LuaJIT 学习（4）—— FFI 语义

Given that the FFI library is designed to interface with C code and that declarations can be written in plain C syntax, it closely follows the C language semantics, wherever possible. Some minor concessions【轻微的让步】 are needed for smoother interoperation with Lua language semantics.

C Language Support

【忽略复数和向量类型】

The FFI library has a built-in C parser with a minimal memory footprint. It’s used by the ffi.* library functions to declare C types or external symbols.

Its only purpose is to parse C declarations, as found e.g. in C header files. Although it does evaluate constant expressions, it’s not a C compiler. The body of inline C function definitions is simply ignored.

Also, this is not a validating C parser. It expects and accepts correctly formed C declarations, but it may choose to ignore bad declarations or show rather generic error messages. If in doubt, please check the input against your favorite C compiler.

The C parser complies to the C99 language standard plus the following extensions:

• The '\e' escape in character and string literals.
• The C99/C++ boolean type, declared with the keywords bool or _Bool.
• Unnamed (‘transparent’) struct/union fields inside a struct/union.
• Incomplete enum declarations, handled like incomplete struct declarations.
• Unnamed enum fields inside a struct/union. This is similar to a scoped C++ enum, except that declared constants are visible in the global namespace, too.
• Scoped static const declarations inside a struct/union (from C++).
• Zero-length arrays ([0]), empty struct/union, variable-length arrays (VLA, [?]) and variable-length structs (VLS, with a trailing VLA).
• C++ reference types (int &x).
• Alternate GCC keywords with ‘__’, e.g. __const__.
• GCC __attribute__ with the following attributes: aligned, packed, mode, vector_size, cdecl, fastcall, stdcall, thiscall.
• The GCC __extension__ keyword and the GCC __alignof__ operator.
• GCC __asm__("symname") symbol name redirection for function declarations.
• MSVC keywords for fixed-length types: __int8, __int16, __int32 and __int64.
• MSVC __cdecl, __fastcall, __stdcall, __thiscall, __ptr32, __ptr64, __declspec(align(n)) and #pragma pack.
• All other GCC/MSVC-specific attributes are ignored.

The following C types are predefined by the C parser (like a typedef, except re-declarations will be ignored):

• Vararg handling: va_list, __builtin_va_list, __gnuc_va_list.
• From <stddef.h>: ptrdiff_t, size_t, wchar_t.
• From <stdint.h>: int8_t, int16_t, int32_t, int64_t, uint8_t, uint16_t, uint32_t, uint64_t, intptr_t, uintptr_t.
• From <unistd.h> (POSIX): ssize_t.

You’re encouraged to use these types in preference to compiler-specific extensions or target-dependent standard types. E.g. char differs in signedness and long differs in size, depending on the target architecture and platform ABI.

The following C features are not supported:

• A declaration must always have a type specifier; it doesn’t default to an int type.
• Old-style empty function declarations (K&R) are not allowed. All C functions must have a proper prototype declaration. A function declared without parameters (int foo();) is treated as a function taking zero arguments, like in C++.
• The long double C type is parsed correctly, but there’s no support for the related conversions, accesses or arithmetic operations.
• Wide character strings and character literals are not supported.
• See below for features that are currently not implemented.

C Type Conversion Rules

Conversions from C types to Lua objects

These conversion rules apply for read accesses to C types: indexing pointers, arrays or struct/union types; reading external variables or constant values; retrieving return values from C calls:

Input	Conversion	Output
int8_t, int16_t	→sign-ext int32_t → double	number
uint8_t, uint16_t	→zero-ext int32_t → double	number
int32_t, uint32_t	→ double	number
int64_t, uint64_t	boxed value	64 bit int cdata
double, float	→ double	number
bool	0 → false, otherwise true	boolean
enum	boxed value	enum cdata
Pointer	boxed value	pointer cdata
Array	boxed reference	reference cdata
struct/union	boxed reference	reference cdata

Bitfields are treated like their underlying type.

Reference types are dereferenced before a conversion can take place — the conversion is applied to the C type pointed to by the reference.

