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 V1.73.2" /><meta name="keywords" content=" ISO C++ , library " /><link rel="start" href="../spine.html" title="The GNU C++ Library Documentation" /><link rel="up" href="appendix_contributing.html" title="Appendix A. Contributing" /><link rel="prev" href="bk01apas04.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="bk01apas04.html">Prev</a> </td><th width="60%" align="center">Appendix A. Contributing</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>
4 </p><div class="literallayout"><p><br />
9 This paper is covers two major areas:<br />
11 - Features and policies not mentioned in the standard that<br />
12 the quality of the library implementation depends on, including<br />
13 extensions and "implementation-defined" features;<br />
15 - Plans for required but unimplemented library features and<br />
16 optimizations to them.<br />
21 The standard defines a large library, much larger than the standard<br />
22 C library. A naive implementation would suffer substantial overhead<br />
23 in compile time, executable size, and speed, rendering it unusable<br />
24 in many (particularly embedded) applications. The alternative demands<br />
25 care in construction, and some compiler support, but there is no<br />
26 need for library subsets.<br />
28 What are the sources of this overhead? There are four main causes:<br />
30 - The library is specified almost entirely as templates, which<br />
31 with current compilers must be included in-line, resulting in<br />
32 very slow builds as tens or hundreds of thousands of lines<br />
33 of function definitions are read for each user source file.<br />
34 Indeed, the entire SGI STL, as well as the dos Reis valarray,<br />
35 are provided purely as header files, largely for simplicity in<br />
36 porting. Iostream/locale is (or will be) as large again.<br />
38 - The library is very flexible, specifying a multitude of hooks<br />
39 where users can insert their own code in place of defaults.<br />
40 When these hooks are not used, any time and code expended to<br />
41 support that flexibility is wasted.<br />
43 - Templates are often described as causing to "code bloat". In<br />
44 practice, this refers (when it refers to anything real) to several<br />
45 independent processes. First, when a class template is manually<br />
46 instantiated in its entirely, current compilers place the definitions<br />
47 for all members in a single object file, so that a program linking<br />
48 to one member gets definitions of all. Second, template functions<br />
49 which do not actually depend on the template argument are, under<br />
50 current compilers, generated anew for each instantiation, rather<br />
51 than being shared with other instantiations. Third, some of the<br />
52 flexibility mentioned above comes from virtual functions (both in<br />
53 regular classes and template classes) which current linkers add<br />
54 to the executable file even when they manifestly cannot be called.<br />
56 - The library is specified to use a language feature, exceptions,<br />
57 which in the current gcc compiler ABI imposes a run time and<br />
58 code space cost to handle the possibility of exceptions even when<br />
59 they are not used. Under the new ABI (accessed with -fnew-abi),<br />
60 there is a space overhead and a small reduction in code efficiency<br />
61 resulting from lost optimization opportunities associated with<br />
62 non-local branches associated with exceptions.<br />
64 What can be done to eliminate this overhead? A variety of coding<br />
65 techniques, and compiler, linker and library improvements and<br />
66 extensions may be used, as covered below. Most are not difficult,<br />
67 and some are already implemented in varying degrees.<br />
69 Overhead: Compilation Time<br />
70 --------------------------<br />
72 Providing "ready-instantiated" template code in object code archives<br />
73 allows us to avoid generating and optimizing template instantiations<br />
74 in each compilation unit which uses them. However, the number of such<br />
75 instantiations that are useful to provide is limited, and anyway this<br />
76 is not enough, by itself, to minimize compilation time. In particular,<br />
77 it does not reduce time spent parsing conforming headers.<br />
79 Quicker header parsing will depend on library extensions and compiler<br />
80 improvements. One approach is some variation on the techniques<br />
81 previously marketed as "pre-compiled headers", now standardized as<br />
82 support for the "export" keyword. "Exported" template definitions<br />
83 can be placed (once) in a "repository" -- really just a library, but<br />
84 of template definitions rather than object code -- to be drawn upon<br />
85 at link time when an instantiation is needed, rather than placed in<br />
86 header files to be parsed along with every compilation unit.<br />
88 Until "export" is implemented we can put some of the lengthy template<br />
89 definitions in #if guards or alternative headers so that users can skip<br />
90 over the full definitions when they need only the ready-instantiated<br />
91 specializations.<br />
93 To be precise, this means that certain headers which define<br />
94 templates which users normally use only for certain arguments<br />
95 can be instrumented to avoid exposing the template definitions<br />
96 to the compiler unless a macro is defined. For example, in<br />
97 <string>, we might have:<br />
99 template <class _CharT, ... > class basic_string {<br />
100 ... // member declarations<br />
102 ... // operator declarations<br />
104 #ifdef _STRICT_ISO_<br />
105 # if _G_NO_TEMPLATE_EXPORT<br />
106 # include <bits/std_locale.h> // headers needed by definitions<br />
