Here's the final draft. Please consider it formally submitted; someone (Dan? Ask? Adam?) needs to assign it a number, please. Dave. =head1 TITLE Conventions and Guidelines for Perl Source Code =head1 VERSION =head2 CURRENT Maintainer: Dave Mitchell <[EMAIL PROTECTED]> Class: Internals PDD Number: TBD Version: 1 Status: Developing Last Modified: 6 August 2001 PDD Format: 1 Language: English =head2 HISTORY Based on an earlier draft which covered only code comments. =head1 CHANGES None. First version =head1 ABSTRACT This document describes the various rules, guidelines and advice for those wishing to contribute to the source code of Perl, in such areas as code structure, naming conventions, comments etc. =head1 DESCRIPTION One of the criticisms of Perl 5 is that it's source code is impenetrable to newcomers, due to such things as inconsistent or obscure variable naming conventions, lack of comments in the source code, and so on. Hence this document. We define three classes of conventions. Those that say I<must> are mandatory, and code will not be accepted (apart from in exceptional circumstances) unless it follows these rules. Those that say I<should> are strong guidelines that should normally be be followed unless there is a sensible reason to do otherwise. Finally, where it says I<may>, this is tentative suggestion to be used at your discretion. Note this this particular PDD makes some recommendations that are specific to the C programming language. This does not preclude Perl being implemented in other languages, but in this case, additional PDDs may need to be authored for the extra language-specific features. =head1 IMPLEMENTATION =head2 Coding style The following I<must> apply: =over 4 =item * 4-wide indents for code and 2-wide indents for nested CPP #directives. All indentation must consist of spaces, no tabs (for ease of patching). To ensure that tabs aren't inadvertently used for indentation, the following boilerplate code must appear at the bottom of each source file. (This rule may be rescinded if I'm ever threatened with a lynching....) /* * Local variables: * c-indentation-style: bsd * c-basic-offset: 4 * indent-tabs-mode: nil * End * * vim:expandtab shiftwidth=4: */ =item * Any other tabs are assumed to be on an 8-character boundary. =item * ANSI C function prototypes =item * "K&R" style for indenting control constructs: ie the closing C<}> should line up with the opening C<if> etc. =item * When a conditional spans multiple lines, the opening brace must line up with the "if" or "while", or be at the end-of-line otherwise. =item * Uncuddled elses: ie avoid C<} else {> =item * No C++ style comments (C<//>): some C compilers may choke on them =item * Mark places that need to be revisited with XXX (and preferably your initials too), and revisit often! =item * In function definitions, the name starts in column 0, with the return type on the previous line =item * Single space after keywords that are followed by parens, eg C<return (x+y)*2>, but no space between function name and following paren, eg C<z = foo(x+y)*2> =back The following I<should> apply =over 4 =item * Do not exceed 79 columns =item * C<return foo;> rather than C<return (foo);> =item * C<if (!foo) ...> rather than C<if (foo == FALSE) ...> etc. =item * Avoid assignments in conditionals, but if they're unavoidable, use Extra paren, e.g. C<if (a && (b = c)) ...> =item * Avoid double negatives, eg C<#ifndef NO_FEATURE_FOO> =item * binary operators should have a space on either side; parentheses should not have space immediately after the opening paren nor immediately before the closing paren, commas should have space after but not before, eg x = (a + b) * f(c, d / e) =item * Long lines should be split before an operator so that it is immediately obvious when scanning the LHS of the code that this has happened, and two extra levels of indent should be used. =item * and/or split on matching parens, with the content on separate line(s) and with one extra indent: do_arbitrary_function( list_of_parameters_with_long_names, or_complex_subexpression( of_more_params, or_expressions + 1 ) ); =back =head2 Naming conventions =over 4 =item Subsystems and APIs Perl core will be split into a number of subsystems, each with an associated API. For the purposes of naming files, data structures, etc, each subsystem will be assigned a short nickname, eg pmc, gc, io. All code within the core will belong to a subsystem; miscellaneous code with no obvious home will be placed in the special subsystem called misc. =item Filenames Filenames must be assumed to be case-insensitive, in the sense that that you may not have two different files called Foo and foo. Normal source-code filenames should be all lower-case; filenames with upper-case letters in them are reserved for notice-me-first files such as README, and for files which need some sort of pre-processing applied to them or which do the preprocessing - eg a script F<foo.SH> might read F<foo.TEMPLATE> and output F<foo.c>. The characters making up filenames must be chosen from the ASCII set A-Z,a-z,0-9 plus .