todo

Mac OS X 10.4.9 for x86 Source Code gcc 实现源代码

std::allocator

  - switch to mt_allocator with --enable-threads=posix. 

  - persistent allocator

  - global/extern allocator

std::string 

  - Policy-based design incorporating COW
    vs. deep copy issues, MT scalability
    See Andrei Alexandrescu, June 2001, C/C++ Users Journal
    "Generic<Programming>: A Policy-Based basic_string Implementation"
    http://www.cuj.com/documents/s=7994/cujcexp1906alexandr/

  - operator!= and utility/rel_ops operators need to be made safe with
    string and vector iterator classes. basic_string::reverse_iterator may
    be implemented incorrectly, or need things like
    operator==(__normal_iterator, const char*&), and swap(vector)

  - 'do the right thing' ctor fixing needs to be done for string. This
    is still subject to some debate on the library issues list, so I
    suggest punting till the dust clears.

  - fix template members of basic_string<> to overload iterators and
    non-iterators properly.  (This is the infamous hack as in vector<> etc
    23.1.1 para 10.)

std::locale  

  - implement __convert_to_v and __convert_from_v without "C" library
    functions and and LANG environment variable dependencies.

  - use localedata to implement generic named (non-MT-safe) locales?
    Figure out a way to use ICU data, like libjava? Re-package and use
    the glibc localedata, even if we aren't on linux? Need a generic
    locale model that does something besides the "C" locale.

  - make locale::classic() separate from named locale code. This will
    improve the static linkage situation, but will require new
    initialization code. In particular, we need lazy-initialization of
    locale::classic(), and maybe the has_facet/use_facet functions for all
    the required facets. The end goal is a self-contained
    locale_init.cc, or one with transitive closure without the locale
    instantiations (locale-inst.cc) or the named locale bits
    (localename.cc).

  - Jerry(?)/Paolo(?) work on __float_to_char.

  - minimize ctype convertion in data facets, see numpunct/num_put/num_get

std::basic_filebuf, 27_io

  - wfilebuf, get variable-encoding working and tested, including
    positioning and seeking. (I think this may be done now)

  - wfilebuf testsuite (getting there...)

  - look ahead for unbuffered io, so know when multiple putc's can be
    coalesced.

  - unlocked __basic_file + new mutext class

  - optimized the sentries for istream/ostream

  - v2 vs. v3 speed

  - add optimization hooks (esp. whitespace eating) to streambuf
  - add _M_begin() and _M_end() to streambuf
  - add algorithm specializations for [io]streambuf_iterator (copy find etc.)

testsuite

  - valgrind hooks into make check so can tell memory leakage
    Some commentary on the valgrind users list

  - add hooks for qmtest for visual diffs 

  - automatic testing of interactive tests

  - diffing generated output files

  - provide testsuites for numerics.

  - make check-abi needs to have full symbol checking. Scope the LSB
    testsuite, see what's going on with the typeinfo etc. bits.

  - try to do a better job of ABI testing, with instantiations of all
    standard-specified types checked, not just exported symbols.

g++/binutils

  - compression for wide versions of basic types, not just narrow

threads

  - create MT abstraction layer for atomicity to pthreads.

  - solution for threads + C++.

- audit for places where __builtin_expect can be used.

- relocations, work on getting these down

- issues with __builtin_memcpy and std::copy from Jerry Quinn
  http://gcc.gnu.org/ml/libstdc++/2003-02/msg00056.html
  http://gcc.gnu.org/ml/libstdc++/2003-02/msg00302.html
  http://gcc.gnu.org/ml/gcc/2003-10/msg01305.html

- fix dependency tracking for includes (.h, .tcc) during build process.

- coordinate with "C" library people the "C" compatibility headers.

- Think about naming all member data and member functions consistently
  as per
  funtions: _M_verb_adverb
  data: _M_noun_adjective

- A C++STYLE guide that deals with nested namespaces, and that
everybody can live with.

- exception specifications need to be reviewed for all parts of the
library support and utility areas, particularly <new>. Part of this is
a standards issue, where the 27_io bugs/standard is really in an odd
spot.

- C-related issues WRT to io and filepos, mbstate_t. Seeking in wide
streams. May need to define operators for mbstate_t so that
'mbstate_t& == mbstate_t' is something that can be done.

- scoping/linking issues WRT to C structs need to be worked out. See
Nathan's commentary on cantrip, http://www.cantrip.org/cheaders.html

- auto_ptr: seems to be some disagreement on what is
standards-conformant behavior, specially on conversion operators.

- list::assignment operator needs const_cast

- a cleaner division between pointers-to-value_type and true iterators
needs to be drawn throughout the entire STL implementation.

- priority_queue conversions may be non-conformant

- Protect valarray::result_type (not Standard) and make it work with
  the various helper classes.

- Make sure `valarray<bool> & == _Expr<_BinClos<logical_or,_ValArray,_ValArray,double,double>,bool>' 
  is defined

- All of the Library working group closed issues need to be
addressed. Some of them proposed resolutions are already in the v-3
sources, with macro-guards. Also, same with the TR.

- add deprecated features from Annex D
  - add #ifdef wrappers for all of them, including in each
    C <xxx.h> header
  - replace typedef'd non-deprecated types with more type-safe equivalents.

- add feature-test macros for non-standard extensions

- add MT support for locale, string, istream, ostream

- need to think about doing a .texi or DocBook manual, instead of all
these HTML pages. In addition, it would be nice to have a full manual,
instead of a lot of ad-hoc pages. Weaknesses include numerics, locale,
and io.

- add FAQ entries -- improve the install instructions

- add HOWTO entries

- do more doxygen manpages