Carp/docs/CInterop.md
Tim Dévé da25a255e9
feat: Adds flag to always output C id with headerparse (#1353)
Adds -c|--emitcname flag to headerparse to always emit the C identifier
in `register` definitions (useful when the register will be in a
module).

Fixes a bug where the kebab case flag would not output C identifiers
making the emitted C identifiers not match with the ones in the headers.

Adds docs entry about headerparse in CInterop doc.

Makes headerparse emit `CChar` instead of `Char` when encountering a
signature containing `char`.
2021-11-03 09:09:26 +01:00

17 KiB

C Interop

This is an extension of what is covered in the Language Guide.

Content

How Carp generates identifiers

When creating a function or def it might be useful to know what identifier gets generated on the C side. Here are some examples:

(def a-def 100)
; => a_MINUS_def

(defn hello [] (println* "Hello"))
; => hello

(sig true? (Fn [Bool] Bool))
(defn true? [b] b)
; true_QMARK_

(defmodule Reverse
  (defn hello [] (println* "Goodbye"))
  ; => Reverse_hello
  (defmodule ReReverse
    (defn hello [] (println* "Hello"))))
    ; => Reverse_ReReverse_hello

; Generic signature
(sig print-first-and-add (Fn [(Ref (Array a)) b b] b))
(defn print-first-and-add [arr x y]
 (do
   (println* (Array.unsafe-first arr))
   (+ x y)))
; Generates no code until it is called

(print-first-and-add &[1] 1 1)
; => print_MINUS_first_MINUS_and_MINUS_add__int_int
(print-first-and-add &[@"hello"] 2l 40l)
; => print_MINUS_first_MINUS_and_MINUS_add__String_Long

Looking at the examples should be clear enough but let's break it down: Carp will replace illegal characters in C with a string representation of them (- => _MINUS_, ? => _QMARK_, etc...) If in modules it will prefix the identifier with the modules name. When the arguments to a function are generic it will suffix the types to the identifiers, the identifiers are not able to be generated until it is used. If a function is potentially generic but you don't want it to be you can add a non-generic signature to it to make Carp generate your function like in our true? example.

When creating bindings to a library it would be hard to coerce this logic into creating the exact identifiers the library uses, this is why register and register-type accepts an optional argument to specify what identifiers to use:

(defmodule CURL
  (register-type HttpPost "curl_httppost")
  (register form-free (Fn [(Ref HttpPost)] ()) "curl_formfree"))

Managed types

In Carp types like String and Array are managed types in that they are allocated on the Heap and the compiler will automatically free the allocated memory when they go out of scope. We'll see how we can go from these managed type to C and back.

String

To use a managed String with a C function requiring a char* you can use the String.cstr function that will turn your (Ref String) into (Ptr CChar):

(register puts (Fn [(Ptr CChar)] ()))

(let [a-str @"A string."]
  (puts (String.cstr &a-str)))
(puts (String.cstr "Hello"))

You may want to hide the C type from the end-user:

(defmodule MyMod
  (hidden puts-c)
  (private puts-c)
  (register puts-c (Fn [(Ptr CChar)] ()) "puts")
  (defn puts [str-ref] (puts-c (String.cstr str-ref))))

(let [a-str @"A string."]
  (MyMod.puts &a-str))
(MyMod.puts "Hello")

If you are given a char* and want to turn it into a managed String you can use String.from-cstr. It will allocate and copy the content of the C string.

