dyncall(3)
NAME
dyncall – encapsulation of architecture-, OS- and compiler-specific
function call semantics
SYNOPSIS
#include <dyncall.h>
DCCallVM *
dcNewCallVM(DCsize size);
void
dcFree(DCCallVM * vm);
void
dcMode(DCCallVM * vm, DCint mode);
void
dcReset(DCCallVM * vm);
void
dcArgBool(DCCallVM * vm, DCbool arg);
void
dcArgChar(DCCallVM * vm, DCchar arg);
void
dcArgShort(DCCallVM * vm, DCshort arg);
void
dcArgInt(DCCallVM * vm, DCint arg);
void
dcArgLong(DCCallVM * vm, DClong arg);
void
dcArgLongLong(DCCallVM * vm, DClonglong arg);
void
dcArgFloat(DCCallVM * vm, DCfloat arg);
void
dcArgDouble(DCCallVM * vm, DCdouble arg);
void
dcArgPointer(DCCallVM * vm, DCpointer arg);
void
dcArgAggr(DCCallVM * vm, const DCaggr * ag, const void * value);
DCvoid
dcCallVoid(DCCallVM * vm, DCpointer funcptr);
DCbool
dcCallBool(DCCallVM * vm, DCpointer funcptr);
DCchar
dcCallChar(DCCallVM * vm, DCpointer funcptr);
DCshort
dcCallShort(DCCallVM * vm, DCpointer funcptr);
DCint
dcCallInt(DCCallVM * vm, DCpointer funcptr);
DClong
dcCallLong(DCCallVM * vm, DCpointer funcptr);
DClonglong
dcCallLongLong(DCCallVM * vm, DCpointer funcptr);
DCfloat
dcCallFloat(DCCallVM * vm, DCpointer funcptr);
DCdouble
dcCallDouble(DCCallVM * vm, DCpointer funcptr);
DCpointer
dcCallPointer(DCCallVM * vm, DCpointer funcptr);
DCpointer
dcCallAggr(DCCallVM * vm, DCpointer funcptr, const DCaggr * ag,
DCpointer ret);
void
dcBeginCallAggr(DCCallVM * vm, const DCaggr * ag);
void
dcArgF(DCCallVM * vm, const DCsigchar * signature, ...);
void
dcVArgF(DCCallVM * vm, const DCsigchar * signature, va_list args);
void
dcCallF(DCCallVM * vm, DCValue * result, DCpointer funcptr,
const DCsigchar * signature, ...);
void
dcVCallF(DCCallVM * vm, DCValue * result, DCpointer funcptr,
const DCsigchar * signature, va_list args);
DCaggr*
dcNewAggr(DCsize maxFieldCount, DCsize size);
void
dcAggrField(DCaggr* ag, DCsigchar type, DCint offset, DCsize array_len,
...);
void
dcCloseAggr(DCaggr* ag);
void
dcFreeAggr(DCaggr* ag);
DESCRIPTION
The dyncall library encapsulates architecture-, OS- and compiler-specific
function call semantics in a virtual "bind argument parameters from left
to right and then call" interface allowing programmers to call C
functions in a completely dynamic manner.
In other words, instead of calling a function directly, the dyncall
library provides a mechanism to push the function parameters manually and
to issue the call afterwards.
Since the idea behind this concept is similar to call dispatching
mechanisms of virtual machines, the object that can be dynamically loaded
with arguments, and then used to actually invoke the call, is called
CallVM. It is possible to change the calling convention used by the
CallVM at run-time. Due to the fact that nearly every platform comes with
one or more distinct calling conventions, the dyncall library project
intends to be a portable and open-source approach to the variety of
compiler/toolchain/platform-specific binary interfaces subtleties, and so
on...
dcNewCallVM() creates a new CallVM object, where size specifies the max
size of the internal stack that will be allocated and used to bind the
arguments to. Use dcFree() to destroy the CallVM object.
dcMode() sets the calling convention to use. See dyncall.h for a list of
available modes. Note that some mode/platform combinations don't make any
sense (e.g. using a PowerPC calling convention on a MIPS platform) and
are silently ignored.
dcReset() resets the internal stack of arguments and prepares it for a
new call. This function should be called after setting the initial/main
call mode (using dcMode()), but prior to binding arguments to the CallVM
(sometimes dcMode() calls are needed after pushing some args, e.g.