例子：访问结构体成员

local ffi = require("ffi")ffi.cdef [[typedef enum {RED = 1,GREEN = 2,BLUE = 3
} Colors;typedef struct {float x, y;
} point;typedef struct {int8_t a;uint16_t b;uint32_t c;uint64_t d;double e;float f;bool g;Colors h;int *i;char j[100];point k;
} t;
]]local function print_type(v)if type(v) ~= "cdata" thenprint(type(v))elseprint(v)end
endlocal t = ffi.new("t")
print_type(t)   -- cdata<struct 106>: 0x7fa50ecd6228
print_type(t.a) -- int8_t -> number
print_type(t.b) -- uint16_t -> number
print_type(t.c) -- uint32_t -> number
print_type(t.d) -- uint64_t -> 0ULL 【64 bit int cdata】
print_type(t.e) -- double -> number
print_type(t.f) -- float -> number
print_type(t.g) -- bool -> boolean
print_type(t.h) -- enum -> cdata<enum 97>: 0 【boxed value】
print_type(t.i) -- pointer -> cdata<int *>: NULL 【boxed value】
print_type(t.j) -- array -> cdata<char (&)[100]>: 0x7fa50ecd6258 【boxed reference】
print_type(t.k) -- struct -> cdata<struct 102 &>: 0x7fa50ecd62bc 【boxed reference】

会发生 C 类型到 Lua 类型的转换

Conversions from Lua objects to C types

These conversion rules apply for write accesses to C types: indexing pointers, arrays or struct/union types; initializing cdata objects; casts to C types; writing to external variables; passing arguments to C calls:

Input	Conversion	Output
number	→	double
boolean	false → 0, true → 1	bool
nil	NULL →	(void *)
lightuserdata	lightuserdata address →	(void *)
userdata	userdata payload →	(void *)
io.* file	get FILE * handle →	(void *)
string	match against enum constant	enum
string	copy string data + zero-byte	int8_t[], uint8_t[]
string	string data →	const char[]
function	create callback →	C function type
table	table initializer	Array
table	table initializer	struct/union
cdata	cdata payload →	C type

If the result type of this conversion doesn’t match the C type of the destination, the conversion rules between C types are applied.

Conversions between C types

These conversion rules are more or less the same as the standard C conversion rules. Some rules only apply to casts, or require pointer or type compatibility:

Input	Conversion	Output
Signed integer	→narrow or sign-extend	Integer
Unsigned integer	→narrow or zero-extend	Integer
Integer	→round	double, float
double, float	→trunc int32_t →narrow	(u)int8_t, (u)int16_t
double, float	→trunc	(u)int32_t, (u)int64_t
double, float	→round	float, double
Number	n == 0 → 0, otherwise 1	bool
bool	false → 0, true → 1	Number
struct/union	take base address (compat)	Pointer
Array	take base address (compat)	Pointer
Function	take function address	Function pointer
Number	convert via uintptr_t (cast)	Pointer
Pointer	convert address (compat/cast)	Pointer
Pointer	convert address (cast)	Integer
Array	convert base address (cast)	Integer
Array	copy (compat)	Array
struct/union	copy (identical type)	struct/union

Bitfields or enum types are treated like their underlying type.

Conversions not listed above will raise an error. E.g. it’s not possible to convert a pointer to a complex number or vice versa.

例子：修改结构体成员

local ffi = require("ffi")ffi.cdef [[typedef enum {RED = 1,GREEN = 2,BLUE = 3
} Colors;typedef struct {float x, y;
} point;typedef struct {int8_t a;uint16_t b;uint32_t c;uint64_t d;double e;float f;bool g;Colors h;int *i;char j[100];point k;
} t;
]]local t = ffi.new("t")
t.a = 1                            -- number -> double -> int8_t
t.b = 2                            -- number -> double -> uint16_t
t.c = 3                            -- number -> double -> uint16_t
t.d = 4                            -- number -> double -> uint64_t
t.e = 5                            -- number -> double
t.f = 6                            -- number -> double -> float
t.g = 7                            -- number -> double -> 1 【n == 0 → 0, otherwise 1】
t.h = "BLUE"                       -- string -> enmu 【match against `enum` constant】
local i = ffi.new("int [1]", 1)
t.i = i                            -- cdata -> int [1]【数组】 -> pointer 【take base address (compat)】
t.j = ffi.new("char [100]", "abc") -- cdata -> char [100]
t.k = ffi.new("point", 1, 2)       -- cdata -> point

