1 |
20 |
jlechner |
|
2 |
|
|
Standard C++ Library Design Document
|
3 |
|
|
------------------------------------
|
4 |
|
|
|
5 |
|
|
This is an overview of libstdc++-v3, with particular attention
|
6 |
|
|
to projects to be done and how they fit into the whole.
|
7 |
|
|
|
8 |
|
|
The Library
|
9 |
|
|
-----------
|
10 |
|
|
|
11 |
|
|
This paper is covers two major areas:
|
12 |
|
|
|
13 |
|
|
- Features and policies not mentioned in the standard that
|
14 |
|
|
the quality of the library implementation depends on, including
|
15 |
|
|
extensions and "implementation-defined" features;
|
16 |
|
|
|
17 |
|
|
- Plans for required but unimplemented library features and
|
18 |
|
|
optimizations to them.
|
19 |
|
|
|
20 |
|
|
Overhead
|
21 |
|
|
--------
|
22 |
|
|
|
23 |
|
|
The standard defines a large library, much larger than the standard
|
24 |
|
|
C library. A naive implementation would suffer substantial overhead
|
25 |
|
|
in compile time, executable size, and speed, rendering it unusable
|
26 |
|
|
in many (particularly embedded) applications. The alternative demands
|
27 |
|
|
care in construction, and some compiler support, but there is no
|
28 |
|
|
need for library subsets.
|
29 |
|
|
|
30 |
|
|
What are the sources of this overhead? There are four main causes:
|
31 |
|
|
|
32 |
|
|
- The library is specified almost entirely as templates, which
|
33 |
|
|
with current compilers must be included in-line, resulting in
|
34 |
|
|
very slow builds as tens or hundreds of thousands of lines
|
35 |
|
|
of function definitions are read for each user source file.
|
36 |
|
|
Indeed, the entire SGI STL, as well as the dos Reis valarray,
|
37 |
|
|
are provided purely as header files, largely for simplicity in
|
38 |
|
|
porting. Iostream/locale is (or will be) as large again.
|
39 |
|
|
|
40 |
|
|
- The library is very flexible, specifying a multitude of hooks
|
41 |
|
|
where users can insert their own code in place of defaults.
|
42 |
|
|
When these hooks are not used, any time and code expended to
|
43 |
|
|
support that flexibility is wasted.
|
44 |
|
|
|
45 |
|
|
- Templates are often described as causing to "code bloat". In
|
46 |
|
|
practice, this refers (when it refers to anything real) to several
|
47 |
|
|
independent processes. First, when a class template is manually
|
48 |
|
|
instantiated in its entirely, current compilers place the definitions
|
49 |
|
|
for all members in a single object file, so that a program linking
|
50 |
|
|
to one member gets definitions of all. Second, template functions
|
51 |
|
|
which do not actually depend on the template argument are, under
|
52 |
|
|
current compilers, generated anew for each instantiation, rather
|
53 |
|
|
than being shared with other instantiations. Third, some of the
|
54 |
|
|
flexibility mentioned above comes from virtual functions (both in
|
55 |
|
|
regular classes and template classes) which current linkers add
|
56 |
|
|
to the executable file even when they manifestly cannot be called.
|
57 |
|
|
|
58 |
|
|
- The library is specified to use a language feature, exceptions,
|
59 |
|
|
which in the current gcc compiler ABI imposes a run time and
|
60 |
|
|
code space cost to handle the possibility of exceptions even when
|
61 |
|
|
they are not used. Under the new ABI (accessed with -fnew-abi),
|
62 |
|
|
there is a space overhead and a small reduction in code efficiency
|
63 |
|
|
resulting from lost optimization opportunities associated with
|
64 |
|
|
non-local branches associated with exceptions.
|
65 |
|
|
|
66 |
|
|
What can be done to eliminate this overhead? A variety of coding
|
67 |
|
|
techniques, and compiler, linker and library improvements and
|
68 |
|
|
extensions may be used, as covered below. Most are not difficult,
|
69 |
|
|
and some are already implemented in varying degrees.
|
70 |
|
|
|
71 |
|
|
Overhead: Compilation Time
|
72 |
|
|
--------------------------
|
73 |
|
|
|
74 |
|
|
Providing "ready-instantiated" template code in object code archives
|
75 |
|
|
allows us to avoid generating and optimizing template instantiations
|
76 |
|
|
in each compilation unit which uses them. However, the number of such
|
77 |
|
|
instantiations that are useful to provide is limited, and anyway this
|
78 |
|
|
is not enough, by itself, to minimize compilation time. In particular,
|
79 |
|
|
it does not reduce time spent parsing conforming headers.
|
80 |
|
|
|
81 |
|
|
Quicker header parsing will depend on library extensions and compiler
|
82 |
|
|
improvements. One approach is some variation on the techniques
|
83 |
|
|
previously marketed as "pre-compiled headers", now standardized as
|
84 |
|
|
support for the "export" keyword. "Exported" template definitions
|
85 |
|
|
can be placed (once) in a "repository" -- really just a library, but
|
86 |
|
|
of template definitions rather than object code -- to be drawn upon
|
87 |
|
|
at link time when an instantiation is needed, rather than placed in
|
88 |
|
|
header files to be parsed along with every compilation unit.
|
89 |
|
|
|
90 |
|
|
Until "export" is implemented we can put some of the lengthy template
|
91 |
|
|
definitions in #if guards or alternative headers so that users can skip
|
92 |
|
|
over the the full definitions when they need only the ready-instantiated
|
93 |
|
|
specializations.
|
94 |
|
|
|
95 |
|
|
To be precise, this means that certain headers which define
|
96 |
|
|
templates which users normally use only for certain arguments
|
97 |
|
|
can be instrumented to avoid exposing the template definitions
|
98 |
|
|
to the compiler unless a macro is defined. For example, in
|
99 |
|
|
, we might have:
|
100 |
|
|
|
101 |
|
|
template class basic_string {
|
102 |
|
|
... // member declarations
|
103 |
|
|
};
|
104 |
|
|
... // operator declarations
|
105 |
|
|
|
106 |
|
|
#ifdef _STRICT_ISO_
|
107 |
|
|
# if _G_NO_TEMPLATE_EXPORT
|
108 |
|
|
# include // headers needed by definitions
|
109 |
|
|
# ...
|
110 |
|
|
# include // member and global template definitions.
|
111 |
|
|
# endif
|
112 |
|
|
#endif
|
113 |
|
|
|
114 |
|
|
Users who compile without specifying a strict-ISO-conforming flag
|
115 |
|
|
would not see many of the template definitions they now see, and rely
|
116 |
|
|
instead on ready-instantiated specializations in the library. This
|
117 |
|
|
technique would be useful for the following substantial components:
|
118 |
|
|
string, locale/iostreams, valarray. It would *not* be useful or
|
119 |
|
|
usable with the following: containers, algorithms, iterators,
|
120 |
|
|
allocator. Since these constitute a large (though decreasing)
|
121 |
|
|
fraction of the library, the benefit the technique offers is
|
122 |
|
|
limited.
