1 <appendix xmlns="http://docbook.org/ns/docbook" version="5.0"
2 xml:id="appendix.contrib" xreflabel="Contributing">
3 <?dbhtml filename="appendix_contributing.html"?>
8 <primary>Appendix</primary>
9 <secondary>Contributing</secondary>
13 <keyword>ISO C++</keyword>
14 <keyword>library</keyword>
21 The GNU C++ Library is part of GCC and follows the same development model,
22 so the general rules for
23 <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/contribute.html">contributing
24 to GCC</link> apply. Active
25 contributors are assigned maintainership responsibility, and given
26 write access to the source repository. First-time contributors
27 should follow this procedure:
30 <section xml:id="contrib.list" xreflabel="Contributor Checklist"><info><title>Contributor Checklist</title></info>
33 <section xml:id="list.reading"><info><title>Reading</title></info>
39 Get and read the relevant sections of the C++ language
40 specification. Copies of the full ISO 14882 standard are
41 available on line via the ISO mirror site for committee
42 members. Non-members, or those who have not paid for the
43 privilege of sitting on the committee and sustained their
44 two meeting commitment for voting rights, may get a copy of
45 the standard from their respective national standards
46 organization. In the USA, this national standards
48 <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="https://www.ansi.org">ANSI</link>.
49 (And if you've already registered with them you can <link
50 xmlns:xlink="http://www.w3.org/1999/xlink"
51 xlink:href="https://webstore.ansi.org/RecordDetail.aspx?sku=ISO%2fIEC+14882%3a2014">buy
52 the standard on-line</link>.)
58 The library working group bugs, and known defects, can
60 <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://www.open-std.org/jtc1/sc22/wg21/">http://www.open-std.org/jtc1/sc22/wg21</link>
67 the <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://www.gnu.org/prep/standards/">GNU
68 Coding Standards</link>, and chuckle when you hit the part
69 about <quote>Using Languages Other Than C</quote>.
75 Be familiar with the extensions that preceded these
76 general GNU rules. These style issues for libstdc++ can be
77 found in <link linkend="contrib.coding_style">Coding Style</link>.
83 And last but certainly not least, read the
84 library-specific information found in
85 <link linkend="appendix.porting">Porting and Maintenance</link>.
91 <section xml:id="list.copyright"><info><title>Assignment</title></info>
94 See the <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/contribute.html#legal">legal prerequisites</link> for all GCC contributions.
98 Historically, the libstdc++ assignment form added the following
104 Which Belgian comic book character is better, Tintin or Asterix, and
110 While not strictly necessary, humoring the maintainers and answering
111 this question would be appreciated.
116 Paolo Carlini at <email>paolo.carlini@oracle.com</email>
118 Jonathan Wakely at <email>jwakely+assign@redhat.com</email>
119 if you are confused about the assignment or have general licensing
120 questions. When requesting an assignment form from
121 <email>assign@gnu.org</email>, please CC the libstdc++
122 maintainers above so that progress can be monitored.
126 <section xml:id="list.getting"><info><title>Getting Sources</title></info>
129 <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/svnwrite.html">Getting write access
130 (look for "Write after approval")</link>
134 <section xml:id="list.patches"><info><title>Submitting Patches</title></info>
138 Every patch must have several pieces of information before it can be
139 properly evaluated. Ideally (and to ensure the fastest possible
140 response from the maintainers) it would have all of these pieces:
146 A description of the bug and how your patch fixes this
147 bug. For new features a description of the feature and your
154 A ChangeLog entry as plain text; see the various
155 ChangeLog files for format and content. If you are
156 using emacs as your editor, simply position the insertion
157 point at the beginning of your change and hit CX-4a to bring
158 up the appropriate ChangeLog entry. See--magic! Similar
159 functionality also exists for vi.
165 A testsuite submission or sample program that will
166 easily and simply show the existing error or test new
173 The patch itself. If you are accessing the SVN
174 repository use <command>svn update; svn diff NEW</command>;
175 else, use <command>diff -cp OLD NEW</command> ... If your
176 version of diff does not support these options, then get the
177 latest version of GNU
178 diff. The <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/wiki/SvnTricks">SVN
179 Tricks</link> wiki page has information on customising the
180 output of <code>svn diff</code>.
186 When you have all these pieces, bundle them up in a
187 mail message and send it to libstdc++@gcc.gnu.org. All
188 patches and related discussion should be sent to the
189 libstdc++ mailing list. In common with the rest of GCC,
190 patches should also be sent to the gcc-patches mailing list.
199 <section xml:id="contrib.organization" xreflabel="Source Organization"><info><title>Directory Layout and Source Conventions</title></info>
200 <?dbhtml filename="source_organization.html"?>
204 The <filename class="directory">libstdc++-v3</filename> directory in the
205 GCC sources contains the files needed to create the GNU C++ Library.
209 It has subdirectories:
214 <term><filename class="directory">doc</filename></term>
216 Files in HTML and text format that document usage, quirks of the
217 implementation, and contributor checklists.
222 <term><filename class="directory">include</filename></term>
224 All header files for the C++ library are within this directory,
225 modulo specific runtime-related files that are in the libsupc++
230 <term><filename class="directory">include/std</filename></term>
232 Files meant to be found by <code>#include <name></code> directives
233 in standard-conforming user programs.
238 <term><filename class="directory">include/c</filename></term>
240 Headers intended to directly include standard C headers.
241 [NB: this can be enabled via <option>--enable-cheaders=c</option>]
246 <term><filename class="directory">include/c_global</filename></term>
248 Headers intended to include standard C headers in
249 the global namespace, and put select names into the <code>std::</code>
250 namespace. [NB: this is the default, and is the same as
251 <option>--enable-cheaders=c_global</option>]
256 <term><filename class="directory">include/c_std</filename></term>
258 Headers intended to include standard C headers
259 already in namespace std, and put select names into the <code>std::</code>
260 namespace. [NB: this is the same as
261 <option>--enable-cheaders=c_std</option>]
266 <term><filename class="directory">include/bits</filename></term>
268 Files included by standard headers and by other files in
274 <term><filename class="directory">include/backward</filename></term>
276 Headers provided for backward compatibility, such as
277 <filename class="headerfile"><backward/hash_map></filename>.
