11 This document attempts to describe a few coding standards that are being used in
12 the LLVM source tree. Although no coding standards should be regarded as
13 absolute requirements to be followed in all instances, coding standards are
14 particularly important for large-scale code bases that follow a library-based
17 While this document may provide guidance for some mechanical formatting issues,
18 whitespace, or other "microscopic details", these are not fixed standards.
19 Always follow the golden rule:
23 **If you are extending, enhancing, or bug fixing already implemented code,
24 use the style that is already being used so that the source is uniform and
27 Note that some code bases (e.g. ``libc++``) have really good reasons to deviate
28 from the coding standards. In the case of ``libc++``, this is because the
29 naming and other conventions are dictated by the C++ standard. If you think
30 there is a specific good reason to deviate from the standards here, please bring
31 it up on the LLVMdev mailing list.
33 There are some conventions that are not uniformly followed in the code base
34 (e.g. the naming convention). This is because they are relatively new, and a
35 lot of code was written before they were put in place. Our long term goal is
36 for the entire codebase to follow the convention, but we explicitly *do not*
37 want patches that do large-scale reformating of existing code. On the other
38 hand, it is reasonable to rename the methods of a class if you're about to
39 change it in some other way. Just do the reformating as a separate commit from
40 the functionality change.
42 The ultimate goal of these guidelines is the increase readability and
43 maintainability of our common source base. If you have suggestions for topics to
44 be included, please mail them to `Chris <mailto:sabre@nondot.org>`_.
46 Languages, Libraries, and Standards
47 ===================================
49 Most source code in LLVM and other LLVM projects using these coding standards
50 is C++ code. There are some places where C code is used either due to
51 environment restrictions, historical restrictions, or due to third-party source
52 code imported into the tree. Generally, our preference is for standards
53 conforming, modern, and portable C++ code as the implementation language of
59 LLVM and Clang are currently written using C++98/03 conforming code, with
60 selective use of C++11 features when they are present in the toolchain.
61 Projects like LLD and LLDB are already heavily using C++11 features.
63 However, LLVM and Clange are also in the process of switching to use C++11 as
64 the base line for standards conformance. Once completed, the same standard
65 baseline will be used for LLVM, Clang, and LLD. LLDB is pushing forward much
66 more aggressively and has their own baseline.
71 Use the C++ standard library facilities whenever they are available for
72 a particular task. LLVM and related projects emphasize and rely on the standard
73 library facilities for as much as possible. Common support libraries providing
74 functionality missing from the standard library for which there are standard
75 interfaces or active work on adding standard interfaces will often be
76 implemented in the LLVM namespace following the expected standard interface.
78 There are some exceptions such as the standard I/O streams library which are
79 avoided. Also, there is much more detailed information on these subjects in the
80 `Programmer's Manual`_.
82 .. _Programmer's Manual:
83 http://llvm.org/docs/ProgrammersManual.html
85 Supported C++11 Language and Library Features
86 -------------------------------------------
89 This section is written to reflect the expected state **AFTER** the
90 transition to C++11 is complete for the LLVM source tree.
92 While LLVM, Clang, and LLD use C++11, not all features are available in all of
93 the toolchains which we support. The set of features supported for use in LLVM
94 is the intersection of those supported in MSVC 2012, GCC 4.7, and Clang 3.1.
95 The ultimate definition of this set is what build bots with those respective
96 toolchains accept. Don't argue with the build bots.
98 Each toolchain provides a good reference for what it accepts:
99 * Clang: http://clang.llvm.org/cxx_status.html
100 * GCC: http://gcc.gnu.org/projects/cxx0x.html
101 * MSVC: http://msdn.microsoft.com/en-us/library/hh567368.aspx
103 In most cases, the MSVC list will be the dominating factor. Here is a summary
104 of the features that are expected to work. Features not on this list are
105 unlikely to be supported by our host compilers.
107 * Rvalue references: N2118_
108 * But *not* Rvalue references for ``*this`` or member qualifiers (N2439_)
109 * Static assert: N1720_
110 * ``auto`` type deduction: N1984_, N1737_
111 * Trailing return types: N2541_
113 * ``decltype``: N2343_
114 * Nested closing right angle brackets: N1757_
115 * Extern templates: N1987_
116 * ``nullptr``: N2431_
117 * Strongly-typed and forward declarable enums: N2347_, N2764_
118 * Local and unnamed types as template arguments: N2657_
119 * Range-based for-loop: N2930_
120 * ``override`` and ``final``: N2928_, N3206_, N3272_
121 * Atomic operations and the C++11 memory model: N2429_
123 .. _N2118: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n2118.html
124 .. _N2439: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2439.htm
125 .. _N1720: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1720.html
126 .. _N1984: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n1984.pdf
127 .. _N1737: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1737.pdf
128 .. _N2541: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2541.htm
129 .. _N2927: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2927.pdf
130 .. _N2343: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2343.pdf
131 .. _N1757: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1757.html
132 .. _N1987: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n1987.htm
133 .. _N2431: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2431.pdf
134 .. _N2347: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2347.pdf
135 .. _N2764: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2764.pdf
136 .. _N2657: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm
137 .. _N2930: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2930.html
138 .. _N2928: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2928.htm
139 .. _N3206: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2010/n3206.htm
140 .. _N3272: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3272.htm
141 .. _N2429: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2429.htm
143 The supported features in the C++11 standard libraries are less well tracked,
144 but also much greater. Most of the standard libraries implement most of C++11's
145 library. The most likely lowest common denominator is Linux support. For
146 libc++, the support is just poorly tested and undocumented but expected to be
147 largely complete. YMMV. For libstdc++, the support is documented in detail in
148 `the libstdc++ manual`_. There are some very minor missing facilities that are
149 unlikely to be common problems, and there are a few larger gaps that are worth
152 * Not all of the type traits are implemented
153 * No regular expression library.
