System Library

This document describes the requirements, design, and implementation details of LLVM's System Library. The library is composed of the header files in llvm/include/llvm/System and the source files in llvm/lib/System. The goal of this library is to completely shield LLVM from the variations in operating system interfaces. By centralizing LLVM's use of operating system interfaces, we make it possible for the LLVM tool chain and runtime libraries to be more easily ported to new platforms since (theoretically) only llvm/lib/System needs to be ported. This library also unclutters the rest of LLVM from #ifdef use and special cases for specific operating systems. Such uses are replaced with simple calls to the interfaces provided in llvm/include/llvm/System.

Note that lib/System is not intended to be a complete operating system wrapper (such as the Adaptive Communications Environment (ACE) or Apache Portable Runtime (APR)), but only to provide the functionality necessary to support LLVM.

The System Library was written by Reid Spencer who formulated the design based on similar original work as part of the eXtensible Programming System (XPS).

The System library's requirements are aimed at shielding LLVM from the variations in operating system interfaces. The following sections define the requirements needed to fulfill this objective. Of necessity, these requirements must be strictly followed in order to ensure the library's goal is reached.

The library must shield LLVM from all system libraries. To obtain system level functionality, LLVM must #include "llvm/System/Thing.h" and nothing else. This means that Thing.h cannot expose any system header files. This protects LLVM from accidentally using system specific functionality except through the lib/System interface. Specifically this means that header files like "unistd.h", "windows.h", "stdio.h", and "string.h" are verbotten outside the implementation of lib/System.

The standard C headers (the ones beginning with "c") are allowed to be exposed through the lib/System interface. These headers and the things they declare are considered to be platform agnostic. LLVM source files may include them or obtain their inclusion through lib/System interfaces.

The standard C++ headers from the standard C++ library and standard template library are allowed to be exposed through the lib/System interface. These headers and the things they declare are considered to be platform agnostic. LLVM source files may include them or obtain their inclusion through lib/System interfaces.

Any functions defined by system libraries (i.e. not defined by lib/System) must not be exposed through the lib/System interface, even if the header file for that function is not exposed. This prevents inadvertent use of system specific functionality.

For example, the stat system call is notorious for having variations in the data it provides. lib/System must not declare stat nor allow it to be declared. Instead it should provide its own interface to discovering information about files and directories. Those interfaces may be implemented in terms of stat but that is strictly an implementation detail.

Any data defined by system libraries (i.e. not defined by lib/System) must not be exposed through the lib/System interface, even if the header file for that function is not exposed. As with functions, this prevents inadvertent use of data that might not exist on all platforms.

If an error occurs that lib/System cannot handle, the only action taken by lib/System is to throw an instance of std:string. The contents of the string must explain both what happened and the context in which it happened. The format of the string should be a (possibly empty) list of contexts each terminated with a : and a space, followed by the error message, optionally followed by a reason, and optionally followed by a suggestion.

For example, failure to open a file named "foo" could result in a message like:

foo: Unable to open file because it doesn't exist."

The "foo:" part is the context. The "Unable to open file" part is the error message. The "because it doesn't exist." part is the reason. This message has no suggestion. Where possible, the implementation of lib/System should use operating system specific facilities for converting the error code returned by a system call into an error message. This will help to make the error message more familiar to users of that type of operating system.

Note that this requirement precludes the throwing of any other exceptions. For example, various C++ standard library functions can cause exceptions to be thrown (e.g. out of memory situation). In all cases, if there is a possibility that non-string exceptions could be thrown, the lib/System library must ensure that the exceptions are translated to std::string form.

None of the lib/System interface functions may be declared with C++ throw() specifications on them. This requirement makes sure that the compiler does not insert additional exception handling code into the interface functions. This is a performance consideration: lib/System functions are at the bottom of the many call chains and as such can be frequently called. We need them to be as efficient as possible.

The implementation of a function for a given platform must be written exactly once. This implies that it must be possible to apply a function's implementation to multiple operating systems if those operating systems can share the same implementation.

In order to fulfill the requirements of the system library, strict design objectives must be maintained in the library as it evolves. The goal here is to provide interfaces to operating system concepts (files, memory maps, sockets, signals, locking, etc) efficiently and in such a way that the remainder of LLVM is completely operating system agnostic.

There must be no functionality specified in the interface of lib/System that isn't actually used by LLVM. We're not writing a general purpose operating system wrapper here, just enough to satisfy LLVM's needs. And, LLVM doesn't need much. This design goal aims to keep the lib/System interface small and understandable which should foster its actual use and adoption.

The entry points specified in the interface of lib/System must be aimed at completing some reasonably high level task needed by LLVM. We do not want to simply wrap each operating system call. It would be preferable to wrap several operating system calls that are always used in conjunction with one another by LLVM.

