I am creating a small C++ wrapper shared library around a Fortran 95 library. Since the Fortran symbols contain . in the symbol name, I have to use dlsym to load the Fortran function into a C++ function pointer.
Currently, I have a bunch of global function pointers in header files:
// test.h
extern void (*f)(int* arg);
and I populate them in the corresponding C++ file:
// test.cc
void (*f))(int* = reinterpret_cast<void(*)(int*>(dlsym(RTLD_DEFAULT, "real_f.symbol_name_");
Questions:
If I do it this way, when are these pointers populated?
Can I assume them to be loaded in my executable that loads this library?
In particular, can I use these functions in statically created objects in my executable or other libraries? Or does this suffer from the static initalization order fiasco?
If the above way is not correct, what is the most elegant way of populating these pointers such that they can be used in static objects in executables and other libraries?
I am using the Sun Studio compiler on Solaris, if that makes a difference, but I would also be interested in a solution for GCC on Linux.
Where does the line
f = reinterpret_cast<void(*)(int*)>(dlsym(RTLD_DEFAULT, "real_f.symbol_name_"));
occur in test.cc? The pointer will be initialized when the line is
executed (which of course depends on when the function which contains it
is called). Or did you mean to write
void (*f)(int* ) = reinterpret_cast<void(*)(int*>(dlsym(RTLD_DEFAULT, "real_f.symbol_name_");
? In this case, the pointer will be initialized during static
initialization. Which means that you still have order of initialization
issues if you try to use the pointers in the constructor of a static
object.
The classical solution for this would be to use some sort of singleton:
struct LibraryPointers
{
void (*f)(int* );
// ...
static LibraryPointers const& instance()
private:
LibraryPointers();
};
LibraryPointers const&
LibraryPointers::instance()
{
static LibraryPointers theOneAndOnly;
return theOneAndOnly;
}
LibraryPointers::LibraryPointers()
: f( reinterpret_cast<void(*)(int*)>(dlsym(RTLD_DEFAULT, "real_f.symbol_name_")) )
, // initialization of other pointers...
{
}
Then wrap the library in a C++ class which uses this structure to get
the addresses of the pointers.
And one last remark: the reinterpret_cast you are trying to do isn't
legal, at least not formally. (I think that both Sun CC and g++ will
accept it, however.) According to Posix, the correct way to get a
pointer to function from dlsym would be:
void (*f)(int* );
*reinterpret_cast<void**>(&f) = dlsym(...);
This doesn't lend itself to initializations, however.
Related
Suppose I have C++ object, like std::function. Is it safe in every way to pass such an object to another dynamically loaded shared library like this:
// lib
extern "C"
{
void call( void* f )
{
auto f_callable = (std::function<void()>*)f;
f_callable->operator()();
}
}
// executable
auto call = ( void (*) ( void* ) ) dlsym( lib_, "call" );
std::function<void()> f = []{
printf( "called" );
};
call( ( void* )&f );
What if the library and executable are compiled by different compilers (like clang and GCC)? Or by different versions of the same compiler?
This is defined behavior, provided that both parts of the C++ code gets generated by the same exact compiler. Casting the same thing to/from a void * is defined behavior. Presuming that the second C++ code sees the same C linkage, this is also defined behavior.
Whether or not it is safe when different compilers or different versions of the compilers are involved depends on whatever ABI guarantees these compilers provide. You will need to check their respective documentation to determine that.
I am experiencing the following issue, in my DLL project:
At the DLL side :
Inside the DLL I have declared a static vector as follows :
static std::vector<FilterProcessor::FilterInfo*> TableOfContents;
At DLL’s initialization time of static members, I am adding some entries to the above vector.
I have defined an extern “C” global function (getTocPointer()) which is returning a pointer to the vector, when it called from the client program.
extern "C" __declspec(dllexport) std::vector<FilterProcessor::FilterInfo*>* __cdecl getLibraryTOC();
At the client’s program side :
The DLL library is loaded without any problem
The address of getTocPointer() function is returned correctly to the client program, when the getProcAddress() function is called.
