gcc/libstdc++-v3/docs/17_intro/DESIGN

Standard C++ Library Design Document
------------------------------------

This is an overview of libstdc++-v3, with particular attention
to projects to be done and how they fit into the whole.

The Library
-----------

This paper is covers two major areas:

 - Features and policies not mentioned in the standard that
   the quality of the library implementation depends on, including
   extensions and "implementation-defined" features;

 - Plans for required but unimplemented library features and
   optimizations to them.

Overhead
--------

The standard defines a large library, much larger than the standard
C library.  A naive implementation would suffer substantial overhead
in compile time, executable size, and speed, rendering it unusable
in many (particularly embedded) applications.  The alternative demands
care in construction, and some compiler support, but there is no
need for library subsets.

What are the sources of this overhead?  There are four main causes:

 - The library is specified almost entirely as templates, which
   with current compilers must be included in-line, resulting in
   very slow builds as tens or hundreds of thousands of lines
   of function definitions are read for each user source file.
   Indeed, the entire SGI STL, as well as the dos Reis valarray,
   are provided purely as header files, largely for simplicity in
   porting.  Iostream/locale is (or will be) as large again.

 - The library is very flexible, specifying a multitude of hooks
   where users can insert their own code in place of defaults.
   When these hooks are not used, any time and code expended to
   support that flexibility is wasted.

 - Templates are often described as causing to "code bloat".  In
   practice, this refers (when it refers to anything real) to several
   independent processes.  First, when a class template is manually
   instantiated in its entirely, current compilers place the definitions
   for all members in a single object file, so that a program linking
   to one member gets definitions of all.  Second, template functions
   which do not actually depend on the template argument are, under
   current compilers, generated anew for each instantiation, rather
   than being shared with other instantiations.  Third, some of the
   flexibility mentioned above comes from virtual functions (both in
   regular classes and template classes) which current linkers add
   to the executable file even when they manifestly cannot be called.

 - The library is specified to use a language feature, exceptions,
   which in the current gcc compiler ABI imposes a run time and
   code space cost to handle the possibility of exceptions even when
   they are not used.  Under the new ABI (accessed with -fnew-abi),
   there is a space overhead and a small reduction in code efficiency
   resulting from lost optimization opportunities associated with
   non-local branches associated with exceptions.

What can be done to eliminate this overhead?  A variety of coding
techniques, and compiler, linker and library improvements and
extensions may be used, as covered below.  Most are not difficult,
and some are already implemented in varying degrees.

Overhead: Compilation Time
--------------------------

Providing "ready-instantiated" template code in object code archives
allows us to avoid generating and optimizing template instantiations
in each compilation unit which uses them.  However, the number of such
instantiations that are useful to provide is limited, and anyway this
is not enough, by itself, to minimize compilation time.  In particular,
it does not reduce time spent parsing conforming headers.

Quicker header parsing will depend on library extensions and compiler
improvements.   One approach is some variation on the techniques
previously marketed as "pre-compiled headers", now standardized as
support for the "export" keyword.  "Exported" template definitions
can be placed (once) in a "repository" -- really just a library, but
of template definitions rather than object code -- to be drawn upon
at link time when an instantiation is needed, rather than placed in
header files to be parsed along with every compilation unit.

Until "export" is implemented we can put some of the lengthy template
definitions in #if guards or alternative headers so that users can skip
over the the full definitions when they need only the ready-instantiated
specializations.

To be precise, this means that certain headers which define
templates which users normally use only for certain arguments
can be instrumented to avoid exposing the template definitions
to the compiler unless a macro is defined.  For example, in
<string>, we might have:

   template <class _CharT, ... > class basic_string {
     ... // member declarations
   };
   ... // operator declarations

   #ifdef _STRICT_ISO_
   # if _G_NO_TEMPLATE_EXPORT
   #   include <bits/std_locale.h>  // headers needed by definitions
   #   ...
   #   include <bits/string.tcc>  // member and global template definitions.
   # endif
   #endif

Users who compile without specifying a strict-ISO-conforming flag
would not see many of the template definitions they now see, and rely
instead on ready-instantiated specializations in the library.  This
technique would be useful for the following substantial components:
string, locale/iostreams, valarray.  It would *not* be useful or
usable with the following: containers, algorithms, iterators,
allocator.  Since these constitute a large (though decreasing)
fraction of the library, the benefit the technique offers is
limited.