108 # include <bits/string.tcc> // member and global template definitions.<br />
112 Users who compile without specifying a strict-ISO-conforming flag<br />
113 would not see many of the template definitions they now see, and rely<br />
114 instead on ready-instantiated specializations in the library. This<br />
115 technique would be useful for the following substantial components:<br />
116 string, locale/iostreams, valarray. It would *not* be useful or<br />
117 usable with the following: containers, algorithms, iterators,<br />
118 allocator. Since these constitute a large (though decreasing)<br />
119 fraction of the library, the benefit the technique offers is<br />
122 The language specifies the semantics of the "export" keyword, but<br />
123 the gcc compiler does not yet support it. When it does, problems<br />
124 with large template inclusions can largely disappear, given some<br />
125 minor library reorganization, along with the need for the apparatus<br />
126 described above.<br />
128 Overhead: Flexibility Cost<br />
129 --------------------------<br />
131 The library offers many places where users can specify operations<br />
132 to be performed by the library in place of defaults. Sometimes<br />
133 this seems to require that the library use a more-roundabout, and<br />
134 possibly slower, way to accomplish the default requirements than<br />
135 would be used otherwise.<br />
137 The primary protection against this overhead is thorough compiler<br />
138 optimization, to crush out layers of inline function interfaces.<br />
139 Kuck & Associates has demonstrated the practicality of this kind<br />
140 of optimization.<br />
142 The second line of defense against this overhead is explicit<br />
143 specialization. By defining helper function templates, and writing<br />
144 specialized code for the default case, overhead can be eliminated<br />
145 for that case without sacrificing flexibility. This takes full<br />
146 advantage of any ability of the optimizer to crush out degenerate<br />
149 The library specifies many virtual functions which current linkers<br />
150 load even when they cannot be called. Some minor improvements to the<br />
151 compiler and to ld would eliminate any such overhead by simply<br />
152 omitting virtual functions that the complete program does not call.<br />
153 A prototype of this work has already been done. For targets where<br />
154 GNU ld is not used, a "pre-linker" could do the same job.<br />
156 The main areas in the standard interface where user flexibility<br />
157 can result in overhead are:<br />
159 - Allocators: Containers are specified to use user-definable<br />
160 allocator types and objects, making tuning for the container<br />
161 characteristics tricky.<br />
163 - Locales: the standard specifies locale objects used to implement<br />
164 iostream operations, involving many virtual functions which use<br />
165 streambuf iterators.<br />
167 - Algorithms and containers: these may be instantiated on any type,<br />
168 frequently duplicating code for identical operations.<br />
170 - Iostreams and strings: users are permitted to use these on their<br />
171 own types, and specify the operations the stream must use on these<br />
174 Note that these sources of overhead are _avoidable_. The techniques<br />
175 to avoid them are covered below.<br />
180 In the SGI STL, and in some other headers, many of the templates<br />
181 are defined "inline" -- either explicitly or by their placement<br />
182 in class definitions -- which should not be inline. This is a<br />
183 source of code bloat. Matt had remarked that he was relying on<br />
184 the compiler to recognize what was too big to benefit from inlining,<br />
185 and generate it out-of-line automatically. However, this also can<br />
186 result in code bloat except where the linker can eliminate the extra<br />
189 Fixing these cases will require an audit of all inline functions<br />
190 defined in the library to determine which merit inlining, and moving<br />
191 the rest out of line. This is an issue mainly in chapters 23, 25, and<br />
192 27. Of course it can be done incrementally, and we should generally<br />
193 accept patches that move large functions out of line and into ".tcc"<br />
194 files, which can later be pulled into a repository. Compiler/linker<br />
195 improvements to recognize very large inline functions and move them<br />
196 out-of-line, but shared among compilation units, could make this<br />
197 work unnecessary.<br />
199 Pre-instantiating template specializations currently produces large<br />
200 amounts of dead code which bloats statically linked programs. The<br />
201 current state of the static library, libstdc++.a, is intolerable on<br />
202 this account, and will fuel further confused speculation about a need<br />
203 for a library "subset". A compiler improvement that treats each<br />
204 instantiated function as a separate object file, for linking purposes,<br />
205 would be one solution to this problem. An alternative would be to<br />
206 split up the manual instantiation files into dozens upon dozens of<br />
207 little files, each compiled separately, but an abortive attempt at<br />
208 this was done for <string> and, though it is far from complete, it<br />
209 is already a nuisance. A better interim solution (just until we have<br />
210 "export") is badly needed.<br />
212 When building a shared library, the current compiler/linker cannot<br />
213 automatically generate the instantiatiations needed. This creates a<br />
214 miserable situation; it means any time something is changed in the<br />
215 library, before a shared library can be built someone must manually<br />
216 copy the declarations of all templates that are needed by other parts<br />