278 They are not used in this library.
283 <term><filename class="directory">include/ext</filename></term>
285 Headers that define extensions to the standard library. No
286 standard header refers to any of them, in theory (there are some
293 <filename class="directory">include/debug</filename>,
294 <filename class="directory">include/parallel</filename>, and
295 <filename class="directory">include/profile</filename>
298 Headers that implement the Debug Mode, Parallel Mode, and Profile Mode
307 <term><filename class="directory">scripts</filename></term>
309 Scripts that are used during the configure, build, make, or test
315 <term><filename class="directory">src</filename></term>
317 Files that are used in constructing the library, but are not
322 <term><filename class="directory">src/c++98</filename></term>
324 Source files compiled using <option>-std=gnu++98</option>.
329 <term><filename class="directory">src/c++11</filename></term>
331 Source files compiled using <option>-std=gnu++11</option>.
336 <term><filename class="directory">src/filesystem</filename></term>
338 Source files for the Filesystem TS.
343 <term><filename class="directory">src/shared</filename></term>
345 Source code included by other files under both
346 <filename class="directory">src/c++98</filename> and
347 <filename class="directory">src/c++11</filename>
355 <term><filename class="directory">testsuites/[backward, demangle, ext, performance, thread, 17_* to 30_*]</filename></term>
357 Test programs are here, and may be used to begin to exercise the
358 library. Support for "make check" and "make check-install" is
359 complete, and runs through all the subdirectories here when this
360 command is issued from the build directory. Please note that
361 "make check" requires DejaGNU 1.4 or later to be installed.
367 Other subdirectories contain variant versions of certain files
368 that are meant to be copied or linked by the configure script.
370 <literallayout><filename class="directory">config/abi</filename>
371 <filename class="directory">config/allocator</filename>
372 <filename class="directory">config/cpu</filename>
373 <filename class="directory">config/io</filename>
374 <filename class="directory">config/locale</filename>
375 <filename class="directory">config/os</filename>
380 In addition, a subdirectory holds the convenience library libsupc++.
385 <term><filename class="directory">libsupc++</filename></term>
387 Contains the runtime library for C++, including exception
388 handling and memory allocation and deallocation, RTTI, terminate
395 Note that glibc also has a <filename class="directory">bits/</filename>
396 subdirectory. We need to be careful not to collide with names in its
397 <filename class="directory">bits/</filename> directory. For example
398 <filename class="headerfile"><bits/std_mutex.h></filename> has to be
399 renamed from <filename class="headerfile"><bits/mutex.h></filename>.
400 Another solution would be to rename <filename class="directory">bits</filename>
401 to (e.g.) <filename class="directory">cppbits</filename>.
405 In files throughout the system, lines marked with an "XXX" indicate
406 a bug or incompletely-implemented feature. Lines marked "XXX MT"
407 indicate a place that may require attention for multi-thread safety.
412 <section xml:id="contrib.coding_style" xreflabel="Coding Style"><info><title>Coding Style</title></info>
413 <?dbhtml filename="source_code_style.html"?>
417 <section xml:id="coding_style.bad_identifiers"><info><title>Bad Identifiers</title></info>
420 Identifiers that conflict and should be avoided.
423 <literallayout class="normal">
424 This is the list of names <quote>reserved to the
425 implementation</quote> that have been claimed by certain
426 compilers and system headers of interest, and should not be used
427 in the library. It will grow, of course. We generally are
428 interested in names that are not all-caps, except for those like
471 [Note that this list is out of date. It applies to the old
472 name-mangling; in G++ 3.0 and higher a different name-mangling is
473 used. In addition, many of the bugs relating to G++ interpreting
474 these names as operators have been fixed.]
476 The full set of __* identifiers (combined from gcc/cp/lex.c and
477 gcc/cplus-dem.c) that are either old or new, but are definitely
478 recognized by the demangler, is:
606 // long double conversion members mangled as __opr
607 // http://gcc.gnu.org/ml/libstdc++/1999-q4/msg00060.html
612 <section xml:id="coding_style.example"><info><title>By Example</title></info>
614 <literallayout class="normal">
615 This library is written to appropriate C++ coding standards. As such,
616 it is intended to precede the recommendations of the GNU Coding
617 Standard, which can be referenced in full here:
619 <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://www.gnu.org/prep/standards/standards.html#Formatting">http://www.gnu.org/prep/standards/standards.html#Formatting</link>
621 The rest of this is also interesting reading, but skip the "Design
624 The GCC coding conventions are here, and are also useful:
625 <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/codingconventions.html">http://gcc.gnu.org/codingconventions.html</link>
627 In addition, because it doesn't seem to be stated explicitly anywhere
628 else, there is an 80 column source limit.
630 <filename>ChangeLog</filename> entries for member functions should use the
631 classname::member function name syntax as follows:
634 1999-04-15 Dennis Ritchie <dr@att.com>
636 * src/basic_file.cc (__basic_file::open): Fix thinko in
637 _G_HAVE_IO_FILE_OPEN bits.
640 Notable areas of divergence from what may be previous local practice
641 (particularly for GNU C) include:
643 01. Pointers and references
648 char *p = "flop"; // wrong
649 char &c = *p; // wrong
652 Reason: In C++, definitions are mixed with executable code. Here,
653 <code>p</code> is being initialized, not <code>*p</code>. This is near-universal
654 practice among C++ programmers; it is normal for C hackers
655 to switch spontaneously as they gain experience.
657 02. Operator names and parentheses
661 operator == (type) // wrong
664 Reason: The <code>==</code> is part of the function name. Separating
665 it makes the declaration look like an expression.
667 03. Function names and parentheses
671 void mangle () // wrong
674 Reason: no space before parentheses (except after a control-flow
675 keyword) is near-universal practice for C++. It identifies the
676 parentheses as the function-call operator or declarator, as
677 opposed to an expression or other overloaded use of parentheses.
679 04. Template function indentation
681 template<typename T>
683 template_function(args)
686 template<class T>
687 void template_function(args) {};
690 Reason: In class definitions, without indentation whitespace is
691 needed both above and below the declaration to distinguish
692 it visually from other members. (Also, re: "typename"
693 rather than "class".) <code>T</code> often could be <code>int</code>, which is
694 not a class. ("class", here, is an anachronism.)
696 05. Template class indentation
698 template<typename _CharT, typename _Traits>
699 class basic_ios : public ios_base
705 template<class _CharT, class _Traits>
706 class basic_ios : public ios_base
712 template<class _CharT, class _Traits>
713 class basic_ios : public ios_base
729 enum { space = _ISspace, print = _ISprint, cntrl = _IScntrl };
732 07. Member initialization lists
733 All one line, separate from class name.
737 : _M_private_data(0), _M_more_stuff(0), _M_helper(0)
740 gribble::gribble() : _M_private_data(0), _M_more_stuff(0), _M_helper(0)
762 09. Member functions declarations and definitions
763 Keywords such as extern, static, export, explicit, inline, etc
764 go on the line above the function name. Thus
773 Reason: GNU coding conventions dictate return types for functions
774 are on a separate line than the function name and parameter list
775 for definitions. For C++, where we have member functions that can
776 be either inline definitions or declarations, keeping to this
777 standard allows all member function names for a given class to be
778 aligned to the same margin, increasing readability.