|
123 |
|
|
|
124 |
|
|
The language specifies the semantics of the "export" keyword, but
|
125 |
|
|
the gcc compiler does not yet support it. When it does, problems
|
126 |
|
|
with large template inclusions can largely disappear, given some
|
127 |
|
|
minor library reorganization, along with the need for the apparatus
|
128 |
|
|
described above.
|
129 |
|
|
|
130 |
|
|
Overhead: Flexibility Cost
|
131 |
|
|
--------------------------
|
132 |
|
|
|
133 |
|
|
The library offers many places where users can specify operations
|
134 |
|
|
to be performed by the library in place of defaults. Sometimes
|
135 |
|
|
this seems to require that the library use a more-roundabout, and
|
136 |
|
|
possibly slower, way to accomplish the default requirements than
|
137 |
|
|
would be used otherwise.
|
138 |
|
|
|
139 |
|
|
The primary protection against this overhead is thorough compiler
|
140 |
|
|
optimization, to crush out layers of inline function interfaces.
|
141 |
|
|
Kuck & Associates has demonstrated the practicality of this kind
|
142 |
|
|
of optimization.
|
143 |
|
|
|
144 |
|
|
The second line of defense against this overhead is explicit
|
145 |
|
|
specialization. By defining helper function templates, and writing
|
146 |
|
|
specialized code for the default case, overhead can be eliminated
|
147 |
|
|
for that case without sacrificing flexibility. This takes full
|
148 |
|
|
advantage of any ability of the optimizer to crush out degenerate
|
149 |
|
|
code.
|
150 |
|
|
|
151 |
|
|
The library specifies many virtual functions which current linkers
|
152 |
|
|
load even when they cannot be called. Some minor improvements to the
|
153 |
|
|
compiler and to ld would eliminate any such overhead by simply
|
154 |
|
|
omitting virtual functions that the complete program does not call.
|
155 |
|
|
A prototype of this work has already been done. For targets where
|
156 |
|
|
GNU ld is not used, a "pre-linker" could do the same job.
|
157 |
|
|
|
158 |
|
|
The main areas in the standard interface where user flexibility
|
159 |
|
|
can result in overhead are:
|
160 |
|
|
|
161 |
|
|
- Allocators: Containers are specified to use user-definable
|
162 |
|
|
allocator types and objects, making tuning for the container
|
163 |
|
|
characteristics tricky.
|
164 |
|
|
|
165 |
|
|
- Locales: the standard specifies locale objects used to implement
|
166 |
|
|
iostream operations, involving many virtual functions which use
|
167 |
|
|
streambuf iterators.
|
168 |
|
|
|
169 |
|
|
- Algorithms and containers: these may be instantiated on any type,
|
170 |
|
|
frequently duplicating code for identical operations.
|
171 |
|
|
|
172 |
|
|
- Iostreams and strings: users are permitted to use these on their
|
173 |
|
|
own types, and specify the operations the stream must use on these
|
174 |
|
|
types.
|
175 |
|
|
|
176 |
|
|
Note that these sources of overhead are _avoidable_. The techniques
|
177 |
|
|
to avoid them are covered below.
|
178 |
|
|
|
179 |
|
|
Code Bloat
|
180 |
|
|
----------
|
181 |
|
|
|
182 |
|
|
In the SGI STL, and in some other headers, many of the templates
|
183 |
|
|
are defined "inline" -- either explicitly or by their placement
|
184 |
|
|
in class definitions -- which should not be inline. This is a
|
185 |
|
|
source of code bloat. Matt had remarked that he was relying on
|
186 |
|
|
the compiler to recognize what was too big to benefit from inlining,
|
187 |
|
|
and generate it out-of-line automatically. However, this also can
|
188 |
|
|
result in code bloat except where the linker can eliminate the extra
|
189 |
|
|
copies.
|
190 |
|
|
|
191 |
|
|
Fixing these cases will require an audit of all inline functions
|
192 |
|
|
defined in the library to determine which merit inlining, and moving
|
193 |
|
|
the rest out of line. This is an issue mainly in chapters 23, 25, and
|
194 |
|
|
27. Of course it can be done incrementally, and we should generally
|
195 |
|
|
accept patches that move large functions out of line and into ".tcc"
|
196 |
|
|
files, which can later be pulled into a repository. Compiler/linker
|
197 |
|
|
improvements to recognize very large inline functions and move them
|
198 |
|
|
out-of-line, but shared among compilation units, could make this
|
199 |
|
|
work unnecessary.
|
200 |
|
|
|
201 |
|
|
Pre-instantiating template specializations currently produces large
|
202 |
|
|
amounts of dead code which bloats statically linked programs. The
|
203 |
|
|
current state of the static library, libstdc++.a, is intolerable on
|
204 |
|
|
this account, and will fuel further confused speculation about a need
|
205 |
|
|
for a library "subset". A compiler improvement that treats each
|
206 |
|
|
instantiated function as a separate object file, for linking purposes,
|
207 |
|
|
would be one solution to this problem. An alternative would be to
|
208 |
|
|
split up the manual instantiation files into dozens upon dozens of
|
209 |
|
|
little files, each compiled separately, but an abortive attempt at
|
210 |
|
|
this was done for and, though it is far from complete, it
|
211 |
|
|
is already a nuisance. A better interim solution (just until we have
|
212 |
|
|
"export") is badly needed.
|
213 |
|
|
|
214 |
|
|
When building a shared library, the current compiler/linker cannot
|
215 |
|
|
automatically generate the instantiatiations needed. This creates a
|
216 |
|
|
miserable situation; it means any time something is changed in the
|
217 |
|
|
library, before a shared library can be built someone must manually
|
218 |
|
|
copy the declarations of all templates that are needed by other parts
|
219 |
|
|
of the library to an "instantiation" file, and add it to the build
|
220 |
|
|
system to be compiled and linked to the library. This process is
|
221 |
|
|
readily automated, and should be automated as soon as possible.