Indeed, when I am performing the debugging process in the DLL-side, the client program calls the above function and the execution process enters to it.
However, the vector has a zero size and, has no any contents which were added to it at initialization time. It seems it points to another vector object. . .
I can’t really understand what exactly goes wrong here.
The way of adding entries to this vector at initialization time, is the proper way?
If yes, what probably goes wrong when the client program calls the getLibraryTOC() function?
Thanks in advance
George
If that static global definition of the vector appears in a header file, then yes you do have multiple different vectors. Change the keyword static to extern to make the header file declare the vector rather than defining it, and then add exactly one definition in an implementation file.
Then, you may encounter the static initialization order fiasco. If the vector is defined in a different compilation unit than the code attempting to add entries to it, there's no guarantee that the vector object is alive yet. Attempting to use a vector whose constructor hasn't run is undefined behavior -- it might easily manifest as the constructor running afterward and setting the contents to zero length (as a default constructor should), but many other problems are possible.
You can avoid the SIOF by using a local static.
std::vector<FilterProcessor::FilterInfo*>& table_of_contents()
{
static std::vector<FilterProcessor::FilterInfo*> singleton;
return singleton;
}
In every location that would have accessed the global, including the initialization logic that fills the vector, and also your getLibraryTOC() exported function, call the accessor function instead.
That all is applicable to any C++ software having multiple compilation units. When you have a DLL, things get even more complicated, because the DLL and EXE are compiled and linked separately from each other, possibly with different settings, different compilers, or even entirely different languages. Sharing of complex objects across DLL boundaries is real trouble. If the DLL and EXE are always recompiled at the same time, it can work. But if you're trying to distribute the DLL for use by another party who writes the EXE code, the strong coupling will quickly become intolerable.
A better approach is to hide the library objects from the DLL boundary, and pass only primitive or OS-managed types across. For example:
#define DLLAPI __declspec(dllexport) __cdecl
extern "C" DLLAPI int32_t getLibraryTocCount()
{ return table_of_contents.size(); }
extern "C" DLLAPI BSTR getLibraryTocName(int index)
{ return ::SysAllocString(table_of_contents[index].name.c_str(); } // assumes std::wstring
// etc
The library I have implemented contains the following code (in a brief description) :
An Index class which implements the Table of contents of the library
A collection of audio filters named Filter01, Filter02 etc.
Index.h
struct LIB_SPECS Library_TableOfContents
{
static bool addTOCEntry(FilterInfo* Filter_Info); // add an entry to the TOC
static std::vector<FilterInfo*> TableOfContents; // TOC
};
/*-------------------------------------------------------------------
Called from the client program to return the pointer to TOC */
extern "C" LIB_SPECS std::vector<FlterInfo*>* __cdecl getLibraryTOC();
Index.cpp
/* Define / Initialize static variables */
std::vector<FilterInfo*> Library_TableOfContents::TableOfContents = {};
//=====================================================================
bool Library_TableOfContents::addTOCEntry(FilterInfo* Filter_Info)
{
Library_TableOfContents::TableOfContents.push_back(Filter_Info);
return false;
}
//======================================================================
std::vector<FilterInfo*>* getLibraryTOC()
{
return &Library_TableOfContents::TableOfContents;
}
For each Audio Filter in the library :
Filterxx.h
class LIB_SPECS Filterxx
{
public:
static struct FilterInfo
{
public:
std::string filterName;
std::string filterDescription;
// other filter info
FilterInfo(); // FilterInfo constructor
} Filter_Info;
virtual String doSomeWork(int AvatarId);
virtual void deleteFilter() = 0;
};
Filterxx.cpp
Filterxx::FilterInfo Filterxx::Filter_Info("Filterxx", “A filter description e.g. Low pass Filter ” ); //
FilterInfo::FilterInfo(std::string name, std::string description)
{
Filter_Info.filterName = name;
Filter_Info.filterDescription = description;
Library_TableOfContents::addTOCEntry(&Filter_Info);
}
// other filter functions
The getLibraryTOC() function, is called from the client program to get the table of contents in order to show it to the user.