// static-str.h
char* returns_a_static_str() {
  return "Hello";
}
(relative-include "static-str.h")

(register returns-a-static-str (Fn [] (Ptr CChar)) "returns_a_static_str")

(let [a-str (String.from-cstr (returns-a-static-str))]
  (println* (String.concat &[a-str @" " @"Carp"])))

The function you're consuming might be allocating the string on the Heap for you. In that case you can declare the function as returning a managed String. However this might be unsafe, you need to ensure that the string is actually Heap-allocated and that the allocator is the same as the one that Carp is using.

char* returns_a_heap_string() {
  char *hello = "Hello from the heap";
  char *str = malloc((strlen(hello)+1));
  strcpy(str, hello);
  return str;
}
(relative-include "heap-string.h")

(register returns-a-heap-str (Fn [] String) "returns_a_heap_string")

(let [a-str (returns-a-heap-str)]
  (println* a-str))

If you are the one writing the C code, you can use the CARP_MALLOC macro to ensure you are using the same allocator as the Carp compiler:

char* returns_a_heap_string() {
  char *hello = "Hello from the heap";
  char *str = CARP_MALLOC((strlen(hello)+1));
  strcpy(str, hello);
  return str;
}

Array

Array.unsafe-raw can be used in case you have a function taking an C array as a parameter.

int sum(int *arr, int len) {
  int acc = 0;
  for (int i = 0; i < len; i++) {
    acc += arr[i];
  }
  return acc;
}
(relative-include "sum.h")

(register sum-c (Fn [(Ptr Int) Int] Int) "sum")

(let [ints [1 2 3]]
  (println* (sum-c (Array.unsafe-raw &ints) (Array.length &ints))))

Again, you might want to wrap the bare C function in more Carp-esque interface.

(relative-include "sum.h")

(defmodule MyMod
  (hidden sum-c)
  (private sum-c)
  (register sum-c (Fn [(Ptr Int) Int] Int) "sum")

  (sig sum (Fn [(Ref (Array Int))] Int))
  (defn sum [ints] (sum-c (Array.unsafe-raw ints) (Array.length ints))))

(MyMod.sum &[1 2 3])

In cases where the consuming function takes ownership over the data, Array.raw can be used. It becomes the responsibility of the consuming function to call free on the pointer and any managed types it contains.

// printall.h
void println_all(char **arr, int len) {
  for (int i = 0; i < len; i++) {
    printf("%s\n", arr[i]);
    CARP_FREE(arr[i]);
  }
  CARP_FREE(arr);
}
(relative-include "printall.h")

(register println-all (Fn [(Ptr String) Int] ()) "println_all")

(let [lines [@"One" @"Two" @"Three"]
      len (Array.length &lines)]
  (println-all (Array.raw lines) len))

Embedding C code in Carp

When interfacing C libraries it is sometimes beneficial to wrap the libraries function with some custom C code. An entirely valid method is the write your code in a header file, include it from the Carp side and register it:

// print.h
// String is a carp core alias for char*
void print_that_takes_ownership(String str) {
  printf("%s", str);
  CARP_FREE(str);
}
(relative-include "print.h")

(register print (Fn [String] ()) "print_that_takes_ownership")

(print @"Print this!")

However you might prefer to keep your C code close to your Carp code, enter deftemplate...

deftemplate

Basic example

We can instead define the previous example like so:

(deftemplate print (Fn [String] ())
                   "void $NAME(String str)"
                   "$DECL {
                     printf(\"%s\", str);
                     CARP_FREE(str);
                   }")

(print @"Print this!")

Let's break down what's going on here: The first argument to deftemplate is the name we'll use to refer to the function. The second is a type signature and is identical to the one found in our previous register call. The third is our function declaration, it'll be injected at the top of the generated C file. The last argument represent the function definition.

Two more things to look at: $NAME is a variable that will be derived from the name you've given the function plus any module it's defined in, so no need to worry about name clashes with other print functions in other modules. $DECL will be replaced with the declaration passed as a third argument when the function is defined

So we've seen how deftemplate can be used to keep Carp and C code close to each other and help you write less code in general but it's real power lies somewhere else...

Generics

Let's say one would like to write a function that adds two numbers, it would be tedious to write a version for every type of number, let's see how deftemplate can help us with that.