781 10. Invocation of member functions with "this->"
782 For non-uglified names, use <code>this->name</code> to call the function.
790 Reason: Koenig lookup.
806 12. Spacing under protected and private in class declarations:
807 space above, none below
820 13. Spacing WRT return statements.
821 no extra spacing before returns, no parenthesis
840 14. Location of global variables.
841 All global variables of class type, whether in the "user visible"
842 space (e.g., <code>cin</code>) or the implementation namespace, must be defined
843 as a character array with the appropriate alignment and then later
844 re-initialized to the correct value.
846 This is due to startup issues on certain platforms, such as AIX.
847 For more explanation and examples, see <filename>src/globals.cc</filename>. All such
848 variables should be contained in that file, for simplicity.
850 15. Exception abstractions
851 Use the exception abstractions found in <filename class="headerfile">functexcept.h</filename>, which allow
852 C++ programmers to use this library with <literal>-fno-exceptions</literal>. (Even if
853 that is rarely advisable, it's a necessary evil for backwards
856 16. Exception error messages
857 All start with the name of the function where the exception is
858 thrown, and then (optional) descriptive text is added. Example:
861 __throw_logic_error(__N("basic_string::_S_construct NULL not valid"));
864 Reason: The verbose terminate handler prints out <code>exception::what()</code>,
865 as well as the typeinfo for the thrown exception. As this is the
866 default terminate handler, by putting location info into the
867 exception string, a very useful error message is printed out for
868 uncaught exceptions. So useful, in fact, that non-programmers can
869 give useful error messages, and programmers can intelligently
870 speculate what went wrong without even using a debugger.
872 17. The doxygen style guide to comments is a separate document,
875 The library currently has a mixture of GNU-C and modern C++ coding
876 styles. The GNU C usages will be combed out gradually.
880 For nonstandard names appearing in Standard headers, we are constrained
881 to use names that begin with underscores. This is called "uglification".
884 Local and argument names: <literal>__[a-z].*</literal>
886 Examples: <code>__count __ix __s1</code>
888 Type names and template formal-argument names: <literal>_[A-Z][^_].*</literal>
890 Examples: <code>_Helper _CharT _N</code>
892 Member data and function names: <literal>_M_.*</literal>
894 Examples: <code>_M_num_elements _M_initialize ()</code>
896 Static data members, constants, and enumerations: <literal>_S_.*</literal>
898 Examples: <code>_S_max_elements _S_default_value</code>
900 Don't use names in the same scope that differ only in the prefix,
901 e.g. _S_top and _M_top. See BADNAMES for a list of forbidden names.
902 (The most tempting of these seem to be and "_T" and "__sz".)
904 Names must never have "__" internally; it would confuse name
905 unmanglers on some targets. Also, never use "__[0-9]", same reason.
907 --------------------------
922 gribble(const gribble&);
925 gribble(int __howmany);
928 operator=(const gribble&);
933 // Start with a capital letter, end with a period.
935 public_member(const char* __arg) const;
937 // In-class function definitions should be restricted to one-liners.
939 one_line() { return 0 }
942 two_lines(const char* arg)
943 { return strchr(arg, 'a'); }
946 three_lines(); // inline, but defined below.
949 template<typename _Formal_argument>
951 public_template() const throw();
953 template<typename _Iterator>
963 int _M_private_function();
972 _S_initialize_library();
975 // More-or-less-standard language features described by lack, not presence.
976 # ifndef _G_NO_LONGLONG
977 extern long long _G_global_with_a_good_long_name; // avoid globals!
980 // Avoid in-class inline definitions, define separately;
981 // likewise for member class definitions:
983 gribble::public_member() const
984 { int __local = 0; return __local; }
986 class gribble::_Helper
990 friend class gribble;
994 // Names beginning with "__": only for arguments and
995 // local variables; never use "__" in a type name, or
996 // within any name; never use "__[0-9]".
998 #endif /* _HEADER_ */
1003 template<typename T> // notice: "typename", not "class", no space
1004 long_return_value_type<with_many, args>
1005 function_name(char* pointer, // "char *pointer" is wrong.
1007 const Reference& ref)
1009 // int a_local; /* wrong; see below. */
1015 int a_local = 0; // declare variable at first use.
1017 // char a, b, *p; /* wrong */
1020 char* c = "abc"; // each variable goes on its own line, always.
1022 // except maybe here...
1023 for (unsigned i = 0, mask = 1; mask; ++i, mask <<= 1) {
1029 : _M_private_data(0), _M_more_stuff(0), _M_helper(0)
1033 gribble::three_lines()
1035 // doesn't fit in one line.
1043 <section xml:id="contrib.design_notes" xreflabel="Design Notes"><info><title>Design Notes</title></info>
1044 <?dbhtml filename="source_design_notes.html"?>
1049 <literallayout class="normal">
1054 This paper is covers two major areas:
1056 - Features and policies not mentioned in the standard that
1057 the quality of the library implementation depends on, including
1058 extensions and "implementation-defined" features;
1060 - Plans for required but unimplemented library features and
1061 optimizations to them.
1066 The standard defines a large library, much larger than the standard
1067 C library. A naive implementation would suffer substantial overhead
1068 in compile time, executable size, and speed, rendering it unusable
1069 in many (particularly embedded) applications. The alternative demands
1070 care in construction, and some compiler support, but there is no
1071 need for library subsets.
1073 What are the sources of this overhead? There are four main causes:
1075 - The library is specified almost entirely as templates, which
1076 with current compilers must be included in-line, resulting in
1077 very slow builds as tens or hundreds of thousands of lines
1078 of function definitions are read for each user source file.
1079 Indeed, the entire SGI STL, as well as the dos Reis valarray,
1080 are provided purely as header files, largely for simplicity in
1081 porting. Iostream/locale is (or will be) as large again.
1083 - The library is very flexible, specifying a multitude of hooks
1084 where users can insert their own code in place of defaults.
1085 When these hooks are not used, any time and code expended to
1086 support that flexibility is wasted.
1088 - Templates are often described as causing to "code bloat". In
1089 practice, this refers (when it refers to anything real) to several
1090 independent processes. First, when a class template is manually
1091 instantiated in its entirely, current compilers place the definitions
1092 for all members in a single object file, so that a program linking
1093 to one member gets definitions of all. Second, template functions
1094 which do not actually depend on the template argument are, under
1095 current compilers, generated anew for each instantiation, rather
1096 than being shared with other instantiations. Third, some of the
1097 flexibility mentioned above comes from virtual functions (both in
1098 regular classes and template classes) which current linkers add
1099 to the executable file even when they manifestly cannot be called.