As I said, indeed it is called by the client but, at the time of call, the table of contents seems to have a zero size.
I want to intercept application's calls to dlsym. I have tried declaring inside the .so that I am preloading dlsym , and using dlsym itself to get it's real address, but that for quite obvious reasons didn't work.
Is there a way easier than taking process' memory maps, and using libelf to find the real location of dlsym inside loaded libdl.so?
WARNING:
I have to explicitely warn everyone who tries to do this. The general premise of having a shared library hooking dlsym has several significant drawbacks. The biggest issue issue is that the original dlsym implementation if glibc will internally use stack unwinding techniques to find out from which loaded module the function was called. If the intercepting shared library then calls the original dlsym on behalf of the original application, this will break lookups using stuff like RTLD_NEXT, as now the current module isn't the originally calling one, but your hook library.
It might be possible to implement this the correct way, but it requires a lot more work. Without having tried it, I think that using dlinfo to get to the chained list of linket maps, you could individually walk through all modules, and do a separate dlsym for each one, to get the RTLD_NEXT behavior right. You still need to get the address of your caller for that, which you might get via the old backtrace(3) family of functions.
MY OLD ANSWER FROM 2013
I stumbled across the same problem with hdante's answer as the commenter: calling __libc_dlsym() directly crashes with a segfault. After reading some glibc sources, I came up with the following hack as a workaround:
extern void *_dl_sym(void *, const char *, void *);
extern void *dlsym(void *handle, const char *name)
{
/* my target binary is even asking for dlsym() via dlsym()... */
if (!strcmp(name,"dlsym"))
return (void*)dlsym;
return _dl_sym(handle, name, dlsym);
}
NOTE two things with this "solution":
This code bypasses the locking which is done internally by (__libc_)dlsym(), so to make this threadsafe, you should add some locking.
The thrid argument of _dl_sym() is the address of the caller, glibc seems to reconstruct this value by stack unwinding, but I just use the address of the function itself. The caller address is used internally to find the link map the caller is in to get things like RTLD_NEXT right (and, using NULL as thrid argument will make the call fail with an error when using RTLD_NEXT). However, I have not looked at glibc's unwindind functionality, so I'm not 100% sure that the above code will do the right thing, and it may happen to work just by chance alone...
The solution presented so far has some significant drawbacks: _dl_sym() acts quite differently than the intended dlsym() in some situations. For example, trying to resolve a symbol which does not exist does exit the program instead of just returning NULL. To work around that, one can use _dl_sym() to just get the pointer to the original dlsym() and use that for everything else (like in the "standard" LD_PRELOAD hook approch without hooking dlsym at all):
extern void *_dl_sym(void *, const char *, void *);
extern void *dlsym(void *handle, const char *name)
{
static void * (*real_dlsym)(void *, const char *)=NULL;
if (real_dlsym == NULL)
real_dlsym=_dl_sym(RTLD_NEXT, "dlsym", dlsym);
/* my target binary is even asking for dlsym() via dlsym()... */
if (!strcmp(name,"dlsym"))
return (void*)dlsym;
return real_dlsym(handle,name);
}
UPDATE FOR 2021 / glibc-2.34
Beginning with glibc 2.34, the function _dl_sym() is no longer publicly exported. Another approach I can suggest is to use dlvsym() instead, which is offically part of the glibc API and ABI. The only downside is that you now need the exact version to ask for the dlsym symbol. Fortunately, that is also part of the glibc ABI, unfortunately, it varies per architecture. However, a grep 'GLIBC_.*\bdlsym\b' -r sysdeps in the root folder of the glibc sources will tell you what you need:
[...]
sysdeps/unix/sysv/linux/i386/libc.abilist:GLIBC_2.0 dlsym F
sysdeps/unix/sysv/linux/i386/libc.abilist:GLIBC_2.34 dlsym F
[...]
sysdeps/unix/sysv/linux/x86_64/64/libc.abilist:GLIBC_2.2.5 dlsym F
sysdeps/unix/sysv/linux/x86_64/64/libc.abilist:GLIBC_2.34 dlsym F
Glibc-2.34 actually introduced new versions of this function, but the old versions are still be kept around for backwards compatibilty.
