1 <?xml version="1.0" encoding="UTF-8" standalone="no"?>
2 <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
3 <html xmlns="http://www.w3.org/1999/xhtml"><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8" /><title>Design Notes</title><meta name="generator" content="DocBook XSL Stylesheets V1.74.0" /><meta name="keywords" content=" ISO C++ , library " /><link rel="home" href="../spine.html" title="The GNU C++ Library Documentation" /><link rel="up" href="appendix_contributing.html" title="Appendix A. Contributing" /><link rel="prev" href="documentation_style.html" title="Documentation Style" /><link rel="next" href="appendix_porting.html" title="Appendix B. Porting and Maintenance" /></head><body><div class="navheader"><table width="100%" summary="Navigation header"><tr><th colspan="3" align="center">Design Notes</th></tr><tr><td width="20%" align="left"><a accesskey="p" href="documentation_style.html">Prev</a> </td><th width="60%" align="center">Appendix A.
6 </th><td width="20%" align="right"> <a accesskey="n" href="appendix_porting.html">Next</a></td></tr></table><hr /></div><div class="sect1" lang="en" xml:lang="en"><div class="titlepage"><div><div><h2 class="title" style="clear: both"><a id="contrib.design_notes"></a>Design Notes</h2></div></div></div><p>
7 </p><div class="literallayout"><p><br />
12 This paper is covers two major areas:<br />
14 - Features and policies not mentioned in the standard that<br />
15 the quality of the library implementation depends on, including<br />
16 extensions and "implementation-defined" features;<br />
18 - Plans for required but unimplemented library features and<br />
19 optimizations to them.<br />
24 The standard defines a large library, much larger than the standard<br />
25 C library. A naive implementation would suffer substantial overhead<br />
26 in compile time, executable size, and speed, rendering it unusable<br />
27 in many (particularly embedded) applications. The alternative demands<br />
28 care in construction, and some compiler support, but there is no<br />
29 need for library subsets.<br />
31 What are the sources of this overhead? There are four main causes:<br />
33 - The library is specified almost entirely as templates, which<br />
34 with current compilers must be included in-line, resulting in<br />
35 very slow builds as tens or hundreds of thousands of lines<br />
36 of function definitions are read for each user source file.<br />
37 Indeed, the entire SGI STL, as well as the dos Reis valarray,<br />
38 are provided purely as header files, largely for simplicity in<br />
39 porting. Iostream/locale is (or will be) as large again.<br />
41 - The library is very flexible, specifying a multitude of hooks<br />
42 where users can insert their own code in place of defaults.<br />
43 When these hooks are not used, any time and code expended to<br />
44 support that flexibility is wasted.<br />
46 - Templates are often described as causing to "code bloat". In<br />
47 practice, this refers (when it refers to anything real) to several<br />
48 independent processes. First, when a class template is manually<br />
49 instantiated in its entirely, current compilers place the definitions<br />
50 for all members in a single object file, so that a program linking<br />
51 to one member gets definitions of all. Second, template functions<br />
52 which do not actually depend on the template argument are, under<br />
53 current compilers, generated anew for each instantiation, rather<br />
54 than being shared with other instantiations. Third, some of the<br />
55 flexibility mentioned above comes from virtual functions (both in<br />
56 regular classes and template classes) which current linkers add<br />
57 to the executable file even when they manifestly cannot be called.<br />
59 - The library is specified to use a language feature, exceptions,<br />
60 which in the current gcc compiler ABI imposes a run time and<br />
61 code space cost to handle the possibility of exceptions even when<br />
62 they are not used. Under the new ABI (accessed with -fnew-abi),<br />
63 there is a space overhead and a small reduction in code efficiency<br />
64 resulting from lost optimization opportunities associated with<br />
65 non-local branches associated with exceptions.<br />
67 What can be done to eliminate this overhead? A variety of coding<br />
68 techniques, and compiler, linker and library improvements and<br />
69 extensions may be used, as covered below. Most are not difficult,<br />
70 and some are already implemented in varying degrees.<br />
72 Overhead: Compilation Time<br />
73 --------------------------<br />
75 Providing "ready-instantiated" template code in object code archives<br />
76 allows us to avoid generating and optimizing template instantiations<br />
77 in each compilation unit which uses them. However, the number of such<br />
78 instantiations that are useful to provide is limited, and anyway this<br />
79 is not enough, by itself, to minimize compilation time. In particular,<br />
80 it does not reduce time spent parsing conforming headers.<br />
82 Quicker header parsing will depend on library extensions and compiler<br />
83 improvements. One approach is some variation on the techniques<br />
84 previously marketed as "pre-compiled headers", now standardized as<br />
85 support for the "export" keyword. "Exported" template definitions<br />
86 can be placed (once) in a "repository" -- really just a library, but<br />
87 of template definitions rather than object code -- to be drawn upon<br />
88 at link time when an instantiation is needed, rather than placed in<br />
89 header files to be parsed along with every compilation unit.<br />
91 Until "export" is implemented we can put some of the lengthy template<br />
92 definitions in #if guards or alternative headers so that users can skip<br />
93 over the full definitions when they need only the ready-instantiated<br />