DC_SIGCHAR_CC_ELLIPSIS_VARARGS, which is used prior to binding varargs of
variadic functions). Use it also when reusing a CallVM, as arguments
don't get flushed automatically after a function call invocation. Note:
you should also call this function after initial creation of the a CallVM
object, as dcNewCallVM doesn't do this, implicitly.
dcArgBool(), dcArgChar(), dcArgShort(), dcArgInt(), dcArgLong(),
dcArgLongLong(), dcArgFloat(), dcArgDouble(), dcArgPointer() and
dcArgAggr() are used to bind arguments of the named types to the CallVM
object. Arguments should be bound in left to right order regarding the C
function prototype.
dcCallVoid(), dcCallBool(), dcCallChar(), dcCallShort(), dcCallInt(),
dcCallLong(), dcCallLongLong(), dcCallFloat(), dcCallDouble(),
dcCallPointer() and dcCallAggr() call the function with the previously
bound arguments and return the named type, where funcptr is a pointer to
the function to call. After the invocation of the function call, the
argument values are still bound to the CallVM and a second call using the
same arguments can be issued. Call dcReset() (as described above) to
clear the internal argument stack.
The interfaces for passing and/or returning aggregates (struct, union) by
value need to be explained as they are a bit more complex. Every such
argument or return type needs some extra info describing its layout via a
DCaggr structure (except for non-trivial C++ aggregates, see AGGREGATE
DESCRIPTION for more information, below). Passing such arguments is then
done by using dcArgAggr(), where ag is a pointer to the description and
value is a pointer to the aggregate in question. Calling a function that
returns an aggregate by value is done via two functions,
dcBeginCallAggr(), which handles special cases to facilitate the
implementation and must be called before pushing any arguments, and
finally dcCallAggr() where ag is a pointer to the description (for both
calls) and ret points to memory large enough to hold the to be returned
aggregate. dcCallAggr() returns a pointer to ret.
NOTE: C++ non-trivial aggregates (check with the std::is_trivial type
trait) need some special handling. First of all, no aggregate description
is needed and NULL must be passed wherever a DCaggr* argument is needed.
Also, as dyncall is oblivious to how to do any custom/non-trivial
construction or copy, and thus cannot do the copy of the aggregate,
passed by-value, itself, the user has to provide such copies, manually,
where needed (e.g. when passing such an aggregate as an argument by-
value, using dcArgAggr(), in order to preserver the call's by-value
semantics).
dcArgF(), dcVArgF(), dcCallF() and dcVCallF() can be used to bind
arguments in a printf-style call, using a signature string encoding the
argument and return types. The former 2 only bind the arguments to the vm
object (and ignore return types specified in the signature), whereas the
latter two issue a call to the given function pointer, afterwards. The
return value will be stored in result. The signature string also
features calling convention mode selection. For information about the
signature format, refer to dyncall_signature.h or the dyncall manual.
For passing aggregates using dc*F() functions, pass two varargs for each
aggregate, first a pointer to DCaggr, then a pointer to the aggregate in
question. For returning aggregates using those functions, pass two final
extra arguments, first a pointer to DCaggr describing the return value,
then a pointer to memory large enough to hold it. An explicit call do
dcBeginCallAggr() is not needed in those cases, and a pointer to the to
be returned aggregate is returned via result.
AGGREGATE DESCRIPTION
In order to describe an aggregate (except for C++ non-trivial aggregates,
as mentioned above), create a DCaggr object using dcNewAggr(), where
maxFieldCount is greater or equal to the number of fields the aggregate
has (a nested aggregate or an array is counted as one field), and size is
the size of the aggregate (e.g. as determined by sizeof()).
dcFreeAggr() destroys the DCaggr object.
dcAggrField() is used to describe the aggregate, field-by-field (in
order), with type being a DC_SIGCHAR_* (see dyncall_signature.h), offset
being the offset of the field from the beginning of the aggregate (use
C's offsetof(3)), and array_len being the number of array elements, iff
the field is an array, otherwise use 1. For nested aggregates (when using
DC_SIGCHAR_AGGREGATE as type), one needs to pass the pointer to the
nested aggregate's DCaggr object as last argument (in ...).
Call dcCloseAggr() after having described all fields of an aggregate.
Note that c99 flexible array members do not count as a field, and must be
omitted, as passing aggregates with a flexible array member by value in C
would also omit it.
EXAMPLES
Note: none of the examples below perform any error checking for
simplicity of the example.
Let's start with a simple example, making a call to the function sqrt(3).