会发生 Lua 类型到 C 类型的转换。如果按照转换规则无法转换，则运行时会报错。

Conversions for vararg C function arguments

The following default conversion rules apply when passing Lua objects to the variable argument part of vararg C functions:

Input	Conversion	Output
number	→	double
boolean	false → 0, true → 1	bool
nil	NULL →	(void *)
userdata	userdata payload →	(void *)
lightuserdata	lightuserdata address →	(void *)
string	string data →	const char *
float cdata	→	double
Array cdata	take base address	Element pointer
struct/union cdata	take base address	struct/union pointer
Function cdata	take function address	Function pointer
Any other cdata	no conversion	C type

To pass a Lua object, other than a cdata object, as a specific type, you need to override the conversion rules: create a temporary cdata object with a constructor or a cast and initialize it with the value to pass:

Assuming x is a Lua number, here’s how to pass it as an integer to a vararg function:

ffi.cdef[[
int printf(const char *fmt, ...);
]]
ffi.C.printf("integer value: %d\n", ffi.new("int", x))

If you don’t do this, the default Lua number → double conversion rule applies. A vararg C function expecting an integer will see a garbled or uninitialized value.

Note: this is the only place where creating a boxed scalar number type is actually useful. Never use ffi.new("int"), ffi.new("float") etc. anywhere else!【只有在作为可变参数 C 函数的参数时才用得到这种写法！】

Ditto for ffi.cast(). Explicitly boxing scalars does not improve performance or force int or float arithmetic! It just adds costly boxing, unboxing and conversions steps. And it may lead to surprise results, because cdata arithmetic on scalar numbers is always performed on 64 bit integers.

Initializers

Creating a cdata object with ffi.new() or the equivalent constructor syntax always initializes its contents, too. Different rules apply, depending on the number of optional initializers and the C types involved:

• If no initializers are given, the object is filled with zero bytes.
• Scalar types (numbers and pointers) accept a single initializer. The Lua object is converted to the scalar C type.
• Aggregate types (arrays and structs) accept either a single cdata initializer of the same type (copy constructor), a single table initializer, or a flat list of initializers.

例子：结构体初始化

local ffi = require("ffi")ffi.cdef [[typedef struct {int a,b,c;
} t;]]local t1 = ffi.new("t", { 1 })  -- 1       0       0local t2 = ffi.new("t", { 1, 2 }) -- 1       2       0local t3 = ffi.new("t", 1)      -- 1       0       0local t4 = ffi.new("t", 1, 2)   -- 1       2       0local t5 = ffi.new("t", t4)     -- 1       2       0-- local t6 = ffi.new("t", 1, 2, 3, 4) -- 报错local t7 = ffi.new("t", {1, 2, 3, 4}) -- 不会报错，忽略多余的元素

• The elements of an array are initialized, starting at index zero. If a single initializer is given for an array, it’s repeated for all remaining elements. This doesn’t happen if two or more initializers are given: all remaining uninitialized elements are filled with zero bytes.

例子：固定大小数组初始化

local a1 = ffi.new("int [3]", 1) -- 1 1 1
print_array(a1, 3)
local a2 = ffi.new("int [3]", 1, 2) -- 1 2 0
print_array(a2, 3)local a3 = ffi.new("int [3]", {1}) -- 1 1 1
print_array(a3, 3)local a4 = ffi.new("int [3]", {1, 2}) -- 1 2 0
print_array(a4, 3)

• Byte arrays may also be initialized with a Lua string. This copies the whole string plus a terminating zero-byte. The copy stops early only if the array has a known, fixed size.

例子：变长数组初始化

local a1 = ffi.new("int [?]", 3, 1) -- 1 1 1local a2 = ffi.new("int [?]", 3, 1, 2) -- 1 2 0local a3 = ffi.new("int [?]", 3, {1}) -- 1 ? ?local a4 = ffi.new("int [?]", 3, {1, 2}) -- 1 2 ?

例子：字节数组初始化

local s = ffi.new("char [100]", "hello world!")
print(ffi.string(s)) -- hello world!

• The fields of a struct are initialized in the order of their declaration. Uninitialized fields are filled with zero bytes.
• Only the first field of a union can be initialized with a flat initializer.
• Elements or fields which are aggregates themselves are initialized with a single initializer, but this may be a table initializer or a compatible aggregate.