(deftemplate add (Fn [a a] a)
                 "$a $NAME($a x, $a y)"
                 "$DECL {
                   return x + y;
                 }")

(add 1 2)
(add 20l 22l)
(add 2.0f 5.0f)

; Can't do that as they're different types
; (add 2.0f 22l)

Carp allows us to use generic type in type signatures, a in that example. You can use $ plus the generic name you used in your signature to refer to that type in your C code. Carp will then generate a separate function everytime the template is used with a different type.

Warning! You'll need to be careful when calling that function as you've lost all type safety the Carp compiler guarantees. You will have to hope the C compiler will catch it.

(deftemplate add (Fn [a a] a)
                 "$a $NAME($a x, $a y)"
                 "$DECL {
                   return x + y;
                 }")

(add @"A string" @" another string")

This thankfully result in this Clang error, but it's probably good not to rely on it.

out/main.c:9153:29: error: invalid operands to binary expression ('String' (aka 'char *') and 'String')
                   return x + y;
                          ~ ^ ~
1 error generated.

Unsafe.emit-c

While deftemplate is flexible and sufficient for most use cases, there are certain scenarios in which it won't accomplish what you need. For example, some C macros, such as c11's static_assert require a string literal argument. deftemplate can't accomplish this. In such cases, you can use Unsafe.emit-c to emit a literal string in the Carp compiler's C output. emit-c is perfect for scenarios like static_assert calls. Assuming static_assert is registered as static-assert, we can use emit-c in the following way to ensure it is passed a string literal in the compiler's emitted C code:

(register static-assert (Fn [a C] ()))

(static-assert 0 (Unsafe.emit-c "\"foo\""))

which will emit the corresponding C:

static_assert(0, "foo")

emit-C returns values of the C type, a special type that represents literal C code in Carp.

Unsafe.preproc

The Carp compiler emits C code in an order that ensures the dependencies of functions are available before functions are called. Sometimes, you may want to include C code before the Carp compiler's output. For instance, you might want to provide some preprocessor directives to a C compiler. The Unsafe.preproc function was designed with this use case in mind. You can use preproc to inject arbitrary C code prior to the Carp compiler's normal C output. Any code passed to preproc will be emitted after file includes but before any other emitted C code.

preproc takes a value of type C as an argument, so it must be used in combination with Unsafe.emit-c. The C code you pass to preproc isn't checked at all, so be careful!

If you do define C symbols using preproc, you'll still need to call register to reference them in Carp code. For example, the following snippet uses preproc to make a C macro and function available in the Carp compiler's output and then calls register to reference these symbols in the main function in the Carp source:

(Unsafe.preproc (Unsafe.emit-c "#define FOO 0"))
(Unsafe.preproc (Unsafe.emit-c "void foo() { printf(\"%d\\n\", 1); }"))

(register FOO Int)
(register foo (Fn [] ()))

(defn main []
  (do (foo)
      (IO.println &(fmt "%d" FOO))))

You can use this technique to add provisional definitions you need to reference in compiler output. If your helper functions, macros, or preprocessor directives are lengthy or complex, you may want to define them in a separate h file and relative-include it in your Carp source instead.

Registering Types

Carp supports a few different ways of registering types defined in C. You can register types using the register-type function. Calling register-type with only a symbol argument registers the C type with a name corresponding to the symbol. For example, the following code registers the C type A as the type A in Carp.

typedef int A;
(register-type A)

After this call to register-type, you can use the type A anywhere type names are valid in Carp code. For example, you can use it in function signatures:

(sig a-prn (Fn [A] String))

The prior type registration only registers the type name in Carp. In other words, the type is entirely "opaque" from the perspective of your Carp program. Carp knows the type exists, but it knows nothing about its implementation or how to construct values of the type--all of that is left up to your C code.

If you want to construct values of this type from Carp code, you have two options:

  1. You can define your own initializers for the type in C and register them in Carp.
  2. You can use register-type to generate initializers for the type in Carp.

If you define an initializer for the type in C, you can access it from Carp by using register:

typedef int A;

A initializer() {
  return 0;
}
(register-type A)
(register initializer (Fn [] A))
;; returns a value of type A
(initializer)

Alternatively, you can add a non-empty array of type members in your register-type call to have Carp generate initializers, getters and setters, and printing functions for the external type. The initializer Carp generates will only initialize the fields you specify. If you omit or misname a field, the generated initializer might cause errors.

