is interned exactly, so if you specify lowercase here, you will get a symbol with lowercase characters.

There is also a more complex cl-ppcre-based match and replace facility:

(c-include "file.h"
           :symbol-regex (("^MATCH_string" (PARAMS)
                            (lambda (string matches regex) ..
                               NEW-STRING))))

Using this facility, you may specify regex-function pairs. PARAMS specifies further parameters to PPCRE:CREATE-SCANNER, e.g., :case-insensitive-mode. If a symbol matches the given regex, the function will be called with the string, any substring matches, and the original regex (in case you want to further apply it). You must return a string, which will then be converted by the above rules into a final string.

This should usually be unnecessary. The use case for its creation was handling names that vary unpredictably only by case:

CLUTTER_KEY_OMEGA
CLUTTER_KEY_omega
CLUTTER_KEY_THORN
CLUTTER_KEY_Thorn
CLUTTER_KEY_Adiaeresis
CLUTTER_KEY_adiaeresis

In this situation, the more complicated regex-function matching is necessary.

Alternatively, as was actually decided for the above clutter case, since there was "no rhyme or reason" to the naming scheme of the #define'd constants, one may filter constant names to be interned, opting, instead, for referencing them through a separate constant-accessor macro:

(c-include "file.h"
            :exclude-constants (".*")
            :constant-accessor clutter-constant)
;; Access constants like this:
(clutter-constant "CLUTTER_Z")
(clutter-constant "CLUTTER_z")

By default all "known" architectures (at the time of writing, windows, mac, linux on i686 and x86_64) are generated by default. This may not always work; for instance, one architecture may require header files your system lacks. You can exclude it using the following:

(c-include "file.h"
           :exclude-arch ("i686-pc-win32" ...))

This will exclude that target triple from being generated and causing a warning or output if it fails.

You can also specify individual packages for symbol exports. This can be useful if, for instance, you wish to import all accessors, or all functions, or similar, while not necessarily importing everything:

(c-include "file.h"
           :definition-package PACKAGE
           :function-package PACKAGE
           :wrapper-package PACKAGE
           :accessor-package PACKAGE
           :constant-package PACKAGE
           :extern-package PACKAGE)
  • :definition-package PACKAGE: All "definition" symbols, which include type names and function names (not to be confused with function macros which you use in your code)
  • :function-package PACKAGE: All "function" symbols, which are all macros expanding to foreign calls
  • :wrapper-package PACKAGE: All "wrapper" symbols, which are all structs generated to wrap foreign record types
  • :accessor-package PACKAGE: All "accessor" symbols, which are all functions generated to access record fields
  • :constant-package PACKAGE: All "constant" symbols, which are all +symbols+ representing C constants
  • :extern-package PACKAGE: All "extern" symbols, which are all symbols (which are symbol-macros) representing C extern symbols

Wrappers and FFI

At this point you probably have definitions generated (or are hopefully submitting a question or bug report!). But how to use them?

While cl-autowrap uses CFFI, it almost exclusively uses the low-level CFFI-SYS interface. It does not use the high-level type translation interface, or even cffi:defcfun. Pointers are still whatever your Lisp provides.

Instead, cl-autowrap defines a "new" higher-level interface I call SFFI, for "simplified FFI". While CFFI's high-level interface is nice for manually defining types and functions, it proves difficult when trying to automatically generate things or exercise precise control over various things like field layout.

You should never have to deal with SFFI directly, but all the fine-grained type information is available should you require access. This is occasionally useful. See below in the SFFI section for details.

However, you cannot use CFFI constructs from another wrapper directly with SFFI-defined functions, or vice versa, but you can always use pointers between the two.

Functions

cl-autowrap defines macros which wrap C calls with a few helpful features:

  • Wrappers (see below) or pointers are accepted for any pointer-to-struct (or union)
  • Symbols or integers are accepted for any enum
  • Lisp strings will be temporarily converted to C strings, then freed, for char* or unsigned char*. (If you need these to persist, you must provide your own pointer!)
  • For char* and unsigned char* returns, both a lisp string and a pointer are returned as VALUES, so you can free the pointer if necessary. You may prevent this conversion, and receive only the pointer, if you wrap the call in INHIBIT-STRING-CONVERSION:
(inhibit-string-conversion (function-returning-string ...))
  ;; => pointer

Otherwise, the call will be like any C call; there is no other type translation. In my experience, all but the most trivial C functions benefit from some wrapping, so this shouldn't be a big issue.