217 of the library to an "instantiation" file, and add it to the build<br />
218 system to be compiled and linked to the library. This process is<br />
219 readily automated, and should be automated as soon as possible.<br />
220 Users building their own shared libraries experience identical<br />
223 Sharing common aspects of template definitions among instantiations<br />
224 can radically reduce code bloat. The compiler could help a great<br />
225 deal here by recognizing when a function depends on nothing about<br />
226 a template parameter, or only on its size, and giving the resulting<br />
227 function a link-name "equate" that allows it to be shared with other<br />
228 instantiations. Implementation code could take advantage of the<br />
229 capability by factoring out code that does not depend on the template<br />
230 argument into separate functions to be merged by the compiler.<br />
232 Until such a compiler optimization is implemented, much can be done<br />
233 manually (if tediously) in this direction. One such optimization is<br />
234 to derive class templates from non-template classes, and move as much<br />
235 implementation as possible into the base class. Another is to partial-<br />
236 specialize certain common instantiations, such as vector<T*>, to share<br />
237 code for instantiations on all types T. While these techniques work,<br />
238 they are far from the complete solution that a compiler improvement<br />
241 Overhead: Expensive Language Features<br />
242 -------------------------------------<br />
244 The main "expensive" language feature used in the standard library<br />
245 is exception support, which requires compiling in cleanup code with<br />
246 static table data to locate it, and linking in library code to use<br />
247 the table. For small embedded programs the amount of such library<br />
248 code and table data is assumed by some to be excessive. Under the<br />
249 "new" ABI this perception is generally exaggerated, although in some<br />
250 cases it may actually be excessive.<br />
252 To implement a library which does not use exceptions directly is<br />
253 not difficult given minor compiler support (to "turn off" exceptions<br />
254 and ignore exception constructs), and results in no great library<br />
255 maintenance difficulties. To be precise, given "-fno-exceptions",<br />
256 the compiler should treat "try" blocks as ordinary blocks, and<br />
257 "catch" blocks as dead code to ignore or eliminate. Compiler<br />
258 support is not strictly necessary, except in the case of "function<br />
259 try blocks"; otherwise the following macros almost suffice:<br />
261 #define throw(X)<br />
262 #define try if (true)<br />
263 #define catch(X) else if (false)<br />
265 However, there may be a need to use function try blocks in the<br />
266 library implementation, and use of macros in this way can make<br />
267 correct diagnostics impossible. Furthermore, use of this scheme<br />
268 would require the library to call a function to re-throw exceptions<br />
269 from a try block. Implementing the above semantics in the compiler<br />
272 Given the support above (however implemented) it only remains to<br />
273 replace code that "throws" with a call to a well-documented "handler"<br />
274 function in a separate compilation unit which may be replaced by<br />
275 the user. The main source of exceptions that would be difficult<br />
276 for users to avoid is memory allocation failures, but users can<br />
277 define their own memory allocation primitives that never throw.<br />
278 Otherwise, the complete list of such handlers, and which library<br />
279 functions may call them, would be needed for users to be able to<br />
280 implement the necessary substitutes. (Fortunately, they have the<br />
286 The template capabilities of C++ offer enormous opportunities for<br />
287 optimizing common library operations, well beyond what would be<br />
288 considered "eliminating overhead". In particular, many operations<br />
289 done in Glibc with macros that depend on proprietary language<br />
290 extensions can be implemented in pristine Standard C++. For example,<br />
291 the chapter 25 algorithms, and even C library functions such as strchr,<br />
292 can be specialized for the case of static arrays of known (small) size.<br />
294 Detailed optimization opportunities are identified below where<br />
295 the component where they would appear is discussed. Of course new<br />
296 opportunities will be identified during implementation.<br />
298 Unimplemented Required Library Features<br />
299 ---------------------------------------<br />
301 The standard specifies hundreds of components, grouped broadly by<br />
302 chapter. These are listed in excruciating detail in the CHECKLIST<br />
316 Annex D backward compatibility<br />
318 Anyone participating in implementation of the library should obtain<br />
319 a copy of the standard, ISO 14882. People in the U.S. can obtain an<br />
320 electronic copy for US$18 from ANSI's web site. Those from other<br />
321 countries should visit http://www.iso.ch/ to find out the location<br />
322 of their country's representation in ISO, in order to know who can<br />
323 sell them a copy.<br />
325 The emphasis in the following sections is on unimplemented features<br />
326 and optimization opportunities.<br />
328 Chapter 17 General<br />
329 -------------------<br />
331 Chapter 17 concerns overall library requirements.<br />
333 The standard doesn't mention threads. A multi-thread (MT) extension<br />
334 primarily affects operators new and delete (18), allocator (20),<br />
335 string (21), locale (22), and iostreams (27). The common underlying<br />
336 support needed for this is discussed under chapter 20.<br />