|
222 |
|
|
Users building their own shared libraries experience identical
|
223 |
|
|
frustrations.
|
224 |
|
|
|
225 |
|
|
Sharing common aspects of template definitions among instantiations
|
226 |
|
|
can radically reduce code bloat. The compiler could help a great
|
227 |
|
|
deal here by recognizing when a function depends on nothing about
|
228 |
|
|
a template parameter, or only on its size, and giving the resulting
|
229 |
|
|
function a link-name "equate" that allows it to be shared with other
|
230 |
|
|
instantiations. Implementation code could take advantage of the
|
231 |
|
|
capability by factoring out code that does not depend on the template
|
232 |
|
|
argument into separate functions to be merged by the compiler.
|
233 |
|
|
|
234 |
|
|
Until such a compiler optimization is implemented, much can be done
|
235 |
|
|
manually (if tediously) in this direction. One such optimization is
|
236 |
|
|
to derive class templates from non-template classes, and move as much
|
237 |
|
|
implementation as possible into the base class. Another is to partial-
|
238 |
|
|
specialize certain common instantiations, such as vector, to share
|
239 |
|
|
code for instantiations on all types T. While these techniques work,
|
240 |
|
|
they are far from the complete solution that a compiler improvement
|
241 |
|
|
would afford.
|
242 |
|
|
|
243 |
|
|
Overhead: Expensive Language Features
|
244 |
|
|
-------------------------------------
|
245 |
|
|
|
246 |
|
|
The main "expensive" language feature used in the standard library
|
247 |
|
|
is exception support, which requires compiling in cleanup code with
|
248 |
|
|
static table data to locate it, and linking in library code to use
|
249 |
|
|
the table. For small embedded programs the amount of such library
|
250 |
|
|
code and table data is assumed by some to be excessive. Under the
|
251 |
|
|
"new" ABI this perception is generally exaggerated, although in some
|
252 |
|
|
cases it may actually be excessive.
|
253 |
|
|
|
254 |
|
|
To implement a library which does not use exceptions directly is
|
255 |
|
|
not difficult given minor compiler support (to "turn off" exceptions
|
256 |
|
|
and ignore exception constructs), and results in no great library
|
257 |
|
|
maintenance difficulties. To be precise, given "-fno-exceptions",
|
258 |
|
|
the compiler should treat "try" blocks as ordinary blocks, and
|
259 |
|
|
"catch" blocks as dead code to ignore or eliminate. Compiler
|
260 |
|
|
support is not strictly necessary, except in the case of "function
|
261 |
|
|
try blocks"; otherwise the following macros almost suffice:
|
262 |
|
|
|
263 |
|
|
#define throw(X)
|
264 |
|
|
#define try if (true)
|
265 |
|
|
#define catch(X) else if (false)
|
266 |
|
|
|
267 |
|
|
However, there may be a need to use function try blocks in the
|
268 |
|
|
library implementation, and use of macros in this way can make
|
269 |
|
|
correct diagnostics impossible. Furthermore, use of this scheme
|
270 |
|
|
would require the library to call a function to re-throw exceptions
|
271 |
|
|
from a try block. Implementing the above semantics in the compiler
|
272 |
|
|
is preferable.
|
273 |
|
|
|
274 |
|
|
Given the support above (however implemented) it only remains to
|
275 |
|
|
replace code that "throws" with a call to a well-documented "handler"
|
276 |
|
|
function in a separate compilation unit which may be replaced by
|
277 |
|
|
the user. The main source of exceptions that would be difficult
|
278 |
|
|
for users to avoid is memory allocation failures, but users can
|
279 |
|
|
define their own memory allocation primitives that never throw.
|
280 |
|
|
Otherwise, the complete list of such handlers, and which library
|
281 |
|
|
functions may call them, would be needed for users to be able to
|
282 |
|
|
implement the necessary substitutes. (Fortunately, they have the
|
283 |
|
|
source code.)
|
284 |
|
|
|
285 |
|
|
Opportunities
|
286 |
|
|
-------------
|
287 |
|
|
|
288 |
|
|
The template capabilities of C++ offer enormous opportunities for
|
289 |
|
|
optimizing common library operations, well beyond what would be
|
290 |
|
|
considered "eliminating overhead". In particular, many operations
|
291 |
|
|
done in Glibc with macros that depend on proprietary language
|
292 |
|
|
extensions can be implemented in pristine Standard C++. For example,
|
293 |
|
|
the chapter 25 algorithms, and even C library functions such as strchr,
|
294 |
|
|
can be specialized for the case of static arrays of known (small) size.
|
295 |
|
|
|
296 |
|
|
Detailed optimization opportunities are identified below where
|
297 |
|
|
the component where they would appear is discussed. Of course new
|
298 |
|
|
opportunities will be identified during implementation.
|
299 |
|
|
|
300 |
|
|
Unimplemented Required Library Features
|
301 |
|
|
---------------------------------------
|
302 |
|
|
|
303 |
|
|
The standard specifies hundreds of components, grouped broadly by
|
304 |
|
|
chapter. These are listed in excruciating detail in the CHECKLIST
|
305 |
|
|
file.
|
306 |
|
|
|
307 |
|
|
17 general
|
308 |
|
|
18 support
|
309 |
|
|
19 diagnostics
|
310 |
|
|
20 utilities
|
311 |
|
|
21 string
|
312 |
|
|
22 locale
|
313 |
|
|
23 containers
|
314 |
|
|
24 iterators
|
315 |
|
|
25 algorithms
|
316 |
|
|
26 numerics
|
317 |
|
|
27 iostreams
|
318 |
|
|
Annex D backward compatibility
|
319 |
|
|
|
320 |
|
|
Anyone participating in implementation of the library should obtain
|
321 |
|
|
a copy of the standard, ISO 14882. People in the U.S. can obtain an
|
322 |
|
|
electronic copy for US$18 from ANSI's web site. Those from other
|
323 |
|
|
countries should visit http://www.iso.ch/ to find out the location
|
324 |
|
|
of their country's representation in ISO, in order to know who can
|
325 |
|
|
sell them a copy.
|
326 |
|
|
|
327 |
|
|
The emphasis in the following sections is on unimplemented features
|
328 |
|
|
and optimization opportunities.
|
329 |
|
|
|
330 |
|
|
Chapter 17 General
|
331 |
|
|
-------------------
|
332 |
|
|
|
333 |
|
|
Chapter 17 concerns overall library requirements.