The language specifies the semantics of the "export" keyword, but
the gcc compiler does not yet support it.  When it does, problems
with large template inclusions can largely disappear, given some
minor library reorganization, along with the need for the apparatus
described above.

Overhead: Flexibility Cost
--------------------------

The library offers many places where users can specify operations
to be performed by the library in place of defaults.  Sometimes
this seems to require that the library use a more-roundabout, and
possibly slower, way to accomplish the default requirements than
would be used otherwise.

The primary protection against this overhead is thorough compiler
optimization, to crush out layers of inline function interfaces.
Kuck & Associates has demonstrated the practicality of this kind
of optimization.

The second line of defense against this overhead is explicit
specialization.  By defining helper function templates, and writing
specialized code for the default case, overhead can be eliminated
for that case without sacrificing flexibility.  This takes full
advantage of any ability of the optimizer to crush out degenerate
code.

The library specifies many virtual functions which current linkers
load even when they cannot be called.  Some minor improvements to the
compiler and to ld would eliminate any such overhead by simply
omitting virtual functions that the complete program does not call.
A prototype of this work has already been done.  For targets where
GNU ld is not used, a "pre-linker" could do the same job.

The main areas in the standard interface where user flexibility
can result in overhead are:

 - Allocators:  Containers are specified to use user-definable
   allocator types and objects, making tuning for the container
   characteristics tricky.

 - Locales: the standard specifies locale objects used to implement
   iostream operations, involving many virtual functions which use
   streambuf iterators.

 - Algorithms and containers: these may be instantiated on any type,
   frequently duplicating code for identical operations.

 - Iostreams and strings: users are permitted to use these on their
   own types, and specify the operations the stream must use on these
   types.

Note that these sources of overhead are _avoidable_.  The techniques
to avoid them are covered below.

Code Bloat
----------

In the SGI STL, and in some other headers, many of the templates
are defined "inline" -- either explicitly or by their placement
in class definitions -- which should not be inline.  This is a
source of code bloat.  Matt had remarked that he was relying on
the compiler to recognize what was too big to benefit from inlining,
and generate it out-of-line automatically.  However, this also can
result in code bloat except where the linker can eliminate the extra
copies.

Fixing these cases will require an audit of all inline functions
defined in the library to determine which merit inlining, and moving
the rest out of line.  This is an issue mainly in chapters 23, 25, and
27.  Of course it can be done incrementally, and we should generally
accept patches that move large functions out of line and into ".tcc"
files, which can later be pulled into a repository.  Compiler/linker
improvements to recognize very large inline functions and move them
out-of-line, but shared among compilation units, could make this
work unnecessary.

Pre-instantiating template specializations currently produces large
amounts of dead code which bloats statically linked programs.  The
current state of the static library, libstdc++.a, is intolerable on
this account, and will fuel further confused speculation about a need
for a library "subset".  A compiler improvement that treats each
instantiated function as a separate object file, for linking purposes,
would be one solution to this problem.  An alternative would be to
split up the manual instantiation files into dozens upon dozens of
little files, each compiled separately, but an abortive attempt at
this was done for <string> and, though it is far from complete, it
is already a nuisance.  A better interim solution (just until we have
"export") is badly needed.

When building a shared library, the current compiler/linker cannot
automatically generate the instantiatiations needed.  This creates a
miserable situation; it means any time something is changed in the
library, before a shared library can be built someone must manually
copy the declarations of all templates that are needed by other parts
of the library to an "instantiation" file, and add it to the build
system to be compiled and linked to the library.  This process is
readily automated, and should be automated as soon as possible.
Users building their own shared libraries experience identical
frustrations.

Sharing common aspects of template definitions among instantiations
can radically reduce code bloat.  The compiler could help a great
deal here by recognizing when a function depends on nothing about
a template parameter, or only on its size, and giving the resulting
function a link-name "equate" that allows it to be shared with other
instantiations.  Implementation code could take advantage of the
capability by factoring out code that does not depend on the template
argument into separate functions to be merged by the compiler.