However, see "Other Features" below for some other helpful features, such as bitmasks.

Wrappers

Instead of merely returning pointers, cl-autowrap defines very thin wrappers for non-atomic named types. Wrappers are structs which contain two things:

  • A pointer, which is accessible with AUTOWRAP:PTR
  • A VALID-P field, which is used for storing pointer validity, and can be checked by AUTOWRAP:VALID-P

Wrappers are extremely useful for "safely" managing pointers, and are meant to be safe and "pretty" enough for users of your wrapper to use directly. Any dereference using PTR automatically checks validity, and you can use finalizers to clean them up. Note however that this is up to you: cl-autowrap merely provides the facility, nothing else. See "Garbage Collection and Wrappers" below.

Additionally, cl-autowrap generates a correct "type hierarchy", as much as such applies to C:

struct x { ... };
typedef struct x y;

Results in:

(defstruct (x (:include wrapper)))
(defstruct (y (:include x)))

This ensures type compatibility where the C side may arbitrarily specify compatible type aliases.

You may also obtain a "child" wrapper for a struct which is a field in another struct, using accessors:

struct foo_t {
   :
   struct { int a, b; } x;
};
(let* ((foo (get-foo-somehow))
       (x (foo-t.x foo)))
  :
  :
  ... )

This keeps a reference to the parent. These may also be safely dereferenced using AUTOWRAP:PTR, and checked using AUTOWRAP:VALID-P. Because there is a reference is kept to the parent, even if a reference is discarded by the user, the child is still safe to use.

Garbage Collection and Wrappers

One of the primary motivators behind wrappers is the ability to easily garbage collect C data. However, this still requires some care. To this end, the AUTOCOLLECT macro has been added; see below.

First, nothing besides checking is done automatically. Pointers are assumed valid when they are returned and made into wrappers. Any further invalidation and garbage collection must be handled by the one writing the wrapper.

Important: Absolutely no effort is made to keep wrappers unique or manage duplicates. Again: YOU CAN HAVE DUPLICATE WRAPPERS AND THIS CAN LEAD TO BAD THINGS. Generally this should only occur if you obtain the same pointer from a C API multiple times, such as a function which returns a global context pointer. It is up to you to handle this. Beware.

Once you are aware of this, you can use something like trivial-garbage to free pointers when you need:

(defun lispy-get-thing ()
  (let* ((thing (get-thing))
         (ptr (autowrap:ptr thing)))
    (tg:finalize thing (lambda () (free-thing ptr)))
    thing))

Note as as always to never reference the object, only the pointer, in the finalizer, or it will never be collected.

It is often useful to free things when you still have a reference. In this case, the pointer becomes invalid, and this is also handled by WRAPPER:

(defun lispy-free-thing (thing)
  (unwind-protect (free-thing thing)
    (tg:cancel-finalization thing)
    (autowrap:invalidate thing)))

In this case, further attempts to dereference THING via AUTOWRAP:PTR will result in an INVALID-WRAPPER error.

You may be tempted to do this:

(defun bad-free-thing (thing)
  (tg:cancel-finalization thing)
  (autowrap:invalidate thing)
  (free-thing thing))

Unfortunately, since you invalidated THING, when you pass it to FREE-THING, it will be invalid ... resulting in an error.

Never manage "child" wrapper objects. This probably goes without saying, but they're tied to the parent object, and not meant to be managed separately.

Also, you may be tempted to do this, to avoid "dangling pointers":

(defun terrible-get-thing ()
  (let* ((thing (get-thing))
         (ptr (autowrap:ptr thing)))
    (tg:finalize thing
      (lambda ()
        (free-thing ptr)
        (setf (autowrap:wrapper-ptr thing)
              (cffi:null-pointer))))
    thing))

This is both wrong and silly: there is a reference to THING in the finalizer, so it will never get freed. And if you had gotten here normally, there would be no references, so nothing would have the dangling pointer!

To facilitate doing this correctly, the AUTOCOLLECT macro has been added:

(autocollect (&optional PTR) WRAPPER-FORM &body) => WRAPPER-FORM-RESULT

If you are using trivial-garbage, this will extract the pointer from WRAPPER-FORM and call tg:finalize on the wrapper. The body forms should use POINTER to free the object. If you are not using trivial-garbage, it will produce an error.