94 specializations.<br />
96 To be precise, this means that certain headers which define<br />
97 templates which users normally use only for certain arguments<br />
98 can be instrumented to avoid exposing the template definitions<br />
99 to the compiler unless a macro is defined. For example, in<br />
100 <string>, we might have:<br />
102 template <class _CharT, ... > class basic_string {<br />
103 ... // member declarations<br />
105 ... // operator declarations<br />
107 #ifdef _STRICT_ISO_<br />
108 # if _G_NO_TEMPLATE_EXPORT<br />
109 # include <bits/std_locale.h> // headers needed by definitions<br />
111 # include <bits/string.tcc> // member and global template definitions.<br />
115 Users who compile without specifying a strict-ISO-conforming flag<br />
116 would not see many of the template definitions they now see, and rely<br />
117 instead on ready-instantiated specializations in the library. This<br />
118 technique would be useful for the following substantial components:<br />
119 string, locale/iostreams, valarray. It would *not* be useful or<br />
120 usable with the following: containers, algorithms, iterators,<br />
121 allocator. Since these constitute a large (though decreasing)<br />
122 fraction of the library, the benefit the technique offers is<br />
125 The language specifies the semantics of the "export" keyword, but<br />
126 the gcc compiler does not yet support it. When it does, problems<br />
127 with large template inclusions can largely disappear, given some<br />
128 minor library reorganization, along with the need for the apparatus<br />
129 described above.<br />
131 Overhead: Flexibility Cost<br />
132 --------------------------<br />
134 The library offers many places where users can specify operations<br />
135 to be performed by the library in place of defaults. Sometimes<br />
136 this seems to require that the library use a more-roundabout, and<br />
137 possibly slower, way to accomplish the default requirements than<br />
138 would be used otherwise.<br />
140 The primary protection against this overhead is thorough compiler<br />
141 optimization, to crush out layers of inline function interfaces.<br />
142 Kuck & Associates has demonstrated the practicality of this kind<br />
143 of optimization.<br />
145 The second line of defense against this overhead is explicit<br />
146 specialization. By defining helper function templates, and writing<br />
147 specialized code for the default case, overhead can be eliminated<br />
148 for that case without sacrificing flexibility. This takes full<br />
149 advantage of any ability of the optimizer to crush out degenerate<br />
152 The library specifies many virtual functions which current linkers<br />
153 load even when they cannot be called. Some minor improvements to the<br />
154 compiler and to ld would eliminate any such overhead by simply<br />
155 omitting virtual functions that the complete program does not call.<br />
156 A prototype of this work has already been done. For targets where<br />
157 GNU ld is not used, a "pre-linker" could do the same job.<br />
159 The main areas in the standard interface where user flexibility<br />
160 can result in overhead are:<br />
162 - Allocators: Containers are specified to use user-definable<br />
163 allocator types and objects, making tuning for the container<br />
164 characteristics tricky.<br />
166 - Locales: the standard specifies locale objects used to implement<br />
167 iostream operations, involving many virtual functions which use<br />
168 streambuf iterators.<br />
170 - Algorithms and containers: these may be instantiated on any type,<br />
171 frequently duplicating code for identical operations.<br />
173 - Iostreams and strings: users are permitted to use these on their<br />
174 own types, and specify the operations the stream must use on these<br />
177 Note that these sources of overhead are _avoidable_. The techniques<br />
178 to avoid them are covered below.<br />
183 In the SGI STL, and in some other headers, many of the templates<br />
184 are defined "inline" -- either explicitly or by their placement<br />
185 in class definitions -- which should not be inline. This is a<br />
186 source of code bloat. Matt had remarked that he was relying on<br />
187 the compiler to recognize what was too big to benefit from inlining,<br />
188 and generate it out-of-line automatically. However, this also can<br />
189 result in code bloat except where the linker can eliminate the extra<br />
192 Fixing these cases will require an audit of all inline functions<br />
193 defined in the library to determine which merit inlining, and moving<br />
194 the rest out of line. This is an issue mainly in chapters 23, 25, and<br />
195 27. Of course it can be done incrementally, and we should generally<br />
196 accept patches that move large functions out of line and into ".tcc"<br />
197 files, which can later be pulled into a repository. Compiler/linker<br />
198 improvements to recognize very large inline functions and move them<br />
199 out-of-line, but shared among compilation units, could make this<br />
200 work unnecessary.<br />
202 Pre-instantiating template specializations currently produces large<br />
203 amounts of dead code which bloats statically linked programs. The<br />
204 current state of the static library, libstdc++.a, is intolerable on<br />
205 this account, and will fuel further confused speculation about a need<br />
206 for a library "subset". A compiler improvement that treats each<br />
207 instantiated function as a separate object file, for linking purposes,<br />
208 would be one solution to this problem. An alternative would be to<br />
209 split up the manual instantiation files into dozens upon dozens of<br />
210 little files, each compiled separately, but an abortive attempt at<br />
211 this was done for <string> and, though it is far from complete, it<br />
