diff options
| author | nga <nga@yandex-team.ru> | 2022-02-10 16:48:09 +0300 |
|---|---|---|
| committer | Daniil Cherednik <dcherednik@yandex-team.ru> | 2022-02-10 16:48:09 +0300 |
| commit | c2a1af049e9deca890e9923abe64fe6c59060348 (patch) | |
| tree | b222e5ac2e2e98872661c51ccceee5da0d291e13 /contrib/libs/sparsehash | |
| parent | 1f553f46fb4f3c5eec631352cdd900a0709016af (diff) | |
| download | ydb-c2a1af049e9deca890e9923abe64fe6c59060348.tar.gz | |
Restoring authorship annotation for <nga@yandex-team.ru>. Commit 2 of 2.
Diffstat (limited to 'contrib/libs/sparsehash')
20 files changed, 7486 insertions, 7486 deletions
diff --git a/contrib/libs/sparsehash/AUTHORS b/contrib/libs/sparsehash/AUTHORS index 5bd5a287f9..d8c24c64ca 100644 --- a/contrib/libs/sparsehash/AUTHORS +++ b/contrib/libs/sparsehash/AUTHORS @@ -1,2 +1,2 @@ -google-sparsehash@googlegroups.com - +google-sparsehash@googlegroups.com + diff --git a/contrib/libs/sparsehash/COPYING b/contrib/libs/sparsehash/COPYING index ef86b820ed..e4956cfd9f 100644 --- a/contrib/libs/sparsehash/COPYING +++ b/contrib/libs/sparsehash/COPYING @@ -1,28 +1,28 @@ -Copyright (c) 2005, Google Inc. -All rights reserved. - -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions are -met: - - * Redistributions of source code must retain the above copyright -notice, this list of conditions and the following disclaimer. - * Redistributions in binary form must reproduce the above -copyright notice, this list of conditions and the following disclaimer -in the documentation and/or other materials provided with the -distribution. - * Neither the name of Google Inc. nor the names of its -contributors may be used to endorse or promote products derived from -this software without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +Copyright (c) 2005, Google Inc. +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: + + * Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above +copyright notice, this list of conditions and the following disclaimer +in the documentation and/or other materials provided with the +distribution. + * Neither the name of Google Inc. nor the names of its +contributors may be used to endorse or promote products derived from +this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/contrib/libs/sparsehash/ChangeLog b/contrib/libs/sparsehash/ChangeLog index ce24746e9a..fd53c6f8fd 100644 --- a/contrib/libs/sparsehash/ChangeLog +++ b/contrib/libs/sparsehash/ChangeLog @@ -3,277 +3,277 @@ Mon Oct 12 21:00:00 2015 Google Inc. <google-sparsehash@googlegroups.com> * sparsehash: version 2.0.3 * Fix compilation on modern compilers and operating systems -Thu Feb 23 23:47:18 2012 Google Inc. <google-sparsehash@googlegroups.com> - - * sparsehash: version 2.0.2 - * BUGFIX: Fix backwards compatibility for <google> include folders - -Wed Feb 01 02:57:48 2012 Google Inc. <google-sparsehash@googlegroups.com> - - * sparsehash: version 2.0.1 - * BUGFIX: Fix path to malloc_extension.h in time_hash_map.cc - -Tue Jan 31 11:33:04 2012 Google Inc. <google-sparsehash@googlegroups.com> - - * sparsehash: version 2.0 - * Renamed include directory from google/ to sparsehash/ (csilvers) - * Changed the 'official' sparsehash email in setup.py/etc - * Renamed google-sparsehash.sln to sparsehash.sln - * Changed copyright text to reflect Google's relinquished ownership - -Tue Dec 20 21:04:04 2011 Google Inc. <opensource@google.com> - - * sparsehash: version 1.12 release - * Add support for serializing/unserializing dense_hash_map/set to disk - * New simpler and more flexible serialization API - * Be more consistent about clearing on unserialize() even if it fails - * Quiet some compiler warnings about unused variables - * Add a timing test for iterating (suggested by google code issue 77) - * Add offset_to_pos, the opposite of pos_to_offset, to sparsetable - * PORTING: Add some missing #includes, needed on some systems - * Die at configure-time when g++ isn't installed - * Successfully make rpm's even when dpkg is missing - * Improve deleted key test in util/gtl/{dense,sparse}hashtable - * Update automake to 1.10.1, and autoconf to 2.62 - -Thu Jun 23 21:12:58 2011 Google Inc. <opensource@google.com> - - * sparsehash: version 1.11 release - * Improve performance on pointer keys by ignoring always-0 low bits - * Fix missing $(top_srcdir) in Makefile.am, which broke some compiles - * BUGFIX: Fix a crashing typo-bug in swap() - * PORTING: Remove support for old compilers that do not use 'std' - * Add some new benchmarks to test for a place dense_hash_* does badly - * Some cosmetic changes due to a switch to a new releasing tool - -Thu Jan 20 16:07:39 2011 Google Inc. <opensource@google.com> - - * sparsehash: version 1.10 release - * Follow ExtractKey return type, allowing it to return a reference - * PORTING: fix MSVC 10 warnings (constifying result_type, placement-new) - * Update from autoconf 2.61 to autoconf 2.65 - -Fri Sep 24 11:37:50 2010 Google Inc. <opensource@google.com> - - * sparsehash: version 1.9 release - * Add is_enum; make all enums PODs by default (romanp) - * Make find_or_insert() usable directly (dawidk) - * Use zero-memory trick for allocators to reduce space use (guilin) - * Fix some compiler warnings (chandlerc, eraman) - * BUGFIX: int -> size_type in one function we missed (csilvers) - * Added sparsehash.pc, for pkg-config (csilvers) - -Thu Jul 29 15:01:29 2010 Google Inc. <opensource@google.com> - - * sparsehash: version 1.8.1 release - * Remove -Werror from Makefile: gcc 4.3 gives spurious warnings - -Thu Jul 29 09:53:26 2010 Google Inc. <opensource@google.com> - - * sparsehash: version 1.8 release - * More support for Allocator, including allocator ctor arg (csilvers) - * Repack hasthable vars to reduce container size *more* (giao) - * Speed up clear() (csilvers) - * Change HT_{OCCUPANCY,SHRINK}_FLT from float to int (csilvers) - * Revamp test suite for more complete code & timing coverage (csilvers) - * BUGFIX: Enforce max_size for dense/sparse_hashtable (giao, csilvers) - * BUGFIX: Raise exception instead of crashing on overflow (csilvers) - * BUGFIX: Allow extraneous const in key type (csilvers) - * BUGFIX: Allow same functor for both hasher and key_equals (giao) - * PORTING: remove is_convertible, which gives AIX cc fits (csilvers) - * PORTING: Renamed README.windows to README_windows.txt (csilvers) - * Created non-empty NEWS file (csilvers) - -Wed Mar 31 12:32:03 2010 Google Inc. <opensource@google.com> - - * sparsehash: version 1.7 release - * Add support for Allocator (guilin) - * Add libc_allocator_with_realloc as the new default allocator (guilin) - * Repack {sparse,dense}hashtable vars to reduce container size (giao) - * BUGFIX: operator== no longer requires same table ordering (csilvers) - * BUGFIX: fix dense_hash_*(it,it) by requiring empty-key too (csilvers) - * PORTING: fix language bugs that gcc allowed (csilvers, chandlerc) - * Update from autoconf 2.61 to autoconf 2.64 - -Fri Jan 8 14:47:55 2010 Google Inc. <opensource@google.com> - - * sparsehash: version 1.6 release - * New accessor methods for deleted_key, empty_key (sjackman) - * Use explicit hash functions in sparsehash tests (csilvers) - * BUGFIX: Cast resize to fix SUNWspro bug (csilvers) - * Check for sz overflow in min_size (csilvers) - * Speed up clear() for dense and sparse hashtables (jeff) - * Avoid shrinking in all cases when min-load is 0 (shaunj, csilvers) - * Improve densehashtable code for the deleted key (gpike) - * BUGFIX: Fix operator= when the 2 empty-keys differ (andreidam) - * BUGFIX: Fix ht copying when empty-key isn't set (andreidam) - * PORTING: Use TmpFile() instead of /tmp on MinGW (csilvers) - * PORTING: Use filenames that work with Stratus VOS. - -Tue May 12 14:16:38 2009 Google Inc. <opensource@google.com> - - * sparsehash: version 1.5.2 release - * Fix compile error: not initializing set_key in all constructors - -Fri May 8 15:23:44 2009 Google Inc. <opensource@google.com> - - * sparsehash: version 1.5.1 release - * Fix broken equal_range() for all the hash-classes (csilvers) - -Wed May 6 11:28:49 2009 Google Inc. <opensource@google.com> - - * sparsehash: version 1.5 release - * Support the tr1 unordered_map (and unordered_set) API (csilvers) - * Store only key for delkey; reduces need for 0-arg c-tor (csilvers) - * Prefer unordered_map to hash_map for the timing test (csilvers) - * PORTING: update the resource use for 64-bit machines (csilvers) - * PORTING: fix MIN/MAX collisions by un-#including windows.h (csilvers) - * Updated autoconf version to 2.61 and libtool version to 1.5.26 - -Wed Jan 28 17:11:31 2009 Google Inc. <opensource@google.com> - - * sparsehash: version 1.4 release - * Allow hashtables to be <32 buckets (csilvers) - * Fix initial-sizing bug: was sizing tables too small (csilvers) - * Add asserts that clients don't abuse deleted/empty key (csilvers) - * Improve determination of 32/64 bit for C code (csilvers) - * Small fix for doc files in rpm (csilvers) - -Thu Nov 6 15:06:09 2008 Google Inc. <opensource@google.com> - - * sparsehash: version 1.3 release - * Add an interface to change the parameters for resizing (myl) - * Document another potentially good hash function (csilvers) - -Thu Sep 18 13:53:20 2008 Google Inc. <opensource@google.com> - - * sparsehash: version 1.2 release - * Augment documentation to better describe namespace issues (csilvers) - * BUG FIX: replace hash<> with SPARSEHASH_HASH, for windows (csilvers) - * Add timing test to unittest to test repeated add+delete (csilvers) - * Do better picking a new size when resizing (csilvers) - * Use ::google instead of google as a namespace (csilvers) - * Improve threading test at config time (csilvers) - -Mon Feb 11 16:30:11 2008 Google Inc. <opensource@google.com> - - * sparsehash: version 1.1 release - * Fix brown-paper-bag bug in some constructors (rafferty) - * Fix problem with variables shadowing member vars, add -Wshadow - -Thu Nov 29 11:44:38 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 1.0.2 release - * Fix a final reference to hash<> to use SPARSEHASH_HASH<> instead. - -Wed Nov 14 08:47:48 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 1.0.1 release :-( - * Remove an unnecessary (harmful) "#define hash" in windows' config.h - -Tue Nov 13 15:15:46 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 1.0 release! We are now out of beta. - * Clean up Makefile awk script to be more readable (csilvers) - * Namespace fixes: use fewer #defines, move typedefs into namespace - -Fri Oct 12 12:35:24 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 0.9.1 release - * Fix Makefile awk script to work on more architectures (csilvers) - * Add test to test code in more 'real life' situations (csilvers) - -Tue Oct 9 14:15:21 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 0.9 release - * More type-hygiene improvements, especially for 64-bit (csilvers) - * Some configure improvements to improve portability, utility (austern) - * Small bugfix for operator== for dense_hash_map (jeff) - -Tue Jul 3 12:55:04 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 0.8 release - * Minor type-hygiene improvements: size_t for int, etc. (csilvers) - * Porting improvements: tests pass on OS X, FreeBSD, Solaris (csilvers) - * Full windows port! VS solution provided for all unittests (csilvers) - -Mon Jun 11 11:33:41 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 0.7 release - * Syntax fixes to better support gcc 4.3 and VC++ 7 (mec, csilvers) - * Improved windows/VC++ support (see README.windows) (csilvers) - * Config improvements: better tcmalloc support and config.h (csilvers) - * More robust with missing hash_map + nix 'trampoline' .h's (csilvers) - * Support for STLport's hash_map/hash_fun locations (csilvers) - * Add .m4 files to distribution; now all source is there (csilvers) - * Tiny modification of shrink-threshhold to allow never-shrinking (amc) - * Protect timing tests against aggressive optimizers (csilvers) - * Extend time_hash_map to test bigger objects (csilvers) - * Extend type-trait support to work with const objects (csilvers) - * USER VISIBLE: speed up all code by replacing memmove with memcpy - (csilvers) - -Tue Mar 20 17:29:34 2007 Google Inc. <opensource@google.com> - - * sparsehash: version 0.6 release - * Some improvement to type-traits (jyasskin) - * Better timing results