Until such a compiler optimization is implemented, much can be done
manually (if tediously) in this direction.  One such optimization is
to derive class templates from non-template classes, and move as much
implementation as possible into the base class.  Another is to partial-
specialize certain common instantiations, such as vector<T*>, to share
code for instantiations on all types T.  While these techniques work,
they are far from the complete solution that a compiler improvement
would afford.

Overhead: Expensive Language Features
-------------------------------------

The main "expensive" language feature used in the standard library
is exception support, which requires compiling in cleanup code with
static table data to locate it, and linking in library code to use
the table.  For small embedded programs the amount of such library
code and table data is assumed by some to be excessive.  Under the
"new" ABI this perception is generally exaggerated, although in some
cases it may actually be excessive.

To implement a library which does not use exceptions directly is
not difficult given minor compiler support (to "turn off" exceptions
and ignore exception contructs), and results in no great library
maintenance difficulties.  To be precise, given "-fno-exceptions",
the compiler should treat "try" blocks as ordinary blocks, and
"catch" blocks as dead code to ignore or eliminate.  Compiler
support is not strictly necessary, except in the case of "function
try blocks"; otherwise the following macros almost suffice:

  #define throw(X)
  #define try      if (true)
  #define catch(X) else if (false)

However, there may be a need to use function try blocks in the
library implementation, and use of macros in this way can make
correct diagnostics impossible.  Furthermore, use of this scheme
would require the library to call a function to re-throw exceptions
from a try block.  Implementing the above semantics in the compiler
is preferable.

Given the support above (however implemented) it only remains to
replace code that "throws" with a call to a well-documented "handler"
function in a separate compilation unit which may be replaced by
the user.  The main source of exceptions that would be difficult
for users to avoid is memory allocation failures, but users can
define their own memory allocation primitives that never throw.
Otherwise, the complete list of such handlers, and which library
functions may call them, would be needed for users to be able to
implement the necessary substitutes.  (Fortunately, they have the
source code.)

Opportunities
-------------

The template capabilities of C++ offer enormous opportunities for
optimizing common library operations, well beyond what would be
considered "eliminating overhead".  In particular, many operations
done in Glibc with macros that depend on proprietary language
extensions can be implemented in pristine Standard C++.  For example,
the chapter 25 algorithms, and even C library functions such as strchr,
can be specialized for the case of static arrays of known (small) size.

Detailed optimization opportunities are identified below where
the component where they would appear is discussed.  Of course new
opportunities will be identified during implementation.

Unimplemented Required Library Features
---------------------------------------

The standard specifies hundreds of components, grouped broadly by
chapter.  These are listed in excruciating detail in the CHECKLIST
file.

  17 general
  18 support
  19 diagnostics
  20 utilities
  21 string
  22 locale
  23 containers
  24 iterators
  25 algorithms
  26 numerics
  27 iostreams
  Annex D  backward compatibility

Anyone participating in implementation of the library should obtain
a copy of the standard, ISO 14882.   People in the U.S. can obtain an
electronic copy for US$18 from ANSI's web site.  Those from other
countries should visit http://www.iso.ch/ to find out the location
of their country's representation in ISO, in order to know who can
sell them a copy.

The emphasis in the following sections is on unimplemented features
and optimization opportunities.

Chapter 17  General
-------------------

Chapter 17 concerns overall library requirements.

The standard doesn't mention threads.  A multi-thread (MT) extension
primarily affects operators new and delete (18), allocator (20),
string (21), locale (22), and iostreams (27).  The common underlying
support needed for this is discussed under chapter 20.

The standard requirements on names from the C headers create a
lot of work, mostly done.  Names in the C headers must be visible
in the std:: and sometimes the global namespace; the names in the
two scopes must refer to the same object.  More stringent is that
Koenig lookup implies that any types specified as defined in std::
really are defined in std::.  Names optionally implemented as
macros in C cannot be macros in C++.  (An overview may be read at
<http://www.cantrip.org/cheaders.html>).  The scripts "inclosure"
and "mkcshadow", and the directories shadow/ and cshadow/, are the
beginning of an effort to conform in this area.