1101 - The library is specified to use a language feature, exceptions,
1102 which in the current gcc compiler ABI imposes a run time and
1103 code space cost to handle the possibility of exceptions even when
1104 they are not used. Under the new ABI (accessed with -fnew-abi),
1105 there is a space overhead and a small reduction in code efficiency
1106 resulting from lost optimization opportunities associated with
1107 non-local branches associated with exceptions.
1109 What can be done to eliminate this overhead? A variety of coding
1110 techniques, and compiler, linker and library improvements and
1111 extensions may be used, as covered below. Most are not difficult,
1112 and some are already implemented in varying degrees.
1114 Overhead: Compilation Time
1115 --------------------------
1117 Providing "ready-instantiated" template code in object code archives
1118 allows us to avoid generating and optimizing template instantiations
1119 in each compilation unit which uses them. However, the number of such
1120 instantiations that are useful to provide is limited, and anyway this
1121 is not enough, by itself, to minimize compilation time. In particular,
1122 it does not reduce time spent parsing conforming headers.
1124 Quicker header parsing will depend on library extensions and compiler
1125 improvements. One approach is some variation on the techniques
1126 previously marketed as "pre-compiled headers", now standardized as
1127 support for the "export" keyword. "Exported" template definitions
1128 can be placed (once) in a "repository" -- really just a library, but
1129 of template definitions rather than object code -- to be drawn upon
1130 at link time when an instantiation is needed, rather than placed in
1131 header files to be parsed along with every compilation unit.
1133 Until "export" is implemented we can put some of the lengthy template
1134 definitions in #if guards or alternative headers so that users can skip
1135 over the full definitions when they need only the ready-instantiated
1138 To be precise, this means that certain headers which define
1139 templates which users normally use only for certain arguments
1140 can be instrumented to avoid exposing the template definitions
1141 to the compiler unless a macro is defined. For example, in
1142 <string>, we might have:
1144 template <class _CharT, ... > class basic_string {
1145 ... // member declarations
1147 ... // operator declarations
1150 # if _G_NO_TEMPLATE_EXPORT
1151 # include <bits/std_locale.h> // headers needed by definitions
1153 # include <bits/string.tcc> // member and global template definitions.
1157 Users who compile without specifying a strict-ISO-conforming flag
1158 would not see many of the template definitions they now see, and rely
1159 instead on ready-instantiated specializations in the library. This
1160 technique would be useful for the following substantial components:
1161 string, locale/iostreams, valarray. It would *not* be useful or
1162 usable with the following: containers, algorithms, iterators,
1163 allocator. Since these constitute a large (though decreasing)
1164 fraction of the library, the benefit the technique offers is
1167 The language specifies the semantics of the "export" keyword, but
1168 the gcc compiler does not yet support it. When it does, problems
1169 with large template inclusions can largely disappear, given some
1170 minor library reorganization, along with the need for the apparatus
1173 Overhead: Flexibility Cost
1174 --------------------------
1176 The library offers many places where users can specify operations
1177 to be performed by the library in place of defaults. Sometimes
1178 this seems to require that the library use a more-roundabout, and
1179 possibly slower, way to accomplish the default requirements than
1180 would be used otherwise.
1182 The primary protection against this overhead is thorough compiler
1183 optimization, to crush out layers of inline function interfaces.
1184 Kuck & Associates has demonstrated the practicality of this kind
1187 The second line of defense against this overhead is explicit
1188 specialization. By defining helper function templates, and writing
1189 specialized code for the default case, overhead can be eliminated
1190 for that case without sacrificing flexibility. This takes full
1191 advantage of any ability of the optimizer to crush out degenerate
1194 The library specifies many virtual functions which current linkers
1195 load even when they cannot be called. Some minor improvements to the
1196 compiler and to ld would eliminate any such overhead by simply
1197 omitting virtual functions that the complete program does not call.
1198 A prototype of this work has already been done. For targets where
1199 GNU ld is not used, a "pre-linker" could do the same job.
1201 The main areas in the standard interface where user flexibility
1202 can result in overhead are:
1204 - Allocators: Containers are specified to use user-definable
1205 allocator types and objects, making tuning for the container
1206 characteristics tricky.
1208 - Locales: the standard specifies locale objects used to implement
1209 iostream operations, involving many virtual functions which use
1210 streambuf iterators.
1212 - Algorithms and containers: these may be instantiated on any type,
1213 frequently duplicating code for identical operations.
1215 - Iostreams and strings: users are permitted to use these on their
1216 own types, and specify the operations the stream must use on these
1219 Note that these sources of overhead are _avoidable_. The techniques
1220 to avoid them are covered below.
1225 In the SGI STL, and in some other headers, many of the templates
1226 are defined "inline" -- either explicitly or by their placement
1227 in class definitions -- which should not be inline. This is a
1228 source of code bloat. Matt had remarked that he was relying on
1229 the compiler to recognize what was too big to benefit from inlining,
1230 and generate it out-of-line automatically. However, this also can
1231 result in code bloat except where the linker can eliminate the extra
1234 Fixing these cases will require an audit of all inline functions
1235 defined in the library to determine which merit inlining, and moving
1236 the rest out of line. This is an issue mainly in clauses 23, 25, and
1237 27. Of course it can be done incrementally, and we should generally
1238 accept patches that move large functions out of line and into ".tcc"
1239 files, which can later be pulled into a repository. Compiler/linker
1240 improvements to recognize very large inline functions and move them
1241 out-of-line, but shared among compilation units, could make this
1244 Pre-instantiating template specializations currently produces large
1245 amounts of dead code which bloats statically linked programs. The
1246 current state of the static library, libstdc++.a, is intolerable on
1247 this account, and will fuel further confused speculation about a need
1248 for a library "subset". A compiler improvement that treats each
1249 instantiated function as a separate object file, for linking purposes,
1250 would be one solution to this problem. An alternative would be to
1251 split up the manual instantiation files into dozens upon dozens of
1252 little files, each compiled separately, but an abortive attempt at
1253 this was done for <string> and, though it is far from complete, it
1254 is already a nuisance. A better interim solution (just until we have
1255 "export") is badly needed.
1257 When building a shared library, the current compiler/linker cannot
1258 automatically generate the instantiations needed. This creates a
1259 miserable situation; it means any time something is changed in the
1260 library, before a shared library can be built someone must manually
1261 copy the declarations of all templates that are needed by other parts
1262 of the library to an "instantiation" file, and add it to the build
1263 system to be compiled and linked to the library. This process is
1264 readily automated, and should be automated as soon as possible.
1265 Users building their own shared libraries experience identical
1268 Sharing common aspects of template definitions among instantiations
1269 can radically reduce code bloat. The compiler could help a great
1270 deal here by recognizing when a function depends on nothing about
1271 a template parameter, or only on its size, and giving the resulting
1272 function a link-name "equate" that allows it to be shared with other
1273 instantiations. Implementation code could take advantage of the
1274 capability by factoring out code that does not depend on the template
1275 argument into separate functions to be merged by the compiler.