when google-perftools is installed (sanjay) - * Updates and fixes to html documentation and README (csilvers) - * A bit more careful about #includes (csilvers) - * Fix for typo that broken compilation on some systems (csilvers) - * USER VISIBLE: New clear_no_resize() method added to dense_hash_map - (uszkoreit) - -Sat Oct 21 13:47:47 2006 Google Inc. <opensource@google.com> - - * sparsehash: version 0.5 release - * Support uint16_t (SunOS) in addition to u_int16_t (BSD) (csilvers) - * Get rid of UNDERSTANDS_ITERATOR_TAGS; everyone understands (csilvers) - * Test that empty-key and deleted-key differ (rbayardo) - * Fix example docs: strcmp needs to test for NULL (csilvers) - -Sun Apr 23 22:42:35 2006 Google Inc. <opensource@google.com> - - * sparsehash: version 0.4 release - * Remove POD requirement for keys and values! (austern) - * Add tr1-compatible type-traits system to speed up POD ops. (austern) - * Fixed const-iterator bug where postfix ++ didn't compile. (csilvers) - * Fixed iterator comparison bugs where <= was incorrect. (csilvers) - * Clean up config.h to keep its #defines from conflicting. (csilvers) - * Big documentation sweep and cleanup. (csilvers) - * Update documentation to talk more about good hash fns. (csilvers) - * Fixes to compile on MSVC (working around some MSVC bugs). (rennie) - * Avoid resizing hashtable on operator[] lookups (austern) - -Thu Nov 3 20:12:31 2005 Google Inc. <opensource@google.com> - - * sparsehash: version 0.3 release - * Quiet compiler warnings on some compilers. (csilvers) - * Some documentation fixes: example code for dense_hash_map. (csilvers) - * Fix a bug where swap() wasn't swapping delete_key(). (csilvers) - * set_deleted_key() and set_empty_key() now take a key only, - allowing hash-map values to be forward-declared. (csilvers) - * support for std::insert_iterator (and std::inserter). (csilvers) - -Mon May 2 07:04:46 2005 Google Inc. <opensource@google.com> - - * sparsehash: version 0.2 release - * Preliminary support for msvc++ compilation. (csilvers) - * Documentation fixes -- some example code was incomplete! (csilvers) - * Minimize size of config.h to avoid other-package conflicts (csilvers) - * Contribute a C-based version of sparsehash that served as the - inspiration for this code. One day, I hope to clean it up and - support it, but for now it's just in experimental/, for playing - around with. (csilvers) - * Change default namespace from std to google. (csilvers) - -Fri Jan 14 16:53:32 2005 Google Inc. <opensource@google.com> - - * sparsehash: initial release: - The sparsehash package contains several hash-map implementations, - similar in API to SGI's hash_map class, but with different - performance characteristics. sparse_hash_map uses very little - space overhead: 1-2 bits per entry. dense_hash_map is typically - faster than the default SGI STL implementation. This package - also includes hash-set analogues of these classes. - +Thu Feb 23 23:47:18 2012 Google Inc. <google-sparsehash@googlegroups.com> + + * sparsehash: version 2.0.2 + * BUGFIX: Fix backwards compatibility for <google> include folders + +Wed Feb 01 02:57:48 2012 Google Inc. <google-sparsehash@googlegroups.com> + + * sparsehash: version 2.0.1 + * BUGFIX: Fix path to malloc_extension.h in time_hash_map.cc + +Tue Jan 31 11:33:04 2012 Google Inc. <google-sparsehash@googlegroups.com> + + * sparsehash: version 2.0 + * Renamed include directory from google/ to sparsehash/ (csilvers) + * Changed the 'official' sparsehash email in setup.py/etc + * Renamed google-sparsehash.sln to sparsehash.sln + * Changed copyright text to reflect Google's relinquished ownership + +Tue Dec 20 21:04:04 2011 Google Inc. <opensource@google.com> + + * sparsehash: version 1.12 release + * Add support for serializing/unserializing dense_hash_map/set to disk + * New simpler and more flexible serialization API + * Be more consistent about clearing on unserialize() even if it fails + * Quiet some compiler warnings about unused variables + * Add a timing test for iterating (suggested by google code issue 77) + * Add offset_to_pos, the opposite of pos_to_offset, to sparsetable + * PORTING: Add some missing #includes, needed on some systems + * Die at configure-time when g++ isn't installed + * Successfully make rpm's even when dpkg is missing + * Improve deleted key test in util/gtl/{dense,sparse}hashtable + * Update automake to 1.10.1, and autoconf to 2.62 + +Thu Jun 23 21:12:58 2011 Google Inc. <opensource@google.com> + + * sparsehash: version 1.11 release + * Improve performance on pointer keys by ignoring always-0 low bits + * Fix missing $(top_srcdir) in Makefile.am, which broke some compiles + * BUGFIX: Fix a crashing typo-bug in swap() + * PORTING: Remove support for old compilers that do not use 'std' + * Add some new benchmarks to test for a place dense_hash_* does badly + * Some cosmetic changes due to a switch to a new releasing tool + +Thu Jan 20 16:07:39 2011 Google Inc. <opensource@google.com> + + * sparsehash: version 1.10 release + * Follow ExtractKey return type, allowing it to return a reference + * PORTING: fix MSVC 10 warnings (constifying result_type, placement-new) + * Update from autoconf 2.61 to autoconf 2.65 + +Fri Sep 24 11:37:50 2010 Google Inc. <opensource@google.com> + + * sparsehash: version 1.9 release + * Add is_enum; make all enums PODs by default (romanp) + * Make find_or_insert() usable directly (dawidk) + * Use zero-memory trick for allocators to reduce space use (guilin) + * Fix some compiler warnings (chandlerc, eraman) + * BUGFIX: int -> size_type in one function we missed (csilvers) + * Added sparsehash.pc, for pkg-config (csilvers) + +Thu Jul 29 15:01:29 2010 Google Inc. <opensource@google.com> + + * sparsehash: version 1.8.1 release + * Remove -Werror from Makefile: gcc 4.3 gives spurious warnings + +Thu Jul 29 09:53:26 2010 Google Inc. <opensource@google.com> + + * sparsehash: version 1.8 release + * More support for Allocator, including allocator ctor arg (csilvers) + * Repack hasthable vars to reduce container size *more* (giao) + * Speed up clear() (csilvers) + * Change HT_{OCCUPANCY,SHRINK}_FLT from float to int (csilvers) + * Revamp test suite for more complete code & timing coverage (csilvers) + * BUGFIX: Enforce max_size for dense/sparse_hashtable (giao, csilvers) + * BUGFIX: Raise exception instead of crashing on overflow (csilvers) + * BUGFIX: Allow extraneous const in key type (csilvers) + * BUGFIX: Allow same functor for both hasher and key_equals (giao) + * PORTING: remove is_convertible, which gives AIX cc fits (csilvers) + * PORTING: Renamed README.windows to README_windows.txt (csilvers) + * Created non-empty NEWS file (csilvers) + +Wed Mar 31 12:32:03 2010 Google Inc. <opensource@google.com> + + * sparsehash: version 1.7 release + * Add support for Allocator (guilin) + * Add libc_allocator_with_realloc as the new default allocator (guilin) + * Repack {sparse,dense}hashtable vars to reduce container size (giao) + * BUGFIX: operator== no longer requires same table ordering (csilvers) + * BUGFIX: fix dense_hash_*(it,it) by requiring empty-key too (csilvers) + * PORTING: fix language bugs that gcc allowed (csilvers, chandlerc) + * Update from autoconf 2.61 to autoconf 2.64 + +Fri Jan 8 14:47:55 2010 Google Inc. <opensource@google.com> + + * sparsehash: version 1.6 release + * New accessor methods for deleted_key, empty_key (sjackman) + * Use explicit hash functions in sparsehash tests (csilvers) + * BUGFIX: Cast resize to fix SUNWspro bug (csilvers) + * Check for sz overflow in min_size (csilvers) + * Speed up clear() for dense and sparse hashtables (jeff) + * Avoid shrinking in all cases when min-load is 0 (shaunj, csilvers) + * Improve densehashtable code for the deleted key (gpike) + * BUGFIX: Fix operator= when the 2 empty-keys differ (andreidam) + * BUGFIX: Fix ht copying when empty-key isn't set (andreidam) + * PORTING: Use TmpFile() instead of /tmp on MinGW (csilvers) + * PORTING: Use filenames that work with Stratus VOS. + +Tue May 12 14:16:38 2009 Google Inc. <opensource@google.com> + + * sparsehash: version 1.5.2 release + * Fix compile error: not initializing set_key in all constructors + +Fri May 8 15:23:44 2009 Google Inc. <opensource@google.com> + + * sparsehash: version 1.5.1 release + * Fix broken equal_range() for all the hash-classes (csilvers) + +Wed May 6 11:28:49 2009 Google Inc. <opensource@google.com> + + * sparsehash: version 1.5 release + * Support the tr1 unordered_map (and unordered_set) API (csilvers) + * Store only key for delkey; reduces need for 0-arg c-tor (csilvers) + * Prefer unordered_map to hash_map for the timing test (csilvers) + * PORTING: update the resource use for 64-bit machines (csilvers) + * PORTING: fix MIN/MAX collisions by un-#including windows.h (csilvers) + * Updated autoconf version to 2.61 and libtool version to 1.5.26 + +Wed Jan 28 17:11:31 2009 Google Inc. <opensource@google.com> + + * sparsehash: version 1.4 release + * Allow hashtables to be <32 buckets (csilvers) + * Fix initial-sizing bug: was sizing tables too small (csilvers) + * Add asserts that clients don't abuse deleted/empty key (csilvers) + * Improve determination of 32/64 bit for C code (csilvers) + * Small fix for doc files in rpm (csilvers) + +Thu Nov 6 15:06:09 2008 Google Inc. <opensource@google.com> + + * sparsehash: version 1.3 release + * Add an interface to change the parameters for resizing (myl) + * Document another potentially good hash function (csilvers) + +Thu Sep 18 13:53:20 2008 Google Inc. <opensource@google.com> + + * sparsehash: version 1.2 release + * Augment documentation to better describe namespace issues (csilvers) + * BUG FIX: replace hash<> with SPARSEHASH_HASH, for windows (csilvers) + * Add timing test to unittest to test repeated add+delete (csilvers) + * Do better picking a new size when resizing (csilvers) + * Use ::google instead of google as a namespace (csilvers) + * Improve threading test at config time (csilvers) + +Mon Feb 11 16:30:11 2008 Google Inc. <opensource@google.com> + + * sparsehash: version 1.1 release + * Fix brown-paper-bag bug in some constructors (rafferty) + * Fix problem with variables shadowing member vars, add -Wshadow + +Thu Nov 29 11:44:38 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 1.0.2 release + * Fix a final reference to hash<> to use SPARSEHASH_HASH<> instead. + +Wed Nov 14 08:47:48 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 1.0.1 release :-( + * Remove an unnecessary (harmful) "#define hash" in windows' config.h + +Tue Nov 13 15:15:46 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 1.0 release! We are now out of beta. + * Clean up Makefile awk script to be more readable (csilvers) + * Namespace fixes: use fewer #defines, move typedefs into namespace + +Fri Oct 12 12:35:24 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 0.9.1 release + * Fix Makefile awk script to work on more architectures (csilvers) + * Add test to test code in more 'real life' situations (csilvers) + +Tue Oct 9 14:15:21 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 0.9 release + * More type-hygiene improvements, especially for 64-bit (csilvers) + * Some configure improvements to improve portability, utility (austern) + * Small bugfix for operator== for dense_hash_map (jeff) + +Tue Jul 3 12:55:04 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 0.8 release + * Minor type-hygiene improvements: size_t for int, etc. (csilvers) + * Porting improvements: tests pass on OS X, FreeBSD, Solaris (csilvers) + * Full windows port! VS solution provided for all unittests (csilvers) + +Mon Jun 11 11:33:41 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 0.7 release + * Syntax fixes to better support gcc 4.3 and VC++ 7 (mec, csilvers) + * Improved windows/VC++ support (see README.windows) (csilvers) + * Config improvements: better tcmalloc support and config.h (csilvers) + * More robust with missing hash_map + nix 'trampoline' .h's (csilvers) + * Support for STLport's hash_map/hash_fun locations (csilvers) + * Add .m4 files to distribution; now all source is there (csilvers) + * Tiny modification of shrink-threshhold to allow never-shrinking (amc) + * Protect timing tests against aggressive optimizers (csilvers) + * Extend time_hash_map to test bigger objects (csilvers) + * Extend type-trait support