(register-type B [])
:i B
=> B : Type
     init : (Fn [] B)
     prn  : (Fn [(Ref B q)] String)
     str  : (Fn [(Ref B q)] String)
}
(register-type C [x Int])
:i C
=> C : Type
   C : Module {
     init     : (Fn [Int] C)
     prn      : (Fn [(Ref C q) String])
     str      : (Fn [(Ref C q) String])
     set-x    : (Fn [C, Int] C)
     set-x!   : (Fn [(Ref C q), Int] ())
     update-x : (Fn [C, (Ref (Fn [Int] Int) q)] C)
     x        : (Fn [(Ref C q)] (Ref Int q))
}

The prn and str functions for the type will also automatically implement their corresponding interfaces.

Be mindful that Carp does not manage the memory associated with external types by default! Unlike types defined in Carp, Carp will not generate copy and delete functions for registered types. If you use generated initializers for a registered type for convenience, remember that you still need to manage the memory associated with values of the type manually. If you want Carp to manage the memory for a registered type, you can provide implementations of the copy and delete interfaces.

If needed, you can override the name Carp emits for a registered type by providing an additional string argument. This comes in handy when the type's name in C does not follow lisp or Carp naming conventions. For example, the type in C might begin with a lowercase letter, while Carp requires all types to begin with uppercase letters:

;; Emitted in C code as "A"
(register-type A)
;; Emitted in C code a "a_type"
(register-type A "a_type")
;; Emitted in C code as "b_type"
(register-type B "b_type" [x Int])

Callbacks

Some C APIs rely on callbacks, let's define a C function that accepts a callback and an argument and returns the result of calling that function as an example:

(deftemplate runner (Fn [(Ptr ()) (Ptr ())] a)
                    "$a $NAME(void* fnptr, void* args)"
                    "$DECL {
                       return (($a(*)(void*))fnptr)(args);
                    }")

; Using a lambda capturing variables from its environment
(let [x 20 y 22 fnfn (fn [] (+ @&x @&y))]
  (= (runner (Function.unsafe-ptr &fnfn) (Function.unsafe-env-ptr &fnfn))
     42))

; Using a static function
(defn double [x] (Int.* @x 2))

(let [x 42]
  (= (runner (Function.unsafe-ptr &double) (Unsafe.coerce &x))
     84))

In the first example we want to use a lambda capturing some variable, we can use Function.unsafe-ptr to get a void* to the function and in the case of lambdas capturing environment the first argument to that function is the environment so we have to use Function.unsafe-env-ptr to pass in that environment.

In the second example we want to use a static function so we can use Function.unsafe-ptr again, and the argument we pass in needs to be coerced from a Ref into a (Ptr ()).

Because everything gets turned into a void pointer all type safety is lost so it is the responsibility of the caller to ensure the operation is safe. It is also important to ensure the lifetime of the Ptr doesn't not exceed the lifetime of the function/env it represents.

Headerparse

headerparse is a Haskell script to aid in writing C bindings by parsing a C header and generating register and register-type for you. It resides in the ./headersparse folder in Carp source repo and can be used in the following way:

stack runhaskell ./headerparse/Main.hs -- ../path/to/c/header.h

The script accepts the following flags:

  • [-p|--prefixtoremove thePrefix] Removes a prefix from the C identifiers
  • [-f|--kebabcase] Converts identifiers to kebab-case
  • [-c|--emitcname] Always emit the C identifier name after the binding

Example

Invoking the script on this C header:

stack runhaskell ./headerparse/Main.hs -- -p "MyModule_" -f ../path/to/aheader.h
// aheader.h
bool MyModule_runThisFile(const char *file);

Will output the following:

(register run-this-file (λ [(Ptr CChar)] Bool) "MyModule_runThisFile")