For x86_64, you could use:
real_dlsym=dlvsym(RTLD_NEXT, "dlsym", "GLIBC_2.2.5");
And, if you both like to get the newest version, as well as a potentially one of another interceptor in the same process, you can use that version to do an unversioned query again:
real_dlsym=real_dlsym(RTLD_NEXT, "dlsym");
If you actually need to hook both dlsym and dlvsym in your shared object, this approach of course won't work either.
UPDATE: hooking both dlsym() and dlvsym() at the same time
Out of curiosity, I thought about some approach to hook both of the glibc symbol query methods, and I came up with a solution using an additional wrapper library which links to libdl. The idea is that the interceptor library can dynamically load this library at runtime using dlopen() with the RTLD_LOCAL | RTLD_DEEPBIND flags, which will create a separate linker scope for this object, also containing the libdl, so that the dlsym and dlvsym will be resolved to the original methods, and not the one in the interceptor library. The problem now is that our interceptor library can not directly call any function inside the wrapper library, because we can not use dlsym, which is our original problem.
However, the shared library can have an initialization function, which the linker will call before the dlopen() returns. We just need to pass some information from the initialization function of the wrapper library to the interceptor library. Since both are in the same process, we can use the environment block for that.
This is the code I came up with:
dlsym_wrapper.h:
#ifndef DLSYM_WRAPPER_H
#define DLSYM_WRAPPER_H
#define DLSYM_WRAPPER_ENVNAME "DLSYM_WRAPPER_ORIG_FPTR"
#define DLSYM_WRAPPER_NAME "dlsym_wrapper.so"
typedef void* (*DLSYM_PROC_T)(void*, const char*);
#endif
dlsym_wrapper.c, compiled to dlsym_wrapper.so:
#include <dlfcn.h>
#include <stdio.h>
#include <stdlib.h>
#include "dlsym_wrapper.h"
__attribute__((constructor))
static void dlsym_wrapper_init()
{
if (getenv(DLSYM_WRAPPER_ENVNAME) == NULL) {
/* big enough to hold our pointer as hex string, plus a NUL-terminator */
char buf[sizeof(DLSYM_PROC_T)*2 + 3];
DLSYM_PROC_T dlsym_ptr=dlsym;
if (snprintf(buf, sizeof(buf), "%p", dlsym_ptr) < (int)sizeof(buf)) {
buf[sizeof(buf)-1] = 0;
if (setenv(DLSYM_WRAPPER_ENVNAME, buf, 1)) {
// error, setenv failed ...
}
} else {
// error, writing pointer hex string failed ...
}
} else {
// error: environment variable already set ...
}
}
And one function in the interceptor library to get the pointer to the
original dlsym() (should be called only once, guared by a mutex):
static void *dlsym_wrapper_get_dlsym
{
char dlsym_wrapper_name = DLSYM_WRAPPER_NAME;
void *wrapper;
const char * ptr_str;
void *res = NULL;
void *ptr = NULL;
if (getenv(DLSYM_WRAPPER_ENVNAME)) {
// error: already defined, shoudn't be...
}
wrapper = dlopen(dlsym_wrapper_name, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND | RTLD_NOLOAD);
if (wrapper) {
// error: dlsym_wrapper.so already loaded ...
// it is important that we load it by ourselves to a sepearte linker scope
}
wrapper = dlopen(dlsym_wrapper_name, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
if (!wrapper) {
// error: dlsym_wrapper.so can't be loaded
}
ptr_str = getenv(DLSYM_WRAPPER_ENVNAME);
if (!ptr_str) {
// error: dlsym_wrapper.so failed...