to work with const objects (csilvers) + * USER VISIBLE: speed up all code by replacing memmove with memcpy + (csilvers) + +Tue Mar 20 17:29:34 2007 Google Inc. <opensource@google.com> + + * sparsehash: version 0.6 release + * Some improvement to type-traits (jyasskin) + * Better timing results when google-perftools is installed (sanjay) + * Updates and fixes to html documentation and README (csilvers) + * A bit more careful about #includes (csilvers) + * Fix for typo that broken compilation on some systems (csilvers) + * USER VISIBLE: New clear_no_resize() method added to dense_hash_map + (uszkoreit) + +Sat Oct 21 13:47:47 2006 Google Inc. <opensource@google.com> + + * sparsehash: version 0.5 release + * Support uint16_t (SunOS) in addition to u_int16_t (BSD) (csilvers) + * Get rid of UNDERSTANDS_ITERATOR_TAGS; everyone understands (csilvers) + * Test that empty-key and deleted-key differ (rbayardo) + * Fix example docs: strcmp needs to test for NULL (csilvers) + +Sun Apr 23 22:42:35 2006 Google Inc. <opensource@google.com> + + * sparsehash: version 0.4 release + * Remove POD requirement for keys and values! (austern) + * Add tr1-compatible type-traits system to speed up POD ops. (austern) + * Fixed const-iterator bug where postfix ++ didn't compile. (csilvers) + * Fixed iterator comparison bugs where <= was incorrect. (csilvers) + * Clean up config.h to keep its #defines from conflicting. (csilvers) + * Big documentation sweep and cleanup. (csilvers) + * Update documentation to talk more about good hash fns. (csilvers) + * Fixes to compile on MSVC (working around some MSVC bugs). (rennie) + * Avoid resizing hashtable on operator[] lookups (austern) + +Thu Nov 3 20:12:31 2005 Google Inc. <opensource@google.com> + + * sparsehash: version 0.3 release + * Quiet compiler warnings on some compilers. (csilvers) + * Some documentation fixes: example code for dense_hash_map. (csilvers) + * Fix a bug where swap() wasn't swapping delete_key(). (csilvers) + * set_deleted_key() and set_empty_key() now take a key only, + allowing hash-map values to be forward-declared. (csilvers) + * support for std::insert_iterator (and std::inserter). (csilvers) + +Mon May 2 07:04:46 2005 Google Inc. <opensource@google.com> + + * sparsehash: version 0.2 release + * Preliminary support for msvc++ compilation. (csilvers) + * Documentation fixes -- some example code was incomplete! (csilvers) + * Minimize size of config.h to avoid other-package conflicts (csilvers) + * Contribute a C-based version of sparsehash that served as the + inspiration for this code. One day, I hope to clean it up and + support it, but for now it's just in experimental/, for playing + around with. (csilvers) + * Change default namespace from std to google. (csilvers) + +Fri Jan 14 16:53:32 2005 Google Inc. <opensource@google.com> + + * sparsehash: initial release: + The sparsehash package contains several hash-map implementations, + similar in API to SGI's hash_map class, but with different + performance characteristics. sparse_hash_map uses very little + space overhead: 1-2 bits per entry. dense_hash_map is typically + faster than the default SGI STL implementation. This package + also includes hash-set analogues of these classes. + diff --git a/contrib/libs/sparsehash/NEWS b/contrib/libs/sparsehash/NEWS index c97f503bff..4af929c394 100644 --- a/contrib/libs/sparsehash/NEWS +++ b/contrib/libs/sparsehash/NEWS @@ -1,193 +1,193 @@ == 12 October 2015 == - + Various small fixes to ensure compilation on modern compilers and operating systems. Tagged as 2.0.3 == 23 February 2012 == -A backwards incompatibility arose from flattening the include headers -structure for the <google> folder. - -This is now fixed in 2.0.2. You only need to upgrade if you had previously -included files from the <google/sparsehash> folder. - -== 1 February 2012 == - -A minor bug related to the namespace switch from google to sparsehash -stopped the build from working when perftools is also installed. - -This is now fixed in 2.0.1. You only need to upgrade if you have perftools -installed. - -== 31 January 2012 == - -I've just released sparsehash 2.0. - -The `google-sparsehash` project has been renamed to `sparsehash`. I -(csilvers) am stepping down as maintainer, to be replaced by the team -of Donovan Hide and Geoff Pike. Welcome to the team, Donovan and -Geoff! Donovan has been an active contributor to sparsehash bug -reports and discussions in the past, and Geoff has been closely -involved with sparsehash inside Google (in addition to writing the -[http://code.google.com/p/cityhash CityHash hash function]). The two -of them together should be a formidable force. For good. - -I bumped the major version number up to 2 to reflect the new community -ownership of the project. All the -[http://sparsehash.googlecode.com/svn/tags/sparsehash-2.0/ChangeLog changes] -are related to the renaming. - -The only functional change from sparsehash 1.12 is that I've renamed -the `google/` include-directory to be `sparsehash/` instead. New code -should `#include <sparsehash/sparse_hash_map>`/etc. I've kept the old -names around as forwarding headers to the new, so `#include -<google/sparse_hash_map>` will continue to work. - -Note that the classes and functions remain in the `google` C++ -namespace (I didn't change that to `sparsehash` as well); I think -that's a trickier transition, and can happen in a future release. - - -=== 18 January 2011 === - -The `google-sparsehash` Google Code page has been renamed to -`sparsehash`, in preparation for the project being renamed to -`sparsehash`. In the coming weeks, I'll be stepping down as -maintainer for the sparsehash project, and as part of that Google is -relinquishing ownership of the project; it will now be entirely -community run. The name change reflects that shift. - - -=== 20 December 2011 === - -I've just released sparsehash 1.12. This release features improved -I/O (serialization) support. Support is finally added to serialize -and unserialize `dense_hash_map`/`set`, paralleling the existing code -for `sparse_hash_map`/`set`. In addition, the serialization API has -gotten simpler, with a single `serialize()` method to write to disk, -and an `unserialize()` method to read from disk. Finally, support has -gotten more generic, with built-in support for both C `FILE*`s and C++ -streams, and an extension mechanism to support arbitrary sources and -sinks. - -There are also more minor changes, including minor bugfixes, an -improved deleted-key test, and a minor addition to the `sparsetable` -API. See the [http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.12/ChangeLog ChangeLog] -for full details. - -=== 23 June 2011 === - -I've just released sparsehash 1.11. The major user-visible change is -that the default behavior is improved -- using the hash_map/set is -faster -- for hashtables where the key is a pointer. We now notice -that case and ignore the low 2-3 bits (which are almost always 0 for -pointers) when hashing. - -Another user-visible change is we've removed the tests for whether the -STL (vector, pair, etc) is defined in the 'std' namespace. gcc 2.95 -is the most recent compiler I know of to put STL types and functions -in the global namespace. If you need to use such an old compiler, do -not update to the latest sparsehash release. - -We've also changed the internal tools we use to integrate -Googler-supplied patches to sparsehash into the opensource release. -These new tools should result in more frequent updates with better -change descriptions. They will also result in future ChangeLog -entries being much more verbose (for better or for worse). - -A full list of changes is described in -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.11/ChangeLog ChangeLog]. - -=== 21 January 2011 === - -I've just released sparsehash 1.10. This fixes a performance -regression in sparsehash 1.8, where sparse_hash_map would copy -hashtable keys by value even when the key was explicitly a reference. -It also fixes compiler warnings from MSVC 10, which uses some c++0x -features that did not interact well with sparsehash. - -There is no reason to upgrade unless you use references for your -hashtable keys, or compile with MSVC 10. A full list of changes is -described in -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.10/ChangeLog ChangeLog]. - - -=== 24 September 2010 === - -I've just released sparsehash 1.9. This fixes a size regression in -sparsehash 1.8, where the new allocator would take up space in -`sparse_hash_map`, doubling the sparse_hash_map overhead (from 1-2 -bits per bucket to 3 or so). All users are encouraged to upgrade. - -This change also marks enums as being Plain Old Data, which can speed -up hashtables with enum keys and/or values. A full list of changes is -described in -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.9/ChangeLog ChangeLog]. - -=== 29 July 2010 === - -I've just released sparsehash 1.8. This includes improved support for -`Allocator`, including supporting the allocator constructor arg and -`get_allocator()` access method. - -To work around a bug in gcc 4.0.x, I've renamed the static variables -`HT_OCCUPANCY_FLT` and `HT_SHRINK_FLT` to `HT_OCCUPANCY_PCT` and -`HT_SHRINK_PCT`, and changed their type from float to int. This -should not be a user-visible change, since these variables are only -used in the internal hashtable classes (sparsehash clients should use -`max_load_factor()` and `min_load_factor()` instead of modifying these -static variables), but if you do access these constants, you will need -to change your code. - -Internally, the biggest change is a revamp of the test suite. It now -has more complete coverage, and a more capable timing tester. There -are other, more minor changes as well. A full list of changes is -described in the -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.8/ChangeLog ChangeLog]. - -=== 31 March 2010 === - -I've just released sparsehash 1.7. The major news here is the -addition of `Allocator` support. Previously, these hashtable classes -would just ignore the `Allocator` template parameter. They now -respect it, and even inherit `size_type`, `pointer`, etc. from the -allocator class. By default, they use a special allocator we provide -that uses libc `malloc` and `free` to allocate. The hash classes -notice when this special allocator is being used, and use `realloc` -when it can. This means that the default allocator is significantly -faster than custom allocators are likely to be (since realloc-like -functionality is not supported by STL allocators). - -There are a few more minor changes as well. A full list of changes is -described in the -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.7/ChangeLog ChangeLog]. - -=== 11 January 2010 === - -I've just released sparsehash 1.6. The API has widened a bit with the -addition of `deleted_key()` and `empty_key()`, which let you query -what values these keys have. A few rather obscure bugs have been -fixed (such as an error when copying one hashtable into another when -the empty_keys differ). A full list of changes is described in the -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.6/ChangeLog ChangeLog]. - -=== 9 May 2009 === - -I've just released sparsehash 1.5.1. Hot on the heels of sparsehash -1.5, this release fixes a longstanding bug in the sparsehash code, -where `equal_range` would always return an empty range. It now works -as documented. All sparsehash users are encouraged to upgrade. - -=== 7 May 2009 === - -I've just released sparsehash 1.5. This release introduces tr1 -compatibility: I've added `rehash`, `begin(i)`, and other methods that -are expected to be part of the `unordered_map` API once `tr1` in -introduced. This allows `sparse_hash_map`, `dense_hash_map`, -`sparse_hash_set`, and `dense_hash_set` to be (almost) drop-in -replacements for `unordered_map` and `unordered_set`. - -There is no need to upgrade unless you need this functionality, or -need one of the other, more minor, changes described in the -[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.5/ChangeLog ChangeLog]. - +A backwards incompatibility arose from flattening the include headers +structure for the <google> folder. + +This is now fixed in 2.0.2. You only need to upgrade if you had previously +included files from the <google/sparsehash> folder. + +== 