例子：嵌套结构体初始化

local ffi = require("ffi")ffi.cdef [[typedef struct {float x, y;
} point;typedef struct {int a,b,c;point d;
} t;]]local t1 = ffi.new("t", 1, 2, 3, {1, 2})
local t2 = ffi.new("t", 1, 2, 3, ffi.new("point", 1, 2))
local t3 = ffi.new("t", {a = 1, d = {1, 2}})

• Excess initializers cause an error.

Table Initializers

The following rules apply if a Lua table is used to initialize an Array or a struct/union:

• If the table index [0] is non-nil, then the table is assumed to be zero-based. Otherwise it’s assumed to be one-based.
• Array elements, starting at index zero, are initialized one-by-one with the consecutive table elements, starting at either index [0] or [1]. This process stops at the first nil table element.
• If exactly one array element was initialized, it’s repeated for all the remaining elements. Otherwise all remaining uninitialized elements are filled with zero bytes.
• The above logic only applies to arrays with a known fixed size. A VLA is only initialized with the element(s) given in the table. Depending on the use case, you may need to explicitly add a NULL or 0 terminator to a VLA.

参考上面的数组初始化例子。

• A struct/union can be initialized in the order of the declaration of its fields. Each field is initialized with consecutive table elements, starting at either index [0] or [1]. This process stops at the first nil table element.
• Otherwise, if neither index [0] nor [1] is present, a struct/union is initialized by looking up each field name (as a string key) in the table. Each non-nil value is used to initialize the corresponding field.

参考上面的嵌套结构体初始化例子。

• Uninitialized fields of a struct are filled with zero bytes, except for the trailing VLA of a VLS.
• Initialization of a union stops after one field has been initialized. If no field has been initialized, the union is filled with zero bytes.
• Elements or fields which are aggregates themselves are initialized with a single initializer, but this may be a nested table initializer (or a compatible aggregate).

参考上面的嵌套结构体初始化例子。

• Excess initializers for an array cause an error. Excess initializers for a struct/union are ignored. Unrelated table entries are ignored, too.

例子：使用 table 初始化数组或结构体

local ffi = require("ffi")ffi.cdef[[
struct foo { int a, b; };
union bar { int i; double d; };
struct nested { int x; struct foo y; };
]]ffi.new("int[3]", {})            --> 0, 0, 0
ffi.new("int[3]", {1})           --> 1, 1, 1
ffi.new("int[3]", {1,2})         --> 1, 2, 0
ffi.new("int[3]", {1,2,3})       --> 1, 2, 3
ffi.new("int[3]", {[0]=1})       --> 1, 1, 1
ffi.new("int[3]", {[0]=1,2})     --> 1, 2, 0
ffi.new("int[3]", {[0]=1,2,3})   --> 1, 2, 3
ffi.new("int[3]", {[0]=1,2,3,4}) --> error: too many initializersffi.new("struct foo", {})            --> a = 0, b = 0
ffi.new("struct foo", {1})           --> a = 1, b = 0
ffi.new("struct foo", {1,2})         --> a = 1, b = 2
ffi.new("struct foo", {[0]=1,2})     --> a = 1, b = 2
ffi.new("struct foo", {b=2})         --> a = 0, b = 2
ffi.new("struct foo", {a=1,b=2,c=3}) --> a = 1, b = 2  'c' is ignoredffi.new("union bar", {})        --> i = 0, d = 0.0
ffi.new("union bar", {1})       --> i = 1, d = ?
ffi.new("union bar", {[0]=1,2}) --> i = 1, d = ?    '2' is ignored
ffi.new("union bar", {d=2})     --> i = ?, d = 2.0ffi.new("struct nested", {1,{2,3}})     --> x = 1, y.a = 2, y.b = 3
ffi.new("struct nested", {x=1,y={2,3}}) --> x = 1, y.a = 2, y.b = 3

Operations on cdata Objects

All standard Lua operators can be applied to cdata objects or a mix of a cdata object and another Lua object. The following list shows the predefined operations.

Reference types are dereferenced before performing each of the operations below — the operation is applied to the C type pointed to by the reference.

The predefined operations are always tried first before deferring to a metamethod or index table (if any) for the corresponding ctype (except for __new). An error is raised if the metamethod lookup or index table lookup fails.