154 * While most of the atomics library is well implemented, the fences are
155 missing. Fortunately, they are rarely needed.
156 * The locale support is incomplete.
158 .. _the libstdc++ manual:
159 http://gcc.gnu.org/onlinedocs/gcc-4.7.3/libstdc++/manual/manual/status.html#status.iso.2011
161 Mechanical Source Issues
162 ========================
164 Source Code Formatting
165 ----------------------
170 Comments are one critical part of readability and maintainability. Everyone
171 knows they should comment their code, and so should you. When writing comments,
172 write them as English prose, which means they should use proper capitalization,
173 punctuation, etc. Aim to describe what the code is trying to do and why, not
174 *how* it does it at a micro level. Here are a few critical things to document:
176 .. _header file comment:
181 Every source file should have a header on it that describes the basic purpose of
182 the file. If a file does not have a header, it should not be checked into the
183 tree. The standard header looks like this:
187 //===-- llvm/Instruction.h - Instruction class definition -------*- C++ -*-===//
189 // The LLVM Compiler Infrastructure
191 // This file is distributed under the University of Illinois Open Source
192 // License. See LICENSE.TXT for details.
194 //===----------------------------------------------------------------------===//
197 /// \brief This file contains the declaration of the Instruction class, which is
198 /// the base class for all of the VM instructions.
200 //===----------------------------------------------------------------------===//
202 A few things to note about this particular format: The "``-*- C++ -*-``" string
203 on the first line is there to tell Emacs that the source file is a C++ file, not
204 a C file (Emacs assumes ``.h`` files are C files by default).
208 This tag is not necessary in ``.cpp`` files. The name of the file is also
209 on the first line, along with a very short description of the purpose of the
210 file. This is important when printing out code and flipping though lots of
213 The next section in the file is a concise note that defines the license that the
214 file is released under. This makes it perfectly clear what terms the source
215 code can be distributed under and should not be modified in any way.
217 The main body is a ``doxygen`` comment describing the purpose of the file. It
218 should have a ``\brief`` command that describes the file in one or two
219 sentences. Any additional information should be separated by a blank line. If
220 an algorithm is being implemented or something tricky is going on, a reference
221 to the paper where it is published should be included, as well as any notes or
222 *gotchas* in the code to watch out for.
227 Classes are one fundamental part of a good object oriented design. As such, a
228 class definition should have a comment block that explains what the class is
229 used for and how it works. Every non-trivial class is expected to have a
230 ``doxygen`` comment block.
235 Methods defined in a class (as well as any global functions) should also be
236 documented properly. A quick note about what it does and a description of the
237 borderline behaviour is all that is necessary here (unless something
238 particularly tricky or insidious is going on). The hope is that people can
239 figure out how to use your interfaces without reading the code itself.
241 Good things to talk about here are what happens when something unexpected
242 happens: does the method return null? Abort? Format your hard disk?
247 In general, prefer C++ style (``//``) comments. They take less space, require
248 less typing, don't have nesting problems, etc. There are a few cases when it is
249 useful to use C style (``/* */``) comments however:
251 #. When writing C code: Obviously if you are writing C code, use C style
254 #. When writing a header file that may be ``#include``\d by a C source file.
256 #. When writing a source file that is used by a tool that only accepts C style
259 To comment out a large block of code, use ``#if 0`` and ``#endif``. These nest
260 properly and are better behaved in general than C style comments.
262 Doxygen Use in Documentation Comments
263 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
265 Use the ``\file`` command to turn the standard file header into a file-level
268 Include descriptive ``\brief`` paragraphs for all public interfaces (public
269 classes, member and non-member functions). Explain API use and purpose in
270 ``\brief`` paragraphs, don't just restate the information that can be inferred
271 from the API name. Put detailed discussion into separate paragraphs.
273 To refer to parameter names inside a paragraph, use the ``\p name`` command.
274 Don't use the ``\arg name`` command since it starts a new paragraph that
275 contains documentation for the parameter.
277 Wrap non-inline code examples in ``\code ... \endcode``.
279 To document a function parameter, start a new paragraph with the
280 ``\param name`` command. If the parameter is used as an out or an in/out
281 parameter, use the ``\param [out] name`` or ``\param [in,out] name`` command,
284 To describe function return value, start a new paragraph with the ``\returns``
287 A minimal documentation comment:
291 /// \brief Does foo and bar.
292 void fooBar(bool Baz);
294 A documentation comment that uses all Doxygen features in a preferred way:
298 /// \brief Does foo and bar.
300 /// Does not do foo the usual way if \p Baz is true.
304 /// fooBar(false, "quux", Res);
307 /// \param Quux kind of foo to do.
308 /// \param [out] Result filled with bar sequence on foo success.
310 /// \returns true on success.
311 bool fooBar(bool Baz, StringRef Quux, std::vector<int> &Result);
313 Don't duplicate the documentation comment in the header file and in the
314 implementation file. Put the documentation comments for public APIs into the
315 header file. Documentation comments for private APIs can go to the
316 implementation file. In any case, implementation files can include additional
317 comments (not necessarily in Doxygen markup) to explain implementation details
320 Don't duplicate function or class name at the beginning of the comment.
321 For humans it is obvious which function or class is being documented;
322 automatic documentation processing tools are smart enough to bind the comment
323 to the correct declaration.
331 /// Something - An abstraction for some complicated thing.
334 /// fooBar - Does foo and bar.
340 /// fooBar - Does foo and bar.