}
if (sscanf(ptr_str, "%p", &ptr) == 1) {
if (ptr) {
// success!
res = ptr;
} else {
// error: got invalid pointer ...
}
} else {
// error: failed to parse pointer...
}
// this is a bit evil: close the wrapper. we can be sure
// that libdl still is used, as this mosule uses it (dlopen)
dlclose(wrapper);
return res;
}
This of course assumes that dlsym_wrapper.so is in the library search path. However, you may prefer to just inject the interceptor library via LD_PRELOAD using a full path, and not modifying LD_LIBRARY_PATH at all. To do so, you can add dladdr(dlsym_wrapper_get_dlsym,...) to find the path of the injector library itself, and use that for searching the wrapper library, too.
http://www.linuxforu.com/2011/08/lets-hook-a-library-function/
From the text:
Do beware of functions that themselves call dlsym(), when you need to call __libc_dlsym (handle, symbol) in the hook.
extern void *__libc_dlsym (void *, const char *);
void *dlsym(void *handle, const char *symbol)
{
printf("Ha Ha...dlsym() Hooked\n");
void* result = __libc_dlsym(handle, symbol); /* now, this will call dlsym() library function */
return result;
}
I'm trying to initialize a global array of function pointers at compile-time, in either C or C++. Something like this:
module.h
typedef int16_t (*myfunc_t)(void);
extern myfunc_array[];
module.cpp
#include "module.h"
int16_t myfunc_1();
int16_t myfunc_2();
...
int16_t myfunc_N();
// the ordering of functions is not that important
myfunc_array[] = { myfunc_1, myfunc_2, ... , myfunc_N };
func1.cpp, func2.cpp, ... funcN.cpp (symbolic links to a single func.cpp file, so that different object files are created: func1.o, func2.o, func3.o, ... , funcN.o. NUMBER is defined using g++ -DNUMBER=N)
#include "module.h"
#define CONCAT2(x, y) x ## y
#define CONCAT(x, y) CONCAT2(x, y)
int16_t CONCAT(myfunc_, NUMBER)() { ... }
When compiled using g++ -DNUMBER=N, after preprocessing becomes:
func1.cpp
...
int16_t myfunc_1() { ... }
func2.cpp
...
int16_t myfunc_2() { ... }
and so on.
The declarations of myfunc_N() and the initialization of myfunc_array[] are not cool, since N changes often and could be between 10 to 200. I prefer not to use a script or Makefile to generate them either. The ordering of functions is not that important, i can work around that. Is there a neater/smarter way to do this?
How To Make a Low-Level Function Registry
First you create a macro to place pointers to your functions in a special section:
/* original typedef from question: */
typedef int16_t (*myfunc)(void);
#define myfunc_register(N) \
static myfunc registered_##myfunc_##N \
__attribute__((__section__(".myfunc_registry"))) = myfunc_##N
The static variable name is arbitrary (it will never be used) but it's nice to choose an expressive name. You use it by placing the registration just below your function:
myfunc_register(NUMBER);
Now when you compile your file (each time) it will have a pointer to your function in the section .myfunc_registry. This will all compile as-is but it won't do you any good without a linker script. Thanks to caf for pointing out the relatively new INSERT AFTER feature:
SECTIONS
{
.rel.rodata.myfunc_registry : {
PROVIDE(myfunc_registry_start = .);
*(.myfunc_registry)
PROVIDE(myfunc_registry_end = .);
}
}
INSERT AFTER .text;
The hardest part of this scheme is creating the entire linker script: You need to embed that snippet in the actual linker script for your host which is probably only available by building binutils by hand and examining the compile tree or via strings ld. It's a shame because I quite like linker script tricks.
Link with gcc -Wl,-Tlinkerscript.ld ... The -T option will enhance (rather than replace) the existing linker script.