A correct conforming definition of C header names based on underlying
C library headers, and practical linking of conforming namespaced
customer code with third-party C libraries depends ultimately on
an ABI change, allowing namespaced C type names to be mangled into
type names as if they were global, somewhat as C function names in a
namespace, or C++ global variable names, are left unmangled.  Perhaps
another "extern" mode, such as 'extern "C-global"' would be an
appropriate place for such type definitions.  Such a type would
affect mangling as follows:

  namespace A {
    struct X {};
    extern "C-global" {  // or maybe just 'extern "C"'
      struct Y {};
    };
  }
  void f(A::X*);  // mangles to f__FPQ21A1X
  void f(A::Y*);  // mangles to f__FP1Y

(It may be that this is really the appropriate semantics for regular
'extern "C"', and 'extern "C-global"', as an extension, would not be
necessary.) This would allow functions declared in non-standard C headers
(and thus fixable by neither us nor users) to link properly with functions
declared using C types defined in properly-namespaced headers.  The
problem this solves is that C headers (which C++ programmers do persist
in using) frequently forward-declare C struct tags without including
the header where the type is defined, as in

  struct tm;
  void munge(tm*);

Without some compiler accommodation, munge cannot be called by correct
C++ code using a pointer to a correctly-scoped tm* value.

The current C headers use the preprocessor extension "#include_next",
which the compiler complains about when run "-pedantic".
(Incidentally, it appears that "-fpedantic" is currently ignored,
probably a bug.)  The solution in the C compiler is to use
"-isystem" rather than "-I", but unfortunately in g++ this seems
also to wrap the whole header in an 'extern "C"' block, so it's
unusable for C++ headers.  The correct solution appears to be to
allow the various special include-directory options, if not given
an argument, to affect subsequent include-directory options additively,
so that if one said

  -pedantic -iprefix $(prefix) \
  -idirafter -ino-pedantic -ino-extern-c -iwithprefix -I g++-v3 \
  -iwithprefix -I g++-v3/ext

the compiler would search $(prefix)/g++-v3 and not report
pedantic warnings for files found there, but treat files in
$(prefix)/g++-v3/ext pedantically.  (The undocumented semantics
of "-isystem" in g++ stink.  Can they be rescinded?  If not it
must be replaced with something more rationally behaved.)

All the C headers need the treatment above; in the standard these
headers are mentioned in various chapters.  Below, I have only
mentioned those that present interesting implementation issues.

The components identified as "mostly complete", below, have not been
audited for conformance.  In many cases where the library passes
conformance tests we have non-conforming extensions that must be
wrapped in #if guards for "pedantic" use, and in some cases renamed
in a conforming way for continued use in the implementation regardless
of conformance flags.

The STL portion of the library still depends on a header
stl/bits/stl_config.h full of #ifdef clauses.  This apparatus
should be replaced with autoconf/automake machinery.

The SGI STL defines a type_traits<> template, specialized for
many types in their code including the built-in numeric and
pointer types and some library types, to direct optimizations of
standard functions.  The SGI compiler has been extended to generate
specializations of this template automatically for user types,
so that use of STL templates on user types can take advantage of
these optimizations.  Specializations for other, non-STL, types
would make more optimizations possible, but extending the gcc
compiler in the same way would be much better.  Probably the next
round of standardization will ratify this, but probably with
changes, so it probably should be renamed to place it in the
implementation namespace.

The SGI STL also defines a large number of extensions visible in
standard headers.  (Other extensions that appear in separate headers
have been sequestered in subdirectories ext/ and backward/.)  All
these extensions should be moved to other headers where possible,
and in any case wrapped in a namespace (not std!), and (where kept
in a standard header) girded about with macro guards.  Some cannot be
moved out of standard headers because they are used to implement
standard features.   The canonical method for accommodating these
is to use a protected name, aliased in macro guards to a user-space
name.  Unfortunately C++ offers no satisfactory template typedef
mechanism, so very ad-hoc and unsatisfactory aliasing must be used
instead.