|
334 |
|
|
|
335 |
|
|
The standard doesn't mention threads. A multi-thread (MT) extension
|
336 |
|
|
primarily affects operators new and delete (18), allocator (20),
|
337 |
|
|
string (21), locale (22), and iostreams (27). The common underlying
|
338 |
|
|
support needed for this is discussed under chapter 20.
|
339 |
|
|
|
340 |
|
|
The standard requirements on names from the C headers create a
|
341 |
|
|
lot of work, mostly done. Names in the C headers must be visible
|
342 |
|
|
in the std:: and sometimes the global namespace; the names in the
|
343 |
|
|
two scopes must refer to the same object. More stringent is that
|
344 |
|
|
Koenig lookup implies that any types specified as defined in std::
|
345 |
|
|
really are defined in std::. Names optionally implemented as
|
346 |
|
|
macros in C cannot be macros in C++. (An overview may be read at
|
347 |
|
|
). The scripts "inclosure"
|
348 |
|
|
and "mkcshadow", and the directories shadow/ and cshadow/, are the
|
349 |
|
|
beginning of an effort to conform in this area.
|
350 |
|
|
|
351 |
|
|
A correct conforming definition of C header names based on underlying
|
352 |
|
|
C library headers, and practical linking of conforming namespaced
|
353 |
|
|
customer code with third-party C libraries depends ultimately on
|
354 |
|
|
an ABI change, allowing namespaced C type names to be mangled into
|
355 |
|
|
type names as if they were global, somewhat as C function names in a
|
356 |
|
|
namespace, or C++ global variable names, are left unmangled. Perhaps
|
357 |
|
|
another "extern" mode, such as 'extern "C-global"' would be an
|
358 |
|
|
appropriate place for such type definitions. Such a type would
|
359 |
|
|
affect mangling as follows:
|
360 |
|
|
|
361 |
|
|
namespace A {
|
362 |
|
|
struct X {};
|
363 |
|
|
extern "C-global" { // or maybe just 'extern "C"'
|
364 |
|
|
struct Y {};
|
365 |
|
|
};
|
366 |
|
|
}
|
367 |
|
|
void f(A::X*); // mangles to f__FPQ21A1X
|
368 |
|
|
void f(A::Y*); // mangles to f__FP1Y
|
369 |
|
|
|
370 |
|
|
(It may be that this is really the appropriate semantics for regular
|
371 |
|
|
'extern "C"', and 'extern "C-global"', as an extension, would not be
|
372 |
|
|
necessary.) This would allow functions declared in non-standard C headers
|
373 |
|
|
(and thus fixable by neither us nor users) to link properly with functions
|
374 |
|
|
declared using C types defined in properly-namespaced headers. The
|
375 |
|
|
problem this solves is that C headers (which C++ programmers do persist
|
376 |
|
|
in using) frequently forward-declare C struct tags without including
|
377 |
|
|
the header where the type is defined, as in
|
378 |
|
|
|
379 |
|
|
struct tm;
|
380 |
|
|
void munge(tm*);
|
381 |
|
|
|
382 |
|
|
Without some compiler accommodation, munge cannot be called by correct
|
383 |
|
|
C++ code using a pointer to a correctly-scoped tm* value.
|
384 |
|
|
|
385 |
|
|
The current C headers use the preprocessor extension "#include_next",
|
386 |
|
|
which the compiler complains about when run "-pedantic".
|
387 |
|
|
(Incidentally, it appears that "-fpedantic" is currently ignored,
|
388 |
|
|
probably a bug.) The solution in the C compiler is to use
|
389 |
|
|
"-isystem" rather than "-I", but unfortunately in g++ this seems
|
390 |
|
|
also to wrap the whole header in an 'extern "C"' block, so it's
|
391 |
|
|
unusable for C++ headers. The correct solution appears to be to
|
392 |
|
|
allow the various special include-directory options, if not given
|
393 |
|
|
an argument, to affect subsequent include-directory options additively,
|
394 |
|
|
so that if one said
|
395 |
|
|
|
396 |
|
|
-pedantic -iprefix $(prefix) \
|
397 |
|
|
-idirafter -ino-pedantic -ino-extern-c -iwithprefix -I g++-v3 \
|
398 |
|
|
-iwithprefix -I g++-v3/ext
|
399 |
|
|
|
400 |
|
|
the compiler would search $(prefix)/g++-v3 and not report
|
401 |
|
|
pedantic warnings for files found there, but treat files in
|
402 |
|
|
$(prefix)/g++-v3/ext pedantically. (The undocumented semantics
|
403 |
|
|
of "-isystem" in g++ stink. Can they be rescinded? If not it
|
404 |
|
|
must be replaced with something more rationally behaved.)
|
405 |
|
|
|
406 |
|
|
All the C headers need the treatment above; in the standard these
|
407 |
|
|
headers are mentioned in various chapters. Below, I have only
|
408 |
|
|
mentioned those that present interesting implementation issues.
|
409 |
|
|
|
410 |
|
|
The components identified as "mostly complete", below, have not been
|
411 |
|
|
audited for conformance. In many cases where the library passes
|
412 |
|
|
conformance tests we have non-conforming extensions that must be
|
413 |
|
|
wrapped in #if guards for "pedantic" use, and in some cases renamed
|
414 |
|
|
in a conforming way for continued use in the implementation regardless
|
415 |
|
|
of conformance flags.
|
416 |
|
|
|
417 |
|
|
The STL portion of the library still depends on a header
|
418 |
|
|
stl/bits/stl_config.h full of #ifdef clauses. This apparatus
|
419 |
|
|
should be replaced with autoconf/automake machinery.
|
420 |
|
|
|
421 |
|
|
The SGI STL defines a type_traits<> template, specialized for
|
422 |
|
|
many types in their code including the built-in numeric and
|
423 |
|
|
pointer types and some library types, to direct optimizations of
|
424 |
|
|
standard functions. The SGI compiler has been extended to generate
|
425 |
|
|
specializations of this template automatically for user types,
|
426 |
|
|
so that use of STL templates on user types can take advantage of
|
427 |
|
|
these optimizations. Specializations for other, non-STL, types
|
428 |
|
|
would make more optimizations possible, but extending the gcc
|
429 |
|
|
compiler in the same way would be much better. Probably the next
|
430 |
|
|
round of standardization will ratify this, but probably with
|
431 |
|
|
changes, so it probably should be renamed to place it in the
|
432 |
|
|
implementation namespace.