Indexing a cdata object

• Indexing a pointer/array: a cdata pointer/array can be indexed by a cdata number or a Lua number. The element address is computed as the base address plus the number value multiplied by the element size in bytes. A read access loads the element value and converts it to a Lua object. A write access converts a Lua object to the element type and stores the converted value to the element. An error is raised if the element size is undefined or a write access to a constant element is attempted.
• Dereferencing a struct/union field: a cdata struct/union or a pointer to a struct/union can be dereferenced by a string key, giving the field name. The field address is computed as the base address plus the relative offset of the field. A read access loads the field value and converts it to a Lua object. A write access converts a Lua object to the field type and stores the converted value to the field. An error is raised if a write access to a constant struct/union or a constant field is attempted. Scoped enum constants or static constants are treated like a constant field.

A ctype object can be indexed with a string key, too. The only predefined operation is reading scoped constants of struct/union types. All other accesses defer to the corresponding metamethods or index tables (if any).

Note: since there’s (deliberately) no address-of operator, a cdata object holding a value type is effectively immutable after initialization. The JIT compiler benefits from this fact when applying certain optimizations.

As a consequence, the elements of complex numbers and vectors are immutable. But the elements of an aggregate holding these types may be modified, of course. I.e. you cannot assign to foo.c.im, but you can assign a (newly created) complex number to foo.c.

The JIT compiler implements strict aliasing rules: accesses to different types do not alias, except for differences in signedness (this applies even to char pointers, unlike C99). Type punning through unions is explicitly detected and allowed.

Calling a cdata object

• Constructor: a ctype object can be called and used as a constructor. This is equivalent to ffi.new(ct, ...), unless a __new metamethod is defined. The __new metamethod is called with the ctype object plus any other arguments passed to the constructor. Note that you have to use ffi.new inside the metamethod, since calling ct(...) would cause infinite recursion.
• C function call: a cdata function or cdata function pointer can be called. The passed arguments are converted to the C types of the parameters given by the function declaration. Arguments passed to the variable argument part of vararg C function use special conversion rules. This C function is called and the return value (if any) is converted to a Lua object.
  On Windows/x86 systems, __stdcall functions are automatically detected, and a function declared as __cdecl (the default) is silently fixed up after the first call.

Arithmetic on cdata objects

• Pointer arithmetic: a cdata pointer/array and a cdata number or a Lua number can be added or subtracted. The number must be on the right-hand side for a subtraction. The result is a pointer of the same type with an address plus or minus the number value multiplied by the element size in bytes. An error is raised if the element size is undefined.
• Pointer difference: two compatible cdata pointers/arrays can be subtracted. The result is the difference between their addresses, divided by the element size in bytes. An error is raised if the element size is undefined or zero.
• 64 bit integer arithmetic: the standard arithmetic operators (+ - * / % ^ and unary minus) can be applied to two cdata numbers, or a cdata number and a Lua number. If one of them is an uint64_t, the other side is converted to an uint64_t and an unsigned arithmetic operation is performed. Otherwise, both sides are converted to an int64_t and a signed arithmetic operation is performed. The result is a boxed 64 bit cdata object.
  If one of the operands is an enum and the other operand is a string, the string is converted to the value of a matching enum constant before the above conversion.
  These rules ensure that 64 bit integers are “sticky”. Any expression involving at least one 64 bit integer operand results in another one. The undefined cases for the division, modulo and power operators return 2LL ^ 63 or 2ULL ^ 63.
  You’ll have to explicitly convert a 64 bit integer to a Lua number (e.g. for regular floating-point calculations) with tonumber(). But note this may incur a precision loss.
• 64 bit bitwise operations: the rules for 64 bit arithmetic operators apply analogously.
  Unlike the other bit.* operations, bit.tobit() converts a cdata number via int64_t to int32_t and returns a Lua number.
  For bit.band(), bit.bor() and bit.bxor(), the conversion to int64_t or uint64_t applies to all arguments, if any argument is a cdata number.
  For all other operations, only the first argument is used to determine the output type. This implies that a cdata number as a shift count for shifts and rotates is accepted, but that alone does not cause a cdata number output.