Now the linker will gather all of your pointers with the section attribute together and helpfully provide a symbol pointing before and after your list:
extern myfunc myfunc_registry_start[], myfunc_registry_end[];
Now you can access your array:
/* this cannot be static because it is not know at compile time */
size_t myfunc_registry_size = (myfunc_registry_end - myfunc_registry_start);
int i;
for (i = 0; i < myfunc_registry_size); ++i)
(*myfunc_registry_start[i])();
They will not be in any particular order. You could number them by putting them in __section__(".myfunc_registry." #N) and then in the linker gathering *(.myfunc_registry.*), but the sorting would be lexographic instead of numeric.
I have tested this out with gcc 4.3.0 (although the gcc parts have been available for a long time) and ld 2.18.50 (you need a fairly recent ld for the INSERT AFTER magic).
This is very similar to the way the compiler and linker conspire to execute your global ctors, so it would be a whole lot easier to use a static C++ class constructor to register your functions and vastly more portable.
You can find examples of this in the Linux kernel, for example __initcall is very similar to this.
I was going to suggest this question is more about C, but on second thoughts, what you want is a global container of function pointers, and to register available functions into it. I believe this is called a Singleton (shudder).
You could make myfunc_array a vector, or wrap up a C equivalent, and provide a function to push myfuncs into it. Now finally, you can create a class (again you can do this in C), that takes a myfunc and pushes it into the global array. This will all occur immediately prior to main being called. Here are some code snippets to get you thinking:
// a header
extern vector<myfunc> myfunc_array;
struct _register_myfunc {
_register_myfunc(myfunc lolz0rs) {
myfunc_array.push_back(lolz0rs);
}
}
#define register_myfunc(lolz0rs) static _register_myfunc _unique_name(lolz0rs);
// a source
vector<myfunc> myfunc_array;
// another source
int16_t myfunc_1() { ... }
register_myfunc(myfunc_1);
// another source
int16_t myfunc_2() { ... }
register_myfunc(myfunc_2);
Keep in mind the following:
You can control the order the functions are registered by manipulating your link step.
The initialization of your translation unit-scoped variables occurs before main is called, i.e. the registering will be completed.
You can generate unique names using some macro magic and __COUNTER__. There may be other sneaky ways that I don't know about. See these useful questions:
Unnamed parameters in C
Unexpected predefined macro behaviour when pasting tokens
How to generate random variable names in C++ using macros?
Your solution sounds much too complicated and error prone to me.
You go over your project with a script (or probably make) to place the -D options to the compiler, anyhow. So I suppose you are keeping a list of all your functions (resp. the files defining them).
I'd use proper names for all the functions, nothing of your numbering scheme and then I would produce the file "module.cpp" with that script and initialize the table with the names.
For this you just have to keep a list of all your functions (and perhaps filenames) in one place. This could be easier be kept consistent than your actual scheme, I think.
Edit: Thinking of it even this might also be overengineering. If you have to maintain a list of your functions somewhere in any case, why not just inside the file "module.cpp"? Just include all the header files of all your functions, there, and list them in the initializer of the table.
Since you allow C++, the answer is obviously yes, with templates:
template<int N> int16_t myfunc() { /* N is a const int here */ }
myfunc_array[] = { myfunc<0>, myfunc<1>, myfunc<2> }
Now, you might wonder if you can create that variable-length initializer list with some macro. The answer is yes, but the macro's needed are ugly. So I'n not going to write them here, but point you to Boost::Preprocessor
However, do you really need such an array? Do you really need the name myfunc_array[0] for myfunc<0> ? Even if you need a runtime argument (myfunc_array[i]) there are other tricks:
inline template <int Nmax> int16_t myfunc_wrapper(int i) {
assert (i<Nmax);
return (i==Nmax) ? myfunc<Nmax> : myfunc_wrapper(i-1);
}
inline int16_t myfunc_wrapper(int i) {
return myfunc_wrapper<NUMBER>(i); // NUMBER is defined on with g++ -DNUMBER=N
}
Ok I worked out a solution based on Matt Joiner's tip:
module.h
typedef int16_t (*myfunc_t)(void);
extern myfunc_array[];
class FunctionRegistrar {
public:
FunctionRegistrar(myfunc_t fn, int fn_number) {
myfunc_array[fn_number - 1] = fn; // ensures correct ordering of functions (not that important though)
}
}
module.cpp
#include "module.h"
myfunc_array[100]; // The size needs to be #defined by the compiler, probably
func1.cpp, func2.cpp, ... funcN.cpp
#include "module.h"
static int16_t myfunc(void) { ... }
static FunctionRegistrar functionRegistrar(myfunc, NUMBER);
Thanks everyone!
