Linking

The Link Binary With Libraries build phase in Xcode projects links frameworks and libraries with object files to produce a binary file. Source files that use code in a framework or a library must include a reference to the appropriate programming interface contained in them.

Libraries and frameworks are linked to object files when building an executable file. However, this process is slow and can detract from the development experience. Xcode provides a feature, called ZeroLink, that eliminates the link step while you work on a project; see Using ZeroLink for details.

Specifying the Search Order of External Symbols

The order in which frameworks and libraries are listed in the Link Binary With Libraries build phase specifies the order in which external symbols are resolved by the static linker at build time and the dynamic linker at runtime. When either of the linkers encounters an undefined external symbol, they look for the symbol starting with the first framework or library listed in the build phase.

When a program is built, the static linker replaces references to external symbols with the addresses of the symbols in the referenced libraries (this is called prebinding), or tells the dynamic linker to resolve the references when a program is loaded or when a symbol is referenced. Having the dynamic linker resolve references to external symbols at runtime provides the most flexibility, as a program can link with new versions of the symbols as they become available. However, this approach is not recommended for all situations, as linking with newer versions of a method or a function may cause problems during a program’s execution.

In addition, how frameworks and libraries are listed in a Link Binary With Libraries build phase, tells the static linker the approach (or the semantics) to use when binding or resolving references to external symbols defined in libraries.

Placing static libraries after dynamic libraries in the Link Binary With Libraries build phase, ensures that the static linker binds references to external symbols defined in static libraries at build time, even when newer versions of the static libraries the application originally was linked with are present in the user’s system.

When static libraries are listed before a dynamic library in the Link Binary With Libraries build phase, the static linker doesn’t resolve references to symbols in those static libraries. Instead, those symbols are resolved by the dynamic linker at runtime. This may cause problems when the static libraries are updated, as the application links to the new, potentially incompatible versions of the symbols instead of the ones the developer intended.

For details on how symbols are resolved, see “Finding Imported Symbols” in Executing Mach-O Files in Mach-O Runtime Architecture and the ld man page.

Preventing Prebinding

Mac OS X includes a prebinding mechanism used to speed-up application launch in programs that link against dynamic libraries. When a user installs an application or upgrades the operating system, a prebinding agent links the application against new versions of the dynamic libraries. Sometimes, however, you may want to prevent this behavior for specific applications.

To link the binary file, framework, library, or plug-in, so that prebinding is never done on it, you need to add the -nofixprebinding option to the linker invocation. To do this, add -nofixprebinding to the Other Linker Flags (OTHER_LDFLAGS) build setting. See the ld man page for more information.

Linking With System Frameworks

When linking with system frameworks (located in /System/Library/Frameworks), include only the umbrella header files in your source files and link only with the appropriate umbrella framework for your application. For example, in a Carbon application that uses the Address Book framework, you would include the following line in the header files of modules that access the Address Book programming interface:

#include <Carbon/Carbon.h>

You would also add AddressBook.framework to the list of files in the Frameworks & Libraries build phase.

Linking to a Dynamic Library in a Nonstandard Location

When you need to link with a custom version of a dynamic library but don’t want to replace the standard version of the library, you can use the -dylib_file option of the linker to tell it where to find the nonstandard version of the library. Just add -dylib_file standard_library_path:nonstandard_library_path to the Other Linker Flags build setting, where standard_library_path is the path to the standard library and nonstandard_library_path is the path to the custom version of the library.

Reducing the Number of Exported Symbols

By default, Xcode builds binary files that export all their symbols. To reduce the number of symbols you want to export from a binary file, create an export file and set the Exported Symbols File (EXPORTED_SYMBOLS_FILE) build setting to the name of the file. For more information, see Minimizing Your Exported Symbols in Code Size Performance Guidelines.

Reducing Paging Activity

To help reduce your application’s paging activity at runtime, you may specify an order file to the linker. You do this by setting the Symbol Ordering Flags (SECTORDER_FLAGS) build setting to -sectorder __TEXT __text <order_file>. For information on order files, see Improving Locality of Reference in Code Size Performance Guidelines.