Comparisons of cdata objects

• Pointer comparison: two compatible cdata pointers/arrays can be compared. The result is the same as an unsigned comparison of their addresses. nil is treated like a NULL pointer, which is compatible with any other pointer type.
• 64 bit integer comparison: two cdata numbers, or a cdata number and a Lua number can be compared with each other. If one of them is an uint64_t, the other side is converted to an uint64_t and an unsigned comparison is performed. Otherwise, both sides are converted to an int64_t and a signed comparison is performed.
  If one of the operands is an enum and the other operand is a string, the string is converted to the value of a matching enum constant before the above conversion.
• Comparisons for equality/inequality never raise an error. Even incompatible pointers can be compared for equality by address. Any other incompatible comparison (also with non-cdata objects) treats the two sides as unequal.

cdata objects as table keys

Lua tables may be indexed by cdata objects, but this doesn’t provide any useful semantics — cdata objects are unsuitable as table keys! 【cdata 对象不适合作为 table 的 keys】

A cdata object is treated like any other garbage-collected object and is hashed and compared by its address for table indexing. Since there’s no interning for cdata value types, the same value may be boxed in different cdata objects with different addresses. Thus, t[1LL+1LL] and t[2LL] usually do not point to the same hash slot, and they certainly do not point to the same hash slot as t[2].

It would seriously drive up implementation complexity and slow down the common case, if one were to add extra handling for by-value hashing and comparisons to Lua tables. Given the ubiquity of their use inside the VM, this is not acceptable.

There are three viable alternatives, if you really need to use cdata objects as keys:

• If you can get by with the precision of Lua numbers (52 bits), then use tonumber() on a cdata number or combine multiple fields of a cdata aggregate to a Lua number. Then use the resulting Lua number as a key when indexing tables.
  One obvious benefit: t[tonumber(2LL)] does point to the same slot as t[2].
• Otherwise, use either tostring() on 64 bit integers or complex numbers or combine multiple fields of a cdata aggregate to a Lua string (e.g. with ffi.string()). Then use the resulting Lua string as a key when indexing tables.
• Create your own specialized hash table implementation using the C types provided by the FFI library, just like you would in C code. Ultimately, this may give much better performance than the other alternatives or what a generic by-value hash table could possibly provide.

Garbage Collection of cdata Objects

All explicitly (ffi.new(), ffi.cast() etc.) or implicitly (accessors) created cdata objects are garbage collected. You need to ensure to retain valid references to cdata objects somewhere on a Lua stack, an upvalue or in a Lua table while they are still in use. Once the last reference to a cdata object is gone, the garbage collector will automatically free the memory used by it (at the end of the next GC cycle).

Please note, that pointers themselves are cdata objects, however they are not followed by the garbage collector. So e.g. if you assign a cdata array to a pointer, you must keep the cdata object holding the array alive as long as the pointer is still in use:

ffi.cdef[[
typedef struct { int *a; } foo_t;
]]local s = ffi.new("foo_t", ffi.new("int[10]")) -- WRONG!local a = ffi.new("int[10]") -- OK
local s = ffi.new("foo_t", a)
-- Now do something with 's', but keep 'a' alive until you're done.

Similar rules apply for Lua strings which are implicitly converted to "const char *": the string object itself must be referenced somewhere or it’ll be garbage collected eventually. The pointer will then point to stale data, which may have already been overwritten. Note that string literals are automatically kept alive as long as the function containing it (actually its prototype) is not garbage collected.

Objects which are passed as an argument to an external C function are kept alive until the call returns. So it’s generally safe to create temporary cdata objects in argument lists. This is a common idiom for passing specific C types to vararg functions.

Memory areas returned by C functions (e.g. from malloc()) must be manually managed, of course (or use ffi.gc()). Pointers to cdata objects are indistinguishable from pointers returned by C functions (which is one of the reasons why the GC cannot follow them).

Callbacks

The LuaJIT FFI automatically generates special callback functions whenever a Lua function is converted to a C function pointer. This associates the generated callback function pointer with the C type of the function pointer and the Lua function object (closure).

This can happen implicitly due to the usual conversions, e.g. when passing a Lua function to a function pointer argument. Or, you can use ffi.cast() to explicitly cast a Lua function to a C function pointer.

Currently, only certain C function types can be used as callback functions. Neither C vararg functions nor functions with pass-by-value aggregate argument or result types are supported. There are no restrictions on the kind of Lua functions that can be called from the callback — no checks for the proper number of arguments are made. The return value of the Lua function will be converted to the result type, and an error will be thrown for invalid conversions.