341 void Something::fooBar() { ... }
349 /// \brief An abstraction for some complicated thing.
352 /// \brief Does foo and bar.
358 // Builds a B-tree in order to do foo. See paper by...
359 void Something::fooBar() { ... }
361 It is not required to use additional Doxygen features, but sometimes it might
362 be a good idea to do so.
366 * adding comments to any narrow namespace containing a collection of
367 related functions or types;
369 * using top-level groups to organize a collection of related functions at
370 namespace scope where the grouping is smaller than the namespace;
372 * using member groups and additional comments attached to member
373 groups to organize within a class.
380 /// \name Functions that do Foo.
391 Immediately after the `header file comment`_ (and include guards if working on a
392 header file), the `minimal list of #includes`_ required by the file should be
393 listed. We prefer these ``#include``\s to be listed in this order:
395 .. _Main Module Header:
396 .. _Local/Private Headers:
398 #. Main Module Header
399 #. Local/Private Headers
401 #. System ``#include``\s
403 and each category should be sorted lexicographically by the full path.
405 The `Main Module Header`_ file applies to ``.cpp`` files which implement an
406 interface defined by a ``.h`` file. This ``#include`` should always be included
407 **first** regardless of where it lives on the file system. By including a
408 header file first in the ``.cpp`` files that implement the interfaces, we ensure
409 that the header does not have any hidden dependencies which are not explicitly
410 ``#include``\d in the header, but should be. It is also a form of documentation
411 in the ``.cpp`` file to indicate where the interfaces it implements are defined.
413 .. _fit into 80 columns:
418 Write your code to fit within 80 columns of text. This helps those of us who
419 like to print out code and look at your code in an ``xterm`` without resizing
422 The longer answer is that there must be some limit to the width of the code in
423 order to reasonably allow developers to have multiple files side-by-side in
424 windows on a modest display. If you are going to pick a width limit, it is
425 somewhat arbitrary but you might as well pick something standard. Going with 90
426 columns (for example) instead of 80 columns wouldn't add any significant value
427 and would be detrimental to printing out code. Also many other projects have
428 standardized on 80 columns, so some people have already configured their editors
429 for it (vs something else, like 90 columns).
431 This is one of many contentious issues in coding standards, but it is not up for
434 Use Spaces Instead of Tabs
435 ^^^^^^^^^^^^^^^^^^^^^^^^^^
437 In all cases, prefer spaces to tabs in source files. People have different
438 preferred indentation levels, and different styles of indentation that they
439 like; this is fine. What isn't fine is that different editors/viewers expand
440 tabs out to different tab stops. This can cause your code to look completely
441 unreadable, and it is not worth dealing with.
443 As always, follow the `Golden Rule`_ above: follow the style of
444 existing code if you are modifying and extending it. If you like four spaces of
445 indentation, **DO NOT** do that in the middle of a chunk of code with two spaces
446 of indentation. Also, do not reindent a whole source file: it makes for
447 incredible diffs that are absolutely worthless.
449 Indent Code Consistently
450 ^^^^^^^^^^^^^^^^^^^^^^^^
452 Okay, in your first year of programming you were told that indentation is
453 important. If you didn't believe and internalize this then, now is the time.
459 Treat Compiler Warnings Like Errors
460 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
462 If your code has compiler warnings in it, something is wrong --- you aren't
463 casting values correctly, you have "questionable" constructs in your code, or
464 you are doing something legitimately wrong. Compiler warnings can cover up
465 legitimate errors in output and make dealing with a translation unit difficult.
467 It is not possible to prevent all warnings from all compilers, nor is it
468 desirable. Instead, pick a standard compiler (like ``gcc``) that provides a
469 good thorough set of warnings, and stick to it. At least in the case of
470 ``gcc``, it is possible to work around any spurious errors by changing the
471 syntax of the code slightly. For example, a warning that annoys me occurs when
472 I write code like this:
476 if (V = getValue()) {
480 ``gcc`` will warn me that I probably want to use the ``==`` operator, and that I
481 probably mistyped it. In most cases, I haven't, and I really don't want the
482 spurious errors. To fix this particular problem, I rewrite the code like
487 if ((V = getValue())) {
491 which shuts ``gcc`` up. Any ``gcc`` warning that annoys you can be fixed by
492 massaging the code appropriately.
497 In almost all cases, it is possible and within reason to write completely
498 portable code. If there are cases where it isn't possible to write portable
499 code, isolate it behind a well defined (and well documented) interface.
501 In practice, this means that you shouldn't assume much about the host compiler
502 (and Visual Studio tends to be the lowest common denominator). If advanced
503 features are used, they should only be an implementation detail of a library
504 which has a simple exposed API, and preferably be buried in ``libSystem``.
506 Do not use RTTI or Exceptions
507 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
509 In an effort to reduce code and executable size, LLVM does not use RTTI
510 (e.g. ``dynamic_cast<>;``) or exceptions. These two language features violate
511 the general C++ principle of *"you only pay for what you use"*, causing
512 executable bloat even if exceptions are never used in the code base, or if RTTI
513 is never used for a class. Because of this, we turn them off globally in the
516 That said, LLVM does make extensive use of a hand-rolled form of RTTI that use
517 templates like `isa<>, cast<>, and dyn_cast<> <ProgrammersManual.html#isa>`_.
518 This form of RTTI is opt-in and can be
519 :doc:`added to any class <HowToSetUpLLVMStyleRTTI>`. It is also
520 substantially more efficient than ``dynamic_cast<>``.