1 February 2012 == + +A minor bug related to the namespace switch from google to sparsehash +stopped the build from working when perftools is also installed. + +This is now fixed in 2.0.1. You only need to upgrade if you have perftools +installed. + +== 31 January 2012 == + +I've just released sparsehash 2.0. + +The `google-sparsehash` project has been renamed to `sparsehash`. I +(csilvers) am stepping down as maintainer, to be replaced by the team +of Donovan Hide and Geoff Pike. Welcome to the team, Donovan and +Geoff! Donovan has been an active contributor to sparsehash bug +reports and discussions in the past, and Geoff has been closely +involved with sparsehash inside Google (in addition to writing the +[http://code.google.com/p/cityhash CityHash hash function]). The two +of them together should be a formidable force. For good. + +I bumped the major version number up to 2 to reflect the new community +ownership of the project. All the +[http://sparsehash.googlecode.com/svn/tags/sparsehash-2.0/ChangeLog changes] +are related to the renaming. + +The only functional change from sparsehash 1.12 is that I've renamed +the `google/` include-directory to be `sparsehash/` instead. New code +should `#include <sparsehash/sparse_hash_map>`/etc. I've kept the old +names around as forwarding headers to the new, so `#include +<google/sparse_hash_map>` will continue to work. + +Note that the classes and functions remain in the `google` C++ +namespace (I didn't change that to `sparsehash` as well); I think +that's a trickier transition, and can happen in a future release. + + +=== 18 January 2011 === + +The `google-sparsehash` Google Code page has been renamed to +`sparsehash`, in preparation for the project being renamed to +`sparsehash`. In the coming weeks, I'll be stepping down as +maintainer for the sparsehash project, and as part of that Google is +relinquishing ownership of the project; it will now be entirely +community run. The name change reflects that shift. + + +=== 20 December 2011 === + +I've just released sparsehash 1.12. This release features improved +I/O (serialization) support. Support is finally added to serialize +and unserialize `dense_hash_map`/`set`, paralleling the existing code +for `sparse_hash_map`/`set`. In addition, the serialization API has +gotten simpler, with a single `serialize()` method to write to disk, +and an `unserialize()` method to read from disk. Finally, support has +gotten more generic, with built-in support for both C `FILE*`s and C++ +streams, and an extension mechanism to support arbitrary sources and +sinks. + +There are also more minor changes, including minor bugfixes, an +improved deleted-key test, and a minor addition to the `sparsetable` +API. See the [http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.12/ChangeLog ChangeLog] +for full details. + +=== 23 June 2011 === + +I've just released sparsehash 1.11. The major user-visible change is +that the default behavior is improved -- using the hash_map/set is +faster -- for hashtables where the key is a pointer. We now notice +that case and ignore the low 2-3 bits (which are almost always 0 for +pointers) when hashing. + +Another user-visible change is we've removed the tests for whether the +STL (vector, pair, etc) is defined in the 'std' namespace. gcc 2.95 +is the most recent compiler I know of to put STL types and functions +in the global namespace. If you need to use such an old compiler, do +not update to the latest sparsehash release. + +We've also changed the internal tools we use to integrate +Googler-supplied patches to sparsehash into the opensource release. +These new tools should result in more frequent updates with better +change descriptions. They will also result in future ChangeLog +entries being much more verbose (for better or for worse). + +A full list of changes is described in +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.11/ChangeLog ChangeLog]. + +=== 21 January 2011 === + +I've just released sparsehash 1.10. This fixes a performance +regression in sparsehash 1.8, where sparse_hash_map would copy +hashtable keys by value even when the key was explicitly a reference. +It also fixes compiler warnings from MSVC 10, which uses some c++0x +features that did not interact well with sparsehash. + +There is no reason to upgrade unless you use references for your +hashtable keys, or compile with MSVC 10. A full list of changes is +described in +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.10/ChangeLog ChangeLog]. + + +=== 24 September 2010 === + +I've just released sparsehash 1.9. This fixes a size regression in +sparsehash 1.8, where the new allocator would take up space in +`sparse_hash_map`, doubling the sparse_hash_map overhead (from 1-2 +bits per bucket to 3 or so). All users are encouraged to upgrade. + +This change also marks enums as being Plain Old Data, which can speed +up hashtables with enum keys and/or values. A full list of changes is +described in +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.9/ChangeLog ChangeLog]. + +=== 29 July 2010 === + +I've just released sparsehash 1.8. This includes improved support for +`Allocator`, including supporting the allocator constructor arg and +`get_allocator()` access method. + +To work around a bug in gcc 4.0.x, I've renamed the static variables +`HT_OCCUPANCY_FLT` and `HT_SHRINK_FLT` to `HT_OCCUPANCY_PCT` and +`HT_SHRINK_PCT`, and changed their type from float to int. This +should not be a user-visible change, since these variables are only +used in the internal hashtable classes (sparsehash clients should use +`max_load_factor()` and `min_load_factor()` instead of modifying these +static variables), but if you do access these constants, you will need +to change your code. + +Internally, the biggest change is a revamp of the test suite. It now +has more complete coverage, and a more capable timing tester. There +are other, more minor changes as well. A full list of changes is +described in the +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.8/ChangeLog ChangeLog]. + +=== 31 March 2010 === + +I've just released sparsehash 1.7. The major news here is the +addition of `Allocator` support. Previously, these hashtable classes +would just ignore the `Allocator` template parameter. They now +respect it, and even inherit `size_type`, `pointer`, etc. from the +allocator class. By default, they use a special allocator we provide +that uses libc `malloc` and `free` to allocate. The hash classes +notice when this special allocator is being used, and use `realloc` +when it can. This means that the default allocator is significantly +faster than custom allocators are likely to be (since realloc-like +functionality is not supported by STL allocators). + +There are a few more minor changes as well. A full list of changes is +described in the +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.7/ChangeLog ChangeLog]. + +=== 11 January 2010 === + +I've just released sparsehash 1.6. The API has widened a bit with the +addition of `deleted_key()` and `empty_key()`, which let you query +what values these keys have. A few rather obscure bugs have been +fixed (such as an error when copying one hashtable into another when +the empty_keys differ). A full list of changes is described in the +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.6/ChangeLog ChangeLog]. + +=== 9 May 2009 === + +I've just released sparsehash 1.5.1. Hot on the heels of sparsehash +1.5, this release fixes a longstanding bug in the sparsehash code, +where `equal_range` would always return an empty range. It now works +as documented. All sparsehash users are encouraged to upgrade. + +=== 7 May 2009 === + +I've just released sparsehash 1.5. This release introduces tr1 +compatibility: I've added `rehash`, `begin(i)`, and other methods that +are expected to be part of the `unordered_map` API once `tr1` in +introduced. This allows `sparse_hash_map`, `dense_hash_map`, +`sparse_hash_set`, and `dense_hash_set` to be (almost) drop-in +replacements for `unordered_map` and `unordered_set`. + +There is no need to upgrade unless you need this functionality, or +need one of the other, more minor, changes described in the +[http://google-sparsehash.googlecode.com/svn/tags/sparsehash-1.5/ChangeLog ChangeLog]. + diff --git a/contrib/libs/sparsehash/README b/contrib/libs/sparsehash/README index 7c7dc26f77..26bb485008 100644 --- a/contrib/libs/sparsehash/README +++ b/contrib/libs/sparsehash/README @@ -1,147 +1,147 @@ -This directory contains several hash-map implementations, similar in -API to SGI's hash_map class, but with different performance -characteristics. sparse_hash_map uses very little space overhead, 1-2 -bits per entry. dense_hash_map is very fast, particulary on lookup. -(sparse_hash_set and dense_hash_set are the set versions of these -routines.) On the other hand, these classes have requirements that -may not make them appropriate for all applications. - -All these implementation use a hashtable with internal quadratic -probing. This method is space-efficient -- there is no pointer -overhead -- and time-efficient for good hash functions. - -COMPILING ---------- -To compile test applications with these classes, run ./configure -followed by make. To install these header files on your system, run -'make install'. (On Windows, the instructions are different; see -README_windows.txt.) See INSTALL for more details. - -This code should work on any modern C++ system. It has been tested on -Linux (Ubuntu, Fedora, RedHat, Debian), Solaris 10 x86, FreeBSD 6.0, -OS X 10.3 and 10.4, and Windows under both VC++7 and VC++8. - -USING ------ -See the html files in the doc directory for small example programs -that use these classes. It's enough to just include the header file: - - #include <sparsehash/sparse_hash_map> // or sparse_hash_set, dense_hash_map, ... - google::sparse_hash_set<int, int> number_mapper; - -and use the class the way you would other hash-map implementations. -(Though see "API" below for caveats.) - -By default (you can change it via a flag to ./configure), these hash -implementations are defined in the google namespace. - -API ---- -The API for sparse_hash_map, dense_hash_map, sparse_hash_set, and -dense_hash_set, are a superset of the API of SGI's hash_map class. -See doc/sparse_hash_map.html, et al., for more information about the -API. - -The usage of these classes differ from SGI's hash_map, and other -hashtable implementations, in the following major ways: - -1) dense_hash_map requires you to set aside one key value as the - 'empty bucket' value, set via the set_empty_key() method. This - *MUST* be called before you can use the dense_hash_map. It is - illegal to insert any elements into a dense_hash_map whose key is - equal to the empty-key. - -2) For both dense_hash_map and sparse_hash_map, if you wish to delete - elements from the hashtable, you must set aside a key value as the - 'deleted bucket' value, set via the set_deleted_key() method. If - your hash-map is insert-only, there is no need to call this - method. If you call set_deleted_key(), it is illegal to insert any - elements into a dense_hash_map or sparse_hash_map whose key is - equal to the deleted-key. - -3) These hash-map implementation support I/O. See below. - -There are also some smaller differences: - -1) The constructor takes an optional argument that specifies the - number of elements you expect to insert into the hashtable. This - differs from SGI's hash_map implementation, which takes an optional - number of buckets. - -2) erase() does not immediately reclaim memory. As a consequence, - erase() does not invalidate any iterators, making loops like this - correct: - for (it = ht.begin(); it != ht.end(); ++it) - if (...) ht.erase(it); - As another consequence, a series of erase() calls can leave your - hashtable using more memory than it needs to. The hashtable will - automatically compact at the next call to insert(), but to - manually compact a hashtable, you can call - ht.resize(0) - -I/O ---- -In addition to the normal hash-map operations, sparse_hash_map can -read and write hashtables to disk. (dense_hash_map also has the API, -but it has not yet been implemented, and writes will always fail.) - -In the simplest case, writing a hashtable is as easy as calling two -methods on the hashtable: - ht.write_metadata(fp); - ht.write_nopointer_data(fp); - -Reading in this data is equally simple: - google::sparse_hash_map<...