It’s allowed to throw errors across a callback invocation, but it’s not advisable in general. Do this only if you know the C function, that called the callback, copes with the forced stack unwinding and doesn’t leak resources.

One thing that’s not allowed, is to let an FFI call into a C function get JIT-compiled, which in turn calls a callback, calling into Lua again. Usually this attempt is caught by the interpreter first and the C function is blacklisted for compilation.

However, this heuristic may fail under specific circumstances: e.g. a message polling function might not run Lua callbacks right away and the call gets JIT-compiled. If it later happens to call back into Lua (e.g. a rarely invoked error callback), you’ll get a VM PANIC with the message "bad callback". Then you’ll need to manually turn off JIT-compilation with jit.off() for the surrounding Lua function that invokes such a message polling function (or similar).

Callback resource handling

Callbacks take up resources — you can only have a limited number of them at the same time (500 - 1000, depending on the architecture). The associated Lua functions are anchored to prevent garbage collection, too.

Callbacks due to implicit conversions are permanent! There is no way to guess their lifetime, since the C side might store the function pointer for later use (typical for GUI toolkits). The associated resources cannot be reclaimed until termination:

ffi.cdef[[
typedef int (__stdcall *WNDENUMPROC)(void *hwnd, intptr_t l);
int EnumWindows(WNDENUMPROC func, intptr_t l);
]]-- Implicit conversion to a callback via function pointer argument.
local count = 0
ffi.C.EnumWindows(function(hwnd, l)count = count + 1return true
end, 0)
-- The callback is permanent and its resources cannot be reclaimed!
-- Ok, so this may not be a problem, if you do this only once.

Note: this example shows that you must properly declare __stdcall callbacks on Windows/x86 systems. The calling convention cannot be automatically detected, unlike for __stdcall calls to Windows functions.

For some use cases, it’s necessary to free up the resources or to dynamically redirect callbacks. Use an explicit cast to a C function pointer and keep the resulting cdata object. Then use the cb:free() or cb:set() methods on the cdata object:

-- Explicitly convert to a callback via cast.
local count = 0
local cb = ffi.cast("WNDENUMPROC", function(hwnd, l)count = count + 1return true
end)-- Pass it to a C function.
ffi.C.EnumWindows(cb, 0)
-- EnumWindows doesn't need the callback after it returns, so free it.cb:free()
-- The callback function pointer is no longer valid and its resources
-- will be reclaimed. The created Lua closure will be garbage collected.

Callback performance

Callbacks are slow! First, the C to Lua transition itself has an unavoidable cost, similar to a lua_call() or lua_pcall(). Argument and result marshalling add to that cost. And finally, neither the C compiler nor LuaJIT can inline or optimize across the language barrier and hoist repeated computations out of a callback function.

Do not use callbacks for performance-sensitive work: e.g. consider a numerical integration routine which takes a user-defined function to integrate over. It’s a bad idea to call a user-defined Lua function from C code millions of times. The callback overhead will be absolutely detrimental for performance.

It’s considerably faster to write the numerical integration routine itself in Lua — the JIT compiler will be able to inline the user-defined function and optimize it together with its calling context, with very competitive performance.

As a general guideline: use callbacks only when you must, because of existing C APIs. E.g. callback performance is irrelevant for a GUI application, which waits for user input most of the time, anyway.

For new designs avoid push-style APIs: a C function repeatedly calling a callback for each result. Instead, **use pull-style APIs: call a C function repeatedly to get a new result. Calls from Lua to C via the FFI are much faster than the other way round. Most well-designed libraries already use pull-style APIs (read/write, get/put) **.

C Library Namespaces

A C library namespace is a special kind of object which allows access to the symbols contained in shared libraries or the default symbol namespace. The default ffi.C namespace is automatically created when the FFI library is loaded. C library namespaces for specific shared libraries may be created with the ffi.load() API function.

Indexing a C library namespace object with a symbol name (a Lua string) automatically binds it to the library. First, the symbol type is resolved — it must have been declared with ffi.cdef. Then the symbol address is resolved by searching for the symbol name in the associated shared libraries or the default symbol namespace. Finally, the resulting binding between the symbol name, the symbol type and its address is cached. Missing symbol declarations or nonexistent symbol names cause an error.