For example, consider what is needed to execute a program, wait for it to complete, and return its result code. On Unix, this involves the following operating system calls: getenv, fork, execve, and wait. The correct thing for lib/System to provide is a function, say ExecuteProgramAndWait, that implements the functionality completely. what we don't want is wrappers for the operating system calls involved.

There must not be a one-to-one relationship between operating system calls and the System library's interface. Any such interface function will be suspicious.

Operating system interfaces will generally provide errors results for every little thing that could go wrong. In almost all cases, you can divide these error results into two groups: normal/good/soft and abnormal/bad/hard. That is, some of the errors are simply information like "file not found", "insufficient privileges", etc. while other errors are much harder like "out of space", "bad disk sector", or "system call interrupted". Well call the first group "soft" errors and the second group "hard" errors.

lib/System must always attempt to minimize soft errors and always just throw a std::string on hard errors. This is a design requirement because the minimization of soft errors can affect the granularity and the nature of the interface. In general, if you find that you're wanting to throw soft errors, you must review the granularity of the interface because it is likely you're trying to implement something that is too low level. The rule of thumb is to provide interface functions that "can't" fail, except when faced with hard errors.

For a trivial example, suppose we wanted to add an "OpenFileForWriting" function. For many operating systems, if the file doesn't exist, attempting to open the file will produce an error. However, lib/System should not simply throw that error if it occurs because its a soft error. The problem is that the interface function, OpenFileForWriting is too low level. It should be OpenOrCreateFileForWriting. In the case of the soft "doesn't exist" error, this function would just create it and then open it for writing.

This design principle needs to be maintained in lib/System because it avoids the propagation of soft error handling throughout the rest of LLVM. Hard errors will generally just cause a termination for an LLVM tool so don't be bashful about throwing them.

Rules of thumb:

Don't throw soft errors, only hard errors.
If you're tempted to throw a soft error, re-think the interface.
Handle internally the most common normal/good/soft error conditions so the rest of LLVM doesn't have to.


Notes:
10. The implementation of a lib/System interface can vary drastically between
    platforms. That's okay as long as the end result of the interface function is
    the same. For example, a function to create a directory is pretty straight
    forward on all operating system. System V IPC on the other hand isn't even
    supported on all platforms. Instead of "supporting" System V IPC, lib/System
    should provide an interface to the basic concept of inter-process 
    communications. The implementations might use System V IPC if that was
    available or named pipes, or whatever gets the job done effectively for a
    given operating system.

11. Implementations are separated first by the general class of operating system
    as provided by the configure script's $build variable. This variable is used
    to create a link from $BUILD_OBJ_ROOT/lib/System/platform to a directory in
    $BUILD_SRC_ROOT/lib/System directory with the same name as the $build
    variable. This provides a retargetable include mechanism. By using the link's
    name (platform) we can actually include the operating specific
    implementation. For example, support $build is "Darwin" for MacOS X. If we
    place:
      #include "platform/File.cpp"
    into a a file in lib/System, it will actually include
    lib/System/Darwin/File.cpp. What this does is quickly differentiate the basic
    class of operating system that will provide the implementation.
 
12. Implementation files in lib/System need may only do two things: (1) define 
    functions and data that is *TRULY* generic (completely platform agnostic) and
    (2) #include the platform specific implementation with:
 
       #include "platform/Impl.cpp"
 
    where Impl is the name of the implementation files.
 
13. Platform specific implementation files (platform/Impl.cpp) may only #include
    other Impl.cpp files found in directories under lib/System. The order of
    inclusion is very important (from most generic to most specific) so that we
    don't inadvertently place an implementation in the wrong place. For example,
    consider a fictitious implementation file named DoIt.cpp. Here's how the
    #includes should work for a Linux platform
 
    lib/System/DoIt.cpp
      #include "platform/DoIt.cpp"        // platform specific impl. of Doit
      DoIt
 
    lib/System/Linux/DoIt.cpp             // impl that works on all Linux 
      #include "../Unix/DoIt.cpp"         // generic Unix impl. of DoIt
      #include "../Unix/SUS/DoIt.cpp      // SUS specific impl. of DoIt
      #include "../Unix/SUS/v3/DoIt.cpp   // SUSv3 specific impl. of DoIt
 
    Note that the #includes in lib/System/Linux/DoIt.cpp are all optional but
    should be used where the implementation of some functionality can be shared
    across some set of Unix variants. We don't want to duplicate code across
    variants if their implementation could be shared.

no public data

onlyprimitive typed private/protected data

data size is "right" for platform, not max of all platforms

each class corresponds to O/S concept

To be written.

