Why Apple Chose Clang (2008)


Building an entirely new compiler front-end is a big task, and it isn’t
always clear to people why we decided to do this. Here we compare clang
and its goals to other open source compiler front-ends that are
available. We restrict the discussion to very specific objective points
to avoid controversy where possible. Also, software is infinitely
mutable, so we don’t talk about little details that can be fixed with
a reasonable amount of effort: we’ll talk about issues that are
difficult to fix for architectural or political reasons.

The goal of this list is to describe how differences in goals lead to
different strengths and weaknesses, not to make some compiler look bad.
This will hopefully help you to evaluate whether using clang is a good
idea for your personal goals. Because we don’t know specifically what
you want to do, we describe the features of these compilers in
terms of our goals: if you are only interested in static
analysis, you may not care that something lacks codegen support, for

Please email cfe-dev if you think we should add another compiler to this
list or if you think some characterization is unfair here.

Clang vs GCC (GNU Compiler Collection)

Pro’s of GCC vs clang:

  • GCC supports languages that clang does not aim to, such as Java, Ada,
    FORTRAN, etc.
  • GCC front-ends are very mature and already support C++.
    clang’s support for C++ is nowhere near
    what GCC supports.
  • GCC supports more targets than LLVM.
  • GCC is popular and widely adopted.
  • GCC does not require a C++ compiler to build it.

Pro’s of clang vs GCC:

  • The Clang ASTs and design are intended to be easily understandable by
    anyone who is familiar with the languages involved and who has a basic
    understanding of how a compiler works. GCC has a very old codebase
    which presents a steep learning curve to new developers.
  • Clang is designed as an API from its inception, allowing it to be reused
    by source analysis tools, refactoring, IDEs (etc) as well as for code
    generation. GCC is built as a monolithic static compiler, which makes
    it extremely difficult to use as an API and integrate into other tools.
    Further, its historic design and current
    makes it difficult to decouple the front-end from the rest of the
  • Various GCC design decisions make it very difficult to reuse: its build
    system is difficult to modify, you can’t link multiple targets into one
    binary, you can’t link multiple front-ends into one binary, it uses a
    custom garbage collector, uses global variables extensively, is not
    reentrant or multi-threadable, etc. Clang has none of these problems.
  • For every token, clang tracks information about where it was written and
    where it was ultimately expanded into if it was involved in a macro.
    GCC does not track information about macro instantiations when parsing
    source code. This makes it very difficult for source rewriting tools
    (e.g. for refactoring) to work in the presence of (even simple)
  • Clang does not implicitly simplify code as it parses it like GCC does.
    Doing so causes many problems for source analysis tools: as one simple
    example, if you write “x-x” in your source code, the GCC AST will
    contain “0”, with no mention of ‘x’. This is extremely bad for a
    refactoring tool that wants to rename ‘x’.
  • Clang can serialize its AST out to disk and read it back into another
    program, which is useful for whole program analysis. GCC does not have
    this. GCC’s PCH mechanism (which is just a dump of the compiler
    memory image) is related, but is architecturally only
    able to read the dump back into the exact same executable as the one
    that produced it (it is not a structured format).
  • Clang is much faster and uses far
    less memory
    than GCC.
  • Clang aims to provide extremely clear and concise diagnostics (error and
    warning messages), and includes support for expressive diagnostics. GCC’s warnings are
    sometimes acceptable, but are often confusing and it does not support
    expressive diagnostics. Clang also preserves typedefs in diagnostics
    consistently, showing macro expansions and many other features.
  • GCC is licensed under the GPL license. clang uses a BSD license, which
    allows it to be used by projects that do not themselves want to be
  • Clang inherits a number of features from its use of LLVM as a backend,
    including support for a bytecode representation for intermediate code,
    pluggable optimizers, link-time optimization support, Just-In-Time
    compilation, ability to link in multiple code generators, etc.

Clang vs Elsa (Elkhound-based C++ Parser)

Pro’s of Elsa vs clang:

  • Elsa’s support for C++ is far beyond what clang provides. If you need
    C++ support in the next year, Elsa is a great way to get it. That said,
    Elsa is missing important support for templates and other pieces: for
    example, it is not capable of compiling the GCC STL headers from any
    version newer than GCC 3.4.
  • Elsa’s parser and AST is designed to be easily extensible by adding
    grammar rules. Clang has a very simple and easily hackable parser,
    but requires you to write C++ code to do it.

Pro’s of clang vs Elsa:

  • The Elsa community is extremely small and major development work seems
    to have ceased in 2005, though it continues to be used by other small
    (e.g. Oink). Clang has a vibrant community including developers that
    are paid to work on it full time. In practice this means that you can
    file bugs against Clang and they will often be fixed for you. If you
    use Elsa, you are (mostly) on your own for bug fixes and feature
  • Elsa is not built as a stack of reusable libraries like clang is. It is
    very difficult to use part of Elsa without the whole front-end. For
    example, you cannot use Elsa to parse C/ObjC code without building an
    AST. You can do this in Clang and it is much faster than building an
  • Elsa does not have an integrated preprocessor, which makes it extremely
    difficult to accurately map from a source location in the AST back to
    its original position before preprocessing. Like GCC, it does not keep
    track of macro expansions.
  • Elsa is even slower and uses more memory than GCC, which itself requires
    far more space and time than clang.
  • Elsa only does partial semantic analysis. It is intended to work on
    code that is already validated by GCC, so it does not do many semantic
    checks required by the languages it implements.
  • Elsa does not support Objective-C.
  • Elsa does not support native code generation.

Note that there is a fork of Elsa known as “Pork”. It addresses some of
these shortcomings by loosely integrating a preprocessor. This allows it
to map from a source location in the AST to the original position before
preprocessing, providing it better support for static analysis and
refactoring. Note that Pork is in stasis now too.

Clang vs PCC (Portable C Compiler)

Pro’s of PCC vs clang:

  • The PCC source base is very small and builds quickly with just a C

Pro’s of clang vs PCC:

  • PCC dates from the 1970’s and has been dormant for most of that time.
    The clang + llvm communities are very active.
  • PCC doesn’t support C99, Objective-C, and doesn’t aim to support
  • PCC’s code generation is very limited compared to LLVM. It produces very
    inefficient code and does not support many important targets.
  • Like Elsa, PCC’s does not have an integrated preprocessor, making it
    extremely difficult to use it for source analysis tools.


Please enter your comment!
Please enter your name here