338 The standard requirements on names from the C headers create a<br />
339 lot of work, mostly done. Names in the C headers must be visible<br />
340 in the std:: and sometimes the global namespace; the names in the<br />
341 two scopes must refer to the same object. More stringent is that<br />
342 Koenig lookup implies that any types specified as defined in std::<br />
343 really are defined in std::. Names optionally implemented as<br />
344 macros in C cannot be macros in C++. (An overview may be read at<br />
345 <http://www.cantrip.org/cheaders.html>). The scripts "inclosure"<br />
346 and "mkcshadow", and the directories shadow/ and cshadow/, are the<br />
347 beginning of an effort to conform in this area.<br />
349 A correct conforming definition of C header names based on underlying<br />
350 C library headers, and practical linking of conforming namespaced<br />
351 customer code with third-party C libraries depends ultimately on<br />
352 an ABI change, allowing namespaced C type names to be mangled into<br />
353 type names as if they were global, somewhat as C function names in a<br />
354 namespace, or C++ global variable names, are left unmangled. Perhaps<br />
355 another "extern" mode, such as 'extern "C-global"' would be an<br />
356 appropriate place for such type definitions. Such a type would<br />
357 affect mangling as follows:<br />
361 extern "C-global" { // or maybe just 'extern "C"'<br />
365 void f(A::X*); // mangles to f__FPQ21A1X<br />
366 void f(A::Y*); // mangles to f__FP1Y<br />
368 (It may be that this is really the appropriate semantics for regular<br />
369 'extern "C"', and 'extern "C-global"', as an extension, would not be<br />
370 necessary.) This would allow functions declared in non-standard C headers<br />
371 (and thus fixable by neither us nor users) to link properly with functions<br />
372 declared using C types defined in properly-namespaced headers. The<br />
373 problem this solves is that C headers (which C++ programmers do persist<br />
374 in using) frequently forward-declare C struct tags without including<br />
375 the header where the type is defined, as in<br />
378 void munge(tm*);<br />
380 Without some compiler accommodation, munge cannot be called by correct<br />
381 C++ code using a pointer to a correctly-scoped tm* value.<br />
383 The current C headers use the preprocessor extension "#include_next",<br />
384 which the compiler complains about when run "-pedantic".<br />
385 (Incidentally, it appears that "-fpedantic" is currently ignored,<br />
386 probably a bug.) The solution in the C compiler is to use<br />
387 "-isystem" rather than "-I", but unfortunately in g++ this seems<br />
388 also to wrap the whole header in an 'extern "C"' block, so it's<br />
389 unusable for C++ headers. The correct solution appears to be to<br />
390 allow the various special include-directory options, if not given<br />
391 an argument, to affect subsequent include-directory options additively,<br />
392 so that if one said<br />
394 -pedantic -iprefix $(prefix) \<br />
395 -idirafter -ino-pedantic -ino-extern-c -iwithprefix -I g++-v3 \<br />
396 -iwithprefix -I g++-v3/ext<br />
398 the compiler would search $(prefix)/g++-v3 and not report<br />
399 pedantic warnings for files found there, but treat files in<br />
400 $(prefix)/g++-v3/ext pedantically. (The undocumented semantics<br />
401 of "-isystem" in g++ stink. Can they be rescinded? If not it<br />
402 must be replaced with something more rationally behaved.)<br />
404 All the C headers need the treatment above; in the standard these<br />
405 headers are mentioned in various chapters. Below, I have only<br />
406 mentioned those that present interesting implementation issues.<br />
408 The components identified as "mostly complete", below, have not been<br />
409 audited for conformance. In many cases where the library passes<br />
410 conformance tests we have non-conforming extensions that must be<br />
411 wrapped in #if guards for "pedantic" use, and in some cases renamed<br />
412 in a conforming way for continued use in the implementation regardless<br />
413 of conformance flags.<br />
415 The STL portion of the library still depends on a header<br />
416 stl/bits/stl_config.h full of #ifdef clauses. This apparatus<br />
417 should be replaced with autoconf/automake machinery.<br />
419 The SGI STL defines a type_traits<> template, specialized for<br />
420 many types in their code including the built-in numeric and<br />
421 pointer types and some library types, to direct optimizations of<br />
422 standard functions. The SGI compiler has been extended to generate<br />
423 specializations of this template automatically for user types,<br />
424 so that use of STL templates on user types can take advantage of<br />
425 these optimizations. Specializations for other, non-STL, types<br />
426 would make more optimizations possible, but extending the gcc<br />
427 compiler in the same way would be much better. Probably the next<br />
428 round of standardization will ratify this, but probably with<br />
429 changes, so it probably should be renamed to place it in the<br />
430 implementation namespace.<br />
432 The SGI STL also defines a large number of extensions visible in<br />
433 standard headers. (Other extensions that appear in separate headers<br />
434 have been sequestered in subdirectories ext/ and backward/.) All<br />
435 these extensions should be moved to other headers where possible,<br />
436 and in any case wrapped in a namespace (not std!), and (where kept<br />
437 in a standard header) girded about with macro guards. Some cannot be<br />
438 moved out of standard headers because they are used to implement<br />
439 standard features. The canonical method for accommodating these<br />