Implementation of a template typedef mechanism should have the highest
priority among possible extensions, on the same level as implementation
of the template "export" feature.

Chapter 18  Language support
----------------------------

Headers: <limits> <new> <typeinfo> <exception>
C headers: <cstddef> <climits> <cfloat>  <cstdarg> <csetjmp>
           <ctime>   <csignal> <cstdlib> (also 21, 25, 26)

This defines the built-in exceptions, rtti, numeric_limits<>,
operator new and delete.  Much of this is provided by the
compiler in its static runtime library.

Work to do includes defining numeric_limits<> specializations in
separate files for all target architectures.  Values for integer types
except for bool and wchar_t are readily obtained from the C header
<limits.h>, but values for the remaining numeric types (bool, wchar_t,
float, double, long double) must be entered manually.  This is
largely dog work except for those members whose values are not
easily deduced from available documentation.  Also, this involves
some work in target configuration to identify the correct choice of
file to build against and to install.

The definitions of the various operators new and delete must be
made thread-safe, which depends on a portable exclusion mechanism,
discussed under chapter 20.   Of course there is always plenty of
room for improvements to the speed of operators new and delete.

<cstdarg>, in Glibc, defines some macros that gcc does not allow to
be wrapped into an inline function.  Probably this header will demand
attention whenever a new target is chosen.  The functions atexit(),
exit(), and abort() in cstdlib have different semantics in C++, so
must be re-implemented for C++.

Chapter 19  Diagnostics
-----------------------

Headers: <stdexcept>
C headers: <cassert> <cerrno>

This defines the standard exception objects, which are "mostly complete".
Cygnus has a version, and now SGI provides a slightly different one.
It makes little difference which we use.

The C global name "errno", which C allows to be a variable or a macro,
is required in C++ to be a macro.  For MT it must typically result in
a function call.

Chapter 20  Utilities
---------------------
Headers: <utility> <functional> <memory>
C header: <ctime> (also in 18)

SGI STL provides "mostly complete" versions of all the components
defined in this chapter.  However, the auto_ptr<> implementation
is known to be wrong.  Furthermore, the standard definition of it
is known to be unimplementable as written.  A minor change to the
standard would fix it, and auto_ptr<> should be adjusted to match.

Multi-threading affects the allocator implementation, and there must
be configuration/installation choices for different users' MT
requirements.  Anyway, users will want to tune allocator options
to support different target conditions, MT or no.

The primitives used for MT implementation should be exposed, as an
extension, for users' own work.  We need cross-CPU "mutex" support,
multi-processor shared-memory atomic integer operations, and single-
processor uninterruptible integer operations, and all three configurable
to be stubbed out for non-MT use, or to use an appropriately-loaded
dynamic library for the actual runtime environment, or statically
compiled in for cases where the target architecture is known.

Chapter 21  String
------------------
Headers: <string>
C headers: <cctype> <cwctype> <cstring> <cwchar> (also in 27)
           <cstdlib> (also in 18, 25, 26)

We have "mostly-complete" char_traits<> implementations.  Many of the
char_traits<char> operations might be optimized further using existing
proprietary language extensions.

We have a "mostly-complete" basic_string<> implementation.  The work
to manually instantiate char and wchar_t specializations in object
files to improve link-time behavior is extremely unsatisfactory,
literally tripling library-build time with no commensurate improvement
in static program link sizes.  It must be redone.  (Similar work is
needed for some components in chapters 22 and 27.)

Other work needed for strings is MT-safety, as discussed under the
chapter 20 heading.

The standard C type mbstate_t from <cwchar> and used in char_traits<>
must be different in C++ than in C, because in C++ the default constructor
value mbstate_t() must be the "base" or "ground" sequence state.
(According to the likely resolution of a recently raised Core issue,
this may become unnecessary.  However, there are other reasons to
use a state type not as limited as whatever the C library provides.)
If we might want to provide conversions from (e.g.) internally-
represented EUC-wide to externally-represented Unicode, or vice-
versa, the mbstate_t we choose will need to be more accommodating
than what might be provided by an underlying C library.