|
433 |
|
|
|
434 |
|
|
The SGI STL also defines a large number of extensions visible in
|
435 |
|
|
standard headers. (Other extensions that appear in separate headers
|
436 |
|
|
have been sequestered in subdirectories ext/ and backward/.) All
|
437 |
|
|
these extensions should be moved to other headers where possible,
|
438 |
|
|
and in any case wrapped in a namespace (not std!), and (where kept
|
439 |
|
|
in a standard header) girded about with macro guards. Some cannot be
|
440 |
|
|
moved out of standard headers because they are used to implement
|
441 |
|
|
standard features. The canonical method for accommodating these
|
442 |
|
|
is to use a protected name, aliased in macro guards to a user-space
|
443 |
|
|
name. Unfortunately C++ offers no satisfactory template typedef
|
444 |
|
|
mechanism, so very ad-hoc and unsatisfactory aliasing must be used
|
445 |
|
|
instead.
|
446 |
|
|
|
447 |
|
|
Implementation of a template typedef mechanism should have the highest
|
448 |
|
|
priority among possible extensions, on the same level as implementation
|
449 |
|
|
of the template "export" feature.
|
450 |
|
|
|
451 |
|
|
Chapter 18 Language support
|
452 |
|
|
----------------------------
|
453 |
|
|
|
454 |
|
|
Headers:
|
455 |
|
|
C headers:
|
456 |
|
|
(also 21, 25, 26)
|
457 |
|
|
|
458 |
|
|
This defines the built-in exceptions, rtti, numeric_limits<>,
|
459 |
|
|
operator new and delete. Much of this is provided by the
|
460 |
|
|
compiler in its static runtime library.
|
461 |
|
|
|
462 |
|
|
Work to do includes defining numeric_limits<> specializations in
|
463 |
|
|
separate files for all target architectures. Values for integer types
|
464 |
|
|
except for bool and wchar_t are readily obtained from the C header
|
465 |
|
|
, but values for the remaining numeric types (bool, wchar_t,
|
466 |
|
|
float, double, long double) must be entered manually. This is
|
467 |
|
|
largely dog work except for those members whose values are not
|
468 |
|
|
easily deduced from available documentation. Also, this involves
|
469 |
|
|
some work in target configuration to identify the correct choice of
|
470 |
|
|
file to build against and to install.
|
471 |
|
|
|
472 |
|
|
The definitions of the various operators new and delete must be
|
473 |
|
|
made thread-safe, which depends on a portable exclusion mechanism,
|
474 |
|
|
discussed under chapter 20. Of course there is always plenty of
|
475 |
|
|
room for improvements to the speed of operators new and delete.
|
476 |
|
|
|
477 |
|
|
, in Glibc, defines some macros that gcc does not allow to
|
478 |
|
|
be wrapped into an inline function. Probably this header will demand
|
479 |
|
|
attention whenever a new target is chosen. The functions atexit(),
|
480 |
|
|
exit(), and abort() in cstdlib have different semantics in C++, so
|
481 |
|
|
must be re-implemented for C++.
|
482 |
|
|
|
483 |
|
|
Chapter 19 Diagnostics
|
484 |
|
|
-----------------------
|
485 |
|
|
|
486 |
|
|
Headers:
|
487 |
|
|
C headers:
|
488 |
|
|
|
489 |
|
|
This defines the standard exception objects, which are "mostly complete".
|
490 |
|
|
Cygnus has a version, and now SGI provides a slightly different one.
|
491 |
|
|
It makes little difference which we use.
|
492 |
|
|
|
493 |
|
|
The C global name "errno", which C allows to be a variable or a macro,
|
494 |
|
|
is required in C++ to be a macro. For MT it must typically result in
|
495 |
|
|
a function call.
|
496 |
|
|
|
497 |
|
|
Chapter 20 Utilities
|
498 |
|
|
---------------------
|
499 |
|
|
Headers:
|
500 |
|
|
C header: (also in 18)
|
501 |
|
|
|
502 |
|
|
SGI STL provides "mostly complete" versions of all the components
|
503 |
|
|
defined in this chapter. However, the auto_ptr<> implementation
|
504 |
|
|
is known to be wrong. Furthermore, the standard definition of it
|
505 |
|
|
is known to be unimplementable as written. A minor change to the
|
506 |
|
|
standard would fix it, and auto_ptr<> should be adjusted to match.
|
507 |
|
|
|
508 |
|
|
Multi-threading affects the allocator implementation, and there must
|
509 |
|
|
be configuration/installation choices for different users' MT
|
510 |
|
|
requirements. Anyway, users will want to tune allocator options
|
511 |
|
|
to support different target conditions, MT or no.
|
512 |
|
|
|
513 |
|
|
The primitives used for MT implementation should be exposed, as an
|
514 |
|
|
extension, for users' own work. We need cross-CPU "mutex" support,
|
515 |
|
|
multi-processor shared-memory atomic integer operations, and single-
|
516 |
|
|
processor uninterruptible integer operations, and all three configurable
|
517 |
|
|
to be stubbed out for non-MT use, or to use an appropriately-loaded
|
518 |
|
|
dynamic library for the actual runtime environment, or statically
|
519 |
|
|
compiled in for cases where the target architecture is known.
|
520 |
|
|
|
521 |
|
|
Chapter 21 String
|
522 |
|
|
------------------
|
523 |
|
|
Headers:
|
524 |
|
|
C headers: (also in 27)
|
525 |
|
|
(also in 18, 25, 26)
|
526 |
|
|
|
527 |
|
|
We have "mostly-complete" char_traits<> implementations. Many of the
|
528 |
|
|
char_traits operations might be optimized further using existing
|
529 |
|
|
proprietary language extensions.
|
530 |
|
|
|
531 |
|
|
We have a "mostly-complete" basic_string<> implementation. The work
|
532 |
|
|
to manually instantiate char and wchar_t specializations in object
|
533 |
|
|
files to improve link-time behavior is extremely unsatisfactory,
|
534 |
|
|
literally tripling library-build time with no commensurate improvement
|
535 |
|
|
in static program link sizes. It must be redone. (Similar work is
|
536 |
|
|
needed for some components in chapters 22 and 27.)
|
537 |
|
|
|
538 |
|
|
Other work needed for strings is MT-safety, as discussed under the
|
539 |
|
|
chapter 20 heading.
|
540 |
|
|
|
541 |
|
|
The standard C type mbstate_t from and used in char_traits<>
|
542 |
|
|
must be different in C++ than in C, because in C++ the default constructor
|
543 |
|
|
value mbstate_t() must be the "base" or "ground" sequence state.