Using the dyncall library, this function would be called as follows:
double r;
DCCallVM* vm = dcNewCallVM(4096);
dcMode(vm, DC_CALL_C_DEFAULT);
dcReset(vm);
/* call: double sqrt(double x); */
dcArgDouble(vm, 4.2373);
r = dcCallDouble(vm, (DCpointer)&sqrt);
dcFree(vm);
Note that the DCCallVM object can be reused and shouldn't be created and
freed per call, for performance reasons. The following examples will omit
creation and freeing of the DCCallVM, for simplicity.
In a more complicated example, let's call printf(3), which requires a
different initial mode, as well as a mode switch for the varargs part:
int n_written_chars, r;
/* initial callconv mode */
dcMode(vm, DC_CALL_C_ELLIPSIS);
dcReset(vm);
/* int printf(const char * restrict format, ...); */
dcArgPointer(vm, "my printf(%d) %s string%n");
/* switch mode for varargs part */
dcMode(vm, DC_CALL_C_ELLIPSIS_VARARGS);
dcArgInt(vm, 3);
dcArgPointer(vm, "format");
dcArgPointer(vm, &n_written_chars);
r = dcCallInt(vm, (DCpointer)&printf);
C/trivial aggregates by-value
Onto an example passing an aggregate by value (note that this is only
available on platforms where macro DC__Feature_AggrByVal is defined).
E.g. passing the following struct S to f():
struct S { char x[3]; double y; };
void f(int, struct S);
requires a DCaggr description of the fields/layout of struct S, and is
called as follows:
struct S s = { { 56, -23, 0 }, -6.28 };
DCaggr *a = dcNewAggr(2, sizeof(struct S));
dcAggrField(a, DC_SIGCHAR_CHAR, offsetof(struct S, x), 3);
dcAggrField(a, DC_SIGCHAR_DOUBLE, offsetof(struct S, y), 1);
dcCloseAggr(a);
dcMode(vm, DC_CALL_C_DEFAULT);
dcArgInt(vm, 999);
dcArgAggr(vm, a, &s);
dcCallVoid(vm, (DCpointer)&f);
dcFreeAggr(a);
Let's look at an example returning by value the above struct S from
function:
struct S g(int, short);
Omitting creation of the DCaggr *a description, for simplicity:
struct S s;
dcMode(vm, DC_CALL_C_DEFAULT);
/* needed when returning aggrs by value, *before* pushing args */
dcBeginCallAggr(vm, a);
dcArgInt(vm, 9);
dcArgShort(vm, 7);
dcCallAggr(vm, (DCpointer)&g, a, &s);
C++
In our next example, let's look at calling a simple C++ method, with the
method declaration being:
virtual void Klass::Method(float, int);
To keep the example simple, let's assume we have a pointer to this
virtual method in var mptr (e.g. grabbed from the instance's vtable), and
a pointer to the instance in var thisptr:
/* thiscall calling convention */
dcMode(vm, DC_CALL_C_DEFAULT_THIS);
dcReset(vm);
/* C++ methods use this-ptr as first/hidden argument */
dcArgPointer(vm, thisptr);
dcArgFloat(vm, 2.3f);
dcArgInt(vm, -19);
dcCallVoid(vm, (DCpointer)mptr);
Extending the last example to a vararg method would need some more
dcMode(3) calls. E.g.:
virtual void Klass::Method(float, int, ...);
would be called as follows:
/* thiscall calling convention (to pass this-ptr) */
dcMode(vm, DC_CALL_C_DEFAULT_THIS);
dcReset(vm);
/* C++ methods use this-ptr as first/hidden argument */
dcArgPointer(vm, thisptr);
/* fixed part of arguments */
dcMode(vm, DC_CALL_C_ELLIPSIS);
dcArgFloat(vm, 2.3f);
dcArgInt(vm, -19);
/* variable part of arguments */
dcMode(vm, DC_CALL_C_ELLIPSIS_VARARGS);
dcArgInt(vm, 7);
dcArgDouble(vm, 7.99);
dcCallVoid(vm, (DCpointer)mptr);
CONFORMING TO
The dyncall library needs at least a c99 compiler with additional support
for anonymous structs/unions (which were introduced officially in c11).
Given that those are generally supported by pretty much all major c99
conforming compilers (as default extension), it should build fine with a
c99 toolchain. Strictly speaking, dyncall conforms to c11, though.
SEE ALSO
dyncallback(3), dynload(3) and the dyncall manual (available in HTML and
PDF format) for more information.
AUTHORS
Daniel Adler ⟨dadler@uni-goettingen.de⟩
Tassilo Philipp ⟨tphilipp@potion-studios.com⟩
December 6, 2022