There remain some basic_string template-member functions which do
not overload properly with their non-template brethren.  The infamous
hack akin to what was done in vector<> is needed, to conform to
23.1.1 para 10.  The CHECKLIST items for basic_string marked 'X',
or incomplete, are so marked for this reason.

Replacing the string iterators, which currently are simple character
pointers, with class objects would greatly increase the safety of the
client interface, and also permit a "debug" mode in which range,
ownership, and validity are rigorously checked.  The current use of
raw pointers as string iterators is evil.  vector<> iterators need the
same treatment.  Note that the current implementation freely mixes
pointers and iterators, and that must be fixed before safer iterators
can be introduced.

Some of the functions in <cstring> are different from the C version.
generally overloaded on const and non-const argument pointers.  For
example, in <cstring> strchr is overloaded.  The functions isupper
etc.  in <cctype> typically implemented as macros in C are functions
in C++, because they are overloaded with others of the same name
defined in <locale>.

Many of the functions required in <cwctype> and <cwchar> cannot be
implemented using underlying C facilities on intended targets because
such facilities only partly exist.

Chapter 22  Locale
------------------
Headers: <locale>
C headers: <clocale>

We have a "mostly complete" class locale, with the exception of
code for constructing, and handling the names of, named locales.
The ways that locales are named (particularly when categories
(e.g. LC_TIME, LC_COLLATE) are different) varies among all target
environments.  This code must be written in various versions and
chosen by configuration parameters.

Members of many of the facets defined in <locale> are stubs.  Generally,
there are two sets of facets: the base class facets (which are supposed
to implement the "C" locale) and the "byname" facets, which are supposed
to read files to determine their behavior.  The base ctype<>, collate<>,
and numpunct<> facets are "mostly complete", except that the table of
bitmask values used for "is" operations, and corresponding mask values,
are still defined in libio and just included/linked.  (We will need to
implement these tables independently, soon, but should take advantage
of libio where possible.)  The num_put<>::put members for integer types
are "mostly complete".

A complete list of what has and has not been implemented may be
found in CHECKLIST.  However, note that the current definition of
codecvt<wchar_t,char,mbstate_t> is wrong.  It should simply write
out the raw bytes representing the wide characters, rather than
trying to convert each to a corresponding single "char" value.

Some of the facets are more important than others.  Specifically,
the members of ctype<>, numpunct<>, num_put<>, and num_get<> facets
are used by other library facilities defined in <string>, <istream>,
and <ostream>, and the codecvt<> facet is used by basic_filebuf<>
in <fstream>, so a conforming iostream implementation depends on
these.

The "long long" type eventually must be supported, but code mentioning
it should be wrapped in #if guards to allow pedantic-mode compiling.

Performance of num_put<> and num_get<> depend critically on
caching computed values in ios_base objects, and on extensions
to the interface with streambufs.

Specifically: retrieving a copy of the locale object, extracting
the needed facets, and gathering data from them, for each call to
(e.g.) operator<< would be prohibitively slow.   To cache format
data for use by num_put<> and num_get<> we have a _Format_cache<>
object stored in the ios_base::pword() array.  This is constructed
and initialized lazily, and is organized purely for utility.  It
is discarded when a new locale with different facets is imbued.

Using only the public interfaces of the iterator arguments to the
facet functions would limit performance by forbidding "vector-style"
character operations.  The streambuf iterator optimizations are
described under chapter 24, but facets can also bypass the streambuf
iterators via explicit specializations and operate directly on the
streambufs, and use extended interfaces to get direct access to the
streambuf internal buffer arrays.  These extensions are mentioned
under chapter 27.  These optimizations are particularly important
for input parsing.

Unused virtual members of locale facets can be omitted, as mentioned
above, by a smart linker.

Chapter 23  Containers
----------------------
Headers: <deque> <list> <queue> <stack> <vector> <map> <set> <bitset>

All the components in chapter 23 are implemented in the SGI STL.
They are "mostly complete"; they include a large number of
nonconforming extensions which must be wrapped.  Some of these
are used internally and must be renamed or duplicated.