212 is already a nuisance. A better interim solution (just until we have<br />
213 "export") is badly needed.<br />
215 When building a shared library, the current compiler/linker cannot<br />
216 automatically generate the instantiations needed. This creates a<br />
217 miserable situation; it means any time something is changed in the<br />
218 library, before a shared library can be built someone must manually<br />
219 copy the declarations of all templates that are needed by other parts<br />
220 of the library to an "instantiation" file, and add it to the build<br />
221 system to be compiled and linked to the library. This process is<br />
222 readily automated, and should be automated as soon as possible.<br />
223 Users building their own shared libraries experience identical<br />
226 Sharing common aspects of template definitions among instantiations<br />
227 can radically reduce code bloat. The compiler could help a great<br />
228 deal here by recognizing when a function depends on nothing about<br />
229 a template parameter, or only on its size, and giving the resulting<br />
230 function a link-name "equate" that allows it to be shared with other<br />
231 instantiations. Implementation code could take advantage of the<br />
232 capability by factoring out code that does not depend on the template<br />
233 argument into separate functions to be merged by the compiler.<br />
235 Until such a compiler optimization is implemented, much can be done<br />
236 manually (if tediously) in this direction. One such optimization is<br />
237 to derive class templates from non-template classes, and move as much<br />
238 implementation as possible into the base class. Another is to partial-<br />
239 specialize certain common instantiations, such as vector<T*>, to share<br />
240 code for instantiations on all types T. While these techniques work,<br />
241 they are far from the complete solution that a compiler improvement<br />
244 Overhead: Expensive Language Features<br />
245 -------------------------------------<br />
247 The main "expensive" language feature used in the standard library<br />
248 is exception support, which requires compiling in cleanup code with<br />
249 static table data to locate it, and linking in library code to use<br />
250 the table. For small embedded programs the amount of such library<br />
251 code and table data is assumed by some to be excessive. Under the<br />
252 "new" ABI this perception is generally exaggerated, although in some<br />
253 cases it may actually be excessive.<br />
255 To implement a library which does not use exceptions directly is<br />
256 not difficult given minor compiler support (to "turn off" exceptions<br />
257 and ignore exception constructs), and results in no great library<br />
258 maintenance difficulties. To be precise, given "-fno-exceptions",<br />
259 the compiler should treat "try" blocks as ordinary blocks, and<br />
260 "catch" blocks as dead code to ignore or eliminate. Compiler<br />
261 support is not strictly necessary, except in the case of "function<br />
262 try blocks"; otherwise the following macros almost suffice:<br />
264 #define throw(X)<br />
265 #define try if (true)<br />
266 #define catch(X) else if (false)<br />
268 However, there may be a need to use function try blocks in the<br />
269 library implementation, and use of macros in this way can make<br />
270 correct diagnostics impossible. Furthermore, use of this scheme<br />
271 would require the library to call a function to re-throw exceptions<br />
272 from a try block. Implementing the above semantics in the compiler<br />
275 Given the support above (however implemented) it only remains to<br />
276 replace code that "throws" with a call to a well-documented "handler"<br />
277 function in a separate compilation unit which may be replaced by<br />
278 the user. The main source of exceptions that would be difficult<br />
279 for users to avoid is memory allocation failures, but users can<br />
280 define their own memory allocation primitives that never throw.<br />
281 Otherwise, the complete list of such handlers, and which library<br />
282 functions may call them, would be needed for users to be able to<br />
283 implement the necessary substitutes. (Fortunately, they have the<br />
289 The template capabilities of C++ offer enormous opportunities for<br />
290 optimizing common library operations, well beyond what would be<br />
291 considered "eliminating overhead". In particular, many operations<br />
292 done in Glibc with macros that depend on proprietary language<br />
293 extensions can be implemented in pristine Standard C++. For example,<br />
294 the chapter 25 algorithms, and even C library functions such as strchr,<br />
295 can be specialized for the case of static arrays of known (small) size.<br />
297 Detailed optimization opportunities are identified below where<br />
298 the component where they would appear is discussed. Of course new<br />
299 opportunities will be identified during implementation.<br />
301 Unimplemented Required Library Features<br />
302 ---------------------------------------<br />
304 The standard specifies hundreds of components, grouped broadly by<br />
305 chapter. These are listed in excruciating detail in the CHECKLIST<br />
319 Annex D backward compatibility<br />
321 Anyone participating in implementation of the library should obtain<br />
322 a copy of the standard, ISO 14882. People in the U.S. can obtain an<br />
323 electronic copy for US$18 from ANSI's web site. Those from other<br />
324 countries should visit http://www.iso.ch/ to find out the location<br />
325 of their country's representation in ISO, in order to know who can<br />
326 sell them a copy.<br />
328 The emphasis in the following sections is on unimplemented features<br />
329 and optimization opportunities.<br />
331 Chapter 17 General<br />
332 -------------------<br />