Dead-Code Stripping

The static linker (ld) supports the removal of unused code and data blocks from executable files. This process (known as dead-code stripping) helps reduce the overall size of executables, which in turn improves performance by reducing the memory footprint of the executable. It also allows programs to link successfully in the situation where unused code refers to an undefined symbol, something that would normally result in a link error.

Dead-code stripping is not limited to removing only unused functions and executable code from a binary. The linker also removes any unused symbols and data that reside in data blocks. Such symbols might include global variables, static variables, and string data among others.

When dead-code stripping is enabled, the static linker searches for code that is unreachable from an initial set of live symbols and blocks. The initial list of live symbols and blocks may include the following:

Enabling Dead-Code Stripping in Your Project

To enable dead-code stripping in your project, you must pass the appropriate command-line options to the linker. From Xcode, you add these options in the Build pane of the target inspector; otherwise, you must add these options to your makefile or build scripts. Table 28-2 lists the Xcode build settings that control dead-code stripping. Enabling either of these build settings causes Xcode to build with the corresponding linker option.

Table 28-1  Xcode build settings for dead stripping

Linker option

Build setting

-dead_strip

Dead Code Stripping (DEAD_CODE_STRIPPING)

-no_dead_strip_inits_and_terms

Don’t Dead-Strip Inits and Terms (PRESERVE_DEAD_CODE_INITS_AND_TERMS)

Table 28-1 lists the basic dead-code stripping options.

Table 28-2  Linker options for dead stripping

Linker option

Description

-dead_strip

Enables basic dead-code stripping by the linker.

-no_dead_strip_inits_and_terms

Prevents all constructors and destructors from being stripped when the -dead_strip option is in effect, even if they are not live.

You must recompile all object files using the compiler included with Xcode 1.5 or later before dead-code stripping can be performed by the linker. You must also compile the object files with the -gfull option to ensure that the resulting binaries can be properly debugged. In Xcode, change the value of the Level of Debug Symbols (GCC_DEBUGGING_SYMBOLS) build setting to All Symbols (-gfull).

Identifying Stripped Symbols

If you want to know what symbols were stripped by the static linker, you can find out by examining the linker-generated load map. This map lists all of the segments and sections in the linked executable and also lists the dead-stripped symbols. To have the linker generate a load map, add the -M option to your linker command-line options. In Xcode, you can add this option to the Other Linker Flags build setting.

Preventing the Stripping of Unused Symbols

If your executable contains symbols that you know should not be stripped, you need to notify the linker that the symbol is actually used. You must prevent the stripping of symbols in situations where external code (such as plug-ins) use those symbols but local code does not.

There are two ways to tell the linker not to dead strip a symbol. You can include it in an exports file or mark the symbol declaration explicitly. To mark the declaration of a symbol, you include the used attribute as part of its definition. For example, you would mark a function as used by declaring it as follows:

void MyFunction(int param1) __attribute__((used));

Alternatively, you can provide an exports list for your executable that lists any symbols you expect to be used by plug-ins or other external code modules. To specify an exports file from Xcode, use the Exported Symbols File (EXPORTED_SYMBOLS_FILE) build setting; enter the path, relative to the project directory, to the exports file. To specify an exports file from the linker command line use the -exported_symbols_list. option. (You can also specify a list of symbols not to export using the -unexported_symbols_list option.)

If you are using an exports list and building either a shared library, or an executable that will be used with ld's -bundle_loader flag, you need to include the symbols for exception frame information in the exports list for your exported C++ symbols. Otherwise, they may be stripped. These symbols end with .eh; you can view them with the nm tool.

Assembly Language Support

If you are writing assembly language code, the assembler now recognizes some additional directives to preserve or enhance the dead-stripping of code and data. You can use these directives to flag individual symbols or entire sections of assembly code.