I'm having some weird issues with static initalization. I'm using a code generator to generate structs and serialization code for a message passing system I wrote. In order to have a way of easily allocating a message based on it's message id I have my code generator ouput something similar to the following for each message type:
MessageAllocator s_InputPushUserControllerMessageAlloc(INPUT_PUSH_USER_CONTROLLER_MESSAGE_ID, (AllocateMessageFunc)Create_InputPushUserControllerMessage);
The MessageAllocator class basically looks like this:
MessageAllocator::MessageAllocator( uint32_t messageTypeID, AllocateMessageFunc func )
{
if (!s_map) s_map = new std::map<uint32_t, AllocateMessageFunc>();
if (s_map->insert(std::make_pair(messageTypeID, func)).second == false)
{
//duplicate key!
ASSERT(false, L"Nooooo!");
}
s_count++;
}
MessageAllocator::~MessageAllocator()
{
s_count--;
if (s_count == 0) delete s_map;
}
where s_map and s_count are static members of MessageAllocator. This works most of the time but sometimes messages are not added to the map. For example, this particular message is not added unless i call Create_InputPushUserControllerMessage() somewhere in my startup code, however other messages work fine. I thought this might be something to do with the linker incorrectly thinking the type is unreferenced and removing it so I disabled that using the /OPT:NOREF switch (I'm using Visual Studio 2008 SP1) but that had no effect.
I'm aware of the problem of the "static initialization order fiasco" but as far as I know the order in which these objects are created shouldn't alter the result so this seems ok to me.
Any insight here would be appreciated.
Put the static into a class so it is a static member of a class
struct InputPushUserControllerMessageAlloc { static MessageAllocator s_obj; };
MessageAllocator InputPushUserControllerMessageAlloc::s_obj(
INPUT_PUSH_USER_CONTROLLER_MESSAGE_ID,
(AllocateMessageFunc)Create_InputPushUserControllerMessage);
The Standard allows it to delay initialization of objects having namespace scope until any function/object from its translation unit is used. If the initialization has side-effect, it can't be optimized out. But that doesn't forbid delaying it.
Not so of objects having class-scope. So that might forbid it optimizing something there.
I would change s_map from a static class member into a static method member:
std::map<uint32_t,AllocateMessageFunc>& MessageAllocator::getMap()
{
// Initialized on first use and destroyed correctly on program termination.
static std::map<uint32_t,AllocateMessageFunc> s_map;
return s_map;
}
MessageAllocator::MessageAllocator( uint32_t messageTypeID, AllocateMessageFunc func )
{
if (getMap().insert(std::make_pair(messageTypeID, func)).second == false)
{
//duplicate key!
ASSERT(false, L"Nooooo!");
}
}
No need for destructor or a count.
If your global objects are in separate DLL's(or shared libs) that are lazy loaded.
This may cause a problem similar to your description.
You are not setting the pointer back to null.
MessageAllocator::~MessageAllocator()
{
s_count--;
if (s_count == 0)
{
delete s_map;
s_map = 0;
}
}
Turns out that the object files containing the static initializers were not included by the linker because nothing referenced any functions in them. To work around this I extern "C"-ed one of the generated functions so that it would have a predictable non-mangled name and then forced a reference to it using a pragma like this for each message
#pragma comment(linker, "/include:Create_GraphicsDynamicMeshCreationMessage")
which I put in the generated header file that is later included in all the other non-generated files. It's MSVC only and kind of hack but I assume I can do something similar on GCC once I eventually port it.