For instance:

(autocollect (pointer)
    (get-thing)
  (free-thing pointer)) ;; => THING-WRAPPER

This will call GET-THING and finalize the resulting wrapper with the body. POINTER is the pointer; this defaults to the symbol PTR.

This is not fool-proof. Things to watch out for:

  • If you reference the wrapper, and not the pointer, it will never be collected.
  • If you try to autocollect a child wrapper, you will probably crash.
  • If you provide a function to manually free resources, you must use tg:cancel-finalization or this finalizer will still be called, likely double-freeing the memory and crashing.
  • It's still up to you to call something to free the pointer.

Accessors

Having wrappers and functions are nice, but getting at the data is important too. Accessors are generated recursively (up to a depth of 5, barring recursive types) for highly convenient access. From the top:

typedef struct foo {
  int a, b;
  char c[3];

  struct {
    unsigned int b0 : 2, b1 : 3;

    struct {
      char x, y;
    } s;
  } x[2];
} foo_t;

Accessors are named starting with their type name (in this case, FOO and FOO-T), followed by fields, separated by dots. There is no reader magic here: these are functions with dots as part of the name. (Dots were used mostly for disambiguation; if only dashes were used, name collision would be probable, since underscores are converted to dashes by default.)

The following special cases are available:

  • type.foo in the case of foo being a record type (struct or union), will return a child wrapper.
  • type.foo in the case of foo being a pointer will return the pointer
  • type.foo& will return a pointer to the field, not a wrapper, regardless of the type of foo. If foo is a pointer, then you get a pointer-pointer. This is just the same as &(x.foo) in C. This does not exist for bitfields.
  • type.foo* will dereference foo where foo is a pointer to a well-defined type, including record fields. E.g., type.foo*.bar; there is no type.foo.bar.
  • type.foo[] references an array element, where foo is declared as an array (not just as a pointer). Array indices are specified in order after the object: (type.foo[].bar[] obj i0 i1) is the equivalent of C's obj.foo[i0].bar[i1].

Additionally, SETF can set almost any field. The exceptions are any accessor which dereferences a record (i.e. returns a child wrapper), or is suffixed with &.

Bitfields are supported under the assumption that they are packed LSB-to-MSB on little endian and MSB-to-LSB on big endian architectures. If you actually encounter a problem with this, file a bug report with full details: the architecture, OS, lisp, C compiler, and an example struct. Theoretical possibilities are not considered bugs.

Note that bitfield operations cannot be done atomically and may not be done field-atomically (that is, you may have to lock the entire struct). Additionally, you cannot take the address of a bitfield. However, you can get information from SFFI metadata, or simply using the convenience function AUTOWRAP:BITFIELD-MASK.

Other Features

cl-autowrap has a number of other features that have not been discussed:

  • Allocation
  • Enums
  • Bitmasks
  • Callbacks
  • SFFI metadata and functions

Allocation

Since autowrap implements its own higher-level constructs over lower-level CFFI, you can't use CFFI's FOREIGN-ALLOC or similar functions and macros to easily allocate foreign records. Thus there are new constructs for doing so:

(let ((thing (autowrap:alloc 'type)))
  :
  (autowrap:free thing))

As you might expect, ALLOC will allocate memory of sufficient size for TYPE, and FREE will free it (and invalidate the wrapper for you). Note that if you are doing garbage collection as above, this does NOT remove finalizers for you: you MUST take care of this yourself where applicable.

There are also macros which will help with temporary allocation:

(with-alloc (thing 'type)
   :
   :
   )

This will take care of allocation and freeing within the block. You should not use finalizers here. If you try to reference the value outside of the scope of the block, it will be invalid. If you wish to allocate multiple objects and free them, you can use the following:

(with-many-alloc ((thing1 'type1)
                  (thing2 'type2)
                  :
                   )
  :
  )

Note that while any typedef type aliases can be referenced simply by symbol as in C, record types are called (:struct (NAME)) or (:union (NAME)), and also like C, you must write this out if there is no type alias for NAME. For example:

struct X { ... };
typedef struct Y { ... } Y;

int main() {
    struct X foo;    /* No type alias */
    Y bar;           /* Type alias */
}
(with-many-alloc ((foo '(:struct (X))) ;; No type alias
                  (bar 'Y))            ;; Type alias
    :
    )

Arrays

In addition to single objects, autowrap also allows allocation and reference to arrays of objects. This is less safe, however: there are no provisions for bounds-checking, since the data is simply not there. (While in theory, we could add size data on the lisp side, this is a false sense of security, since you will often be dealing arrays from C.)