522 .. _static constructor:
524 Do not use Static Constructors
525 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
527 Static constructors and destructors (e.g. global variables whose types have a
528 constructor or destructor) should not be added to the code base, and should be
529 removed wherever possible. Besides `well known problems
530 <http://yosefk.com/c++fqa/ctors.html#fqa-10.12>`_ where the order of
531 initialization is undefined between globals in different source files, the
532 entire concept of static constructors is at odds with the common use case of
533 LLVM as a library linked into a larger application.
535 Consider the use of LLVM as a JIT linked into another application (perhaps for
536 `OpenGL, custom languages <http://llvm.org/Users.html>`_, `shaders in movies
537 <http://llvm.org/devmtg/2010-11/Gritz-OpenShadingLang.pdf>`_, etc). Due to the
538 design of static constructors, they must be executed at startup time of the
539 entire application, regardless of whether or how LLVM is used in that larger
540 application. There are two problems with this:
542 * The time to run the static constructors impacts startup time of applications
543 --- a critical time for GUI apps, among others.
545 * The static constructors cause the app to pull many extra pages of memory off
546 the disk: both the code for the constructor in each ``.o`` file and the small
547 amount of data that gets touched. In addition, touched/dirty pages put more
548 pressure on the VM system on low-memory machines.
550 We would really like for there to be zero cost for linking in an additional LLVM
551 target or other library into an application, but static constructors violate
554 That said, LLVM unfortunately does contain static constructors. It would be a
555 `great project <http://llvm.org/PR11944>`_ for someone to purge all static
556 constructors from LLVM, and then enable the ``-Wglobal-constructors`` warning
557 flag (when building with Clang) to ensure we do not regress in the future.
559 Use of ``class`` and ``struct`` Keywords
560 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
562 In C++, the ``class`` and ``struct`` keywords can be used almost
563 interchangeably. The only difference is when they are used to declare a class:
564 ``class`` makes all members private by default while ``struct`` makes all
565 members public by default.
567 Unfortunately, not all compilers follow the rules and some will generate
568 different symbols based on whether ``class`` or ``struct`` was used to declare
569 the symbol. This can lead to problems at link time.
571 So, the rule for LLVM is to always use the ``class`` keyword, unless **all**
572 members are public and the type is a C++ `POD
573 <http://en.wikipedia.org/wiki/Plain_old_data_structure>`_ type, in which case
574 ``struct`` is allowed.
579 The High-Level Issues
580 ---------------------
582 A Public Header File **is** a Module
583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
585 C++ doesn't do too well in the modularity department. There is no real
586 encapsulation or data hiding (unless you use expensive protocol classes), but it
587 is what we have to work with. When you write a public header file (in the LLVM
588 source tree, they live in the top level "``include``" directory), you are
589 defining a module of functionality.
591 Ideally, modules should be completely independent of each other, and their
592 header files should only ``#include`` the absolute minimum number of headers
593 possible. A module is not just a class, a function, or a namespace: it's a
594 collection of these that defines an interface. This interface may be several
595 functions, classes, or data structures, but the important issue is how they work
598 In general, a module should be implemented by one or more ``.cpp`` files. Each
599 of these ``.cpp`` files should include the header that defines their interface
600 first. This ensures that all of the dependences of the module header have been
601 properly added to the module header itself, and are not implicit. System
602 headers should be included after user headers for a translation unit.
604 .. _minimal list of #includes:
606 ``#include`` as Little as Possible
607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
609 ``#include`` hurts compile time performance. Don't do it unless you have to,
610 especially in header files.
612 But wait! Sometimes you need to have the definition of a class to use it, or to
613 inherit from it. In these cases go ahead and ``#include`` that header file. Be
614 aware however that there are many cases where you don't need to have the full
615 definition of a class. If you are using a pointer or reference to a class, you
616 don't need the header file. If you are simply returning a class instance from a
617 prototyped function or method, you don't need it. In fact, for most cases, you
618 simply don't need the definition of a class. And not ``#include``\ing speeds up
621 It is easy to try to go too overboard on this recommendation, however. You
622 **must** include all of the header files that you are using --- you can include
623 them either directly or indirectly through another header file. To make sure
624 that you don't accidentally forget to include a header file in your module
625 header, make sure to include your module header **first** in the implementation
626 file (as mentioned above). This way there won't be any hidden dependencies that
627 you'll find out about later.
629 Keep "Internal" Headers Private
630 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
632 Many modules have a complex implementation that causes them to use more than one
633 implementation (``.cpp``) file. It is often tempting to put the internal
634 communication interface (helper classes, extra functions, etc) in the public
635 module header file. Don't do this!
637 If you really need to do something like this, put a private header file in the
638 same directory as the source files, and include it locally. This ensures that
639 your private interface remains private and undisturbed by outsiders.
643 It's okay to put extra implementation methods in a public class itself. Just
644 make them private (or protected) and all is well.
648 Use Early Exits and ``continue`` to Simplify Code
649 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
651 When reading code, keep in mind how much state and how many previous decisions
652 have to be remembered by the reader to understand a block of code. Aim to
653 reduce indentation where possible when it doesn't make it more difficult to
654 understand the code. One great way to do this is by making use of early exits
655 and the ``continue`` keyword in long loops. As an example of using an early
656 exit from a function, consider this "bad" code:
660 Value *doSomething(Instruction *I) {
661 if (!isa<TerminatorInst>(I) &&
662 I->hasOneUse() && doOtherThing(I)) {
663 ... some long code ....
669 This code has several problems if the body of the ``'if'`` is large. When
670 you're looking at the top of the function, it isn't immediately clear that this
671 *only* does interesting things with non-terminator instructions, and only
672 applies to things with the other predicates. Second, it is relatively difficult
673 to describe (in comments) why these predicates are important because the ``if``
674 statement makes it difficult to lay out the comments. Third, when you're deep
675 within the body of the code, it is indented an extra level. Finally, when
676 reading the top of the function, it isn't clear what the result is if the
677 predicate isn't true; you have to read to the end of the function to know that
680 It is much preferred to format the code like this:
684 Value *doSomething(Instruction *I) {
685 // Terminators never need 'something' done to them because ...