1277 Until such a compiler optimization is implemented, much can be done
1278 manually (if tediously) in this direction. One such optimization is
1279 to derive class templates from non-template classes, and move as much
1280 implementation as possible into the base class. Another is to partial-
1281 specialize certain common instantiations, such as vector<T*>, to share
1282 code for instantiations on all types T. While these techniques work,
1283 they are far from the complete solution that a compiler improvement
1286 Overhead: Expensive Language Features
1287 -------------------------------------
1289 The main "expensive" language feature used in the standard library
1290 is exception support, which requires compiling in cleanup code with
1291 static table data to locate it, and linking in library code to use
1292 the table. For small embedded programs the amount of such library
1293 code and table data is assumed by some to be excessive. Under the
1294 "new" ABI this perception is generally exaggerated, although in some
1295 cases it may actually be excessive.
1297 To implement a library which does not use exceptions directly is
1298 not difficult given minor compiler support (to "turn off" exceptions
1299 and ignore exception constructs), and results in no great library
1300 maintenance difficulties. To be precise, given "-fno-exceptions",
1301 the compiler should treat "try" blocks as ordinary blocks, and
1302 "catch" blocks as dead code to ignore or eliminate. Compiler
1303 support is not strictly necessary, except in the case of "function
1304 try blocks"; otherwise the following macros almost suffice:
1307 #define try if (true)
1308 #define catch(X) else if (false)
1310 However, there may be a need to use function try blocks in the
1311 library implementation, and use of macros in this way can make
1312 correct diagnostics impossible. Furthermore, use of this scheme
1313 would require the library to call a function to re-throw exceptions
1314 from a try block. Implementing the above semantics in the compiler
1317 Given the support above (however implemented) it only remains to
1318 replace code that "throws" with a call to a well-documented "handler"
1319 function in a separate compilation unit which may be replaced by
1320 the user. The main source of exceptions that would be difficult
1321 for users to avoid is memory allocation failures, but users can
1322 define their own memory allocation primitives that never throw.
1323 Otherwise, the complete list of such handlers, and which library
1324 functions may call them, would be needed for users to be able to
1325 implement the necessary substitutes. (Fortunately, they have the
1331 The template capabilities of C++ offer enormous opportunities for
1332 optimizing common library operations, well beyond what would be
1333 considered "eliminating overhead". In particular, many operations
1334 done in Glibc with macros that depend on proprietary language
1335 extensions can be implemented in pristine Standard C++. For example,
1336 the chapter 25 algorithms, and even C library functions such as strchr,
1337 can be specialized for the case of static arrays of known (small) size.
1339 Detailed optimization opportunities are identified below where
1340 the component where they would appear is discussed. Of course new
1341 opportunities will be identified during implementation.
1343 Unimplemented Required Library Features
1344 ---------------------------------------
1346 The standard specifies hundreds of components, grouped broadly by
1347 chapter. These are listed in excruciating detail in the CHECKLIST
1361 Annex D backward compatibility
1363 Anyone participating in implementation of the library should obtain
1364 a copy of the standard, ISO 14882. People in the U.S. can obtain an
1365 electronic copy for US$18 from ANSI's web site. Those from other
1366 countries should visit http://www.iso.org/ to find out the location
1367 of their country's representation in ISO, in order to know who can
1370 The emphasis in the following sections is on unimplemented features
1371 and optimization opportunities.
1376 Chapter 17 concerns overall library requirements.
1378 The standard doesn't mention threads. A multi-thread (MT) extension
1379 primarily affects operators new and delete (18), allocator (20),
1380 string (21), locale (22), and iostreams (27). The common underlying
1381 support needed for this is discussed under chapter 20.
1383 The standard requirements on names from the C headers create a
1384 lot of work, mostly done. Names in the C headers must be visible
1385 in the std:: and sometimes the global namespace; the names in the
1386 two scopes must refer to the same object. More stringent is that
1387 Koenig lookup implies that any types specified as defined in std::
1388 really are defined in std::. Names optionally implemented as
1389 macros in C cannot be macros in C++. (An overview may be read at
1390 <http://www.cantrip.org/cheaders.html>). The scripts "inclosure"
1391 and "mkcshadow", and the directories shadow/ and cshadow/, are the
1392 beginning of an effort to conform in this area.
1394 A correct conforming definition of C header names based on underlying
1395 C library headers, and practical linking of conforming namespaced
1396 customer code with third-party C libraries depends ultimately on
1397 an ABI change, allowing namespaced C type names to be mangled into
1398 type names as if they were global, somewhat as C function names in a
1399 namespace, or C++ global variable names, are left unmangled. Perhaps
1400 another "extern" mode, such as 'extern "C-global"' would be an
1401 appropriate place for such type definitions. Such a type would
1402 affect mangling as follows:
1406 extern "C-global" { // or maybe just 'extern "C"'
1410 void f(A::X*); // mangles to f__FPQ21A1X
1411 void f(A::Y*); // mangles to f__FP1Y
1413 (It may be that this is really the appropriate semantics for regular
1414 'extern "C"', and 'extern "C-global"', as an extension, would not be
1415 necessary.) This would allow functions declared in non-standard C headers
1416 (and thus fixable by neither us nor users) to link properly with functions
1417 declared using C types defined in properly-namespaced headers. The
1418 problem this solves is that C headers (which C++ programmers do persist
1419 in using) frequently forward-declare C struct tags without including
1420 the header where the type is defined, as in
1425 Without some compiler accommodation, munge cannot be called by correct
1426 C++ code using a pointer to a correctly-scoped tm* value.
1428 The current C headers use the preprocessor extension "#include_next",
1429 which the compiler complains about when run "-pedantic".
1430 (Incidentally, it appears that "-fpedantic" is currently ignored,
1431 probably a bug.) The solution in the C compiler is to use
1432 "-isystem" rather than "-I", but unfortunately in g++ this seems
1433 also to wrap the whole header in an 'extern "C"' block, so it's
1434 unusable for C++ headers. The correct solution appears to be to
1435 allow the various special include-directory options, if not given
1436 an argument, to affect subsequent include-directory options additively,
1439 -pedantic -iprefix $(prefix) \
1440 -idirafter -ino-pedantic -ino-extern-c -iwithprefix -I g++-v3 \
1441 -iwithprefix -I g++-v3/ext
1443 the compiler would search $(prefix)/g++-v3 and not report
1444 pedantic warnings for files found there, but treat files in
1445 $(prefix)/g++-v3/ext pedantically. (The undocumented semantics
1446 of "-isystem" in g++ stink. Can they be rescinded? If not it
1447 must be replaced with something more rationally behaved.)