Allocation methods all take an optional COUNT parameter:

(alloc x 'type 3)

(with-alloc (x 'type 5) ...)

(with-many-alloc ((x 'type 5)
                  (y 'type 2))
  ...)

To reference these, you can use C-APTR and C-AREF:

(c-aptr x 1) ;; => raw pointer
(c-aref y 2) ;; => wrapper

Unfortunately, this may present some performance issues, since unlike record accessors, the type must be looked up at runtime. In theory, autowrap could generate array accessors for all types, but this would vastly increase the number of accessors generated with little value, since most will not be used.

Instead, you may specify the type explicitly:

(c-aptr x 1 'type) ;; => pointer
(c-aptr y 2 'type) ;; => wrapper

In this case, as long as 'type is constant-p, the compiler macro should expand it at compile-time.

Basic C types (e.g., :int, :char, etc) are also supported; in this case, a wrapper is not returned, but the value itself:

(c-aref x 1 :int) ;; => number

You can also set array members for basic types only:

(setf (c-aref x 1 :int) 10)

In both of these cases, since autowrap does not provide additional wrappers for basic types, you must specify the type explicitly.

Enums

Enums are imported and created as types, but they're typically used by specifying a keyword:

enum E {
  FOO_X, FOO_Y, FOO_Z
};

void fun(E);
(fun :x)

As you can see, common prefixes are eliminated and the symbols are interned as keywords. Additionally, functions taking enum symbols can also take numbers:

(fun 1)

You can also find the value or keyword for an enum as follows:

(autowrap:enum-key '(:enum (enum-name)) :key)
(autowrap:enum-value '(:enum (enum-name)) 1)

An actual AUTOWRAP:FOREIGN-ENUM can be used in place of 'enum-name if desired; otherwise it will be looked up via AUTOWRAP:FIND-TYPE.

Bitmasks

Bitmasks aren't actually a type in C, and are often defined as constants instead of enums or similar. Therefore, there is no real automatic way to determine a bitmask. Thus cl-autowrap provides a number of convenience facilities for doing this:

(autowrap:define-bitmask 'NAME
  '((:key1 . #x0001)
     :
      ...))

This defines a bitmask called NAME, which is separate from other C types, and can be used with the MASK function:

(some-function (autowrap:mask 'NAME :key1 :key5))

This also has a compiler macro which will expand to an integer constant if the value can be determined at compile-time.

Additionally, to aid in converting predefined constants to bitmasks, there is the following macro, which expands to an AUTOWRAP:DEFINE-BITMASK call:

(autowrap:define-bitmask-from-constants (name)
  +some-foo+
  +some-bar+
  +some-baz+)

This essentially expands to the following:

(autowrap:define-bitmask 'name
  (list `(:foo . ,+some-foo+)
        `(:bar . ,+some-bar+)
        `(:baz . ,+some-baz+)))

Callbacks

Autowrap now provides a thin layer on top of CFFI-SYS:%DEFCALLBACK:

(autowrap:defcallback NAME RETURN-TYPE
    ((PARAM TYPE)
     ...)
  ...)

The main difference is that you may specify SFFI type aliases as parameters, since these are not available to the higher-level CFFI:DEFCALLBACK.

Additionally, there is the following:

(autowrap:callback 'name)

This simply expands to CFFI-SYS:%CALLBACK, but is provided for convenience.

SFFI Metadata and Functions

This is not fully-documented at the moment, but full access to metadata and definition functions is available. For instance:

(autowrap:find-type '(:struct (struct-name)))

This will return the object that represents struct struct_name, or nil. If nothing else, it should be fairly easy to inspect this value and look at fields, types, etc. Accessors are exported for all types (or should be); see package.lisp for a complete list.

These values can certainly be useful when doing various tricky things with C data, and it's also certainly possible to manually write definitions for every type and generate lisp functions, though for records (i.e. struct and union), this requires explicitly specifying bit sizes and field layouts. While probably not directly useful (or necessary) for importing C types, these could be useful for generating similar definitions via other means than c2ffi.

Copying

This is licensed under the BSD 2-Clause license.