686 if (isa<TerminatorInst>(I))
689 // We conservatively avoid transforming instructions with multiple uses
690 // because goats like cheese.
694 // This is really just here for example.
695 if (!doOtherThing(I))
698 ... some long code ....
701 This fixes these problems. A similar problem frequently happens in ``for``
702 loops. A silly example is something like this:
706 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
707 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(II)) {
708 Value *LHS = BO->getOperand(0);
709 Value *RHS = BO->getOperand(1);
716 When you have very, very small loops, this sort of structure is fine. But if it
717 exceeds more than 10-15 lines, it becomes difficult for people to read and
718 understand at a glance. The problem with this sort of code is that it gets very
719 nested very quickly. Meaning that the reader of the code has to keep a lot of
720 context in their brain to remember what is going immediately on in the loop,
721 because they don't know if/when the ``if`` conditions will have ``else``\s etc.
722 It is strongly preferred to structure the loop like this:
726 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
727 BinaryOperator *BO = dyn_cast<BinaryOperator>(II);
730 Value *LHS = BO->getOperand(0);
731 Value *RHS = BO->getOperand(1);
732 if (LHS == RHS) continue;
737 This has all the benefits of using early exits for functions: it reduces nesting
738 of the loop, it makes it easier to describe why the conditions are true, and it
739 makes it obvious to the reader that there is no ``else`` coming up that they
740 have to push context into their brain for. If a loop is large, this can be a
741 big understandability win.
743 Don't use ``else`` after a ``return``
744 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
746 For similar reasons above (reduction of indentation and easier reading), please
747 do not use ``'else'`` or ``'else if'`` after something that interrupts control
748 flow --- like ``return``, ``break``, ``continue``, ``goto``, etc. For
749 example, this is *bad*:
755 Type = Context.getsigjmp_bufType();
757 Error = ASTContext::GE_Missing_sigjmp_buf;
763 Type = Context.getjmp_bufType();
765 Error = ASTContext::GE_Missing_jmp_buf;
773 It is better to write it like this:
779 Type = Context.getsigjmp_bufType();
781 Error = ASTContext::GE_Missing_sigjmp_buf;
785 Type = Context.getjmp_bufType();
787 Error = ASTContext::GE_Missing_jmp_buf;
793 Or better yet (in this case) as:
799 Type = Context.getsigjmp_bufType();
801 Type = Context.getjmp_bufType();
804 Error = Signed ? ASTContext::GE_Missing_sigjmp_buf :
805 ASTContext::GE_Missing_jmp_buf;
810 The idea is to reduce indentation and the amount of code you have to keep track
811 of when reading the code.
813 Turn Predicate Loops into Predicate Functions
814 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
816 It is very common to write small loops that just compute a boolean value. There
817 are a number of ways that people commonly write these, but an example of this
822 bool FoundFoo = false;
823 for (unsigned I = 0, E = BarList.size(); I != E; ++I)
824 if (BarList[I]->isFoo()) {
833 This sort of code is awkward to write, and is almost always a bad sign. Instead
834 of this sort of loop, we strongly prefer to use a predicate function (which may
835 be `static`_) that uses `early exits`_ to compute the predicate. We prefer the
836 code to be structured like this:
840 /// \returns true if the specified list has an element that is a foo.
841 static bool containsFoo(const std::vector<Bar*> &List) {
842 for (unsigned I = 0, E = List.size(); I != E; ++I)
843 if (List[I]->isFoo())
849 if (containsFoo(BarList)) {
853 There are many reasons for doing this: it reduces indentation and factors out
854 code which can often be shared by other code that checks for the same predicate.
855 More importantly, it *forces you to pick a name* for the function, and forces
856 you to write a comment for it. In this silly example, this doesn't add much
857 value. However, if the condition is complex, this can make it a lot easier for
858 the reader to understand the code that queries for this predicate. Instead of
859 being faced with the in-line details of how we check to see if the BarList
860 contains a foo, we can trust the function name and continue reading with better
866 Name Types, Functions, Variables, and Enumerators Properly
867 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
869 Poorly-chosen names can mislead the reader and cause bugs. We cannot stress
870 enough how important it is to use *descriptive* names. Pick names that match
871 the semantics and role of the underlying entities, within reason. Avoid
872 abbreviations unless they are well known. After picking a good name, make sure
873 to use consistent capitalization for the name, as inconsistency requires clients
874 to either memorize the APIs or to look it up to find the exact spelling.
876 In general, names should be in camel case (e.g. ``TextFileReader`` and
877 ``isLValue()``). Different kinds of declarations have different rules:
879 * **Type names** (including classes, structs, enums, typedefs, etc) should be
880 nouns and start with an upper-case letter (e.g. ``TextFileReader``).
882 * **Variable names** should be nouns (as they represent state). The name should
883 be camel case, and start with an upper case letter (e.g. ``Leader`` or
886 * **Function names** should be verb phrases (as they represent actions), and
887 command-like function should be imperative. The name should be camel case,
888 and start with a lower case letter (e.g. ``openFile()`` or ``isFoo()``).
890 * **Enum declarations** (e.g. ``enum Foo {...}``) are types, so they should
891 follow the naming conventions for types. A common use for enums is as a
892 discriminator for a union, or an indicator of a subclass. When an enum is
893 used for something like this, it should have a ``Kind`` suffix
894 (e.g. ``ValueKind``).