The SGI components are optimized for large-memory environments.  For
embedded targets, different criteria might be more appropriate.  Users
will want to be able to tune this behavior.  We should provide
ways for users to compile the library with different memory usage
characteristics.

A lot more work is needed on factoring out common code from different
specializations to reduce code size here and in chapter 25.  The
easiest fix for this would be a compiler/ABI improvement that allows
the compiler to recognize when a specialization depends only on the
size (or other gross quality) of a template argument, and allow the
linker to share the code with similar specializations.  In its
absence, many of the algorithms and containers can be partial-
specialized, at least for the case of pointers, but this only solves
a small part of the problem.  Use of a type_traits-style template
allows a few more optimization opportunities, more if the compiler
can generate the specializations automatically.

As an optimization, containers can specialize on the default allocator
and bypass it, or take advantage of details of its implementation
after it has been improved upon.

Replacing the vector iterators, which currently are simple element
pointers, with class objects would greatly increase the safety of the
client interface, and also permit a "debug" mode in which range,
ownership, and validity are rigorously checked.  The current use of
pointers for iterators is evil.

As mentioned for chapter 24, the deque iterator is a good example of
an opportunity to implement a "staged" iterator that would benefit
from specializations of some algorithms.

Chapter 24  Iterators
---------------------
Headers: <iterator>

Standard iterators are "mostly complete", with the exception of
the stream iterators, which are not yet templatized on the
stream type.  Also, the base class template iterator<> appears
to be wrong, so everything derived from it must also be wrong,
currently.

The streambuf iterators (currently located in stl/bits/std_iterator.h,
but should be under bits/) can be rewritten to take advantage of
friendship with the streambuf implementation.

Matt Austern has identified opportunities where certain iterator
types, particularly including streambuf iterators and deque
iterators, have a "two-stage" quality, such that an intermediate
limit can be checked much more quickly than the true limit on
range operations.  If identified with a member of iterator_traits,
algorithms may be specialized for this case.  Of course the
iterators that have this quality can be identified by specializing
a traits class.

Many of the algorithms must be specialized for the streambuf
iterators, to take advantage of block-mode operations, in order
to allow iostream/locale operations' performance not to suffer.
It may be that they could be treated as staged iterators and
take advantage of those optimizations.

Chapter 25  Algorithms
----------------------
Headers: <algorithm>
C headers: <cstdlib> (also in 18, 21, 26))

The algorithms are "mostly complete".  As mentioned above, they
are optimized for speed at the expense of code and data size.

Specializations of many of the algorithms for non-STL types would
give performance improvements, but we must use great care not to
interfere with fragile template overloading semantics for the
standard interfaces.  Conventionally the standard function template
interface is an inline which delegates to a non-standard function
which is then overloaded (this is already done in many places in
the library).  Particularly appealing opportunities for the sake of
iostream performance are for copy and find applied to streambuf
iterators or (as noted elsewhere) for staged iterators, of which
the streambuf iterators are a good example.

The bsearch and qsort functions cannot be overloaded properly as
required by the standard because gcc does not yet allow overloading
on the extern-"C"-ness of a function pointer.

Chapter 26  Numerics
--------------------
Headers: <complex> <valarray> <numeric>
C headers: <cmath>, <cstdlib> (also 18, 21, 25)

Numeric components: Gabriel dos Reis's valarray, Drepper's complex,
and the few algorithms from the STL are "mostly done".   Of course
optimization opportunities abound for the numerically literate.  It
is not clear whether the valarray implementation really conforms
fully, in the assumptions it makes about aliasing (and lack thereof)
in its arguments.

The C div() and ldiv() functions are interesting, because they are the
only case where a C library function returns a class object by value.
Since the C++ type div_t must be different from the underlying C type
(which is in the wrong namespace) the underlying functions div() and
ldiv() cannot be re-used efficiently.  Fortunately they are trivial to
re-implement.

Chapter 27  Iostreams
---------------------
Headers: <iosfwd> <streambuf> <ios> <ostream> <istream> <iostream>
         <iomanip> <sstream> <fstream>
C headers: <cstdio> <cwchar> (also in 21)

Iostream is currently in a very incomplete state.  <iosfwd>, <iomanip>,
ios_base, and basic_ios<> are "mostly complete".  basic_streambuf<> and
basic_ostream<> are well along, but basic_istream<> has had little work
done.  The standard stream objects, <sstream> and <fstream> have been
started; basic_filebuf<> "write" functions have been implemented just
enough to do "hello, world".