440 is to use a protected name, aliased in macro guards to a user-space<br />
441 name. Unfortunately C++ offers no satisfactory template typedef<br />
442 mechanism, so very ad-hoc and unsatisfactory aliasing must be used<br />
445 Implementation of a template typedef mechanism should have the highest<br />
446 priority among possible extensions, on the same level as implementation<br />
447 of the template "export" feature.<br />
449 Chapter 18 Language support<br />
450 ----------------------------<br />
452 Headers: <limits> <new> <typeinfo> <exception><br />
453 C headers: <cstddef> <climits> <cfloat> <cstdarg> <csetjmp><br />
454 <ctime> <csignal> <cstdlib> (also 21, 25, 26)<br />
456 This defines the built-in exceptions, rtti, numeric_limits<>,<br />
457 operator new and delete. Much of this is provided by the<br />
458 compiler in its static runtime library.<br />
460 Work to do includes defining numeric_limits<> specializations in<br />
461 separate files for all target architectures. Values for integer types<br />
462 except for bool and wchar_t are readily obtained from the C header<br />
463 <limits.h>, but values for the remaining numeric types (bool, wchar_t,<br />
464 float, double, long double) must be entered manually. This is<br />
465 largely dog work except for those members whose values are not<br />
466 easily deduced from available documentation. Also, this involves<br />
467 some work in target configuration to identify the correct choice of<br />
468 file to build against and to install.<br />
470 The definitions of the various operators new and delete must be<br />
471 made thread-safe, which depends on a portable exclusion mechanism,<br />
472 discussed under chapter 20. Of course there is always plenty of<br />
473 room for improvements to the speed of operators new and delete.<br />
475 <cstdarg>, in Glibc, defines some macros that gcc does not allow to<br />
476 be wrapped into an inline function. Probably this header will demand<br />
477 attention whenever a new target is chosen. The functions atexit(),<br />
478 exit(), and abort() in cstdlib have different semantics in C++, so<br />
479 must be re-implemented for C++.<br />
481 Chapter 19 Diagnostics<br />
482 -----------------------<br />
484 Headers: <stdexcept><br />
485 C headers: <cassert> <cerrno><br />
487 This defines the standard exception objects, which are "mostly complete".<br />
488 Cygnus has a version, and now SGI provides a slightly different one.<br />
489 It makes little difference which we use.<br />
491 The C global name "errno", which C allows to be a variable or a macro,<br />
492 is required in C++ to be a macro. For MT it must typically result in<br />
493 a function call.<br />
495 Chapter 20 Utilities<br />
496 ---------------------<br />
497 Headers: <utility> <functional> <memory><br />
498 C header: <ctime> (also in 18)<br />
500 SGI STL provides "mostly complete" versions of all the components<br />
501 defined in this chapter. However, the auto_ptr<> implementation<br />
502 is known to be wrong. Furthermore, the standard definition of it<br />
503 is known to be unimplementable as written. A minor change to the<br />
504 standard would fix it, and auto_ptr<> should be adjusted to match.<br />
506 Multi-threading affects the allocator implementation, and there must<br />
507 be configuration/installation choices for different users' MT<br />
508 requirements. Anyway, users will want to tune allocator options<br />
509 to support different target conditions, MT or no.<br />
511 The primitives used for MT implementation should be exposed, as an<br />
512 extension, for users' own work. We need cross-CPU "mutex" support,<br />
513 multi-processor shared-memory atomic integer operations, and single-<br />
514 processor uninterruptible integer operations, and all three configurable<br />
515 to be stubbed out for non-MT use, or to use an appropriately-loaded<br />
516 dynamic library for the actual runtime environment, or statically<br />
517 compiled in for cases where the target architecture is known.<br />
519 Chapter 21 String<br />
520 ------------------<br />
521 Headers: <string><br />
522 C headers: <cctype> <cwctype> <cstring> <cwchar> (also in 27)<br />
523 <cstdlib> (also in 18, 25, 26)<br />
525 We have "mostly-complete" char_traits<> implementations. Many of the<br />
526 char_traits<char> operations might be optimized further using existing<br />
527 proprietary language extensions.<br />
529 We have a "mostly-complete" basic_string<> implementation. The work<br />
530 to manually instantiate char and wchar_t specializations in object<br />
531 files to improve link-time behavior is extremely unsatisfactory,<br />
532 literally tripling library-build time with no commensurate improvement<br />
533 in static program link sizes. It must be redone. (Similar work is<br />
534 needed for some components in chapters 22 and 27.)<br />
536 Other work needed for strings is MT-safety, as discussed under the<br />
537 chapter 20 heading.<br />
539 The standard C type mbstate_t from <cwchar> and used in char_traits<><br />
540 must be different in C++ than in C, because in C++ the default constructor<br />
541 value mbstate_t() must be the "base" or "ground" sequence state.<br />
542 (According to the likely resolution of a recently raised Core issue,<br />
543 this may become unnecessary. However, there are other reasons to<br />
544 use a state type not as limited as whatever the C library provides.)<br />
545 If we might want to provide conversions from (e.g.) internally-<br />
546 represented EUC-wide to externally-represented Unicode, or vice-<br />
547 versa, the mbstate_t we choose will need to be more accommodating<br />
548 than what might be provided by an underlying C library.<br />
