Why is it possible to redefine C library functions? - c

Question

I noticed that if I write a function named getline, this function will be used if I invoke it, even if I #include <stdio.h>, but if I don't write such a function, the one from stdio.h will be used.
I expected instead to get a linker error, the same as if I had done the following:
foo.c:
int f() { return 0; }
main.c:
int f() { return 1; }
int main() { return f(); }
Compile:
$ gcc -c foo.c
$ gcc -c main.c
$ gcc foo.o main.o
/usr/bin/ld: main.o: in function `f':
main.c:(.text+0x0): multiple definition of `f'; foo.o:foo.c:(.text+0x0): first defined here
collect2: error: ld returned 1 exit status
The linker error makes sense to me; when the linker attempts to combine the object files into a single binary, it doesn't know how to resolve the invocation of f(); should it use foo.o's f() or main.o's f()?
But then why don't I get such a linker error when I write my own versions of getline or other C library functions?
This came up because I noticed that when compiling with -std=c99, gcc gives me a implicit-function-declaration warning for using getline. I can make an explicit function prototype, and it works correctly, but this implies that glibc's getline is being linked, so I tested what happens if I write my own getline, and if I do, the linker uses it instead and produces no error... The same appears to be true for other C library functions. Why is this? Why don't I get a linker error instead?

Answer 1

Linkers process library files differently than object files. The following discusses typical behavior for linkers. Details may vary with specific linkers and command-line switches or other settings.
When a linker processes an object file, it includes the entire object file in the output file it is building. As it is doing this, it builds a list of symbols that the object files use (refer to) but that are not defined yet.
A library file consists of multiple object modules inside a containing file. When a linker processes a library file, it examines each module in the library file and compares the symbols that module defines to that list of symbols that are needed but not yet defined. When it finds such a module, the linker includes that module in the output file. (The linker may also go back to earlier modules in the same library file, in case a later module uses a symbol that an earlier one defines.)
Any modules in the library file that do not provide a needed symbol are not needed in the output file, so the linker does not include them.
A consequence of this is that, if a same symbol is defined more than once in the object files, there will be multiple definitions because they are both built into the output file. However, if a symbol is defined once in the object files and once in the library, the one in the library will not be used because, when the linker considers the module it is in, that symbol will not be on the list of needed symbols, and the linker will not include it in the output file. So the output file ends up with just one definition of the symbol, the one from the object modules.
There are some complications to this. Suppose a module in a library defines both sin and cos, and an object module defines sin and uses both sin and cos. When the linker processes the object module, it will note that sin and cos are both used. The reference to sin will be satisfied by the object module, but cos is still needed. Then, when the linker processes the library, it will find cos and include that module. But that module also defines sin, so there will be two definitions of sin in the output file, and the linker will complain. So you can get multiple-definition errors from library modules this way.
Another complication is that the order of processing matters. If the linker first processes an object module that needs getline, and then a library module that defines getline, and then an object module that defines getline, the library module will be included in the output file (because getline was needed when the linker processed the library), and the object module that defines getline will also be included (because the linker includes all object files). So the output will have multiple definitions of getline, and the linker will complain. This is one reason why libraries are generally processed last, so that all object modules are processed first, and only things that are needed from libraries are taken.
In spite of this linker behavior, you cannot rely on defining your own versions of standard C routines. Compilers may have built-in knowledge about how the routines are specified by the C standard, and they may replace calls to those routines with other code. If you do need to provide your own version of a standard routine, the compiler may have a switch to disable its special treatment of that routine. For example, GCC has -fno-builtin-function, where function is replaced with a particular name, to tell it to disable special knowledge of a function.