334 Chapter 17 concerns overall library requirements.<br />
336 The standard doesn't mention threads. A multi-thread (MT) extension<br />
337 primarily affects operators new and delete (18), allocator (20),<br />
338 string (21), locale (22), and iostreams (27). The common underlying<br />
339 support needed for this is discussed under chapter 20.<br />
341 The standard requirements on names from the C headers create a<br />
342 lot of work, mostly done. Names in the C headers must be visible<br />
343 in the std:: and sometimes the global namespace; the names in the<br />
344 two scopes must refer to the same object. More stringent is that<br />
345 Koenig lookup implies that any types specified as defined in std::<br />
346 really are defined in std::. Names optionally implemented as<br />
347 macros in C cannot be macros in C++. (An overview may be read at<br />
348 <http://www.cantrip.org/cheaders.html>). The scripts "inclosure"<br />
349 and "mkcshadow", and the directories shadow/ and cshadow/, are the<br />
350 beginning of an effort to conform in this area.<br />
352 A correct conforming definition of C header names based on underlying<br />
353 C library headers, and practical linking of conforming namespaced<br />
354 customer code with third-party C libraries depends ultimately on<br />
355 an ABI change, allowing namespaced C type names to be mangled into<br />
356 type names as if they were global, somewhat as C function names in a<br />
357 namespace, or C++ global variable names, are left unmangled. Perhaps<br />
358 another "extern" mode, such as 'extern "C-global"' would be an<br />
359 appropriate place for such type definitions. Such a type would<br />
360 affect mangling as follows:<br />
364 extern "C-global" { // or maybe just 'extern "C"'<br />
368 void f(A::X*); // mangles to f__FPQ21A1X<br />
369 void f(A::Y*); // mangles to f__FP1Y<br />
371 (It may be that this is really the appropriate semantics for regular<br />
372 'extern "C"', and 'extern "C-global"', as an extension, would not be<br />
373 necessary.) This would allow functions declared in non-standard C headers<br />
374 (and thus fixable by neither us nor users) to link properly with functions<br />
375 declared using C types defined in properly-namespaced headers. The<br />
376 problem this solves is that C headers (which C++ programmers do persist<br />
377 in using) frequently forward-declare C struct tags without including<br />
378 the header where the type is defined, as in<br />
381 void munge(tm*);<br />
383 Without some compiler accommodation, munge cannot be called by correct<br />
384 C++ code using a pointer to a correctly-scoped tm* value.<br />
386 The current C headers use the preprocessor extension "#include_next",<br />
387 which the compiler complains about when run "-pedantic".<br />
388 (Incidentally, it appears that "-fpedantic" is currently ignored,<br />
389 probably a bug.) The solution in the C compiler is to use<br />
390 "-isystem" rather than "-I", but unfortunately in g++ this seems<br />
391 also to wrap the whole header in an 'extern "C"' block, so it's<br />
392 unusable for C++ headers. The correct solution appears to be to<br />
393 allow the various special include-directory options, if not given<br />
394 an argument, to affect subsequent include-directory options additively,<br />
395 so that if one said<br />
397 -pedantic -iprefix $(prefix) \<br />
398 -idirafter -ino-pedantic -ino-extern-c -iwithprefix -I g++-v3 \<br />
399 -iwithprefix -I g++-v3/ext<br />
401 the compiler would search $(prefix)/g++-v3 and not report<br />
402 pedantic warnings for files found there, but treat files in<br />
403 $(prefix)/g++-v3/ext pedantically. (The undocumented semantics<br />
404 of "-isystem" in g++ stink. Can they be rescinded? If not it<br />
405 must be replaced with something more rationally behaved.)<br />
407 All the C headers need the treatment above; in the standard these<br />
408 headers are mentioned in various chapters. Below, I have only<br />
409 mentioned those that present interesting implementation issues.<br />
411 The components identified as "mostly complete", below, have not been<br />
412 audited for conformance. In many cases where the library passes<br />
413 conformance tests we have non-conforming extensions that must be<br />
414 wrapped in #if guards for "pedantic" use, and in some cases renamed<br />
415 in a conforming way for continued use in the implementation regardless<br />
416 of conformance flags.<br />
418 The STL portion of the library still depends on a header<br />
419 stl/bits/stl_config.h full of #ifdef clauses. This apparatus<br />
420 should be replaced with autoconf/automake machinery.<br />
422 The SGI STL defines a type_traits<> template, specialized for<br />
423 many types in their code including the built-in numeric and<br />
424 pointer types and some library types, to direct optimizations of<br />
425 standard functions. The SGI compiler has been extended to generate<br />
426 specializations of this template automatically for user types,<br />
427 so that use of STL templates on user types can take advantage of<br />
428 these optimizations. Specializations for other, non-STL, types<br />
429 would make more optimizations possible, but extending the gcc<br />
430 compiler in the same way would be much better. Probably the next<br />
431 round of standardization will ratify this, but probably with<br />
432 changes, so it probably should be renamed to place it in the<br />
433 implementation namespace.<br />
435 The SGI STL also defines a large number of extensions visible in<br />
436 standard headers. (Other extensions that appear in separate headers<br />
437 have been sequestered in subdirectories ext/ and backward/.) All<br />
438 these extensions should be moved to other headers where possible,<br />
439 and in any case wrapped in a namespace (not std!), and (where kept<br />
