You want boilerplate-less delay loading. On Windows, MSVC implements delay loading by emitting a stub that resolves the function through a function pointer. You can do the same. First, let's observe that function pointers and functions are interchangeable if all you do is call them. The syntax for invoking a function or a function pointer is the same:
void foo_impl() {}
void (*foo)() = foo_impl;
int main() {
foo_impl();
foo();
}
The idea is to set the function pointer initially to a thunk that will resolve the real function at runtime:
extern void (*foo)();
void foo_thunk() {
foo = QLibrary::resolve("libmylib", "foo");
if (!foo) abort();
return foo();
}
void (*foo)() = foo_thunk;
int main() {
foo(); // calls foo_thunk to resolve foo and calls foo from libmylib
foo(); // calls foo from libmylib
}
When you first call foo, it will really call foo_thunk, resolve the function address, and call real foo implementation.
To do this, you can split the library into two libraries:
- The library implementation. It is unaware of demand-loading.
- A demand-load stub.
The executable will link to the demand-load stub library; that is either static or dynamic. The demand-load stub will automatically resolve the symbols at runtime and call into the implementation.
If you're clever, you can design the header for the implementation such that the header itself can be used to generate all the stubs without having to enter their details twice.
Complete Example
Everything follows, it's also available from https://github.com/KubaO/stackoverflown/tree/master/questions/demand-load-39291032
The top-level project consists of:
lib1 - the dynamic librarylib1_demand - the static demand-load thunk for lib1main - the application that uses lib1_demand
demand-load-39291032.pro
TEMPLATE = subdirs
SUBDIRS = lib1 lib1_demand main
main.depends = lib1_demand
lib1_demand.depends = lib1
We can factor out the cleverness into a separate header. This header allows us to define the library interface so that the thunks can be automatically generated.
The heavy use of preprocessor and a somewhat redundant syntax is needed due to limitations of C. If you wanted to implement this for C++ only, there'd be no need to repeat the argument list.
demand_load.h
// Configuration macros:
// DEMAND_NAME - must be set to a unique identifier of the library
// DEMAND_LOAD - if defined, the functions are declared as function pointers, **or**
// DEMAND_BUILD - if defined, the thunks and function pointers are defined
#if defined(DEMAND_FUN)
#error Multiple inclusion of demand_load.h without undefining DEMAND_FUN first.
#endif
#if !defined(DEMAND_NAME)
#error DEMAND_NAME must be defined
#endif
#if defined(DEMAND_LOAD)
// Interface via a function pointer
#define DEMAND_FUN(ret,name,args,arg_call) \
extern ret (*name)args;
#elif defined(DEMAND_BUILD)
// Implementation of the demand loader stub
#ifndef DEMAND_CAT
#define DEMAND_CAT_(x,y) x##y
#define DEMAND_CAT(x,y) DEMAND_CAT_(x,y)
#endif
void (* DEMAND_CAT(resolve_,DEMAND_NAME)(const char *))();
#if defined(__cplusplus)
#define DEMAND_FUN(ret,name,args,arg_call) \
extern ret (*name)args; \
ret name##_thunk args { \
name = reinterpret_cast<decltype(name)>(DEMAND_CAT(resolve_,DEMAND_NAME)(#name)); \
return name arg_call; \
}\
ret (*name)args = name##_thunk;
#else
#define DEMAND_FUN(ret,name,args,arg_call) \
extern ret (*name)args; \
ret name##_impl args { \
name = (void*)DEMAND_CAT(resolve_,DEMAND_NAME)(#name); \
name arg_call; \
}\
ret (*name)args = name##_impl;
#endif // __cplusplus
#else
// Interface via a function
#define DEMAND_FUN(ret,name,args,arg_call) \
ret name args;
#endif
Then, the dynamic library itself:
lib1/lib1.pro
TEMPLATE = lib
SOURCES = lib1.c
HEADERS = lib1.h
INCLUDEPATH += ..
DEPENDPATH += ..
Instead of declaring the functions directly, we'll use DEMAND_FUN from demand_load.h. If DEMAND_LOAD_LIB1 is defined when the header is included, it will offer a demand-load interface to the library. If DEMAND_BUILD is defined, it'll define the demand-load thunks. If neither is defined, it will offer a normal interface.
We take care to undefine the implementation-specific macros so that the global namespace is not polluted. We can then include multiple libraries the project, each one individually selectable between demand- and non-demand loading.
lib1/lib1.h
#ifndef LIB_H
#define LIB_H
#ifdef __cplusplus
extern "C" {
#endif
#define DEMAND_NAME LIB1
#ifdef DEMAND_LOAD_LIB1
#define DEMAND_LOAD
#endif
#include "demand_load.h"
#undef DEMAND_LOAD
DEMAND_FUN(int, My_Add, (int i, int j), (i,j))
DEMAND_FUN(int, My_Subtract, (int i, int j), (i,j))
#undef DEMAND_FUN
#undef DEMAND_NAME
#ifdef __cplusplus
}
#endif
#endif
The implementation is uncontroversial:
lib1/lib1.c
#include "lib1.h"
int My_Add(int i, int j) {
return i+j;
}
int My_Subtract(int i, int j) {
return i-j;
}
For the user of such a library, demand loading is reduced to defining one macro and using the thunk library lib1_demand instead of the dynamic library lib1.
main/main.pro
if (true) {
# Use demand-loaded lib1
DEFINES += DEMAND_LOAD_LIB1
LIBS += -L../lib1_demand -llib1_demand
} else {
# Use direct-loaded lib1
LIBS += -L../lib1 -llib1
}
QT = core
CONFIG += console c++11
CONFIG -= app_bundle
TARGET = demand-load-39291032
TEMPLATE = app
INCLUDEPATH += ..
DEPENDPATH += ..
SOURCES = main.cpp
main/main.cpp
#include "lib1/lib1.h"
#include <QtCore>
int main() {
auto a = My_Add(1, 2);
Q_ASSERT(a == 3);
auto b = My_Add(3, 4);
Q_ASSERT(b == 7);
auto c = My_Subtract(5, 7);
Q_ASSERT(c == -2);
}
Finally, the implementation of the thunk. Here we have a choice between using dlopen+dlsym or QLibrary. For simplicity, I opted for the latter:
lib1_demand/lib1_demand.pro
QT = core
TEMPLATE = lib
CONFIG += staticlib
INCLUDEPATH += ..
DEPENDPATH += ..
SOURCES = lib1_demand.cpp
HEADERS = ../demand_load.h
lib1_demand/lib1_demand.cpp
#define DEMAND_BUILD
#include "lib1/lib1.h"
#include <QLibrary>
void (* resolve_LIB1(const char * name))() {
auto f = QLibrary::resolve("../lib1/liblib1", name);
return f;
}