550 There remain some basic_string template-member functions which do<br />
551 not overload properly with their non-template brethren. The infamous<br />
552 hack akin to what was done in vector<> is needed, to conform to<br />
553 23.1.1 para 10. The CHECKLIST items for basic_string marked 'X',<br />
554 or incomplete, are so marked for this reason.<br />
556 Replacing the string iterators, which currently are simple character<br />
557 pointers, with class objects would greatly increase the safety of the<br />
558 client interface, and also permit a "debug" mode in which range,<br />
559 ownership, and validity are rigorously checked. The current use of<br />
560 raw pointers as string iterators is evil. vector<> iterators need the<br />
561 same treatment. Note that the current implementation freely mixes<br />
562 pointers and iterators, and that must be fixed before safer iterators<br />
563 can be introduced.<br />
565 Some of the functions in <cstring> are different from the C version.<br />
566 generally overloaded on const and non-const argument pointers. For<br />
567 example, in <cstring> strchr is overloaded. The functions isupper<br />
568 etc. in <cctype> typically implemented as macros in C are functions<br />
569 in C++, because they are overloaded with others of the same name<br />
570 defined in <locale>.<br />
572 Many of the functions required in <cwctype> and <cwchar> cannot be<br />
573 implemented using underlying C facilities on intended targets because<br />
574 such facilities only partly exist.<br />
576 Chapter 22 Locale<br />
577 ------------------<br />
578 Headers: <locale><br />
579 C headers: <clocale><br />
581 We have a "mostly complete" class locale, with the exception of<br />
582 code for constructing, and handling the names of, named locales.<br />
583 The ways that locales are named (particularly when categories<br />
584 (e.g. LC_TIME, LC_COLLATE) are different) varies among all target<br />
585 environments. This code must be written in various versions and<br />
586 chosen by configuration parameters.<br />
588 Members of many of the facets defined in <locale> are stubs. Generally,<br />
589 there are two sets of facets: the base class facets (which are supposed<br />
590 to implement the "C" locale) and the "byname" facets, which are supposed<br />
591 to read files to determine their behavior. The base ctype<>, collate<>,<br />
592 and numpunct<> facets are "mostly complete", except that the table of<br />
593 bitmask values used for "is" operations, and corresponding mask values,<br />
594 are still defined in libio and just included/linked. (We will need to<br />
595 implement these tables independently, soon, but should take advantage<br />
596 of libio where possible.) The num_put<>::put members for integer types<br />
597 are "mostly complete".<br />
599 A complete list of what has and has not been implemented may be<br />
600 found in CHECKLIST. However, note that the current definition of<br />
601 codecvt<wchar_t,char,mbstate_t> is wrong. It should simply write<br />
602 out the raw bytes representing the wide characters, rather than<br />
603 trying to convert each to a corresponding single "char" value.<br />
605 Some of the facets are more important than others. Specifically,<br />
606 the members of ctype<>, numpunct<>, num_put<>, and num_get<> facets<br />
607 are used by other library facilities defined in <string>, <istream>,<br />
608 and <ostream>, and the codecvt<> facet is used by basic_filebuf<><br />
609 in <fstream>, so a conforming iostream implementation depends on<br />
612 The "long long" type eventually must be supported, but code mentioning<br />
613 it should be wrapped in #if guards to allow pedantic-mode compiling.<br />
615 Performance of num_put<> and num_get<> depend critically on<br />
616 caching computed values in ios_base objects, and on extensions<br />
617 to the interface with streambufs.<br />
619 Specifically: retrieving a copy of the locale object, extracting<br />
620 the needed facets, and gathering data from them, for each call to<br />
621 (e.g.) operator<< would be prohibitively slow. To cache format<br />
622 data for use by num_put<> and num_get<> we have a _Format_cache<><br />
623 object stored in the ios_base::pword() array. This is constructed<br />
624 and initialized lazily, and is organized purely for utility. It<br />
625 is discarded when a new locale with different facets is imbued.<br />
627 Using only the public interfaces of the iterator arguments to the<br />
628 facet functions would limit performance by forbidding "vector-style"<br />
629 character operations. The streambuf iterator optimizations are<br />
630 described under chapter 24, but facets can also bypass the streambuf<br />
631 iterators via explicit specializations and operate directly on the<br />
632 streambufs, and use extended interfaces to get direct access to the<br />
633 streambuf internal buffer arrays. These extensions are mentioned<br />
634 under chapter 27. These optimizations are particularly important<br />
635 for input parsing.<br />
637 Unused virtual members of locale facets can be omitted, as mentioned<br />
638 above, by a smart linker.<br />
640 Chapter 23 Containers<br />
641 ----------------------<br />
642 Headers: <deque> <list> <queue> <stack> <vector> <map> <set> <bitset><br />
644 All the components in chapter 23 are implemented in the SGI STL.<br />
645 They are "mostly complete"; they include a large number of<br />
646 nonconforming extensions which must be wrapped. Some of these<br />
647 are used internally and must be renamed or duplicated.<br />
649 The SGI components are optimized for large-memory environments. For<br />
650 embedded targets, different criteria might be more appropriate. Users<br />