1449 All the C headers need the treatment above; in the standard these
1450 headers are mentioned in various clauses. Below, I have only
1451 mentioned those that present interesting implementation issues.
1453 The components identified as "mostly complete", below, have not been
1454 audited for conformance. In many cases where the library passes
1455 conformance tests we have non-conforming extensions that must be
1456 wrapped in #if guards for "pedantic" use, and in some cases renamed
1457 in a conforming way for continued use in the implementation regardless
1458 of conformance flags.
1460 The STL portion of the library still depends on a header
1461 stl/bits/stl_config.h full of #ifdef clauses. This apparatus
1462 should be replaced with autoconf/automake machinery.
1464 The SGI STL defines a type_traits<> template, specialized for
1465 many types in their code including the built-in numeric and
1466 pointer types and some library types, to direct optimizations of
1467 standard functions. The SGI compiler has been extended to generate
1468 specializations of this template automatically for user types,
1469 so that use of STL templates on user types can take advantage of
1470 these optimizations. Specializations for other, non-STL, types
1471 would make more optimizations possible, but extending the gcc
1472 compiler in the same way would be much better. Probably the next
1473 round of standardization will ratify this, but probably with
1474 changes, so it probably should be renamed to place it in the
1475 implementation namespace.
1477 The SGI STL also defines a large number of extensions visible in
1478 standard headers. (Other extensions that appear in separate headers
1479 have been sequestered in subdirectories ext/ and backward/.) All
1480 these extensions should be moved to other headers where possible,
1481 and in any case wrapped in a namespace (not std!), and (where kept
1482 in a standard header) girded about with macro guards. Some cannot be
1483 moved out of standard headers because they are used to implement
1484 standard features. The canonical method for accommodating these
1485 is to use a protected name, aliased in macro guards to a user-space
1486 name. Unfortunately C++ offers no satisfactory template typedef
1487 mechanism, so very ad-hoc and unsatisfactory aliasing must be used
1490 Implementation of a template typedef mechanism should have the highest
1491 priority among possible extensions, on the same level as implementation
1492 of the template "export" feature.
1494 Chapter 18 Language support
1495 ----------------------------
1497 Headers: <limits> <new> <typeinfo> <exception>
1498 C headers: <cstddef> <climits> <cfloat> <cstdarg> <csetjmp>
1499 <ctime> <csignal> <cstdlib> (also 21, 25, 26)
1501 This defines the built-in exceptions, rtti, numeric_limits<>,
1502 operator new and delete. Much of this is provided by the
1503 compiler in its static runtime library.
1505 Work to do includes defining numeric_limits<> specializations in
1506 separate files for all target architectures. Values for integer types
1507 except for bool and wchar_t are readily obtained from the C header
1508 <limits.h>, but values for the remaining numeric types (bool, wchar_t,
1509 float, double, long double) must be entered manually. This is
1510 largely dog work except for those members whose values are not
1511 easily deduced from available documentation. Also, this involves
1512 some work in target configuration to identify the correct choice of
1513 file to build against and to install.
1515 The definitions of the various operators new and delete must be
1516 made thread-safe, which depends on a portable exclusion mechanism,
1517 discussed under chapter 20. Of course there is always plenty of
1518 room for improvements to the speed of operators new and delete.
1520 <cstdarg>, in Glibc, defines some macros that gcc does not allow to
1521 be wrapped into an inline function. Probably this header will demand
1522 attention whenever a new target is chosen. The functions atexit(),
1523 exit(), and abort() in cstdlib have different semantics in C++, so
1524 must be re-implemented for C++.
1526 Chapter 19 Diagnostics
1527 -----------------------
1529 Headers: <stdexcept>
1530 C headers: <cassert> <cerrno>
1532 This defines the standard exception objects, which are "mostly complete".
1533 Cygnus has a version, and now SGI provides a slightly different one.
1534 It makes little difference which we use.
1536 The C global name "errno", which C allows to be a variable or a macro,
1537 is required in C++ to be a macro. For MT it must typically result in
1540 Chapter 20 Utilities
1541 ---------------------
1542 Headers: <utility> <functional> <memory>
1543 C header: <ctime> (also in 18)
1545 SGI STL provides "mostly complete" versions of all the components
1546 defined in this chapter. However, the auto_ptr<> implementation
1547 is known to be wrong. Furthermore, the standard definition of it
1548 is known to be unimplementable as written. A minor change to the
1549 standard would fix it, and auto_ptr<> should be adjusted to match.
1551 Multi-threading affects the allocator implementation, and there must
1552 be configuration/installation choices for different users' MT
1553 requirements. Anyway, users will want to tune allocator options
1554 to support different target conditions, MT or no.
1556 The primitives used for MT implementation should be exposed, as an
1557 extension, for users' own work. We need cross-CPU "mutex" support,
1558 multi-processor shared-memory atomic integer operations, and single-
1559 processor uninterruptible integer operations, and all three configurable
1560 to be stubbed out for non-MT use, or to use an appropriately-loaded
1561 dynamic library for the actual runtime environment, or statically
1562 compiled in for cases where the target architecture is known.
1566 Headers: <string>
1567 C headers: <cctype> <cwctype> <cstring> <cwchar> (also in 27)
1568 <cstdlib> (also in 18, 25, 26)
1570 We have "mostly-complete" char_traits<> implementations. Many of the
1571 char_traits<char> operations might be optimized further using existing
1572 proprietary language extensions.
1574 We have a "mostly-complete" basic_string<> implementation. The work
1575 to manually instantiate char and wchar_t specializations in object
1576 files to improve link-time behavior is extremely unsatisfactory,
1577 literally tripling library-build time with no commensurate improvement
1578 in static program link sizes. It must be redone. (Similar work is
1579 needed for some components in clauses 22 and 27.)
1581 Other work needed for strings is MT-safety, as discussed under the
1584 The standard C type mbstate_t from <cwchar> and used in char_traits<>
1585 must be different in C++ than in C, because in C++ the default constructor
1586 value mbstate_t() must be the "base" or "ground" sequence state.
1587 (According to the likely resolution of a recently raised Core issue,
1588 this may become unnecessary. However, there are other reasons to
1589 use a state type not as limited as whatever the C library provides.)
1590 If we might want to provide conversions from (e.g.) internally-
1591 represented EUC-wide to externally-represented Unicode, or vice-
1592 versa, the mbstate_t we choose will need to be more accommodating
1593 than what might be provided by an underlying C library.
1595 There remain some basic_string template-member functions which do
1596 not overload properly with their non-template brethren. The infamous
1597 hack akin to what was done in vector<> is needed, to conform to
1598 23.1.1 para 10. The CHECKLIST items for basic_string marked 'X',
1599 or incomplete, are so marked for this reason.