-_ An underscore should be used to separate words rather than a hyphen (-). A file should not normally have more than a single '.' in it, and this should be used to denote a suffix of some description. The filename must still be unique if the main part is truncated to 8 characters and any suffix truncated to 3 characters. Ideally, filenames should restricted to 8.3 in the first place, but this is not essential. Each subsystem I<foo> should supply the following files. This arrangement is based on the assumption that each subsystem will - as far as is practical - present an opaque interface to all other subsystems within the core, as well as to extensions and embeddings. =over 4 =item foo.h This contains all the declarations needed for external users of that API (and nothing more), ie it defines the API. It is permissible for the API to include different or extra functionality when used by other parts of the core, compared with its use in extensions and embeddings. In this case, the extra stuff within the file is enabled by testing for the macro PERL_IN_CORE. =item foo_private.h This contains declarations used internally by that subsystem, and which must only be included within source files associated the subsystem. This file defines the macro PERL_IN_FOO so that code knows when it is being used within that subsystem. The file will also contain all the 'convenience' macros used to define shorter working names for functions without the perl prefix (see below). =item foo_globals.h This file contains the declaration of a single structure containing the private global variables used by the subsystem (see the section on globals below for more details). =item foo.sym This file (format and contents TBD) contains information about global symbols associated with the subsystem, and may be used by scripts to auto-generate such stuff as the include files mentioned above, linker map tables, documentation etc, based upon portability and extensibility requirements. =item foo_bar.[ch] etc All other source files associated with the subsystem will have the prefix foo_ =back The top-level structure of the Perl source tarball should be as follows: /README, etc a few top-level documents /doc/ Assorted miscellaneous documentation /pdd/ The current PDDs /perl/ The source code for Perl itself /perl/os/foo/ OS-specific source code for operating system foo /foo/ The source code for other families of binaries (eg /x2p/) /hints/ per-OS build hints files /scripts/ scripts needed during the building process /t/ scripts used by make test /lib/ perl modules ready for installation /ext/ perl modules that need compiling /pod/ src of the Perl man pages etc plus others as it becomes necessary. =item Names of code entities Code entities such as variables, functions, macros etc (apart from strictly local ones) should all follow these general guidelines. =over 4 =item * Multiple words or components should be separated with underscores rather than using tricks such as capitalisation, eg C<new_foo_bar> rather than C<NewFooBar> or (gasp) C<newfoobar>. =item * The names of entities should err on the side of verbosity, eg C<create_foo_from_bar()> in preference to C<ct_foo_bar()>. Avoid cryptic abbreviations wherever possible. =item * All entities should be prefixed with the name of the subsystem they appear in, eg C<pmc_foo()>, C<struct io_bar>. They should be further prefixed with the word 'perl' if they have external visibility or linkage, namely, non-static functions, plus macros and typedefs etc which appear in public header files. (Global variables are handled specially; see below.) For example: perlpmc_foo() struct perlio_bar typedef struct perlio_bar Perlio_bar #define PERLPMC_readonly_TEST ... In the specific case of the use of global variables and functions within a subsystem, convenience macros will be defined (in foo_private.h) that allow use of the shortened name in the case of functions (ie C<pmc_foo()> instead of C<perlpmc_foo()>), and hide the real representation in the case of global variables. =item * Variables and structure names should be all lower-case, eg C<pmc_foo>. =item * structure elements should be all lower-case, and the first component of the name should incorporate the structure's name or an abbreviation of it. =item * Typedef names should be lower-case except for the first letter, eg C<Foo_bar>. The exception