651 will want to be able to tune this behavior. We should provide<br />
652 ways for users to compile the library with different memory usage<br />
653 characteristics.<br />
655 A lot more work is needed on factoring out common code from different<br />
656 specializations to reduce code size here and in chapter 25. The<br />
657 easiest fix for this would be a compiler/ABI improvement that allows<br />
658 the compiler to recognize when a specialization depends only on the<br />
659 size (or other gross quality) of a template argument, and allow the<br />
660 linker to share the code with similar specializations. In its<br />
661 absence, many of the algorithms and containers can be partial-<br />
662 specialized, at least for the case of pointers, but this only solves<br />
663 a small part of the problem. Use of a type_traits-style template<br />
664 allows a few more optimization opportunities, more if the compiler<br />
665 can generate the specializations automatically.<br />
667 As an optimization, containers can specialize on the default allocator<br />
668 and bypass it, or take advantage of details of its implementation<br />
669 after it has been improved upon.<br />
671 Replacing the vector iterators, which currently are simple element<br />
672 pointers, with class objects would greatly increase the safety of the<br />
673 client interface, and also permit a "debug" mode in which range,<br />
674 ownership, and validity are rigorously checked. The current use of<br />
675 pointers for iterators is evil.<br />
677 As mentioned for chapter 24, the deque iterator is a good example of<br />
678 an opportunity to implement a "staged" iterator that would benefit<br />
679 from specializations of some algorithms.<br />
681 Chapter 24 Iterators<br />
682 ---------------------<br />
683 Headers: <iterator><br />
685 Standard iterators are "mostly complete", with the exception of<br />
686 the stream iterators, which are not yet templatized on the<br />
687 stream type. Also, the base class template iterator<> appears<br />
688 to be wrong, so everything derived from it must also be wrong,<br />
691 The streambuf iterators (currently located in stl/bits/std_iterator.h,<br />
692 but should be under bits/) can be rewritten to take advantage of<br />
693 friendship with the streambuf implementation.<br />
695 Matt Austern has identified opportunities where certain iterator<br />
696 types, particularly including streambuf iterators and deque<br />
697 iterators, have a "two-stage" quality, such that an intermediate<br />
698 limit can be checked much more quickly than the true limit on<br />
699 range operations. If identified with a member of iterator_traits,<br />
700 algorithms may be specialized for this case. Of course the<br />
701 iterators that have this quality can be identified by specializing<br />
702 a traits class.<br />
704 Many of the algorithms must be specialized for the streambuf<br />
705 iterators, to take advantage of block-mode operations, in order<br />
706 to allow iostream/locale operations' performance not to suffer.<br />
707 It may be that they could be treated as staged iterators and<br />
708 take advantage of those optimizations.<br />
710 Chapter 25 Algorithms<br />
711 ----------------------<br />
712 Headers: <algorithm><br />
713 C headers: <cstdlib> (also in 18, 21, 26))<br />
715 The algorithms are "mostly complete". As mentioned above, they<br />
716 are optimized for speed at the expense of code and data size.<br />
718 Specializations of many of the algorithms for non-STL types would<br />
719 give performance improvements, but we must use great care not to<br />
720 interfere with fragile template overloading semantics for the<br />
721 standard interfaces. Conventionally the standard function template<br />
722 interface is an inline which delegates to a non-standard function<br />
723 which is then overloaded (this is already done in many places in<br />
724 the library). Particularly appealing opportunities for the sake of<br />
725 iostream performance are for copy and find applied to streambuf<br />
726 iterators or (as noted elsewhere) for staged iterators, of which<br />
727 the streambuf iterators are a good example.<br />
729 The bsearch and qsort functions cannot be overloaded properly as<br />
730 required by the standard because gcc does not yet allow overloading<br />
731 on the extern-"C"-ness of a function pointer.<br />
733 Chapter 26 Numerics<br />
734 --------------------<br />
735 Headers: <complex> <valarray> <numeric><br />
736 C headers: <cmath>, <cstdlib> (also 18, 21, 25)<br />
738 Numeric components: Gabriel dos Reis's valarray, Drepper's complex,<br />
739 and the few algorithms from the STL are "mostly done". Of course<br />
740 optimization opportunities abound for the numerically literate. It<br />
741 is not clear whether the valarray implementation really conforms<br />
742 fully, in the assumptions it makes about aliasing (and lack thereof)<br />
743 in its arguments.<br />
745 The C div() and ldiv() functions are interesting, because they are the<br />
746 only case where a C library function returns a class object by value.<br />
747 Since the C++ type div_t must be different from the underlying C type<br />
748 (which is in the wrong namespace) the underlying functions div() and<br />
749 ldiv() cannot be re-used efficiently. Fortunately they are trivial to<br />
752 Chapter 27 Iostreams<br />
753 ---------------------<br />
754 Headers: <iosfwd> <streambuf> <ios> <ostream> <istream> <iostream><br />
755 <iomanip> <sstream> <fstream><br />
756 C headers: <cstdio> <cwchar> (also in 21)<br />
758 Iostream is currently in a very incomplete state. <iosfwd>, <iomanip>,<br />
759 ios_base, and basic_ios<> are "mostly complete". basic_streambuf<> and<br />