440 in a standard header) girded about with macro guards. Some cannot be<br />
441 moved out of standard headers because they are used to implement<br />
442 standard features. The canonical method for accommodating these<br />
443 is to use a protected name, aliased in macro guards to a user-space<br />
444 name. Unfortunately C++ offers no satisfactory template typedef<br />
445 mechanism, so very ad-hoc and unsatisfactory aliasing must be used<br />
448 Implementation of a template typedef mechanism should have the highest<br />
449 priority among possible extensions, on the same level as implementation<br />
450 of the template "export" feature.<br />
452 Chapter 18 Language support<br />
453 ----------------------------<br />
455 Headers: <limits> <new> <typeinfo> <exception><br />
456 C headers: <cstddef> <climits> <cfloat> <cstdarg> <csetjmp><br />
457 <ctime> <csignal> <cstdlib> (also 21, 25, 26)<br />
459 This defines the built-in exceptions, rtti, numeric_limits<>,<br />
460 operator new and delete. Much of this is provided by the<br />
461 compiler in its static runtime library.<br />
463 Work to do includes defining numeric_limits<> specializations in<br />
464 separate files for all target architectures. Values for integer types<br />
465 except for bool and wchar_t are readily obtained from the C header<br />
466 <limits.h>, but values for the remaining numeric types (bool, wchar_t,<br />
467 float, double, long double) must be entered manually. This is<br />
468 largely dog work except for those members whose values are not<br />
469 easily deduced from available documentation. Also, this involves<br />
470 some work in target configuration to identify the correct choice of<br />
471 file to build against and to install.<br />
473 The definitions of the various operators new and delete must be<br />
474 made thread-safe, which depends on a portable exclusion mechanism,<br />
475 discussed under chapter 20. Of course there is always plenty of<br />
476 room for improvements to the speed of operators new and delete.<br />
478 <cstdarg>, in Glibc, defines some macros that gcc does not allow to<br />
479 be wrapped into an inline function. Probably this header will demand<br />
480 attention whenever a new target is chosen. The functions atexit(),<br />
481 exit(), and abort() in cstdlib have different semantics in C++, so<br />
482 must be re-implemented for C++.<br />
484 Chapter 19 Diagnostics<br />
485 -----------------------<br />
487 Headers: <stdexcept><br />
488 C headers: <cassert> <cerrno><br />
490 This defines the standard exception objects, which are "mostly complete".<br />
491 Cygnus has a version, and now SGI provides a slightly different one.<br />
492 It makes little difference which we use.<br />
494 The C global name "errno", which C allows to be a variable or a macro,<br />
495 is required in C++ to be a macro. For MT it must typically result in<br />
496 a function call.<br />
498 Chapter 20 Utilities<br />
499 ---------------------<br />
500 Headers: <utility> <functional> <memory><br />
501 C header: <ctime> (also in 18)<br />
503 SGI STL provides "mostly complete" versions of all the components<br />
504 defined in this chapter. However, the auto_ptr<> implementation<br />
505 is known to be wrong. Furthermore, the standard definition of it<br />
506 is known to be unimplementable as written. A minor change to the<br />
507 standard would fix it, and auto_ptr<> should be adjusted to match.<br />
509 Multi-threading affects the allocator implementation, and there must<br />
510 be configuration/installation choices for different users' MT<br />
511 requirements. Anyway, users will want to tune allocator options<br />
512 to support different target conditions, MT or no.<br />
514 The primitives used for MT implementation should be exposed, as an<br />
515 extension, for users' own work. We need cross-CPU "mutex" support,<br />
516 multi-processor shared-memory atomic integer operations, and single-<br />
517 processor uninterruptible integer operations, and all three configurable<br />
518 to be stubbed out for non-MT use, or to use an appropriately-loaded<br />
519 dynamic library for the actual runtime environment, or statically<br />
520 compiled in for cases where the target architecture is known.<br />
522 Chapter 21 String<br />
523 ------------------<br />
524 Headers: <string><br />
525 C headers: <cctype> <cwctype> <cstring> <cwchar> (also in 27)<br />
526 <cstdlib> (also in 18, 25, 26)<br />
528 We have "mostly-complete" char_traits<> implementations. Many of the<br />
529 char_traits<char> operations might be optimized further using existing<br />
530 proprietary language extensions.<br />
532 We have a "mostly-complete" basic_string<> implementation. The work<br />
533 to manually instantiate char and wchar_t specializations in object<br />
534 files to improve link-time behavior is extremely unsatisfactory,<br />
535 literally tripling library-build time with no commensurate improvement<br />
536 in static program link sizes. It must be redone. (Similar work is<br />
537 needed for some components in chapters 22 and 27.)<br />
539 Other work needed for strings is MT-safety, as discussed under the<br />
540 chapter 20 heading.<br />
542 The standard C type mbstate_t from <cwchar> and used in char_traits<><br />
543 must be different in C++ than in C, because in C++ the default constructor<br />
544 value mbstate_t() must be the "base" or "ground" sequence state.<br />
545 (According to the likely resolution of a recently raised Core issue,<br />
546 this may become unnecessary. However, there are other reasons to<br />
547 use a state type not as limited as whatever the C library provides.)<br />
548 If we might want to provide conversions from (e.g.) internally-<br />
549 represented EUC-wide to externally-represented Unicode, or vice-<br />
550 versa, the mbstate_t we choose will need to be more accommodating<br />