to this is when the first component is a short abbreviation, in which case the whole first component may be made uppercase for readability purposes, eg C<IO_foo> rather than C<Io_foo>. Structures should generally be typedefed. =item * Macros should have their first component uppercase, and the majority of the remaining components should be likewise. Where there is a family of macros, the variable part can be indicated in lowercase, eg C<PMC_foo_FLAG>, C<PMC_bar_FLAG>, .... =item * A macro which defines a flag bit should be suffixed with C<_FLAG>, eg C<PMC_readonly_FLAG> (although you probably want to use an C<enum> instead.) =item * A macro which tests a flag bit should be suffixed with C<_TEST>, eg C<if (PMC_readonly_TEST(foo)) ...> =item * A macro which sets a flag bit should be suffixed with C<_SET>, eg C<PMC_readonly_SET(foo);> =item * A macro which clears a flag bit should be suffixed with C<_CLEAR>, eg C<PMC_readonly_CLEAR(foo);> =item * A macro defining a mask of flag bits should be suffixed with C<_MASK>, eg C<foo &= ~PMC_STATUS_MASK> (but see notes on extensibility below). =item * Macros can be defined to cover common flag combinations, in which case they should have C<_SETALL>, C<CLEARALL>, C<_TESTALL> or <_TESTANY> suffixes as appropriate, to indicate aggregate bits, eg C<PMC_valid_CLEARALL(foo)> =item * A macro defining an auto-configuration value should be prefixed with C<HAS_>, eg C<HAS_BROKEN_FLOCK>, C<HAS_EBCDIC>. =item * A macro indicating the compilation 'location' should be prefixed with C<IN_>, eg C<PERL_IN_CORE>, C<PERL_IN_PMC>, C<PERL_IN_X2P>. Individual include file visitations should be marked with C<PERL_IN_FOO_H> for file foo.h =item * A macro indicating major compilation switches should be prefixed with C<USE_>, eg C<PERL_USE_STDIO>, C<USE_MULTIPLICITY>. =item * A macro that may declare stuff and thus needs to be at the start of a block should be prefixed with C<DECL_>, eg C<DECL_SAVE_STACK>. Note that macros which implicitly declare and then use variables are strongly discouraged, unless it is essential for portability or extensibility. The following are in decreasing preference style-wise, but increasing preference extensibility-wise. { Stack sp = GETSTACK; x = POPSTACK(sp) ... /* sp is an auto variable */ { DECL_STACK(sp); x = POPSTACK(sp); ... /* sp may or may not be auto */ { DECL_STACK; x = POPSTACK; ... /* anybody's guess */ =back =item Global Variables Global variables must never be accessed directly outside the subsystem in which they are used. Some other method, such as accessor functions, must be provided by that subsystem's API. (For efficiency the 'accessor functions' may occasionally actually be macros, but then the rule still applies in spirit at least). All global variables needed for the internal use of a particular subsystem should all be declared within a single struct called foo_globals for subsystem foo. This structure's declaration is placed in the file foo_globals.h. Then somewhere a single compound structure will be declared which has as members the individual structures from each subsystem. Instances of this structure are then defined as a one-off global variable, or as per-thread instances, or whatever is required. [Actually, three separate structures may be required, for global, per-interpreter and per-thread variables.] Within an individual subsystem, macros are defined for each global variable of the form GLOBAL_foo (the name being deliberately clunky). So we might for example have the following macros: /* perl_core.h or similar */ #ifdef HAS_THREADS # define GLOBALS_BASE (aTHX_->globals) #else # define GLOBALS_BASE (Perl_globals) #endif /* pmc_private.h */ #define GLOBAL_foo GLOBALS_BASE.pmc.foo #define GLOBAL_bar GLOBALS_BASE.pmc.bar ... etc ... =back =head2 Code comments The importance of good code documentation cannot be stressed enough. To make your code understandable by others (and indeed by yourself when you come to make changes a year later :-), the following conventions apply to all source files. =over 4 =item Developer files For each source file (eg a F<foo.c> F<foo.h> pair), there should be an accompanying developer file called F<foo.dev>. This text file contains documentation on all the implementation decisions associated with the source file. (Note that this is in contrast to PDDs, which describe design decisions). This is the place for mini-essays on how to avoid overflows in unsigned arithmetic, or on the pros and cons of differing hash algorithms, and why the current one was chosen, and how it works. In principle, someone coming to a particular source file for the first time should be able to read the F<.dev> file and gain an immediate overview of what the source file is