896 * **Enumerators** (e.g. ``enum { Foo, Bar }``) and **public member variables**
897 should start with an upper-case letter, just like types. Unless the
898 enumerators are defined in their own small namespace or inside a class,
899 enumerators should have a prefix corresponding to the enum declaration name.
900 For example, ``enum ValueKind { ... };`` may contain enumerators like
901 ``VK_Argument``, ``VK_BasicBlock``, etc. Enumerators that are just
902 convenience constants are exempt from the requirement for a prefix. For
912 As an exception, classes that mimic STL classes can have member names in STL's
913 style of lower-case words separated by underscores (e.g. ``begin()``,
914 ``push_back()``, and ``empty()``). Classes that provide multiple
915 iterators should add a singular prefix to ``begin()`` and ``end()``
916 (e.g. ``global_begin()`` and ``use_begin()``).
918 Here are some examples of good and bad names:
924 Factory<Tire> F; // Bad -- abbreviation and non-descriptive.
925 Factory<Tire> Factory; // Better.
926 Factory<Tire> TireFactory; // Even better -- if VehicleMaker has more than one
927 // kind of factories.
930 Vehicle MakeVehicle(VehicleType Type) {
931 VehicleMaker M; // Might be OK if having a short life-span.
932 Tire Tmp1 = M.makeTire(); // Bad -- 'Tmp1' provides no information.
933 Light Headlight = M.makeLight("head"); // Good -- descriptive.
940 Use the "``assert``" macro to its fullest. Check all of your preconditions and
941 assumptions, you never know when a bug (not necessarily even yours) might be
942 caught early by an assertion, which reduces debugging time dramatically. The
943 "``<cassert>``" header file is probably already included by the header files you
944 are using, so it doesn't cost anything to use it.
946 To further assist with debugging, make sure to put some kind of error message in
947 the assertion statement, which is printed if the assertion is tripped. This
948 helps the poor debugger make sense of why an assertion is being made and
949 enforced, and hopefully what to do about it. Here is one complete example:
953 inline Value *getOperand(unsigned I) {
954 assert(I < Operands.size() && "getOperand() out of range!");
958 Here are more examples:
962 assert(Ty->isPointerType() && "Can't allocate a non-pointer type!");
964 assert((Opcode == Shl || Opcode == Shr) && "ShiftInst Opcode invalid!");
966 assert(idx < getNumSuccessors() && "Successor # out of range!");
968 assert(V1.getType() == V2.getType() && "Constant types must be identical!");
970 assert(isa<PHINode>(Succ->front()) && "Only works on PHId BBs!");
974 In the past, asserts were used to indicate a piece of code that should not be
975 reached. These were typically of the form:
979 assert(0 && "Invalid radix for integer literal");
981 This has a few issues, the main one being that some compilers might not
982 understand the assertion, or warn about a missing return in builds where
983 assertions are compiled out.
985 Today, we have something much better: ``llvm_unreachable``:
989 llvm_unreachable("Invalid radix for integer literal");
991 When assertions are enabled, this will print the message if it's ever reached
992 and then exit the program. When assertions are disabled (i.e. in release
993 builds), ``llvm_unreachable`` becomes a hint to compilers to skip generating
994 code for this branch. If the compiler does not support this, it will fall back
995 to the "abort" implementation.
997 Another issue is that values used only by assertions will produce an "unused
998 value" warning when assertions are disabled. For example, this code will warn:
1002 unsigned Size = V.size();
1003 assert(Size > 42 && "Vector smaller than it should be");
1005 bool NewToSet = Myset.insert(Value);
1006 assert(NewToSet && "The value shouldn't be in the set yet");
1008 These are two interesting different cases. In the first case, the call to
1009 ``V.size()`` is only useful for the assert, and we don't want it executed when
1010 assertions are disabled. Code like this should move the call into the assert
1011 itself. In the second case, the side effects of the call must happen whether
1012 the assert is enabled or not. In this case, the value should be cast to void to
1013 disable the warning. To be specific, it is preferred to write the code like
1018 assert(V.size() > 42 && "Vector smaller than it should be");
1020 bool NewToSet = Myset.insert(Value); (void)NewToSet;
1021 assert(NewToSet && "The value shouldn't be in the set yet");
1023 Do Not Use ``using namespace std``
1024 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1026 In LLVM, we prefer to explicitly prefix all identifiers from the standard
1027 namespace with an "``std::``" prefix, rather than rely on "``using namespace
1030 In header files, adding a ``'using namespace XXX'`` directive pollutes the
1031 namespace of any source file that ``#include``\s the header. This is clearly a
1034 In implementation files (e.g. ``.cpp`` files), the rule is more of a stylistic
1035 rule, but is still important. Basically, using explicit namespace prefixes
1036 makes the code **clearer**, because it is immediately obvious what facilities
1037 are being used and where they are coming from. And **more portable**, because
1038 namespace clashes cannot occur between LLVM code and other namespaces. The
1039 portability rule is important because different standard library implementations
1040 expose different symbols (potentially ones they shouldn't), and future revisions
1041 to the C++ standard will add more symbols to the ``std`` namespace. As such, we
1042 never use ``'using namespace std;'`` in LLVM.
1044 The exception to the general rule (i.e. it's not an exception for the ``std``
1045 namespace) is for implementation files. For example, all of the code in the
1046 LLVM project implements code that lives in the 'llvm' namespace. As such, it is
1047 ok, and actually clearer, for the ``.cpp`` files to have a ``'using namespace
1048 llvm;'`` directive at the top, after the ``#include``\s. This reduces
1049 indentation in the body of the file for source editors that indent based on
1050 braces, and keeps the conceptual context cleaner. The general form of this rule
1051 is that any ``.cpp`` file that implements code in any namespace may use that
1052 namespace (and its parents'), but should not use any others.