551 than what might be provided by an underlying C library.<br />
553 There remain some basic_string template-member functions which do<br />
554 not overload properly with their non-template brethren. The infamous<br />
555 hack akin to what was done in vector<> is needed, to conform to<br />
556 23.1.1 para 10. The CHECKLIST items for basic_string marked 'X',<br />
557 or incomplete, are so marked for this reason.<br />
559 Replacing the string iterators, which currently are simple character<br />
560 pointers, with class objects would greatly increase the safety of the<br />
561 client interface, and also permit a "debug" mode in which range,<br />
562 ownership, and validity are rigorously checked. The current use of<br />
563 raw pointers as string iterators is evil. vector<> iterators need the<br />
564 same treatment. Note that the current implementation freely mixes<br />
565 pointers and iterators, and that must be fixed before safer iterators<br />
566 can be introduced.<br />
568 Some of the functions in <cstring> are different from the C version.<br />
569 generally overloaded on const and non-const argument pointers. For<br />
570 example, in <cstring> strchr is overloaded. The functions isupper<br />
571 etc. in <cctype> typically implemented as macros in C are functions<br />
572 in C++, because they are overloaded with others of the same name<br />
573 defined in <locale>.<br />
575 Many of the functions required in <cwctype> and <cwchar> cannot be<br />
576 implemented using underlying C facilities on intended targets because<br />
577 such facilities only partly exist.<br />
579 Chapter 22 Locale<br />
580 ------------------<br />
581 Headers: <locale><br />
582 C headers: <clocale><br />
584 We have a "mostly complete" class locale, with the exception of<br />
585 code for constructing, and handling the names of, named locales.<br />
586 The ways that locales are named (particularly when categories<br />
587 (e.g. LC_TIME, LC_COLLATE) are different) varies among all target<br />
588 environments. This code must be written in various versions and<br />
589 chosen by configuration parameters.<br />
591 Members of many of the facets defined in <locale> are stubs. Generally,<br />
592 there are two sets of facets: the base class facets (which are supposed<br />
593 to implement the "C" locale) and the "byname" facets, which are supposed<br />
594 to read files to determine their behavior. The base ctype<>, collate<>,<br />
595 and numpunct<> facets are "mostly complete", except that the table of<br />
596 bitmask values used for "is" operations, and corresponding mask values,<br />
597 are still defined in libio and just included/linked. (We will need to<br />
598 implement these tables independently, soon, but should take advantage<br />
599 of libio where possible.) The num_put<>::put members for integer types<br />
600 are "mostly complete".<br />
602 A complete list of what has and has not been implemented may be<br />
603 found in CHECKLIST. However, note that the current definition of<br />
604 codecvt<wchar_t,char,mbstate_t> is wrong. It should simply write<br />
605 out the raw bytes representing the wide characters, rather than<br />
606 trying to convert each to a corresponding single "char" value.<br />
608 Some of the facets are more important than others. Specifically,<br />
609 the members of ctype<>, numpunct<>, num_put<>, and num_get<> facets<br />
610 are used by other library facilities defined in <string>, <istream>,<br />
611 and <ostream>, and the codecvt<> facet is used by basic_filebuf<><br />
612 in <fstream>, so a conforming iostream implementation depends on<br />
615 The "long long" type eventually must be supported, but code mentioning<br />
616 it should be wrapped in #if guards to allow pedantic-mode compiling.<br />
618 Performance of num_put<> and num_get<> depend critically on<br />
619 caching computed values in ios_base objects, and on extensions<br />
620 to the interface with streambufs.<br />
622 Specifically: retrieving a copy of the locale object, extracting<br />
623 the needed facets, and gathering data from them, for each call to<br />
624 (e.g.) operator<< would be prohibitively slow. To cache format<br />
625 data for use by num_put<> and num_get<> we have a _Format_cache<><br />
626 object stored in the ios_base::pword() array. This is constructed<br />
627 and initialized lazily, and is organized purely for utility. It<br />
628 is discarded when a new locale with different facets is imbued.<br />
630 Using only the public interfaces of the iterator arguments to the<br />
631 facet functions would limit performance by forbidding "vector-style"<br />
632 character operations. The streambuf iterator optimizations are<br />
633 described under chapter 24, but facets can also bypass the streambuf<br />
634 iterators via explicit specializations and operate directly on the<br />
635 streambufs, and use extended interfaces to get direct access to the<br />
636 streambuf internal buffer arrays. These extensions are mentioned<br />
637 under chapter 27. These optimizations are particularly important<br />
638 for input parsing.<br />
640 Unused virtual members of locale facets can be omitted, as mentioned<br />
641 above, by a smart linker.<br />
643 Chapter 23 Containers<br />
644 ----------------------<br />
645 Headers: <deque> <list> <queue> <stack> <vector> <map> <set> <bitset><br />
647 All the components in chapter 23 are implemented in the SGI STL.<br />
648 They are "mostly complete"; they include a large number of<br />
649 nonconforming extensions which must be wrapped. Some of these<br />
650 are used internally and must be renamed or duplicated.<br />
652 The SGI components are optimized for large-memory environments. For<br />
653 embedded targets, different criteria might be more appropriate. Users<br />
654 will want to be able to tune this behavior. We should provide<br />