This is what happens on a read access for the different kinds of symbols:

• External functions: a cdata object with the type of the function and its address is returned.
• External variables: the symbol address is dereferenced and the loaded value is converted to a Lua object and returned.
• Constant values (static const or enum constants): the constant is converted to a Lua object and returned.

This is what happens on a write access:

• External variables: the value to be written is converted to the C type of the variable and then stored at the symbol address.
• Writing to constant variables or to any other symbol type causes an error, like any other attempted write to a constant location.

C library namespaces themselves are garbage collected objects. If the last reference to the namespace object is gone, the garbage collector will eventually release the shared library reference and remove all memory associated with the namespace. Since this may trigger the removal of the shared library from the memory of the running process, it’s generally not safe to use function cdata objects obtained from a library if the namespace object may be unreferenced.

Performance notice: the JIT compiler specializes to the identity of namespace objects and to the strings used to index it. This effectively turns function cdata objects into constants. It’s not useful and actually counter-productive to explicitly cache these function objects, e.g. local strlen = ffi.C.strlen. OTOH, it is useful to cache the namespace itself, e.g. local C = ffi.C.

No Hand-holding!

The FFI library has been designed as a low-level library. The goal is to interface with C code and C data types with a minimum of overhead. This means you can do anything you can do from C: access all memory, overwrite anything in memory, call machine code at any memory address and so on.

The FFI library provides no memory safety, unlike regular Lua code. It will happily allow you to dereference a NULL pointer, to access arrays out of bounds or to misdeclare C functions. If you make a mistake, your application might crash, just like equivalent C code would.

This behavior is inevitable, since the goal is to provide full interoperability with C code. Adding extra safety measures, like bounds checks, would be futile. There’s no way to detect misdeclarations of C functions, since shared libraries only provide symbol names, but no type information. Likewise, there’s no way to infer the valid range of indexes for a returned pointer.

Again: the FFI library is a low-level library. This implies it needs to be used with care, but it’s flexibility and performance often outweigh this concern. If you’re a C or C++ developer, it’ll be easy to apply your existing knowledge. OTOH, writing code for the FFI library is not for the faint of heart and probably shouldn’t be the first exercise for someone with little experience in Lua, C or C++.

As a corollary of the above, the FFI library is not safe for use by untrusted Lua code. If you’re sandboxing untrusted Lua code, you definitely don’t want to give this code access to the FFI library or to any cdata object (except 64 bit integers or complex numbers). Any properly engineered Lua sandbox needs to provide safety wrappers for many of the standard Lua library functions — similar wrappers need to be written for high-level operations on FFI data types, too.

Current Status

The initial release of the FFI library has some limitations and is missing some features. Most of these will be fixed in future releases.

C language support is currently incomplete:

• C declarations are not passed through a C pre-processor, yet.
• The C parser is able to evaluate most constant expressions commonly found in C header files. However, it doesn’t handle the full range of C expression semantics and may fail for some obscure constructs.
• static const declarations only work for integer types up to 32 bits. Neither declaring string constants nor floating-point constants is supported.
• Packed struct bitfields that cross container boundaries are not implemented.
• Native vector types may be defined with the GCC mode or vector_size attribute. But no operations other than loading, storing and initializing them are supported, yet.
• The volatile type qualifier is currently ignored by compiled code.
• ffi.cdef silently ignores most re-declarations. Note: avoid re-declarations which do not conform to C99. The implementation will eventually be changed to perform strict checks.

The JIT compiler already handles a large subset of all FFI operations. It automatically falls back to the interpreter for unimplemented operations (you can check for this with the -jv command line option). The following operations are currently not compiled and may exhibit suboptimal performance, especially when used in inner loops:

• Vector operations.
• Table initializers.
• Initialization of nested struct/union types.
• Non-default initialization of VLA/VLS or large C types (> 128 bytes or > 16 array elements).
• Bitfield initializations.
• Pointer differences for element sizes that are not a power of two.
• Calls to C functions with aggregates passed or returned by value.
• Calls to ctype metamethods which are not plain functions.
• ctype __newindex tables and non-string lookups in ctype __index tables.
• tostring() for cdata types.
• Calls to ffi.cdef(), ffi.load() and ffi.metatype().

Other missing features:

• Arithmetic for complex numbers.
• Passing structs by value to vararg C functions.
• C++ exception interoperability does not extend to C functions called via the FFI, if the call is compiled.