|
544 |
|
|
(According to the likely resolution of a recently raised Core issue,
|
545 |
|
|
this may become unnecessary. However, there are other reasons to
|
546 |
|
|
use a state type not as limited as whatever the C library provides.)
|
547 |
|
|
If we might want to provide conversions from (e.g.) internally-
|
548 |
|
|
represented EUC-wide to externally-represented Unicode, or vice-
|
549 |
|
|
versa, the mbstate_t we choose will need to be more accommodating
|
550 |
|
|
than what might be provided by an underlying C library.
|
551 |
|
|
|
552 |
|
|
There remain some basic_string template-member functions which do
|
553 |
|
|
not overload properly with their non-template brethren. The infamous
|
554 |
|
|
hack akin to what was done in vector<> is needed, to conform to
|
555 |
|
|
23.1.1 para 10. The CHECKLIST items for basic_string marked 'X',
|
556 |
|
|
or incomplete, are so marked for this reason.
|
557 |
|
|
|
558 |
|
|
Replacing the string iterators, which currently are simple character
|
559 |
|
|
pointers, with class objects would greatly increase the safety of the
|
560 |
|
|
client interface, and also permit a "debug" mode in which range,
|
561 |
|
|
ownership, and validity are rigorously checked. The current use of
|
562 |
|
|
raw pointers as string iterators is evil. vector<> iterators need the
|
563 |
|
|
same treatment. Note that the current implementation freely mixes
|
564 |
|
|
pointers and iterators, and that must be fixed before safer iterators
|
565 |
|
|
can be introduced.
|
566 |
|
|
|
567 |
|
|
Some of the functions in are different from the C version.
|
568 |
|
|
generally overloaded on const and non-const argument pointers. For
|
569 |
|
|
example, in strchr is overloaded. The functions isupper
|
570 |
|
|
etc. in typically implemented as macros in C are functions
|
571 |
|
|
in C++, because they are overloaded with others of the same name
|
572 |
|
|
defined in .
|
573 |
|
|
|
574 |
|
|
Many of the functions required in and cannot be
|
575 |
|
|
implemented using underlying C facilities on intended targets because
|
576 |
|
|
such facilities only partly exist.
|
577 |
|
|
|
578 |
|
|
Chapter 22 Locale
|
579 |
|
|
------------------
|
580 |
|
|
Headers:
|
581 |
|
|
C headers:
|
582 |
|
|
|
583 |
|
|
We have a "mostly complete" class locale, with the exception of
|
584 |
|
|
code for constructing, and handling the names of, named locales.
|
585 |
|
|
The ways that locales are named (particularly when categories
|
586 |
|
|
(e.g. LC_TIME, LC_COLLATE) are different) varies among all target
|
587 |
|
|
environments. This code must be written in various versions and
|
588 |
|
|
chosen by configuration parameters.
|
589 |
|
|
|
590 |
|
|
Members of many of the facets defined in are stubs. Generally,
|
591 |
|
|
there are two sets of facets: the base class facets (which are supposed
|
592 |
|
|
to implement the "C" locale) and the "byname" facets, which are supposed
|
593 |
|
|
to read files to determine their behavior. The base ctype<>, collate<>,
|
594 |
|
|
and numpunct<> facets are "mostly complete", except that the table of
|
595 |
|
|
bitmask values used for "is" operations, and corresponding mask values,
|
596 |
|
|
are still defined in libio and just included/linked. (We will need to
|
597 |
|
|
implement these tables independently, soon, but should take advantage
|
598 |
|
|
of libio where possible.) The num_put<>::put members for integer types
|
599 |
|
|
are "mostly complete".
|
600 |
|
|
|
601 |
|
|
A complete list of what has and has not been implemented may be
|
602 |
|
|
found in CHECKLIST. However, note that the current definition of
|
603 |
|
|
codecvt is wrong. It should simply write
|
604 |
|
|
out the raw bytes representing the wide characters, rather than
|
605 |
|
|
trying to convert each to a corresponding single "char" value.
|
606 |
|
|
|
607 |
|
|
Some of the facets are more important than others. Specifically,
|
608 |
|
|
the members of ctype<>, numpunct<>, num_put<>, and num_get<> facets
|
609 |
|
|
are used by other library facilities defined in , ,
|
610 |
|
|
and , and the codecvt<> facet is used by basic_filebuf<>
|
611 |
|
|
in , so a conforming iostream implementation depends on
|
612 |
|
|
these.
|
613 |
|
|
|
614 |
|
|
The "long long" type eventually must be supported, but code mentioning
|
615 |
|
|
it should be wrapped in #if guards to allow pedantic-mode compiling.
|
616 |
|
|
|
617 |
|
|
Performance of num_put<> and num_get<> depend critically on
|
618 |
|
|
caching computed values in ios_base objects, and on extensions
|
619 |
|
|
to the interface with streambufs.
|
620 |
|
|
|
621 |
|
|
Specifically: retrieving a copy of the locale object, extracting
|
622 |
|
|
the needed facets, and gathering data from them, for each call to
|
623 |
|
|
(e.g.) operator<< would be prohibitively slow. To cache format
|
624 |
|
|
data for use by num_put<> and num_get<> we have a _Format_cache<>
|
625 |
|
|
object stored in the ios_base::pword() array. This is constructed
|
626 |
|
|
and initialized lazily, and is organized purely for utility. It
|
627 |
|
|
is discarded when a new locale with different facets is imbued.
|
628 |
|
|
|
629 |
|
|
Using only the public interfaces of the iterator arguments to the
|
630 |
|
|
facet functions would limit performance by forbidding "vector-style"
|
631 |
|
|
character operations. The streambuf iterator optimizations are
|
632 |
|
|
described under chapter 24, but facets can also bypass the streambuf
|
633 |
|
|
iterators via explicit specializations and operate directly on the
|
634 |
|
|
streambufs, and use extended interfaces to get direct access to the
|
635 |
|
|
streambuf internal buffer arrays. These extensions are mentioned
|
636 |
|
|
under chapter 27. These optimizations are particularly important
|
637 |
|
|
for input parsing.
|
638 |
|
|
|
639 |
|
|
Unused virtual members of locale facets can be omitted, as mentioned
|
640 |
|
|
above, by a smart linker.
|
641 |
|
|
|
642 |
|
|
Chapter 23 Containers
|
643 |
|
|
----------------------
|
644 |
|
|
Headers:
|
645 |
|
|
|
646 |
|
|
All the components in chapter 23 are implemented in the SGI STL.