Most of the istream and ostream operators << and >> (with the exception
of the op<<(integer) ones) have not been changed to use locale primitives,
sentry objects, or char_traits members.

All these templates should be manually instantiated for char and
wchar_t in a way that links only used members into user programs.

Streambuf is fertile ground for optimization extensions.  An extended
interface giving iterator access to its internal buffer would be very
useful for other library components.

Iostream operations (primarily operators << and >>) can take advantage
of the case where user code has not specified a locale, and bypass locale
operations entirely.  The current implementation of op<</num_put<>::put,
for the integer types, demonstrates how they can cache encoding details
from the locale on each operation.  There is lots more room for
optimization in this area.

The definition of the relationship between the standard streams
cout et al. and stdout et al. requires something like a "stdiobuf".
The SGI solution of using double-indirection to actually use a
stdio FILE object for buffering is unsatisfactory, because it
interferes with peephole loop optimizations.

The <sstream> header work has begun.  stringbuf can benefit from
friendship with basic_string<> and basic_string<>::_Rep to use
those objects directly as buffers, and avoid allocating and making
copies.

The basic_filebuf<> template is a complex beast.  It is specified to
use the locale facet codecvt<> to translate characters between native
files and the locale character encoding.  In general this involves
two buffers, one of "char" representing the file and another of
"char_type", for the stream, with codecvt<> translating.  The process
is complicated by the variable-length nature of the translation, and
the need to seek to corresponding places in the two representations.
For the case of basic_filebuf<char>, when no translation is needed,
a single buffer suffices.  A specialized filebuf can be used to reduce
code space overhead when no locale has been imbued.  Matt Austern's
work at SGI will be useful, perhaps directly as a source of code, or
at least as an example to draw on.

Filebuf, almost uniquely (cf. operator new), depends heavily on
underlying environmental facilities.  In current releases iostream
depends fairly heavily on libio constant definitions, but it should
be made independent.   It also depends on operating system primitives
for file operations.  There is immense room for optimizations using
(e.g.) mmap for reading.  The shadow/ directory wraps, besides the
standard C headers, the libio.h and unistd.h headers, for use mainly
by filebuf.  These wrappings have not been completed, though there
is scaffolding in place.

The encapulation of certain C header <cstdio> names presents an
interesting problem.  It is possible to define an inline std::fprintf()
implemented in terms of the 'extern "C"' vfprintf(), but there is no
standard vfscanf() to use to implement std::fscanf().  It appears that
vfscanf but be re-implemented in C++ for targets where no vfscanf
extension has been defined.  This is interesting in that it seems
to be the only significant case in the C library where this kind of
rewriting is necessary.  (Of course Glibc provides the vfscanf()
extension.)  (The functions related to exit() must be rewritten
for other reasons.)


Annex D
-------
Headers: <strstream>

Annex D defines many non-library features, and many minor
modifications to various headers, and a complete header.
It is "mostly done", except that the libstdc++-2 <strstream>
header has not been adopted into the library, or checked to
verify that it matches the draft in those details that were
clarified by the committee.  Certainly it must at least be
moved into the std namespace.

We still need to wrap all the deprecated features in #if guards
so that pedantic compile modes can detect their use.

Nonstandard Extensions
----------------------
Headers: <iostream.h> <strstream.h> <hash> <rbtree>
         <pthread_alloc> <stdiobuf> (etc.)

User code has come to depend on a variety of nonstandard components
that we must not omit.  Much of this code can be adopted from
libstdc++-v2 or from the SGI STL.  This particularly includes
<iostream.h>, <strstream.h>, and various SGI extensions such
as <hash_map.h>.  Many of these are already placed in the
subdirectories ext/ and backward/.  (Note that it is better to
include them via "<backward/hash_map.h>" or "<ext/hash_map>" than
to search the subdirectory itself via a "-I" directive.