for, the algorithms it implements, etc. Currently no particular format or structure is imposed on the developer file, but it should have as a minimum the following sections: =over 4 =item Overview Explain the purpose of the source file. =item Data structures and algorithms Explain how it all works. =item History Record major changes to the file, eg "we moved from a linked list to a hash table implementation for storing Foos, as it was found to be much faster". =item Notes Anything that may be of interest to your successors, eg benchmarks of differing hash algorithms, essays on how to do integer arithmetic. =item References Links to pages and books that may contain useful info relevant to the stuff going on in the code - eg the book you stole the hash function from. =back =item Top-of-file comments In addition to the copyright message and optional quote, each source file must have a short comment at the top explaining the basic purpose of the file, eg /* pp_hot.c - like pp.c, this file contains functions that operate * on the contents of the stack (pp == 'push & pop'), but in this * case, frequently used ('hot') functions have been moved here * from pp.c to (hopefully) improve CPU cache hit rates. */ =item Per-section comments If there is a collection of functions, structures or whatever which are grouped together and have a common theme or purpose, there should be a general comment at the start of the section briefly explaining their overall purpose. (Detailed essays should be left to the developer file). If there is really only one section, then the top-of-file comment already satisfies this requirement. /* This section deals with 'arenas', which are chunks of PMCs of * a particular type that are allocated in one go. Individual * requests can then be made to grab or release individual PMCs. * For each type foo, there is a pointer called GLOBAL_arena_foo * which blah blah.... */ =item Per-entity comments Every non-local named entity, be it a function, variable, structure, macro or whatever, must have an accompanying comment explaining it's purpose. This comment must be in the special format described below, in order to allow automatic extraction by tools - for example, to generate per API man pages, B<perldoc -f> style utilities and so on. Often the comment need only be a single line explaining its purpose, but sometimes more explanation may be needed. For example, "return an Integer Foo to its allocation pool" may be enough to demystify the function C<del_I_foo()> Each comment should be of the form /*=for api apiname entityname[,entityname..] flags ....(TBC).... comments.... */ where I<apiname> is the API the entity belongs to, eg I<pmc>, and entity name is the actual name of the function or macro or whatever. Where there is a whole family of entities that have the same properties and can be collectively described with a single comment, a list of entity names can be provided. TBC ... =item Optimisations Whenever code has deliberately been written in an odd way for performance reasons, you should point this out - if nothing else, to avoid some poor shmuck trying subsequently to replace it with something 'cleaner'. /* The loop is partially unrolled here as it makes it a lot faster. * See the .dev file for the full details */ =item General comments While there is no need to go mad commenting every line of code, it is immensely helpful to to provide a "running commentary" every 10 or so lines say; if nothing else, this makes it easy to quickly locate a specific chunk of code. Such comments are particularly useful at the top of each major branch, eg if (FOO_bar_BAZ(**p+*q) <= (r-s[FOZ & FAZ_MASK]) || FLOP_2(z99)) { /* we're in foo mode: clean up lexicals */ ... (20 lines of gibberish) ... } else if (...) { /* we're in bar mode: clean up globals */ ... (20 more lines of gibberish) ... } else { /* we're in baz mode: self-destruct */ .... } =back =head2 Extensibility If Perl 5 is anything to go by, the lifetime of Perl 6 will be at least seven years. During this period, the source code will undergo many major changes never envisaged by its original authors - cf threads, unicode in perl 5. To this end, Your code should balance out the assumptions that make things possible, fast or small, with the assumptions that make it difficult to change things in future. This is especially important for parts of the code which are exposed through APIs - the requirements of src or binary compatibility for such things as extensions can make it very hard to change things later on. For example, if you define suitable macros to set/test flags in a struct, then you can later add a second word of flags to the struct without breaking source compatibility. (Although you might