|
647 |
|
|
They are "mostly complete"; they include a large number of
|
648 |
|
|
nonconforming extensions which must be wrapped. Some of these
|
649 |
|
|
are used internally and must be renamed or duplicated.
|
650 |
|
|
|
651 |
|
|
The SGI components are optimized for large-memory environments. For
|
652 |
|
|
embedded targets, different criteria might be more appropriate. Users
|
653 |
|
|
will want to be able to tune this behavior. We should provide
|
654 |
|
|
ways for users to compile the library with different memory usage
|
655 |
|
|
characteristics.
|
656 |
|
|
|
657 |
|
|
A lot more work is needed on factoring out common code from different
|
658 |
|
|
specializations to reduce code size here and in chapter 25. The
|
659 |
|
|
easiest fix for this would be a compiler/ABI improvement that allows
|
660 |
|
|
the compiler to recognize when a specialization depends only on the
|
661 |
|
|
size (or other gross quality) of a template argument, and allow the
|
662 |
|
|
linker to share the code with similar specializations. In its
|
663 |
|
|
absence, many of the algorithms and containers can be partial-
|
664 |
|
|
specialized, at least for the case of pointers, but this only solves
|
665 |
|
|
a small part of the problem. Use of a type_traits-style template
|
666 |
|
|
allows a few more optimization opportunities, more if the compiler
|
667 |
|
|
can generate the specializations automatically.
|
668 |
|
|
|
669 |
|
|
As an optimization, containers can specialize on the default allocator
|
670 |
|
|
and bypass it, or take advantage of details of its implementation
|
671 |
|
|
after it has been improved upon.
|
672 |
|
|
|
673 |
|
|
Replacing the vector iterators, which currently are simple element
|
674 |
|
|
pointers, with class objects would greatly increase the safety of the
|
675 |
|
|
client interface, and also permit a "debug" mode in which range,
|
676 |
|
|
ownership, and validity are rigorously checked. The current use of
|
677 |
|
|
pointers for iterators is evil.
|
678 |
|
|
|
679 |
|
|
As mentioned for chapter 24, the deque iterator is a good example of
|
680 |
|
|
an opportunity to implement a "staged" iterator that would benefit
|
681 |
|
|
from specializations of some algorithms.
|
682 |
|
|
|
683 |
|
|
Chapter 24 Iterators
|
684 |
|
|
---------------------
|
685 |
|
|
Headers:
|
686 |
|
|
|
687 |
|
|
Standard iterators are "mostly complete", with the exception of
|
688 |
|
|
the stream iterators, which are not yet templatized on the
|
689 |
|
|
stream type. Also, the base class template iterator<> appears
|
690 |
|
|
to be wrong, so everything derived from it must also be wrong,
|
691 |
|
|
currently.
|
692 |
|
|
|
693 |
|
|
The streambuf iterators (currently located in stl/bits/std_iterator.h,
|
694 |
|
|
but should be under bits/) can be rewritten to take advantage of
|
695 |
|
|
friendship with the streambuf implementation.
|
696 |
|
|
|
697 |
|
|
Matt Austern has identified opportunities where certain iterator
|
698 |
|
|
types, particularly including streambuf iterators and deque
|
699 |
|
|
iterators, have a "two-stage" quality, such that an intermediate
|
700 |
|
|
limit can be checked much more quickly than the true limit on
|
701 |
|
|
range operations. If identified with a member of iterator_traits,
|
702 |
|
|
algorithms may be specialized for this case. Of course the
|
703 |
|
|
iterators that have this quality can be identified by specializing
|
704 |
|
|
a traits class.
|
705 |
|
|
|
706 |
|
|
Many of the algorithms must be specialized for the streambuf
|
707 |
|
|
iterators, to take advantage of block-mode operations, in order
|
708 |
|
|
to allow iostream/locale operations' performance not to suffer.
|
709 |
|
|
It may be that they could be treated as staged iterators and
|
710 |
|
|
take advantage of those optimizations.
|
711 |
|
|
|
712 |
|
|
Chapter 25 Algorithms
|
713 |
|
|
----------------------
|
714 |
|
|
Headers:
|
715 |
|
|
C headers: (also in 18, 21, 26))
|
716 |
|
|
|
717 |
|
|
The algorithms are "mostly complete". As mentioned above, they
|
718 |
|
|
are optimized for speed at the expense of code and data size.
|
719 |
|
|
|
720 |
|
|
Specializations of many of the algorithms for non-STL types would
|
721 |
|
|
give performance improvements, but we must use great care not to
|
722 |
|
|
interfere with fragile template overloading semantics for the
|
723 |
|
|
standard interfaces. Conventionally the standard function template
|
724 |
|
|
interface is an inline which delegates to a non-standard function
|
725 |
|
|
which is then overloaded (this is already done in many places in
|
726 |
|
|
the library). Particularly appealing opportunities for the sake of
|
727 |
|
|
iostream performance are for copy and find applied to streambuf
|
728 |
|
|
iterators or (as noted elsewhere) for staged iterators, of which
|
729 |
|
|
the streambuf iterators are a good example.
|
730 |
|
|
|
731 |
|
|
The bsearch and qsort functions cannot be overloaded properly as
|
732 |
|
|
required by the standard because gcc does not yet allow overloading
|
733 |
|
|
on the extern-"C"-ness of a function pointer.
|
734 |
|
|
|
735 |
|
|
Chapter 26 Numerics
|
736 |
|
|
--------------------
|
737 |
|
|
Headers:
|
738 |
|
|
C headers: , (also 18, 21, 25)
|
739 |
|
|
|
740 |
|
|
Numeric components: Gabriel dos Reis's valarray, Drepper's complex,
|
741 |
|
|
and the few algorithms from the STL are "mostly done". Of course
|
742 |
|
|
optimization opportunities abound for the numerically literate. It
|
743 |
|
|
is not clear whether the valarray implementation really conforms
|
744 |
|
|
fully, in the assumptions it makes about aliasing (and lack thereof)
|
745 |
|
|
in its arguments.
|
746 |
|
|
|
747 |
|
|
The C div() and ldiv() functions are interesting, because they are the
|
748 |
|
|
only case where a C library function returns a class object by value.
|
749 |
|
|
Since the C++ type div_t must be different from the underlying C type
|
750 |
|
|
(which is in the wrong namespace) the underlying functions div() and
|
751 |
|
|
ldiv() cannot be re-used efficiently. Fortunately they are trivial to
|
752 |
|
|
re-implement.