Preserving Individual Symbols

To prevent the dead stripping of an individual symbol, use the .no_dead_strip directive. For example, the following code prevents the specified string from being dead stripped:

.no_dead_strip _my_version_string
.cstring
_my_version_string:
.ascii "Version 1.1"

Preserving Sections

To prevent an entire section from being dead stripped, add the no_dead_strip attribute to the section declaration. The following example demonstrates the use of this attribute:

.section __OBJC, __image_info, regular, no_dead_strip

You can also add the live_support attribute to a section to prevent it from being dead stripped if it references other code that is live. This attribute prevents the dead stripping of some code that might actually be needed but not referenced in a detectable way. For example, the compiler adds this attribute to C++ exception frame information. In your code, you might use the attribute as follows:

.section __TEXT, __eh_frame, coalesced, no_toc+strip_static_syms+live_support

Dividing Blocks of Symbols

The .subsections_via_symbols directive notifies the assembler that the contents of sections may be safely divided into individual blocks prior to dead-code stripping. This directive makes it possible for individual symbols to be stripped out of a given section if they are not used. This directive applies to all section declarations in your assembly file and should be placed outside of any section declarations, as shown below:

.subsections_via_symbols
 
; Section declarations...

If you use this directive, make sure that each symbol in the section marks the beginning of a separate block of code. Implicit dependencies between blocks of code might result in the removal of needed code from the executable. For example, the following section contains three individual symbols, but execution of the code at _plus_three ends at the blr statement at the bottom of the code block.

.text
.globl _plus_three
_plus_three:
addi r3, r3, 1
.globl _plus_two
_plus_two:
addi r3, r3, 1
.global _plus_one
_plus_one:
addi r3, r3, 1
blr

If you were to use the .subsections_via_symbols directive on this code, the assembler would permit the stripping of the symbols _plus_two and _plus_one if they were not called by any other code. If this occurred, _plus_three would no longer return the correct value because part of its code would be missing. In addition, if _plus_one were dead stripped, the code might crash as it continued executing into the next block.

Using ZeroLink

ZeroLink speeds application development time by eliminating the link process from development builds. Instead, Xcode generates an application stub that contains the full paths to the object files that make up the application. At runtime, each object (.o) file is linked as it’s needed. This works only when running your application within Xcode. You cannot deploy applications using ZeroLink.

To turn ZeroLink on or off, use the ZeroLink (ZERO_LINK) build setting. ZeroLink is enabled by default in the Development build style. If you build with this build style, you automatically get ZeroLink functionality. See Build Styles for more information on using build styles. ZeroLink works only for native targets.

The following sections explain how you can customize ZeroLink to further reduce application launch times and identify issues you must keep in mind when using ZeroLink.

Customizing ZeroLink

ZeroLink postpones the linking of object files until the last moment possible. However, there are some symbols that, by default, are always resolved (that is, the corresponding object files are linked against the application and the code is executed). These symbols are static initializers in C++ and +load methods and categories in Objective-C. You can tell ZeroLink not to search the object files of your application for these symbols to reduce the application’s launch time.

There are situations that require the initialization of objects before a program’s main function is called. For example, a class may declare global variables that can be accessed by other code before the class has a chance to initialize them. In Objective-C, the +initialize method may be executed too late (see Initializing a Class Object in The Objective-C Programming Language. The purpose of static initializers in C++ and +load methods in Objective-C is to provide developers with a mechanism to initialize variables at the earliest possible point during a program’s launch process. In Objective-C–based applications, categories are also loaded before main is called.

When you build an application, the static linker adds the standard entry-point function to the main executable file. This function sets up the runtime environment state for the kernel and the application before calling main, which involves calling static initializers for C++ code and loading categories and invoking +load methods for Objective-C code.

When using ZeroLink, you can further reduce the launch time of the application by postponing the execution of static initializers and +load methods, and the loading of categories. But you must be certain that code in your application doesn’t rely on static initializers or +load methods being called before main or on categories being loaded before main is called. Otherwise, your application may crash or behave unexpectedly.

There are three linker options you can use to customize ZeroLink in your project. To use these options, add them to the Other Linker Flags (OTHER_LDFLAGS) build setting:

  • -no-run-initializers-before-main: If an application contains C++ code, the linker looks for static initializers at launch time before calling main in all the object files that make up the application and, if it finds any, links the object files containing them into the application. This may slow down application launch. If you’re developing an application using C++ and it doesn’t depend on static initializers being run before main, use this option to prevent ZeroLink from scanning object files in search of static initializers before calling main.