1054 Provide a Virtual Method Anchor for Classes in Headers
1055 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1057 If a class is defined in a header file and has a vtable (either it has virtual
1058 methods or it derives from classes with virtual methods), it must always have at
1059 least one out-of-line virtual method in the class. Without this, the compiler
1060 will copy the vtable and RTTI into every ``.o`` file that ``#include``\s the
1061 header, bloating ``.o`` file sizes and increasing link times.
1063 Don't use default labels in fully covered switches over enumerations
1064 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1066 ``-Wswitch`` warns if a switch, without a default label, over an enumeration
1067 does not cover every enumeration value. If you write a default label on a fully
1068 covered switch over an enumeration then the ``-Wswitch`` warning won't fire
1069 when new elements are added to that enumeration. To help avoid adding these
1070 kinds of defaults, Clang has the warning ``-Wcovered-switch-default`` which is
1071 off by default but turned on when building LLVM with a version of Clang that
1072 supports the warning.
1074 A knock-on effect of this stylistic requirement is that when building LLVM with
1075 GCC you may get warnings related to "control may reach end of non-void function"
1076 if you return from each case of a covered switch-over-enum because GCC assumes
1077 that the enum expression may take any representable value, not just those of
1078 individual enumerators. To suppress this warning, use ``llvm_unreachable`` after
1081 Use ``LLVM_DELETED_FUNCTION`` to mark uncallable methods
1082 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1084 Prior to C++11, a common pattern to make a class uncopyable was to declare an
1085 unimplemented copy constructor and copy assignment operator and make them
1086 private. This would give a compiler error for accessing a private method or a
1087 linker error because it wasn't implemented.
1089 With C++11, we can mark methods that won't be implemented with ``= delete``.
1090 This will trigger a much better error message and tell the compiler that the
1091 method will never be implemented. This enables other checks like
1092 ``-Wunused-private-field`` to run correctly on classes that contain these
1095 To maintain compatibility with C++03, ``LLVM_DELETED_FUNCTION`` should be used
1096 which will expand to ``= delete`` if the compiler supports it. These methods
1097 should still be declared private. Example of the uncopyable pattern:
1103 DontCopy(const DontCopy&) LLVM_DELETED_FUNCTION;
1104 DontCopy &operator =(const DontCopy&) LLVM_DELETED_FUNCTION;
1109 Don't evaluate ``end()`` every time through a loop
1110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1112 Because C++ doesn't have a standard "``foreach``" loop (though it can be
1113 emulated with macros and may be coming in C++'0x) we end up writing a lot of
1114 loops that manually iterate from begin to end on a variety of containers or
1115 through other data structures. One common mistake is to write a loop in this
1120 BasicBlock *BB = ...
1121 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I)
1124 The problem with this construct is that it evaluates "``BB->end()``" every time
1125 through the loop. Instead of writing the loop like this, we strongly prefer
1126 loops to be written so that they evaluate it once before the loop starts. A
1127 convenient way to do this is like so:
1131 BasicBlock *BB = ...
1132 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1135 The observant may quickly point out that these two loops may have different
1136 semantics: if the container (a basic block in this case) is being mutated, then
1137 "``BB->end()``" may change its value every time through the loop and the second
1138 loop may not in fact be correct. If you actually do depend on this behavior,
1139 please write the loop in the first form and add a comment indicating that you
1140 did it intentionally.
1142 Why do we prefer the second form (when correct)? Writing the loop in the first
1143 form has two problems. First it may be less efficient than evaluating it at the
1144 start of the loop. In this case, the cost is probably minor --- a few extra
1145 loads every time through the loop. However, if the base expression is more
1146 complex, then the cost can rise quickly. I've seen loops where the end
1147 expression was actually something like: "``SomeMap[X]->end()``" and map lookups
1148 really aren't cheap. By writing it in the second form consistently, you
1149 eliminate the issue entirely and don't even have to think about it.
1151 The second (even bigger) issue is that writing the loop in the first form hints
1152 to the reader that the loop is mutating the container (a fact that a comment
1153 would handily confirm!). If you write the loop in the second form, it is
1154 immediately obvious without even looking at the body of the loop that the
1155 container isn't being modified, which makes it easier to read the code and
1156 understand what it does.
1158 While the second form of the loop is a few extra keystrokes, we do strongly
1161 ``#include <iostream>`` is Forbidden
1162 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1164 The use of ``#include <iostream>`` in library files is hereby **forbidden**,
1165 because many common implementations transparently inject a `static constructor`_
1166 into every translation unit that includes it.
1168 Note that using the other stream headers (``<sstream>`` for example) is not
1169 problematic in this regard --- just ``<iostream>``. However, ``raw_ostream``
1170 provides various APIs that are better performing for almost every use than
1171 ``std::ostream`` style APIs.
1175 New code should always use `raw_ostream`_ for writing, or the
1176 ``llvm::MemoryBuffer`` API for reading files.
1183 LLVM includes a lightweight, simple, and efficient stream implementation in
1184 ``llvm/Support/raw_ostream.h``, which provides all of the common features of
1185 ``std::ostream``. All new code should use ``raw_ostream`` instead of
1188 Unlike ``std::ostream``, ``raw_ostream`` is not a template and can be forward
1189 declared as ``class raw_ostream``. Public headers should generally not include
1190 the ``raw_ostream`` header, but use forward declarations and constant references
1191 to ``raw_ostream`` instances.