|
753 |
|
|
|
754 |
|
|
Chapter 27 Iostreams
|
755 |
|
|
---------------------
|
756 |
|
|
Headers:
|
757 |
|
|
|
758 |
|
|
C headers: (also in 21)
|
759 |
|
|
|
760 |
|
|
Iostream is currently in a very incomplete state. , ,
|
761 |
|
|
ios_base, and basic_ios<> are "mostly complete". basic_streambuf<> and
|
762 |
|
|
basic_ostream<> are well along, but basic_istream<> has had little work
|
763 |
|
|
done. The standard stream objects, and have been
|
764 |
|
|
started; basic_filebuf<> "write" functions have been implemented just
|
765 |
|
|
enough to do "hello, world".
|
766 |
|
|
|
767 |
|
|
Most of the istream and ostream operators << and >> (with the exception
|
768 |
|
|
of the op<<(integer) ones) have not been changed to use locale primitives,
|
769 |
|
|
sentry objects, or char_traits members.
|
770 |
|
|
|
771 |
|
|
All these templates should be manually instantiated for char and
|
772 |
|
|
wchar_t in a way that links only used members into user programs.
|
773 |
|
|
|
774 |
|
|
Streambuf is fertile ground for optimization extensions. An extended
|
775 |
|
|
interface giving iterator access to its internal buffer would be very
|
776 |
|
|
useful for other library components.
|
777 |
|
|
|
778 |
|
|
Iostream operations (primarily operators << and >>) can take advantage
|
779 |
|
|
of the case where user code has not specified a locale, and bypass locale
|
780 |
|
|
operations entirely. The current implementation of op<::put,
|
781 |
|
|
for the integer types, demonstrates how they can cache encoding details
|
782 |
|
|
from the locale on each operation. There is lots more room for
|
783 |
|
|
optimization in this area.
|
784 |
|
|
|
785 |
|
|
The definition of the relationship between the standard streams
|
786 |
|
|
cout et al. and stdout et al. requires something like a "stdiobuf".
|
787 |
|
|
The SGI solution of using double-indirection to actually use a
|
788 |
|
|
stdio FILE object for buffering is unsatisfactory, because it
|
789 |
|
|
interferes with peephole loop optimizations.
|
790 |
|
|
|
791 |
|
|
The header work has begun. stringbuf can benefit from
|
792 |
|
|
friendship with basic_string<> and basic_string<>::_Rep to use
|
793 |
|
|
those objects directly as buffers, and avoid allocating and making
|
794 |
|
|
copies.
|
795 |
|
|
|
796 |
|
|
The basic_filebuf<> template is a complex beast. It is specified to
|
797 |
|
|
use the locale facet codecvt<> to translate characters between native
|
798 |
|
|
files and the locale character encoding. In general this involves
|
799 |
|
|
two buffers, one of "char" representing the file and another of
|
800 |
|
|
"char_type", for the stream, with codecvt<> translating. The process
|
801 |
|
|
is complicated by the variable-length nature of the translation, and
|
802 |
|
|
the need to seek to corresponding places in the two representations.
|
803 |
|
|
For the case of basic_filebuf, when no translation is needed,
|
804 |
|
|
a single buffer suffices. A specialized filebuf can be used to reduce
|
805 |
|
|
code space overhead when no locale has been imbued. Matt Austern's
|
806 |
|
|
work at SGI will be useful, perhaps directly as a source of code, or
|
807 |
|
|
at least as an example to draw on.
|
808 |
|
|
|
809 |
|
|
Filebuf, almost uniquely (cf. operator new), depends heavily on
|
810 |
|
|
underlying environmental facilities. In current releases iostream
|
811 |
|
|
depends fairly heavily on libio constant definitions, but it should
|
812 |
|
|
be made independent. It also depends on operating system primitives
|
813 |
|
|
for file operations. There is immense room for optimizations using
|
814 |
|
|
(e.g.) mmap for reading. The shadow/ directory wraps, besides the
|
815 |
|
|
standard C headers, the libio.h and unistd.h headers, for use mainly
|
816 |
|
|
by filebuf. These wrappings have not been completed, though there
|
817 |
|
|
is scaffolding in place.
|
818 |
|
|
|
819 |
|
|
The encapulation of certain C header names presents an
|
820 |
|
|
interesting problem. It is possible to define an inline std::fprintf()
|
821 |
|
|
implemented in terms of the 'extern "C"' vfprintf(), but there is no
|
822 |
|
|
standard vfscanf() to use to implement std::fscanf(). It appears that
|
823 |
|
|
vfscanf but be re-implemented in C++ for targets where no vfscanf
|
824 |
|
|
extension has been defined. This is interesting in that it seems
|
825 |
|
|
to be the only significant case in the C library where this kind of
|
826 |
|
|
rewriting is necessary. (Of course Glibc provides the vfscanf()
|
827 |
|
|
extension.) (The functions related to exit() must be rewritten
|
828 |
|
|
for other reasons.)
|
829 |
|
|
|
830 |
|
|
|
831 |
|
|
Annex D
|
832 |
|
|
-------
|
833 |
|
|
Headers:
|
834 |
|
|
|
835 |
|
|
Annex D defines many non-library features, and many minor
|
836 |
|
|
modifications to various headers, and a complete header.
|
837 |
|
|
It is "mostly done", except that the libstdc++-2
|
838 |
|
|
header has not been adopted into the library, or checked to
|
839 |
|
|
verify that it matches the draft in those details that were
|
840 |
|
|
clarified by the committee. Certainly it must at least be
|
841 |
|
|
moved into the std namespace.
|
842 |
|
|
|
843 |
|
|
We still need to wrap all the deprecated features in #if guards
|
844 |
|
|
so that pedantic compile modes can detect their use.
|
845 |
|
|
|
846 |
|
|
Nonstandard Extensions
|
847 |
|
|
----------------------
|
848 |
|
|
Headers:
|
849 |
|
|
(etc.)
|
850 |
|
|
|
851 |
|
|
User code has come to depend on a variety of nonstandard components
|
852 |
|
|
that we must not omit. Much of this code can be adopted from
|
853 |
|
|
libstdc++-v2 or from the SGI STL. This particularly includes
|
854 |
|
|
, , and various SGI extensions such
|
855 |
|
|
as . Many of these are already placed in the
|
856 |
|
|
subdirectories ext/ and backward/. (Note that it is better to
|
857 |
|
|
include them via "" or "" than
|
858 |
|
|
to search the subdirectory itself via a "-I" directive.
|
859 |
|
|
|