    Keep in mind that if there’s code that requires that static initializers be run before main, your application could crash. This specially true for applications that use static initializers to register code with a registry. If the sole purpose of the static initializers is to perform the registration (that is, if no other code would ever trigger the execution of the static initializers by accessing a global variable, for example), no registration would take place. For example, when reading a serialized file, the reader reads the name of the class of each serialized object and tells the class to reconstruct an instance from the data. In C++ there is no infrastructure to look up a class by name. Instead, a common idiom is for each class capable of serializing and deserializing instances of itself to register its class name and deserialization method address with the reader using a static initializer.

  • -no-load-categories-before-main: If an application contains Objective-C code, the linker looks for categories at launch time before calling main in all the object files that make up the application and, if it finds any, links the object files containing them into the application. As with -no-run-initializers-before-main, this may slow application launch. If your application doesn’t depend on categories, use this option to prevent ZeroLink from scanning object files in search of categories before calling main.

    Some developers rely on the use of categories in the implementation of their class hierarchies. In such a model, the implementation of a class spans one or more categories of that class. Using this design model, a program may run incorrectly when the categories that implement additional class functionality are not loaded. With ZeroLink, classes can be loaded when need because there’s a direct reference to them, but ZeroLink cannot load categories dynamically. Therefore, if your application depends on categories, it may crash or behave unexpectedly if you use the -no-load-categories-before-main flag.

  • -no-run-load-methods-before-main: Similarly to -no-load-categories-before-main, ZeroLink scans object files for +load methods in Objective-C code before main is invoked to link them to the application. Using this option prevents the scanning of object files for +load methods before ZeroLink calls main. +load methods in Objective-C are similar to static initializers in C++: They are executed early during application launch, before main is called, to perform tasks that must be performed at the earliest possible time. However, the order of execution of +load methods is not guaranteed. If your code depends on +load methods to be run before main, your application may crash or run incorrectly if you use this flag.

When building an application that uses static libraries (.a files), each static library is linked to produce a bundle. At runtime, each bundle is loaded on demand. If your application starts slowly while using ZeroLink and has a large number (100 or more) of object files, you can try adding an intermediate static library target, containing the relatively stable parts of your source code. This gives you the best of both worlds: static (build time) linking for stable code and dynamic (runtime) linking for code that changes frequently.

If you want to view information about the loading and linking of object files as your application runs, set the ZERO_LINK_VERBOSE environment variable to any value. The information appears in run log of the application or stderr.

Caveats When Using ZeroLink

These are some things you should keep in mind when using ZeroLink:

  • ZeroLink doesn’t support the use of private external symbols; that is symbols declared as __private_extern__.

    Private external symbols are visible only to other modules within the same Mach-O file as the modules that contain them. If you use a private external symbol in your project while ZeroLink is turned on, you get an unknown-symbol error when your code tries to access it. For example, if you have the definition __private_extern__ int my_extern = 800; in a source file and the declaration extern int my_extern; in another source file, when the second module accesses my_extern, your application exits with the following log output:

    ZeroLink: unknown symbol '_my_extern'
    MyApplication has exited due to signal 6 (SIGABRT)

    For more information on private external symbols, see “Scope and Treatment of Symbol Definitions” in Executing Mach-O Files in Mach-O Runtime Architecture.

  • When setting breakpoints on methods, you must use full method specifiers, not just selectors. For example, if your project has a class named MyClass with an instance method called myMethod, you must specify a breakpoint on myMethod like this:

    -[MyClass myMethod]
  • If you get an error similar to this one:

    dyld: /Users/user_name/MyApp/build/MyApp.app/Contents/MacOS/MyApp  Undefined symbols:
    Foundation undefined reference to _objc_exception_set_functions  expected to be defined in
    /System/Library/PrivateFrameworks/ZeroLink.framework/Versions/A/Resources/libobjc.A.dylib

    delete /System/Library/PrivateFrameworks/ZeroLink.framework/Versions/A/Resources/libobjc.A.dylib.

    You would get this error if you install the Xcode tools in a system with a prerelease version of Mac OS X. The libobjc.A.dylib file contains a developmental copy of the Objective-C runtime. It’s not necessary for normal development.