See bug 351 for further details on the progress of this work

In order to provide different implementations of the lib/System interface for different platforms, it is necessary for the library to "sense" which operating system is being compiled for and conditionally compile only the applicable parts of the library. While several operating system wrapper libraries (e.g. APR, ACE) choose to use #ifdef preprocessor statements in combination with autoconf variable (HAVE_* family), lib/System chooses an alternate strategy.

To put it succinctly, the lib/System strategy has traded "#ifdef hell" for "#include hell". That is, a given implementation file defines one or more functions for a particular operating system variant. The functions defined in that file have no #ifdef's to disambiguate the platform since the file is only compiled on one kind of platform. While this leads to the same function being implemented differently in different files, it is our contention that this leads to better maintenance and easier portability.

For example, consider a function having different implementations on a variety of platforms. Many wrapper libraries choose to deal with the different implementations by using #ifdef, like this:


      void SomeFunction(void) {
      #if defined __LINUX
        // .. Linux implementation
      #elif defined __WIN32
        // .. Win32 implementation
      #elif defined __SunOS
        // .. SunOS implementation
      #else
      #warning "Don't know how to implement SomeFunction on this platform"
      #endif
      }

The problem with this is that its very messy to read, especially as the number of operating systems and their variants grow. The above example is actually tame compared to what can happen when the implementation depends on specific flavors and versions of the operating system. In that case you end up with multiple levels of nested #if statements. This is what we mean by "#ifdef hell".

To avoid the situation above, we've chosen to locate all functions for a given implementation file for a specific operating system into one place. This has the following advantages:

No "#ifdef hell"
When porting, the strategy is quite straight forward: copy the implementation file from a similar operating system to a new directory and re-implement them.
Correctness is helped during porting because the new operating system's implementation is wholly contained in a separate directory. There's no chance to make an error in the #if statements and affect some other operating system's implementation.

So, given that we have decided to use #include instead of #if to provide platform specific implementations, there are actually three ways we can go about doing this. None of them are perfect, but we believe we've chosen the lesser of the three evils. Given that there is a variable named $OS which names the platform for which we must build, here's a summary of the three approaches we could use to determine the correct directory:

Provide the compiler with a -I$(OS) on the command line. This could be provided in only the lib/System makefile.
Use autoconf to transform #include statements in the implementation files by using substitutions of @OS@. For example, if we had a file, File.cpp.in, that contained "#include <@OS@/File.cpp>" this would get transformed to "#include <actual/File.cpp>" where "actual" is the actual name of the operating system
Create a link from $OBJ_DIR/platform to $SRC_DIR/$OS. This allows us to use a generic directory name to get the correct platform, as in #include <platform/File.cpp>

Let's look at the pitfalls of each approach.

In approach #1, we end up with some confusion as to what gets included. Suppose we have lib/System/File.cpp that includes just File.cpp to get the platform specific part of the implementation. In this case, the include directive with the <> syntax will include the right file but the include directive with the "" syntax will recursively include the same file, lib/System/File.cpp. In the case of #include <File.cpp>, the -I options to the compiler are searched first so it works. But in the #include "File.cpp" case, the current directory is searched first. Furthermore, in both cases, neither include directive documents which File.cpp is getting included.

In approach #2, we have the problem of needing to reconfigure repeatedly. Developer's generally hate that and we don't want lib/System to be a thorn in everyone's side because it will constantly need updating as operating systems change and as new operating systems are added. The problem occurs when a new implementation file is added to the library. First of all, you have to add a file with the .in suffix, then you have to add that file name to the list of configurable files in the autoconf/configure.ac file, then you have to run AutoRegen.sh to rebuild the configure script, then you have to run the configure script. This is deemed to be a pretty large hassle.

In approach #3, we have the problem that not all platforms support links. Fortunately the autoconf macro used to create the link can compensate for this. If a link can't be made, the configure script will copy the correct directory from $BUILD_SRC_DIR to $BUILD_OBJ_DIR under the new name. The only problem with this is that if a copy is made, the copy doesn't get updated if the programmer adds or modifies files in the $BUILD_SRC_DIR. A reconfigure or manual copying is needed to get things to compile.

The approach we have taken in lib/System is #3. Here's why:

Approach #1 is rejected because it doesn't document what's actually getting included and the potential for mistakes with alternate include directive forms is high.
Approach #2 are both viable and only really impact development when new files are added to the library.
However, approach #2 impacts every new file on every platform all the time. With approach #3, only those platforms not supporting links will be affected. The number of platforms not supporting links is very small and they are generally archaic.
Given the above, approach #3 seems to have the least impact.

The linux implementation of the system library will always be the reference implementation. This means that (a) the concepts defined by the linux must be identically replicated in the other implementations and (b) the linux implementation must always be complete (provide implementations for all concepts).