655 ways for users to compile the library with different memory usage<br />
656 characteristics.<br />
658 A lot more work is needed on factoring out common code from different<br />
659 specializations to reduce code size here and in chapter 25. The<br />
660 easiest fix for this would be a compiler/ABI improvement that allows<br />
661 the compiler to recognize when a specialization depends only on the<br />
662 size (or other gross quality) of a template argument, and allow the<br />
663 linker to share the code with similar specializations. In its<br />
664 absence, many of the algorithms and containers can be partial-<br />
665 specialized, at least for the case of pointers, but this only solves<br />
666 a small part of the problem. Use of a type_traits-style template<br />
667 allows a few more optimization opportunities, more if the compiler<br />
668 can generate the specializations automatically.<br />
670 As an optimization, containers can specialize on the default allocator<br />
671 and bypass it, or take advantage of details of its implementation<br />
672 after it has been improved upon.<br />
674 Replacing the vector iterators, which currently are simple element<br />
675 pointers, with class objects would greatly increase the safety of the<br />
676 client interface, and also permit a "debug" mode in which range,<br />
677 ownership, and validity are rigorously checked. The current use of<br />
678 pointers for iterators is evil.<br />
680 As mentioned for chapter 24, the deque iterator is a good example of<br />
681 an opportunity to implement a "staged" iterator that would benefit<br />
682 from specializations of some algorithms.<br />
684 Chapter 24 Iterators<br />
685 ---------------------<br />
686 Headers: <iterator><br />
688 Standard iterators are "mostly complete", with the exception of<br />
689 the stream iterators, which are not yet templatized on the<br />
690 stream type. Also, the base class template iterator<> appears<br />
691 to be wrong, so everything derived from it must also be wrong,<br />
694 The streambuf iterators (currently located in stl/bits/std_iterator.h,<br />
695 but should be under bits/) can be rewritten to take advantage of<br />
696 friendship with the streambuf implementation.<br />
698 Matt Austern has identified opportunities where certain iterator<br />
699 types, particularly including streambuf iterators and deque<br />
700 iterators, have a "two-stage" quality, such that an intermediate<br />
701 limit can be checked much more quickly than the true limit on<br />
702 range operations. If identified with a member of iterator_traits,<br />
703 algorithms may be specialized for this case. Of course the<br />
704 iterators that have this quality can be identified by specializing<br />
705 a traits class.<br />
707 Many of the algorithms must be specialized for the streambuf<br />
708 iterators, to take advantage of block-mode operations, in order<br />
709 to allow iostream/locale operations' performance not to suffer.<br />
710 It may be that they could be treated as staged iterators and<br />
711 take advantage of those optimizations.<br />
713 Chapter 25 Algorithms<br />
714 ----------------------<br />
715 Headers: <algorithm><br />
716 C headers: <cstdlib> (also in 18, 21, 26))<br />
718 The algorithms are "mostly complete". As mentioned above, they<br />
719 are optimized for speed at the expense of code and data size.<br />
721 Specializations of many of the algorithms for non-STL types would<br />
722 give performance improvements, but we must use great care not to<br />
723 interfere with fragile template overloading semantics for the<br />
724 standard interfaces. Conventionally the standard function template<br />
725 interface is an inline which delegates to a non-standard function<br />
726 which is then overloaded (this is already done in many places in<br />
727 the library). Particularly appealing opportunities for the sake of<br />
728 iostream performance are for copy and find applied to streambuf<br />
729 iterators or (as noted elsewhere) for staged iterators, of which<br />
730 the streambuf iterators are a good example.<br />
732 The bsearch and qsort functions cannot be overloaded properly as<br />
733 required by the standard because gcc does not yet allow overloading<br />
734 on the extern-"C"-ness of a function pointer.<br />
736 Chapter 26 Numerics<br />
737 --------------------<br />
738 Headers: <complex> <valarray> <numeric><br />
739 C headers: <cmath>, <cstdlib> (also 18, 21, 25)<br />
741 Numeric components: Gabriel dos Reis's valarray, Drepper's complex,<br />
742 and the few algorithms from the STL are "mostly done". Of course<br />
743 optimization opportunities abound for the numerically literate. It<br />
744 is not clear whether the valarray implementation really conforms<br />
745 fully, in the assumptions it makes about aliasing (and lack thereof)<br />
746 in its arguments.<br />
748 The C div() and ldiv() functions are interesting, because they are the<br />
749 only case where a C library function returns a class object by value.<br />
750 Since the C++ type div_t must be different from the underlying C type<br />
751 (which is in the wrong namespace) the underlying functions div() and<br />
752 ldiv() cannot be re-used efficiently. Fortunately they are trivial to<br />
755 Chapter 27 Iostreams<br />
756 ---------------------<br />
757 Headers: <iosfwd> <streambuf> <ios> <ostream> <istream> <iostream><br />
758 <iomanip> <sstream> <fstream><br />
759 C headers: <cstdio> <cwchar> (also in 21)<br />
761 Iostream is currently in a very incomplete state. <iosfwd>, <iomanip>,<br />
762 ios_base, and basic_ios<> are "mostly complete". basic_streambuf<> and<br />
763 basic_ostream<> are well along, but basic_istream<> has had little work<br />
764 done. The standard stream objects, <sstream> and <fstream> have been<br />
765 started; basic_filebuf<> "write" functions have been implemented just<br />