still break binary compatibility if you're not careful.) Of the following two methods of setting a common combination of flags, the second doesn't assume that all the flags are contained within a single field: foo->flags |= (FOO_int_FLAG | FOO_num_FLAG | FOO_str_FLAG); FOO_valid_value_SETALL(foo); Similarly, avoid using a char* (or {char*,length}) if it is feasible to later use a PMC* at the same point: cf UTF-8 hash keys in Perl 5. Of course, private code hidden behind an API can play more fast and loose than code which gets exposed. =head2 Portability Related to extensibility is portability. Perl runs on many, many platforms, and will no doubt be ported to ever more bizarre and obscure ones over time. You should never assume an operating system, processor architecture, endian-ness, word size, or whatever. In particular, don't fall into the any of the following common traps: Internal data types and their utility functions (especially for strings) should be used over a bare char * whenever possible. Ideally there should be no char * in the source anywhere, and no use of C's standard string library. Dont assume GNU C, and don't use any GNU extensions unless protected by #ifdefs for non-GNU-C builds. TBC ... Any contributions welcome !!! =head2 Performance We want Perl to be fast. Very fast. But we also want it to be portable and extensible. Based on the 90/10 principle, (or 80/20, or 95/5, depending on who you speak to), most performance is gained or lost in a few small but critical areas of code. Concentrate your optimisation efforts there. Note that the most overwhelmingly important factor in performance is in choosing the correct algorithms and data structures in the first place. Any subsequent tweaking of code is secondary to this. Also, any tweaking that is done should as far as possible be platform independent, or at least likely to cause speed-ups in a wide variety of environments, and do no harm elsewhere. Only in exceptional circumstances should assembly ever even be considered, and then only if generic fallback code is made available that can still be used by all other non-optimised platforms. Probably the dominant factor (circa 2001) that effects processor performance is the cache. Processor clock rates have increased far in excess of of main memory access rates, and the only way for the processor to proceed without stalling is for most of the data items it needs to be found to hand in the cache. It is reckoned that even a 2% cache miss rate can cause a slowdown in the region of 50%. It is for this reason that algorithms and data structures must be designed to be 'cache-friendly'. A typical cache may have a block size of anywhere between 4 and 256 bytes. When a program attempts to read a word from memory and the word is already in the cache, then processing continues unaffected. Otherwise, the processor is typically stalled while a whole contiguous chunk of main memory is read in and stored in a cache block. Thus, after incurring the initial time penalty, you then get all the memory adjacent to the initially read data item for free. Algorithms that make use of this fact can experience quite dramatic speedups. For example, the following pathological code ran four times faster on my machine by simply swapping C<i> and C<j>. int a[1000][1000]; ... (a gets populated) ... int i,j,k; for (i=0; i<1000; i++) { for (j=0; j<1000; j++) { k += a[j][i]; } } This all boils down to: keep things near to each other that get accessed at around the same time. (This is why the important optimisations occur in data structure and algorithm design rather than in the detail of the code.) This rule applies both to the layout of different objects relative to each other, and to the relative positioning of individual fields within a single structure. If you do put an optimisation in, time it on as many architectures as you can, and be suspicious of it if it slows down on any of them! Perhaps it will be slow on other architectures too (current and future). Perhaps it wasn't so clever after all? If the optimisation is platform specific, you should probably put it in a platform-specific function in a platform-specific file, rather than cluttering the main source with zillions of #ifdefs. And remember to document it. Loosely speaking, Perl tends to optimise for speed rather than space, So you may want to code for speed first, then tweak to reclaim some space while not affecting performance. =head1 REFERENCES The section on coding style is based on Perl5's F<Porting/patching.pod> by Daniel Grisinger. The section on naming conventions grew from some suggestions by Paolo Molaro <[EMAIL PROTECTED]>. Other snippets came from various P5Pers. The rest of it is probably my fault.