1601 Replacing the string iterators, which currently are simple character
1602 pointers, with class objects would greatly increase the safety of the
1603 client interface, and also permit a "debug" mode in which range,
1604 ownership, and validity are rigorously checked. The current use of
1605 raw pointers as string iterators is evil. vector<> iterators need the
1606 same treatment. Note that the current implementation freely mixes
1607 pointers and iterators, and that must be fixed before safer iterators
1610 Some of the functions in <cstring> are different from the C version.
1611 generally overloaded on const and non-const argument pointers. For
1612 example, in <cstring> strchr is overloaded. The functions isupper
1613 etc. in <cctype> typically implemented as macros in C are functions
1614 in C++, because they are overloaded with others of the same name
1615 defined in <locale>.
1617 Many of the functions required in <cwctype> and <cwchar> cannot be
1618 implemented using underlying C facilities on intended targets because
1619 such facilities only partly exist.
1623 Headers: <locale>
1624 C headers: <clocale>
1626 We have a "mostly complete" class locale, with the exception of
1627 code for constructing, and handling the names of, named locales.
1628 The ways that locales are named (particularly when categories
1629 (e.g. LC_TIME, LC_COLLATE) are different) varies among all target
1630 environments. This code must be written in various versions and
1631 chosen by configuration parameters.
1633 Members of many of the facets defined in <locale> are stubs. Generally,
1634 there are two sets of facets: the base class facets (which are supposed
1635 to implement the "C" locale) and the "byname" facets, which are supposed
1636 to read files to determine their behavior. The base ctype<>, collate<>,
1637 and numpunct<> facets are "mostly complete", except that the table of
1638 bitmask values used for "is" operations, and corresponding mask values,
1639 are still defined in libio and just included/linked. (We will need to
1640 implement these tables independently, soon, but should take advantage
1641 of libio where possible.) The num_put<>::put members for integer types
1642 are "mostly complete".
1644 A complete list of what has and has not been implemented may be
1645 found in CHECKLIST. However, note that the current definition of
1646 codecvt<wchar_t,char,mbstate_t> is wrong. It should simply write
1647 out the raw bytes representing the wide characters, rather than
1648 trying to convert each to a corresponding single "char" value.
1650 Some of the facets are more important than others. Specifically,
1651 the members of ctype<>, numpunct<>, num_put<>, and num_get<> facets
1652 are used by other library facilities defined in <string>, <istream>,
1653 and <ostream>, and the codecvt<> facet is used by basic_filebuf<>
1654 in <fstream>, so a conforming iostream implementation depends on
1657 The "long long" type eventually must be supported, but code mentioning
1658 it should be wrapped in #if guards to allow pedantic-mode compiling.
1660 Performance of num_put<> and num_get<> depend critically on
1661 caching computed values in ios_base objects, and on extensions
1662 to the interface with streambufs.
1664 Specifically: retrieving a copy of the locale object, extracting
1665 the needed facets, and gathering data from them, for each call to
1666 (e.g.) operator<< would be prohibitively slow. To cache format
1667 data for use by num_put<> and num_get<> we have a _Format_cache<>
1668 object stored in the ios_base::pword() array. This is constructed
1669 and initialized lazily, and is organized purely for utility. It
1670 is discarded when a new locale with different facets is imbued.
1672 Using only the public interfaces of the iterator arguments to the
1673 facet functions would limit performance by forbidding "vector-style"
1674 character operations. The streambuf iterator optimizations are
1675 described under chapter 24, but facets can also bypass the streambuf
1676 iterators via explicit specializations and operate directly on the
1677 streambufs, and use extended interfaces to get direct access to the
1678 streambuf internal buffer arrays. These extensions are mentioned
1679 under chapter 27. These optimizations are particularly important
1682 Unused virtual members of locale facets can be omitted, as mentioned
1683 above, by a smart linker.
1685 Chapter 23 Containers
1686 ----------------------
1687 Headers: <deque> <list> <queue> <stack> <vector> <map> <set> <bitset>
1689 All the components in chapter 23 are implemented in the SGI STL.
1690 They are "mostly complete"; they include a large number of
1691 nonconforming extensions which must be wrapped. Some of these
1692 are used internally and must be renamed or duplicated.
1694 The SGI components are optimized for large-memory environments. For
1695 embedded targets, different criteria might be more appropriate. Users
1696 will want to be able to tune this behavior. We should provide
1697 ways for users to compile the library with different memory usage
1700 A lot more work is needed on factoring out common code from different
1701 specializations to reduce code size here and in chapter 25. The
1702 easiest fix for this would be a compiler/ABI improvement that allows
1703 the compiler to recognize when a specialization depends only on the
1704 size (or other gross quality) of a template argument, and allow the
1705 linker to share the code with similar specializations. In its
1706 absence, many of the algorithms and containers can be partial-
1707 specialized, at least for the case of pointers, but this only solves
1708 a small part of the problem. Use of a type_traits-style template
1709 allows a few more optimization opportunities, more if the compiler
1710 can generate the specializations automatically.
1712 As an optimization, containers can specialize on the default allocator
1713 and bypass it, or take advantage of details of its implementation
1714 after it has been improved upon.
1716 Replacing the vector iterators, which currently are simple element
1717 pointers, with class objects would greatly increase the safety of the
1718 client interface, and also permit a "debug" mode in which range,
1719 ownership, and validity are rigorously checked. The current use of
1720 pointers for iterators is evil.
1722 As mentioned for chapter 24, the deque iterator is a good example of
1723 an opportunity to implement a "staged" iterator that would benefit
1724 from specializations of some algorithms.
1726 Chapter 24 Iterators
1727 ---------------------
1728 Headers: <iterator>
1730 Standard iterators are "mostly complete", with the exception of
1731 the stream iterators, which are not yet templatized on the
1732 stream type. Also, the base class template iterator<> appears
1733 to be wrong, so everything derived from it must also be wrong,
1736 The streambuf iterators (currently located in stl/bits/std_iterator.h,
1737 but should be under bits/) can be rewritten to take advantage of
1738 friendship with the streambuf implementation.
1740 Matt Austern has identified opportunities where certain iterator
1741 types, particularly including streambuf iterators and deque
1742 iterators, have a "two-stage" quality, such that an intermediate
1743 limit can be checked much more quickly than the true limit on
1744 range operations. If identified with a member of iterator_traits,
1745 algorithms may be specialized for this case. Of course the
1746 iterators that have this quality can be identified by specializing
1749 Many of the algorithms must be specialized for the streambuf
1750 iterators, to take advantage of block-mode operations, in order
1751 to allow iostream/locale operations' performance not to suffer.
1752 It may be that they could be treated as staged iterators and
1753 take advantage of those optimizations.