766 enough to do "hello, world".<br />
768 Most of the istream and ostream operators << and >> (with the exception<br />
769 of the op<<(integer) ones) have not been changed to use locale primitives,<br />
770 sentry objects, or char_traits members.<br />
772 All these templates should be manually instantiated for char and<br />
773 wchar_t in a way that links only used members into user programs.<br />
775 Streambuf is fertile ground for optimization extensions. An extended<br />
776 interface giving iterator access to its internal buffer would be very<br />
777 useful for other library components.<br />
779 Iostream operations (primarily operators << and >>) can take advantage<br />
780 of the case where user code has not specified a locale, and bypass locale<br />
781 operations entirely. The current implementation of op<</num_put<>::put,<br />
782 for the integer types, demonstrates how they can cache encoding details<br />
783 from the locale on each operation. There is lots more room for<br />
784 optimization in this area.<br />
786 The definition of the relationship between the standard streams<br />
787 cout et al. and stdout et al. requires something like a "stdiobuf".<br />
788 The SGI solution of using double-indirection to actually use a<br />
789 stdio FILE object for buffering is unsatisfactory, because it<br />
790 interferes with peephole loop optimizations.<br />
792 The <sstream> header work has begun. stringbuf can benefit from<br />
793 friendship with basic_string<> and basic_string<>::_Rep to use<br />
794 those objects directly as buffers, and avoid allocating and making<br />
797 The basic_filebuf<> template is a complex beast. It is specified to<br />
798 use the locale facet codecvt<> to translate characters between native<br />
799 files and the locale character encoding. In general this involves<br />
800 two buffers, one of "char" representing the file and another of<br />
801 "char_type", for the stream, with codecvt<> translating. The process<br />
802 is complicated by the variable-length nature of the translation, and<br />
803 the need to seek to corresponding places in the two representations.<br />
804 For the case of basic_filebuf<char>, when no translation is needed,<br />
805 a single buffer suffices. A specialized filebuf can be used to reduce<br />
806 code space overhead when no locale has been imbued. Matt Austern's<br />
807 work at SGI will be useful, perhaps directly as a source of code, or<br />
808 at least as an example to draw on.<br />
810 Filebuf, almost uniquely (cf. operator new), depends heavily on<br />
811 underlying environmental facilities. In current releases iostream<br />
812 depends fairly heavily on libio constant definitions, but it should<br />
813 be made independent. It also depends on operating system primitives<br />
814 for file operations. There is immense room for optimizations using<br />
815 (e.g.) mmap for reading. The shadow/ directory wraps, besides the<br />
816 standard C headers, the libio.h and unistd.h headers, for use mainly<br />
817 by filebuf. These wrappings have not been completed, though there<br />
818 is scaffolding in place.<br />
820 The encapsulation of certain C header <cstdio> names presents an<br />
821 interesting problem. It is possible to define an inline std::fprintf()<br />
822 implemented in terms of the 'extern "C"' vfprintf(), but there is no<br />
823 standard vfscanf() to use to implement std::fscanf(). It appears that<br />
824 vfscanf but be re-implemented in C++ for targets where no vfscanf<br />
825 extension has been defined. This is interesting in that it seems<br />
826 to be the only significant case in the C library where this kind of<br />
827 rewriting is necessary. (Of course Glibc provides the vfscanf()<br />
828 extension.) (The functions related to exit() must be rewritten<br />
829 for other reasons.)<br />
834 Headers: <strstream><br />
836 Annex D defines many non-library features, and many minor<br />
837 modifications to various headers, and a complete header.<br />
838 It is "mostly done", except that the libstdc++-2 <strstream><br />
839 header has not been adopted into the library, or checked to<br />
840 verify that it matches the draft in those details that were<br />
841 clarified by the committee. Certainly it must at least be<br />
842 moved into the std namespace.<br />
844 We still need to wrap all the deprecated features in #if guards<br />
845 so that pedantic compile modes can detect their use.<br />
847 Nonstandard Extensions<br />
848 ----------------------<br />
849 Headers: <iostream.h> <strstream.h> <hash> <rbtree><br />
850 <pthread_alloc> <stdiobuf> (etc.)<br />
852 User code has come to depend on a variety of nonstandard components<br />
853 that we must not omit. Much of this code can be adopted from<br />
854 libstdc++-v2 or from the SGI STL. This particularly includes<br />
855 <iostream.h>, <strstream.h>, and various SGI extensions such<br />
856 as <hash_map.h>. Many of these are already placed in the<br />
857 subdirectories ext/ and backward/. (Note that it is better to<br />
858 include them via "<backward/hash_map.h>" or "<ext/hash_map>" than<br />
859 to search the subdirectory itself via a "-I" directive.<br />
860 </p></div></div><div class="navfooter"><hr /><table width="100%" summary="Navigation footer"><tr><td width="40%" align="left"><a accesskey="p" href="documentation_style.html">Prev</a> </td><td width="20%" align="center"><a accesskey="u" href="appendix_contributing.html">Up</a></td><td width="40%" align="right"> <a accesskey="n" href="appendix_porting.html">Next</a></td></tr><tr><td width="40%" align="left" valign="top">Documentation Style </td><td width="20%" align="center"><a accesskey="h" href="../spine.html">Home</a></td><td width="40%" align="right" valign="top"> Appendix B.
861 Porting and Maintenance
863 </td></tr></table></div></body></html>