760 basic_ostream<> are well along, but basic_istream<> has had little work<br />
761 done. The standard stream objects, <sstream> and <fstream> have been<br />
762 started; basic_filebuf<> "write" functions have been implemented just<br />
763 enough to do "hello, world".<br />
765 Most of the istream and ostream operators << and >> (with the exception<br />
766 of the op<<(integer) ones) have not been changed to use locale primitives,<br />
767 sentry objects, or char_traits members.<br />
769 All these templates should be manually instantiated for char and<br />
770 wchar_t in a way that links only used members into user programs.<br />
772 Streambuf is fertile ground for optimization extensions. An extended<br />
773 interface giving iterator access to its internal buffer would be very<br />
774 useful for other library components.<br />
776 Iostream operations (primarily operators << and >>) can take advantage<br />
777 of the case where user code has not specified a locale, and bypass locale<br />
778 operations entirely. The current implementation of op<</num_put<>::put,<br />
779 for the integer types, demonstrates how they can cache encoding details<br />
780 from the locale on each operation. There is lots more room for<br />
781 optimization in this area.<br />
783 The definition of the relationship between the standard streams<br />
784 cout et al. and stdout et al. requires something like a "stdiobuf".<br />
785 The SGI solution of using double-indirection to actually use a<br />
786 stdio FILE object for buffering is unsatisfactory, because it<br />
787 interferes with peephole loop optimizations.<br />
789 The <sstream> header work has begun. stringbuf can benefit from<br />
790 friendship with basic_string<> and basic_string<>::_Rep to use<br />
791 those objects directly as buffers, and avoid allocating and making<br />
794 The basic_filebuf<> template is a complex beast. It is specified to<br />
795 use the locale facet codecvt<> to translate characters between native<br />
796 files and the locale character encoding. In general this involves<br />
797 two buffers, one of "char" representing the file and another of<br />
798 "char_type", for the stream, with codecvt<> translating. The process<br />
799 is complicated by the variable-length nature of the translation, and<br />
800 the need to seek to corresponding places in the two representations.<br />
801 For the case of basic_filebuf<char>, when no translation is needed,<br />
802 a single buffer suffices. A specialized filebuf can be used to reduce<br />
803 code space overhead when no locale has been imbued. Matt Austern's<br />
804 work at SGI will be useful, perhaps directly as a source of code, or<br />
805 at least as an example to draw on.<br />
807 Filebuf, almost uniquely (cf. operator new), depends heavily on<br />
808 underlying environmental facilities. In current releases iostream<br />
809 depends fairly heavily on libio constant definitions, but it should<br />
810 be made independent. It also depends on operating system primitives<br />
811 for file operations. There is immense room for optimizations using<br />
812 (e.g.) mmap for reading. The shadow/ directory wraps, besides the<br />
813 standard C headers, the libio.h and unistd.h headers, for use mainly<br />
814 by filebuf. These wrappings have not been completed, though there<br />
815 is scaffolding in place.<br />
817 The encapulation of certain C header <cstdio> names presents an<br />
818 interesting problem. It is possible to define an inline std::fprintf()<br />
819 implemented in terms of the 'extern "C"' vfprintf(), but there is no<br />
820 standard vfscanf() to use to implement std::fscanf(). It appears that<br />
821 vfscanf but be re-implemented in C++ for targets where no vfscanf<br />
822 extension has been defined. This is interesting in that it seems<br />
823 to be the only significant case in the C library where this kind of<br />
824 rewriting is necessary. (Of course Glibc provides the vfscanf()<br />
825 extension.) (The functions related to exit() must be rewritten<br />
826 for other reasons.)<br />
831 Headers: <strstream><br />
833 Annex D defines many non-library features, and many minor<br />
834 modifications to various headers, and a complete header.<br />
835 It is "mostly done", except that the libstdc++-2 <strstream><br />
836 header has not been adopted into the library, or checked to<br />
837 verify that it matches the draft in those details that were<br />
838 clarified by the committee. Certainly it must at least be<br />
839 moved into the std namespace.<br />
841 We still need to wrap all the deprecated features in #if guards<br />
842 so that pedantic compile modes can detect their use.<br />
844 Nonstandard Extensions<br />
845 ----------------------<br />
846 Headers: <iostream.h> <strstream.h> <hash> <rbtree><br />
847 <pthread_alloc> <stdiobuf> (etc.)<br />
849 User code has come to depend on a variety of nonstandard components<br />
850 that we must not omit. Much of this code can be adopted from<br />
851 libstdc++-v2 or from the SGI STL. This particularly includes<br />
852 <iostream.h>, <strstream.h>, and various SGI extensions such<br />
853 as <hash_map.h>. Many of these are already placed in the<br />
854 subdirectories ext/ and backward/. (Note that it is better to<br />
855 include them via "<backward/hash_map.h>" or "<ext/hash_map>" than<br />
856 to search the subdirectory itself via a "-I" directive.<br />
857 </p></div></div><div class="navfooter"><hr /><table width="100%" summary="Navigation footer"><tr><td width="40%" align="left"><a accesskey="p" href="bk01apas04.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. Porting and Maintenance</td></tr></table></div></body></html>
OSDN Git Service