1196 The ``std::endl`` modifier, when used with ``iostreams`` outputs a newline to
1197 the output stream specified. In addition to doing this, however, it also
1198 flushes the output stream. In other words, these are equivalent:
1202 std::cout << std::endl;
1203 std::cout << '\n' << std::flush;
1205 Most of the time, you probably have no reason to flush the output stream, so
1206 it's better to use a literal ``'\n'``.
1208 Don't use ``inline`` when defining a function in a class definition
1209 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1211 A member function defined in a class definition is implicitly inline, so don't
1212 put the ``inline`` keyword in this case.
1239 This section describes preferred low-level formatting guidelines along with
1240 reasoning on why we prefer them.
1242 Spaces Before Parentheses
1243 ^^^^^^^^^^^^^^^^^^^^^^^^^
1245 We prefer to put a space before an open parenthesis only in control flow
1246 statements, but not in normal function call expressions and function-like
1247 macros. For example, this is good:
1252 for (I = 0; I != 100; ++I) ...
1253 while (LLVMRocks) ...
1256 assert(3 != 4 && "laws of math are failing me");
1258 A = foo(42, 92) + bar(X);
1265 for(I = 0; I != 100; ++I) ...
1266 while(LLVMRocks) ...
1269 assert (3 != 4 && "laws of math are failing me");
1271 A = foo (42, 92) + bar (X);
1273 The reason for doing this is not completely arbitrary. This style makes control
1274 flow operators stand out more, and makes expressions flow better. The function
1275 call operator binds very tightly as a postfix operator. Putting a space after a
1276 function name (as in the last example) makes it appear that the code might bind
1277 the arguments of the left-hand-side of a binary operator with the argument list
1278 of a function and the name of the right side. More specifically, it is easy to
1279 misread the "``A``" example as:
1283 A = foo ((42, 92) + bar) (X);
1285 when skimming through the code. By avoiding a space in a function, we avoid
1286 this misinterpretation.
1291 Hard fast rule: Preincrement (``++X``) may be no slower than postincrement
1292 (``X++``) and could very well be a lot faster than it. Use preincrementation
1295 The semantics of postincrement include making a copy of the value being
1296 incremented, returning it, and then preincrementing the "work value". For
1297 primitive types, this isn't a big deal. But for iterators, it can be a huge
1298 issue (for example, some iterators contains stack and set objects in them...
1299 copying an iterator could invoke the copy ctor's of these as well). In general,
1300 get in the habit of always using preincrement, and you won't have a problem.
1303 Namespace Indentation
1304 ^^^^^^^^^^^^^^^^^^^^^
1306 In general, we strive to reduce indentation wherever possible. This is useful
1307 because we want code to `fit into 80 columns`_ without wrapping horribly, but
1308 also because it makes it easier to understand the code. To facilitate this and
1309 avoid some insanely deep nesting on occasion, don't indent namespaces. If it
1310 helps readability, feel free to add a comment indicating what namespace is
1311 being closed by a ``}``. For example:
1316 namespace knowledge {
1318 /// This class represents things that Smith can have an intimate
1319 /// understanding of and contains the data associated with it.
1323 explicit Grokable() { ... }
1324 virtual ~Grokable() = 0;
1330 } // end namespace knowledge
1331 } // end namespace llvm
1334 Feel free to skip the closing comment when the namespace being closed is
1335 obvious for any reason. For example, the outer-most namespace in a header file
1336 is rarely a source of confusion. But namespaces both anonymous and named in
1337 source files that are being closed half way through the file probably could use
1342 Anonymous Namespaces
1343 ^^^^^^^^^^^^^^^^^^^^
1345 After talking about namespaces in general, you may be wondering about anonymous
1346 namespaces in particular. Anonymous namespaces are a great language feature
1347 that tells the C++ compiler that the contents of the namespace are only visible
1348 within the current translation unit, allowing more aggressive optimization and
1349 eliminating the possibility of symbol name collisions. Anonymous namespaces are
1350 to C++ as "static" is to C functions and global variables. While "``static``"
1351 is available in C++, anonymous namespaces are more general: they can make entire
1352 classes private to a file.
1354 The problem with anonymous namespaces is that they naturally want to encourage
1355 indentation of their body, and they reduce locality of reference: if you see a
1356 random function definition in a C++ file, it is easy to see if it is marked
1357 static, but seeing if it is in an anonymous namespace requires scanning a big
1360 Because of this, we have a simple guideline: make anonymous namespaces as small
1361 as possible, and only use them for class declarations. For example, this is
1371 bool operator<(const char *RHS) const;
1373 } // end anonymous namespace
1375 static void runHelper() {
1379 bool StringSort::operator<(const char *RHS) const {
1393 bool operator<(const char *RHS) const;
1400 bool StringSort::operator<(const char *RHS) const {
1404 } // end anonymous namespace
1406 This is bad specifically because if you're looking at "``runHelper``" in the middle
1407 of a large C++ file, that you have no immediate way to tell if it is local to
1408 the file. When it is marked static explicitly, this is immediately obvious.
1409 Also, there is no reason to enclose the definition of "``operator<``" in the
1410 namespace just because it was declared there.
1415 A lot of these comments and recommendations have been culled from other sources.
1416 Two particularly important books for our work are:
1419 <http://www.amazon.com/Effective-Specific-Addison-Wesley-Professional-Computing/dp/0321334876>`_
1420 by Scott Meyers. Also interesting and useful are "More Effective C++" and
1421 "Effective STL" by the same author.
1423 #. `Large-Scale C++ Software Design
1424 <http://www.amazon.com/Large-Scale-Software-Design-John-Lakos/dp/0201633620/ref=sr_1_1>`_
1427 If you get some free time, and you haven't read them: do so, you might learn