1755 Chapter 25 Algorithms
1756 ----------------------
1757 Headers: <algorithm>
1758 C headers: <cstdlib> (also in 18, 21, 26))
1760 The algorithms are "mostly complete". As mentioned above, they
1761 are optimized for speed at the expense of code and data size.
1763 Specializations of many of the algorithms for non-STL types would
1764 give performance improvements, but we must use great care not to
1765 interfere with fragile template overloading semantics for the
1766 standard interfaces. Conventionally the standard function template
1767 interface is an inline which delegates to a non-standard function
1768 which is then overloaded (this is already done in many places in
1769 the library). Particularly appealing opportunities for the sake of
1770 iostream performance are for copy and find applied to streambuf
1771 iterators or (as noted elsewhere) for staged iterators, of which
1772 the streambuf iterators are a good example.
1774 The bsearch and qsort functions cannot be overloaded properly as
1775 required by the standard because gcc does not yet allow overloading
1776 on the extern-"C"-ness of a function pointer.
1779 --------------------
1780 Headers: <complex> <valarray> <numeric>
1781 C headers: <cmath>, <cstdlib> (also 18, 21, 25)
1783 Numeric components: Gabriel dos Reis's valarray, Drepper's complex,
1784 and the few algorithms from the STL are "mostly done". Of course
1785 optimization opportunities abound for the numerically literate. It
1786 is not clear whether the valarray implementation really conforms
1787 fully, in the assumptions it makes about aliasing (and lack thereof)
1790 The C div() and ldiv() functions are interesting, because they are the
1791 only case where a C library function returns a class object by value.
1792 Since the C++ type div_t must be different from the underlying C type
1793 (which is in the wrong namespace) the underlying functions div() and
1794 ldiv() cannot be re-used efficiently. Fortunately they are trivial to
1797 Chapter 27 Iostreams
1798 ---------------------
1799 Headers: <iosfwd> <streambuf> <ios> <ostream> <istream> <iostream>
1800 <iomanip> <sstream> <fstream>
1801 C headers: <cstdio> <cwchar> (also in 21)
1803 Iostream is currently in a very incomplete state. <iosfwd>, <iomanip>,
1804 ios_base, and basic_ios<> are "mostly complete". basic_streambuf<> and
1805 basic_ostream<> are well along, but basic_istream<> has had little work
1806 done. The standard stream objects, <sstream> and <fstream> have been
1807 started; basic_filebuf<> "write" functions have been implemented just
1808 enough to do "hello, world".
1810 Most of the istream and ostream operators << and >> (with the exception
1811 of the op<<(integer) ones) have not been changed to use locale primitives,
1812 sentry objects, or char_traits members.
1814 All these templates should be manually instantiated for char and
1815 wchar_t in a way that links only used members into user programs.
1817 Streambuf is fertile ground for optimization extensions. An extended
1818 interface giving iterator access to its internal buffer would be very
1819 useful for other library components.
1821 Iostream operations (primarily operators << and >>) can take advantage
1822 of the case where user code has not specified a locale, and bypass locale
1823 operations entirely. The current implementation of op<</num_put<>::put,
1824 for the integer types, demonstrates how they can cache encoding details
1825 from the locale on each operation. There is lots more room for
1826 optimization in this area.
1828 The definition of the relationship between the standard streams
1829 cout et al. and stdout et al. requires something like a "stdiobuf".
1830 The SGI solution of using double-indirection to actually use a
1831 stdio FILE object for buffering is unsatisfactory, because it
1832 interferes with peephole loop optimizations.
1834 The <sstream> header work has begun. stringbuf can benefit from
1835 friendship with basic_string<> and basic_string<>::_Rep to use
1836 those objects directly as buffers, and avoid allocating and making
1839 The basic_filebuf<> template is a complex beast. It is specified to
1840 use the locale facet codecvt<> to translate characters between native
1841 files and the locale character encoding. In general this involves
1842 two buffers, one of "char" representing the file and another of
1843 "char_type", for the stream, with codecvt<> translating. The process
1844 is complicated by the variable-length nature of the translation, and
1845 the need to seek to corresponding places in the two representations.
1846 For the case of basic_filebuf<char>, when no translation is needed,
1847 a single buffer suffices. A specialized filebuf can be used to reduce
1848 code space overhead when no locale has been imbued. Matt Austern's
1849 work at SGI will be useful, perhaps directly as a source of code, or
1850 at least as an example to draw on.
1852 Filebuf, almost uniquely (cf. operator new), depends heavily on
1853 underlying environmental facilities. In current releases iostream
1854 depends fairly heavily on libio constant definitions, but it should
1855 be made independent. It also depends on operating system primitives
1856 for file operations. There is immense room for optimizations using
1857 (e.g.) mmap for reading. The shadow/ directory wraps, besides the
1858 standard C headers, the libio.h and unistd.h headers, for use mainly
1859 by filebuf. These wrappings have not been completed, though there
1860 is scaffolding in place.
1862 The encapsulation of certain C header <cstdio> names presents an
1863 interesting problem. It is possible to define an inline std::fprintf()
1864 implemented in terms of the 'extern "C"' vfprintf(), but there is no
1865 standard vfscanf() to use to implement std::fscanf(). It appears that
1866 vfscanf but be re-implemented in C++ for targets where no vfscanf
1867 extension has been defined. This is interesting in that it seems
1868 to be the only significant case in the C library where this kind of
1869 rewriting is necessary. (Of course Glibc provides the vfscanf()
1870 extension.) (The functions related to exit() must be rewritten
1876 Headers: <strstream>
1878 Annex D defines many non-library features, and many minor
1879 modifications to various headers, and a complete header.
1880 It is "mostly done", except that the libstdc++-2 <strstream>
1881 header has not been adopted into the library, or checked to
1882 verify that it matches the draft in those details that were
1883 clarified by the committee. Certainly it must at least be
1884 moved into the std namespace.
1886 We still need to wrap all the deprecated features in #if guards
1887 so that pedantic compile modes can detect their use.
1889 Nonstandard Extensions
1890 ----------------------
1891 Headers: <iostream.h> <strstream.h> <hash> <rbtree>
1892 <pthread_alloc> <stdiobuf> (etc.)
1894 User code has come to depend on a variety of nonstandard components
1895 that we must not omit. Much of this code can be adopted from
1896 libstdc++-v2 or from the SGI STL. This particularly includes
1897 <iostream.h>, <strstream.h>, and various SGI extensions such
1898 as <hash_map.h>. Many of these are already placed in the
1899 subdirectories ext/ and backward/. (Note that it is better to
1900 include them via "<backward/hash_map.h>" or "<ext/hash_map>" than
1901 to search the subdirectory itself via a "-I" directive.