> ht; - ht.read_metadata(fp); - ht.read_nopointer_data(fp); - -The above is sufficient if the key and value do not contain any -pointers: they are basic C types or agglomorations of basic C types. -If the key and/or value do contain pointers, you can still store the -hashtable by replacing write_nopointer_data() with a custom writing -routine. See sparse_hash_map.html et al. for more information. - -SPARSETABLE ------------ -In addition to the hash-map and hash-set classes, this package also -provides sparsetable.h, an array implementation that uses space -proportional to the number of elements in the array, rather than the +This directory contains several hash-map implementations, similar in +API to SGI's hash_map class, but with different performance +characteristics. sparse_hash_map uses very little space overhead, 1-2 +bits per entry. dense_hash_map is very fast, particulary on lookup. +(sparse_hash_set and dense_hash_set are the set versions of these +routines.) On the other hand, these classes have requirements that +may not make them appropriate for all applications. + +All these implementation use a hashtable with internal quadratic +probing. This method is space-efficient -- there is no pointer +overhead -- and time-efficient for good hash functions. + +COMPILING +--------- +To compile test applications with these classes, run ./configure +followed by make. To install these header files on your system, run +'make install'. (On Windows, the instructions are different; see +README_windows.txt.) See INSTALL for more details. + +This code should work on any modern C++ system. It has been tested on +Linux (Ubuntu, Fedora, RedHat, Debian), Solaris 10 x86, FreeBSD 6.0, +OS X 10.3 and 10.4, and Windows under both VC++7 and VC++8. + +USING +----- +See the html files in the doc directory for small example programs +that use these classes. It's enough to just include the header file: + + #include <sparsehash/sparse_hash_map> // or sparse_hash_set, dense_hash_map, ... + google::sparse_hash_set<int, int> number_mapper; + +and use the class the way you would other hash-map implementations. +(Though see "API" below for caveats.) + +By default (you can change it via a flag to ./configure), these hash +implementations are defined in the google namespace. + +API +--- +The API for sparse_hash_map, dense_hash_map, sparse_hash_set, and +dense_hash_set, are a superset of the API of SGI's hash_map class. +See doc/sparse_hash_map.html, et al., for more information about the +API. + +The usage of these classes differ from SGI's hash_map, and other +hashtable implementations, in the following major ways: + +1) dense_hash_map requires you to set aside one key value as the + 'empty bucket' value, set via the set_empty_key() method. This + *MUST* be called before you can use the dense_hash_map. It is + illegal to insert any elements into a dense_hash_map whose key is + equal to the empty-key. + +2) For both dense_hash_map and sparse_hash_map, if you wish to delete + elements from the hashtable, you must set aside a key value as the + 'deleted bucket' value, set via the set_deleted_key() method. If + your hash-map is insert-only, there is no need to call this + method. If you call set_deleted_key(), it is illegal to insert any + elements into a dense_hash_map or sparse_hash_map whose key is + equal to the deleted-key. + +3) These hash-map implementation support I/O. See below. + +There are also some smaller differences: + +1) The constructor takes an optional argument that specifies the + number of elements you expect to insert into the hashtable. This + differs from SGI's hash_map implementation, which takes an optional + number of buckets. + +2) erase() does not immediately reclaim memory. As a consequence, + erase() does not invalidate any iterators, making loops like this + correct: + for (it = ht.begin(); it != ht.end(); ++it) + if (...) ht.erase(it); + As another consequence, a series of erase() calls can leave your + hashtable using more memory than it needs to. The hashtable will + automatically compact at the next call to insert(), but to + manually compact a hashtable, you can call + ht.resize(0) + +I/O +--- +In addition to the normal hash-map operations, sparse_hash_map can +read and write hashtables to disk. (dense_hash_map also has the API, +but it has not yet been implemented, and writes will always fail.) + +In the simplest case, writing a hashtable is as easy as calling two +methods on the hashtable: + ht.write_metadata(fp); + ht.write_nopointer_data(fp); + +Reading in this data is equally simple: + google::sparse_hash_map<...> ht; + ht.read_metadata(fp); + ht.read_nopointer_data(fp); + +The above is sufficient if the key and value do not contain any +pointers: they are basic C types or agglomorations of basic C types. +If the key and/or value do contain pointers, you can still store the +hashtable by replacing write_nopointer_data() with a custom writing +routine. See sparse_hash_map.html et al. for more information. + +SPARSETABLE +----------- +In addition to the hash-map and hash-set classes, this package also +provides sparsetable.h, an array implementation that uses space +proportional to the number of elements in the array, rather than the maximum element index. It uses very little space overhead: 2 to 5 bits per entry. See doc/sparsetable.html for the API. - -RESOURCE USAGE --------------- + +RESOURCE USAGE +-------------- * sparse_hash_map has memory overhead of about 4 to 10 bits per hash-map entry, assuming a typical average occupancy of 50%. -* dense_hash_map has a factor of 2-3 memory overhead: if your - hashtable data takes X bytes, dense_hash_map will use 3X-4X memory - total. - -Hashtables tend to double in size when resizing, creating an -additional 50% space overhead. dense_hash_map does in fact have a +* dense_hash_map has a factor of 2-3 memory overhead: if your + hashtable data takes X bytes, dense_hash_map will use 3X-4X memory + total. + +Hashtables tend to double in size when resizing, creating an +additional 50% space overhead. dense_hash_map does in fact have a significant "high water mark" memory use requirement, which is 6 times the size of hash entries in the table when resizing (when reaching 50% occupancy, the table resizes to double the previous size, and the old table (2x) is copied to the new table (4x)). -sparse_hash_map, however, is written to need very little space -overhead when resizing: only a few bits per hashtable entry. - -PERFORMANCE ------------ -You can compile and run the included file time_hash_map.cc to examine -the performance of sparse_hash_map, dense_hash_map, and your native -hash_map implementation on your system. One test against the -SGI hash_map implementation gave the following timing information for -a simple find() call: - SGI hash_map: 22 ns - dense_hash_map: 13 ns - sparse_hash_map: 117 ns - SGI map: 113 ns - -See doc/performance.html for more detailed charts on resource usage -and performance data. - ---- -16 March 2005 -(Last updated: 12 September 2010) +sparse_hash_map, however, is written to need very little space +overhead when resizing: only a few bits per hashtable entry. + +PERFORMANCE +----------- +You can compile and run the included file time_hash_map.cc to examine +the performance of sparse_hash_map, dense_hash_map, and your native +hash_map implementation on your system. One test against the +SGI hash_map implementation gave the following timing information for +a simple find() call: + SGI hash_map: 22 ns + dense_hash_map: 13 ns + sparse_hash_map: 117 ns + SGI map: 113 ns + +See doc/performance.html for more detailed charts on resource usage +and performance data. + +--- +16 March 2005 +(Last updated: 12 September 2010) diff --git a/contrib/libs/sparsehash/README_windows.txt b/contrib/libs/sparsehash/README_windows.txt index 99234e02fa..54df6f8eac 100644 --- a/contrib/libs/sparsehash/README_windows.txt +++ b/contrib/libs/sparsehash/README_windows.txt @@ -1,25 +1,25 @@ -This project has been ported to Windows. A working solution file -exists in this directory: - sparsehash.sln - -You can load this solution file into either VC++ 7.1 (Visual Studio -2003) or VC++ 8.0 (Visual Studio 2005) -- in the latter case, it will -automatically convert the files to the latest format for you. - -When you build the solution, it will create a number of -unittests,which you can run by hand (or, more easily, under the Visual -Studio debugger) to make sure everything is working properly on your -system. The binaries will end up in a directory called "debug" or -"release" in the top-level directory (next to the .sln file). - -Note that these systems are set to build in Debug mode by default. -You may want to change them to Release mode. - -I have little experience with Windows programming, so there may be -better ways to set this up than I've done! If you run across any -problems, please post to the google-sparsehash Google Group, or report -them on the sparsehash Google Code site: - http://groups.google.com/group/google-sparsehash - http://code.google.com/p/sparsehash/issues/list - --- craig +This project has been ported to Windows. A working solution file +exists in this directory: + sparsehash.sln + +You can load this solution file into either VC++ 7.1 (Visual Studio +2003) or VC++ 8.0 (Visual Studio 2005) -- in the latter case, it will +automatically convert the files to the latest format for you. + +When you build the solution, it will create a number of +unittests,which you can run by hand (or, more easily, under the Visual +Studio debugger) to make sure everything is working properly on your +system. The binaries will end up in a directory called "debug" or +"release" in the top-level directory (next to the .sln file). + +Note that these systems are set to build in Debug mode by default. +You may want to change them to Release mode. + +I have little experience with Windows programming, so there may be +better ways to set this up than I've done! If you run across any +problems, please post to the google-sparsehash Google Group, or report +them on the sparsehash Google Code site: + http://groups.google.com/group/google-sparsehash + http://code.google.com/p/sparsehash/issues/list + +-- craig diff --git a/contrib/libs/sparsehash/TODO b/contrib/libs/sparsehash/TODO index 50476bd06f..e9b0263cf5 100644 --- a/contrib/libs/sparsehash/TODO +++ b/contrib/libs/sparsehash/TODO @@ -1,28 +1,28 @@ -1) TODO: I/O implementation in densehashtable.h - -2) TODO: document SPARSEHASH_STAT_UPDATE macro, and also macros that - tweak performance. Perhaps add support to these to the API? - -3) TODO: support exceptions? - -4) BUG: sparsetable's operator[] doesn't work well with printf: you - need to explicitly cast the result to value_type to print it. (It - works fine with streams.) - -5) TODO: consider rewriting dense_hash_map to use a 'groups' scheme, - like sparsetable, but without the sparse-allocation within a - group. This makes resizing have better memory-use properties. The - downside is that probes across groups might take longer since - groups are not contiguous in memory. Making groups the same size - as a cache-line, and ensuring they're loaded on cache-line - boundaries, might help. Needs careful testing to make sure it - doesn't hurt performance. - -6) TODO: Get the C-only version of sparsehash in experimental/ ready - for prime-time. - -7) TODO: use cmake (www.cmake.org) to make it easy to isntall this on - a windows system. - ---- -28 February 2007 +1) TODO: I/O implementation in densehashtable.h + +2) TODO: document SPARSEHASH_STAT_UPDATE macro, and also macros that + tweak performance. Perhaps add support to these to the API? + +3) TODO: support exceptions? + +4) BUG: sparsetable's operator[] doesn't work well with printf: you + need to explicitly cast the result to value_type to print it. (It + works fine with streams.) + +5) TODO: consider rewriting dense_hash_map to use a 'groups' scheme, + like sparsetable, but without the sparse-allocation within a + group. This makes resizing have better memory-use properties. The + downside is that probes across groups might take longer since + groups are not contiguous in memory. Making groups the same size + as a cache-line, and ensuring they're loaded on cache-line + boundaries, might help. Needs careful testing to make sure it + doesn't hurt performance. + +6) TODO: Get the C-only version of sparsehash in experimental/ ready + for prime-time. + +7) TODO: use cmake (www.cmake.org) to make it easy to isntall this on + a windows system. + +--- +28 February 2007 diff --git a/contrib/libs/sparsehash/src/sparsehash/dense_hash_map b/contrib/libs/sparsehash/src/sparsehash/dense_hash_map index 693ada163f..05fd580e64 100644 --- a/contrib/libs/sparsehash/src/sparsehash/dense_hash_map +++ b/contrib/libs/sparsehash/src/sparsehash/dense_hash_map @@ -1,369 +1,369 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// ---- -// -// This is just a very thin wrapper over densehashtable.h, just -// like sgi stl's stl_hash_map is a very thin wrapper over -// stl_hashtable. The major thing we define is operator[], because -// we have a concept of a data_type which stl_hashtable doesn't -// (it only has a key and a value). -// -// NOTE: this is exactly like sparse_hash_map.h, with the word -// "sparse" replaced by "dense", except for the addition of -// set_empty_key(). -// -// YOU MUST CALL SET_EMPTY_KEY() IMMEDIATELY AFTER CONSTRUCTION. -// -// Otherwise your program will die in mysterious ways. (Note if you -// use the constructor that takes an InputIterator range, you pass in -// the empty key in the constructor, rather than after. As a result, -// this constructor differs from the standard STL version.) -// -// In other respects, we adhere mostly to the STL semantics for -// hash-map. One important exception is that insert() may invalidate -// iterators entirely -- STL semantics are that insert() may reorder -// iterators, but they all still refer to something valid in the -// hashtable. Not so for us. Likewise, insert() may invalidate -// pointers into the hashtable. (Whether insert invalidates iterators -// and pointers depends on whether it results in a hashtable resize). -// On the plus side, delete() doesn't invalidate iterators or pointers -// at all, or even change the ordering of elements. -// -// Here are a few "power user" tips: -// -// 1) set_deleted_key(): -// If you want to use erase() you *must* call set_deleted_key(), -// in addition to set_empty_key(), after construction. -// The deleted and empty keys must differ. -// -// 2) resize(0): -// When an item is deleted, its memory isn't freed right -// away. This allows you to iterate over a hashtable, -// and call erase(), without invalidating the iterator. -// To force the memory to be freed, call resize(0). -// For tr1 compatibility, this can also be called as rehash(0). -// -// 3) min_load_factor(0.0) -// Setting the minimum load factor to 0.0 guarantees that -// the hash table will never shrink. -// -// Roughly speaking: -// (1) dense_hash_map: fastest, uses the most memory unless entries are small -// (2) sparse_hash_map: slowest, uses the least memory -// (3) hash_map / unordered_map (STL): in the middle -// -// Typically I use sparse_hash_map when I care about space and/or when -// I need to save the hashtable on disk. I use hash_map otherwise. I -// don't personally use dense_hash_set ever; some people use it for -// small sets with lots of lookups. -// -// - dense_hash_map has, typically, about 78% memory overhead (if your -// data takes up X bytes, the hash_map uses .78X more bytes in overhead). -// - sparse_hash_map has about 4 bits overhead per entry. -// - sparse_hash_map can be 3-7 times slower than the others for lookup and, -// especially, inserts. See time_hash_map.cc for details. -// -// See /usr/(local/)?doc/sparsehash-*/dense_hash_map.html -// for information about how to use this class. - -#ifndef _DENSE_HASH_MAP_H_ -#define _DENSE_HASH_MAP_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <algorithm> // needed by stl_alloc -#include <functional> // for equal_to<>, select1st<>, etc -#include <memory> // for alloc -#include <utility> // for pair<> -#include <sparsehash/internal/densehashtable.h> // IWYU pragma: export -#include <sparsehash/internal/libc_allocator_with_realloc.h> -#include HASH_FUN_H // for hash<> -_START_GOOGLE_NAMESPACE_ - -template <class Key, class T, - class HashFcn = SPARSEHASH_HASH<Key>, // defined in sparseconfig.h - class EqualKey = std::equal_to<Key>, - class Alloc = libc_allocator_with_realloc<std::pair<const Key, T> > > -class dense_hash_map { - private: - // Apparently select1st is not stl-standard, so we define our own - struct SelectKey { - typedef const Key& result_type; - const Key& operator()(const std::pair<const Key, T>& p) const { - return p.first; - } - }; - struct SetKey { - void operator()(std::pair<const Key, T>* value, const Key& new_key) const { - *const_cast<Key*>(&value->first) = new_key; - // It would be nice to clear the rest of value here as well, in - // case it's taking up a lot of memory. We do this by clearing - // the value. This assumes T has a zero-arg constructor! - value->second = T(); - } - }; - // For operator[]. - struct DefaultValue { - std::pair<const Key, T> operator()(const Key& key) { - return std::make_pair(key, T()); - } - }; - - // The actual data - typedef dense_hashtable<std::pair<const Key, T>, Key, HashFcn, SelectKey, - SetKey, EqualKey, Alloc> ht; - ht rep; - - public: - typedef typename ht::key_type key_type; - typedef T data_type; - typedef T mapped_type; - typedef typename ht::value_type value_type; - typedef typename ht::hasher hasher; - typedef typename ht::key_equal key_equal; - typedef Alloc allocator_type; - - typedef typename ht::size_type size_type; - typedef typename ht::difference_type difference_type; - typedef typename ht::pointer pointer; - typedef typename ht::const_pointer const_pointer; - typedef typename ht::reference reference; - typedef typename ht::const_reference const_reference; - - typedef typename ht::iterator iterator; - typedef typename ht::const_iterator const_iterator; - typedef typename ht::local_iterator local_iterator; - typedef typename ht::const_local_iterator const_local_iterator; - - // Iterator functions - iterator begin() { return rep.begin(); } - iterator end() { return rep.end(); } - const_iterator begin() const { return rep.begin(); } - const_iterator end() const { return rep.end(); } - - - // These come from tr1's unordered_map. For us, a bucket has 0 or 1 elements. - local_iterator begin(size_type i) { return rep.begin(i); } - local_iterator end(size_type i) { return rep.end(i); } - const_local_iterator begin(size_type i) const { return rep.begin(i); } - const_local_iterator end(size_type i) const { return rep.end(i); } - - // Accessor functions - allocator_type get_allocator() const { return rep.get_allocator(); } - hasher hash_funct() const { return rep.hash_funct(); } - hasher hash_function() const { return hash_funct(); } - key_equal key_eq() const { return rep.key_eq(); } - - - // Constructors - explicit dense_hash_map(size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { - } - - template <class InputIterator> - dense_hash_map(InputIterator f, InputIterator l, - const key_type& empty_key_val, - size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { - set_empty_key(empty_key_val); - rep.insert(f, l); - } - // We use the default copy constructor - // We use the default operator=() - // We use the default destructor - - void clear() { rep.clear(); } - // This clears the hash map without resizing it down to the minimum - // bucket count, but rather keeps the number of buckets constant - void clear_no_resize() { rep.clear_no_resize(); } - void swap(dense_hash_map& hs) { rep.swap(hs.rep); } - - - // Functions concerning size - size_type size() const { return rep.size(); } - size_type max_size() const { return rep.max_size(); } - bool empty() const { return rep.empty(); } - size_type bucket_count() const { return rep.bucket_count(); } - size_type max_bucket_count() const { return rep.max_bucket_count(); } - - // These are tr1 methods. bucket() is the bucket the key is or would be in. - size_type bucket_size(size_type i) const { return rep.bucket_size(i); } - size_type bucket(const key_type& key) const { return rep.bucket(key); } - float load_factor() const { - return size() * 1.0f / bucket_count(); - } - float max_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return grow; - } - void max_load_factor(float new_grow) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(shrink, new_grow); - } - // These aren't tr1 methods but perhaps ought to be. - float min_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return shrink; - } - void min_load_factor(float new_shrink) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(new_shrink, grow); - } - // Deprecated; use min_load_factor() or max_load_factor() instead. - void set_resizing_parameters(float shrink, float grow) { - rep.set_resizing_parameters(shrink, grow); - } - - void resize(size_type hint) { rep.resize(hint); } - void rehash(size_type hint) { resize(hint); } // the tr1 name - - // Lookup routines - iterator find(const key_type& key) { return rep.find(key); } - const_iterator find(const key_type& key) const { return rep.find(key); } - - data_type& operator[](const key_type& key) { // This is our value-add! - // If key is in the hashtable, returns find(key)->second, - // otherwise returns insert(value_type(key, T()).first->second. - // Note it does not create an empty T unless the find fails. - return rep.template find_or_insert<DefaultValue>(key).second; - } - - size_type count(const key_type& key) const { return rep.count(key); } - - std::pair<iterator, iterator> equal_range(const key_type& key) { - return rep.equal_range(key); - } - std::pair<const_iterator, const_iterator> equal_range(const key_type& key) - const { - return rep.equal_range(key); - } - - - // Insertion routines - std::pair<iterator, bool> insert(const value_type& obj) { - return rep.insert(obj); - } - template <class InputIterator> void insert(InputIterator f, InputIterator l) { - rep.insert(f, l); - } - void insert(const_iterator f, const_iterator l) { - rep.insert(f, l); - } - // Required for std::insert_iterator; the passed-in iterator is ignored. - iterator insert(iterator, const value_type& obj) { - return insert(obj).first; - } - - // Deletion and empty routines - // THESE ARE NON-STANDARD! I make you specify an "impossible" key - // value to identify deleted and empty buckets. You can change the - // deleted key as time goes on, or get rid of it entirely to be insert-only. - void set_empty_key(const key_type& key) { // YOU MUST CALL THIS! - rep.set_empty_key(value_type(key, data_type())); // rep wants a value - } - key_type empty_key() const { - return rep.empty_key().first; // rep returns a value - } - - void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } - void clear_deleted_key() { rep.clear_deleted_key(); } - key_type deleted_key() const { return rep.deleted_key(); } - - // These are standard - size_type erase(const key_type& key) { return rep.erase(key); } - void erase(iterator it) { rep.erase(it); } - void erase(iterator f, iterator l) { rep.erase(f, l); } - - - // Comparison - bool operator==(const dense_hash_map& hs) const { return rep == hs.rep; } - bool operator!=(const dense_hash_map& hs) const { return rep != hs.rep; } - - - // I/O -- this is an add-on for writing hash map to disk - // - // For maximum flexibility, this does not assume a particular - // file type (though it will probably be a FILE *). We just pass - // the fp through to rep. - - // If your keys and values are simple enough, you can pass this - // serializer to serialize()/unserialize(). "Simple enough" means - // value_type is a POD type that contains no pointers. Note, - // however, we don't try to normalize endianness. - typedef typename ht::NopointerSerializer NopointerSerializer; - - // serializer: a class providing operator()(OUTPUT*, const value_type&) - // (writing value_type to OUTPUT). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a - // pointer to a class providing size_t Write(const void*, size_t), - // which writes a buffer into a stream (which fp presumably - // owns) and returns the number of bytes successfully written. - // Note basic_ostream<not_char> is not currently supported. - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT* fp) { - return rep.serialize(serializer, fp); - } - - // serializer: a functor providing operator()(INPUT*, value_type*) - // (reading from INPUT and into value_type). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a - // pointer to a class providing size_t Read(void*, size_t), - // which reads into a buffer from a stream (which fp presumably - // owns) and returns the number of bytes successfully read. - // Note basic_istream<not_char> is not currently supported. - // NOTE: Since value_type is std::pair<const Key, T>, ValueSerializer - // may need to do a const cast in order to fill in the key. - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT* fp) { - return rep.unserialize(serializer, fp); - } -}; - -// We need a global swap as well -template <class Key, class T, class HashFcn, class EqualKey, class Alloc> -inline void swap(dense_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm1, - dense_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm2) { - hm1.swap(hm2); -} - -_END_GOOGLE_NAMESPACE_ - -#endif /* _DENSE_HASH_MAP_H_ */ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// ---- +// +// This is just a very thin wrapper over densehashtable.h, just +// like sgi stl's stl_hash_map is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// NOTE: this is exactly like sparse_hash_map.h, with the word +// "sparse" replaced by "dense", except for the addition of +// set_empty_key(). +// +// YOU MUST CALL SET_EMPTY_KEY() IMMEDIATELY AFTER CONSTRUCTION. +// +// Otherwise your program will die in mysterious ways. (Note if you +// use the constructor that takes an InputIterator range, you pass in +// the empty key in the constructor, rather than after. As a result, +// this constructor differs from the standard STL version.) +// +// In other respects, we adhere mostly to the STL semantics for +// hash-map. One important exception is that insert() may invalidate +// iterators entirely -- STL semantics are that insert() may reorder +// iterators, but they all still refer to something valid in the +// hashtable. Not so for us. Likewise, insert() may invalidate +// pointers into the hashtable. (Whether insert invalidates iterators +// and pointers depends on whether it results in a hashtable resize). +// On the plus side, delete() doesn't invalidate iterators or pointers +// at all, or even change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// If you want to use erase() you *must* call set_deleted_key(), +// in addition to set_empty_key(), after construction. +// The deleted and empty keys must differ. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This allows you to iterate over a hashtable, +// and call erase(), without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_map: fastest, uses the most memory unless entries are small +// (2) sparse_hash_map: slowest, uses the least memory +// (3) hash_map / unordered_map (STL): in the middle +// +// Typically I use sparse_hash_map when I care about space and/or when +// I need to save the hashtable on disk. I use hash_map otherwise. I +// don't personally use dense_hash_set ever; some people use it for +// small sets with lots of lookups. +// +// - dense_hash_map has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_map uses .78X more bytes in overhead). +// - sparse_hash_map has about 4 bits overhead per entry. +// - sparse_hash_map can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/dense_hash_map.html +// for information about how to use this class. + +#ifndef _DENSE_HASH_MAP_H_ +#define _DENSE_HASH_MAP_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <algorithm> // needed by stl_alloc +#include <functional> // for equal_to<>, select1st<>, etc +#include <memory> // for alloc +#include <utility> // for pair<> +#include <sparsehash/internal/densehashtable.h> // IWYU pragma: export +#include <sparsehash/internal/libc_allocator_with_realloc.h> +#include HASH_FUN_H // for hash<> +_START_GOOGLE_NAMESPACE_ + +template <class Key, class T, + class HashFcn = SPARSEHASH_HASH<Key>, // defined in sparseconfig.h + class EqualKey = std::equal_to<Key>, + class Alloc = libc_allocator_with_realloc<std::pair<const Key, T> > > +class dense_hash_map { + private: + // Apparently select1st is not stl-standard, so we define our own + struct SelectKey { + typedef const Key& result_type; + const Key& operator()(const std::pair<const Key, T>& p) const { + return p.first; + } + }; + struct SetKey { + void operator()(std::pair<const Key, T>* value, const Key& new_key) const { + *const_cast<Key*>(&value->first) = new_key; + // It would be nice to clear the rest of value here as well, in + // case it's taking up a lot of memory. We do this by clearing + // the value. This assumes T has a zero-arg constructor! + value->second = T(); + } + }; + // For operator[]. + struct DefaultValue { + std::pair<const Key, T> operator()(const Key& key) { + return std::make_pair(key, T()); + } + }; + + // The actual data + typedef dense_hashtable<std::pair<const Key, T>, Key, HashFcn, SelectKey, + SetKey, EqualKey, Alloc> ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef T data_type; + typedef T mapped_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + // Iterator functions + iterator begin() { return rep.begin(); } + iterator end() { return rep.end(); } + const_iterator begin() const { return rep.begin(); } + const_iterator end() const { return rep.end(); } + + + // These come from tr1's unordered_map. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) { return rep.begin(i); } + local_iterator end(size_type i) { return rep.end(i); } + const_local_iterator begin(size_type i) const { return rep.begin(i); } + const_local_iterator end(size_type i) const { return rep.end(i); } + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit dense_hash_map(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + } + + template <class InputIterator> + dense_hash_map(InputIterator f, InputIterator l, + const key_type& empty_key_val, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + set_empty_key(empty_key_val); + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + // This clears the hash map without resizing it down to the minimum + // bucket count, but rather keeps the number of buckets constant + void clear_no_resize() { rep.clear_no_resize(); } + void swap(dense_hash_map& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) { return rep.find(key); } + const_iterator find(const key_type& key) const { return rep.find(key); } + + data_type& operator[](const key_type& key) { // This is our value-add! + // If key is in the hashtable, returns find(key)->second, + // otherwise returns insert(value_type(key, T()).first->second. + // Note it does not create an empty T unless the find fails. + return rep.template find_or_insert<DefaultValue>(key).second; + } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair<iterator, iterator> equal_range(const key_type& key) { + return rep.equal_range(key); + } + std::pair<const_iterator, const_iterator> equal_range(const key_type& key) + const { + return rep.equal_range(key); + } + + + // Insertion routines + std::pair<iterator, bool> insert(const value_type& obj) { + return rep.insert(obj); + } + template <class InputIterator> void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion and empty routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted and empty buckets. You can change the + // deleted key as time goes on, or get rid of it entirely to be insert-only. + void set_empty_key(const key_type& key) { // YOU MUST CALL THIS! + rep.set_empty_key(value_type(key, data_type())); // rep wants a value + } + key_type empty_key() const { + return rep.empty_key().first; // rep returns a value + } + + void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const dense_hash_map& hs) const { return rep == hs.rep; } + bool operator!=(const dense_hash_map& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing hash map to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream<not_char> is not currently supported. + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream<not_char> is not currently supported. + // NOTE: Since value_type is std::pair<const Key, T>, ValueSerializer + // may need to do a const cast in order to fill in the key. + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } +}; + +// We need a global swap as well +template <class Key, class T, class HashFcn, class EqualKey, class Alloc> +inline void swap(dense_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm1, + dense_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm2) { + hm1.swap(hm2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _DENSE_HASH_MAP_H_ */ diff --git a/contrib/libs/sparsehash/src/sparsehash/dense_hash_set b/contrib/libs/sparsehash/src/sparsehash/dense_hash_set index 5238cd1e86..050b15d1d5 100644 --- a/contrib/libs/sparsehash/src/sparsehash/dense_hash_set +++ b/contrib/libs/sparsehash/src/sparsehash/dense_hash_set @@ -1,338 +1,338 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// This is just a very thin wrapper over densehashtable.h, just -// like sgi stl's stl_hash_set is a very thin wrapper over -// stl_hashtable. The major thing we define is operator[], because -// we have a concept of a data_type which stl_hashtable doesn't -// (it only has a key and a value). -// -// This is more different from dense_hash_map than you might think, -// because all iterators for sets are const (you obviously can't -// change the key, and for sets there is no value). -// -// NOTE: this is exactly like sparse_hash_set.h, with the word -// "sparse" replaced by "dense", except for the addition of -// set_empty_key(). -// -// YOU MUST CALL SET_EMPTY_KEY() IMMEDIATELY AFTER CONSTRUCTION. -// -// Otherwise your program will die in mysterious ways. (Note if you -// use the constructor that takes an InputIterator range, you pass in -// the empty key in the constructor, rather than after. As a result, -// this constructor differs from the standard STL version.) -// -// In other respects, we adhere mostly to the STL semantics for -// hash-map. One important exception is that insert() may invalidate -// iterators entirely -- STL semantics are that insert() may reorder -// iterators, but they all still refer to something valid in the -// hashtable. Not so for us. Likewise, insert() may invalidate -// pointers into the hashtable. (Whether insert invalidates iterators -// and pointers depends on whether it results in a hashtable resize). -// On the plus side, delete() doesn't invalidate iterators or pointers -// at all, or even change the ordering of elements. -// -// Here are a few "power user" tips: -// -// 1) set_deleted_key(): -// If you want to use erase() you must call set_deleted_key(), -// in addition to set_empty_key(), after construction. -// The deleted and empty keys must differ. -// -// 2) resize(0): -// When an item is deleted, its memory isn't freed right -// away. This allows you to iterate over a hashtable, -// and call erase(), without invalidating the iterator. -// To force the memory to be freed, call resize(0). -// For tr1 compatibility, this can also be called as rehash(0). -// -// 3) min_load_factor(0.0) -// Setting the minimum load factor to 0.0 guarantees that -// the hash table will never shrink. -// -// Roughly speaking: -// (1) dense_hash_set: fastest, uses the most memory unless entries are small -// (2) sparse_hash_set: slowest, uses the least memory -// (3) hash_set / unordered_set (STL): in the middle -// -// Typically I use sparse_hash_set when I care about space and/or when -// I need to save the hashtable on disk. I use hash_set otherwise. I -// don't personally use dense_hash_set ever; some people use it for -// small sets with lots of lookups. -// -// - dense_hash_set has, typically, about 78% memory overhead (if your -// data takes up X bytes, the hash_set uses .78X more bytes in overhead). -// - sparse_hash_set has about 4 bits overhead per entry. -// - sparse_hash_set can be 3-7 times slower than the others for lookup and, -// especially, inserts. See time_hash_map.cc for details. -// -// See /usr/(local/)?doc/sparsehash-*/dense_hash_set.html -// for information about how to use this class. - -#ifndef _DENSE_HASH_SET_H_ -#define _DENSE_HASH_SET_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <algorithm> // needed by stl_alloc -#include <functional> // for equal_to<>, select1st<>, etc -#include <memory> // for alloc -#include <utility> // for pair<> -#include <sparsehash/internal/densehashtable.h> // IWYU pragma: export -#include <sparsehash/internal/libc_allocator_with_realloc.h> -#include HASH_FUN_H // for hash<> -_START_GOOGLE_NAMESPACE_ - -template <class Value, - class HashFcn = SPARSEHASH_HASH<Value>, // defined in sparseconfig.h - class EqualKey = std::equal_to<Value>, - class Alloc = libc_allocator_with_realloc<Value> > -class dense_hash_set { - private: - // Apparently identity is not stl-standard, so we define our own - struct Identity { - typedef const Value& result_type; - const Value& operator()(const Value& v) const { return v; } - }; - struct SetKey { - void operator()(Value* value, const Value& new_key) const { - *value = new_key; - } - }; - - // The actual data - typedef dense_hashtable<Value, Value, HashFcn, Identity, SetKey, - EqualKey, Alloc> ht; - ht rep; - - public: - typedef typename ht::key_type key_type; - typedef typename ht::value_type value_type; - typedef typename ht::hasher hasher; - typedef typename ht::key_equal key_equal; - typedef Alloc allocator_type; - - typedef typename ht::size_type size_type; - typedef typename ht::difference_type difference_type; - typedef typename ht::const_pointer pointer; - typedef typename ht::const_pointer const_pointer; - typedef typename ht::const_reference reference; - typedef typename ht::const_reference const_reference; - - typedef typename ht::const_iterator iterator; - typedef typename ht::const_iterator const_iterator; - typedef typename ht::const_local_iterator local_iterator; - typedef typename ht::const_local_iterator const_local_iterator; - - - // Iterator functions -- recall all iterators are const - iterator begin() const { return rep.begin(); } - iterator end() const { return rep.end(); } - - // These come from tr1's unordered_set. For us, a bucket has 0 or 1 elements. - local_iterator begin(size_type i) const { return rep.begin(i); } - local_iterator end(size_type i) const { return rep.end(i); } - - - // Accessor functions - allocator_type get_allocator() const { return rep.get_allocator(); } - hasher hash_funct() const { return rep.hash_funct(); } - hasher hash_function() const { return hash_funct(); } // tr1 name - key_equal key_eq() const { return rep.key_eq(); } - - - // Constructors - explicit dense_hash_set(size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { - } - - template <class InputIterator> - dense_hash_set(InputIterator f, InputIterator l, - const key_type& empty_key_val, - size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { - set_empty_key(empty_key_val); - rep.insert(f, l); - } - // We use the default copy constructor - // We use the default operator=() - // We use the default destructor - - void clear() { rep.clear(); } - // This clears the hash set without resizing it down to the minimum - // bucket count, but rather keeps the number of buckets constant - void clear_no_resize() { rep.clear_no_resize(); } - void swap(dense_hash_set& hs) { rep.swap(hs.rep); } - - - // Functions concerning size - size_type size() const { return rep.size(); } - size_type max_size() const { return rep.max_size(); } - bool empty() const { return rep.empty(); } - size_type bucket_count() const { return rep.bucket_count(); } - size_type max_bucket_count() const { return rep.max_bucket_count(); } - - // These are tr1 methods. bucket() is the bucket the key is or would be in. - size_type bucket_size(size_type i) const { return rep.bucket_size(i); } - size_type bucket(const key_type& key) const { return rep.bucket(key); } - float load_factor() const { - return size() * 1.0f / bucket_count(); - } - float max_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return grow; - } - void max_load_factor(float new_grow) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(shrink, new_grow); - } - // These aren't tr1 methods but perhaps ought to be. - float min_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return shrink; - } - void min_load_factor(float new_shrink) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(new_shrink, grow); - } - // Deprecated; use min_load_factor() or max_load_factor() instead. - void set_resizing_parameters(float shrink, float grow) { - rep.set_resizing_parameters(shrink, grow); - } - - void resize(size_type hint) { rep.resize(hint); } - void rehash(size_type hint) { resize(hint); } // the tr1 name - - // Lookup routines - iterator find(const key_type& key) const { return rep.find(key); } - - size_type count(const key_type& key) const { return rep.count(key); } - - std::pair<iterator, iterator> equal_range(const key_type& key) const { - return rep.equal_range(key); - } - - - // Insertion routines - std::pair<iterator, bool> insert(const value_type& obj) { - std::pair<typename ht::iterator, bool> p = rep.insert(obj); - return std::pair<iterator, bool>(p.first, p.second); // const to non-const - } - template <class InputIterator> void insert(InputIterator f, InputIterator l) { - rep.insert(f, l); - } - void insert(const_iterator f, const_iterator l) { - rep.insert(f, l); - } - // Required for std::insert_iterator; the passed-in iterator is ignored. - iterator insert(iterator, const value_type& obj) { - return insert(obj).first; - } - - // Deletion and empty routines - // THESE ARE NON-STANDARD! I make you specify an "impossible" key - // value to identify deleted and empty buckets. You can change the - // deleted key as time goes on, or get rid of it entirely to be insert-only. - void set_empty_key(const key_type& key) { rep.set_empty_key(key); } - key_type empty_key() const { return rep.empty_key(); } - - void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } - void clear_deleted_key() { rep.clear_deleted_key(); } - key_type deleted_key() const { return rep.deleted_key(); } - - // These are standard - size_type erase(const key_type& key) { return rep.erase(key); } - void erase(iterator it) { rep.erase(it); } - void erase(iterator f, iterator l) { rep.erase(f, l); } - - - // Comparison - bool operator==(const dense_hash_set& hs) const { return rep == hs.rep; } - bool operator!=(const dense_hash_set& hs) const { return rep != hs.rep; } - - - // I/O -- this is an add-on for writing metainformation to disk - // - // For maximum flexibility, this does not assume a particular - // file type (though it will probably be a FILE *). We just pass - // the fp through to rep. - - // If your keys and values are simple enough, you can pass this - // serializer to serialize()/unserialize(). "Simple enough" means - // value_type is a POD type that contains no pointers. Note, - // however, we don't try to normalize endianness. - typedef typename ht::NopointerSerializer NopointerSerializer; - - // serializer: a class providing operator()(OUTPUT*, const value_type&) - // (writing value_type to OUTPUT). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a - // pointer to a class providing size_t Write(const void*, size_t), - // which writes a buffer into a stream (which fp presumably - // owns) and returns the number of bytes successfully written. - // Note basic_ostream<not_char> is not currently supported. - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT* fp) { - return rep.serialize(serializer, fp); - } - - // serializer: a functor providing operator()(INPUT*, value_type*) - // (reading from INPUT and into value_type). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a - // pointer to a class providing size_t Read(void*, size_t), - // which reads into a buffer from a stream (which fp presumably - // owns) and returns the number of bytes successfully read. - // Note basic_istream<not_char> is not currently supported. - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT* fp) { - return rep.unserialize(serializer, fp); - } -}; - -template <class Val, class HashFcn, class EqualKey, class Alloc> -inline void swap(dense_hash_set<Val, HashFcn, EqualKey, Alloc>& hs1, - dense_hash_set<Val, HashFcn, EqualKey, Alloc>& hs2) { - hs1.swap(hs2); -} - -_END_GOOGLE_NAMESPACE_ - -#endif /* _DENSE_HASH_SET_H_ */ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// This is just a very thin wrapper over densehashtable.h, just +// like sgi stl's stl_hash_set is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// This is more different from dense_hash_map than you might think, +// because all iterators for sets are const (you obviously can't +// change the key, and for sets there is no value). +// +// NOTE: this is exactly like sparse_hash_set.h, with the word +// "sparse" replaced by "dense", except for the addition of +// set_empty_key(). +// +// YOU MUST CALL SET_EMPTY_KEY() IMMEDIATELY AFTER CONSTRUCTION. +// +// Otherwise your program will die in mysterious ways. (Note if you +// use the constructor that takes an InputIterator range, you pass in +// the empty key in the constructor, rather than after. As a result, +// this constructor differs from the standard STL version.) +// +// In other respects, we adhere mostly to the STL semantics for +// hash-map. One important exception is that insert() may invalidate +// iterators entirely -- STL semantics are that insert() may reorder +// iterators, but they all still refer to something valid in the +// hashtable. Not so for us. Likewise, insert() may invalidate +// pointers into the hashtable. (Whether insert invalidates iterators +// and pointers depends on whether it results in a hashtable resize). +// On the plus side, delete() doesn't invalidate iterators or pointers +// at all, or even change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// If you want to use erase() you must call set_deleted_key(), +// in addition to set_empty_key(), after construction. +// The deleted and empty keys must differ. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This allows you to iterate over a hashtable, +// and call erase(), without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_set: fastest, uses the most memory unless entries are small +// (2) sparse_hash_set: slowest, uses the least memory +// (3) hash_set / unordered_set (STL): in the middle +// +// Typically I use sparse_hash_set when I care about space and/or when +// I need to save the hashtable on disk. I use hash_set otherwise. I +// don't personally use dense_hash_set ever; some people use it for +// small sets with lots of lookups. +// +// - dense_hash_set has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_set uses .78X more bytes in overhead). +// - sparse_hash_set has about 4 bits overhead per entry. +// - sparse_hash_set can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/dense_hash_set.html +// for information about how to use this class. + +#ifndef _DENSE_HASH_SET_H_ +#define _DENSE_HASH_SET_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <algorithm> // needed by stl_alloc +#include <functional> // for equal_to<>, select1st<>, etc +#include <memory> // for alloc +#include <utility> // for pair<> +#include <sparsehash/internal/densehashtable.h> // IWYU pragma: export +#include <sparsehash/internal/libc_allocator_with_realloc.h> +#include HASH_FUN_H // for hash<> +_START_GOOGLE_NAMESPACE_ + +template <class Value, + class HashFcn = SPARSEHASH_HASH<Value>, // defined in sparseconfig.h + class EqualKey = std::equal_to<Value>, + class Alloc = libc_allocator_with_realloc<Value> > +class dense_hash_set { + private: + // Apparently identity is not stl-standard, so we define our own + struct Identity { + typedef const Value& result_type; + const Value& operator()(const Value& v) const { return v; } + }; + struct SetKey { + void operator()(Value* value, const Value& new_key) const { + *value = new_key; + } + }; + + // The actual data + typedef dense_hashtable<Value, Value, HashFcn, Identity, SetKey, + EqualKey, Alloc> ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::const_pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::const_reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::const_iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::const_local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + + // Iterator functions -- recall all iterators are const + iterator begin() const { return rep.begin(); } + iterator end() const { return rep.end(); } + + // These come from tr1's unordered_set. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) const { return rep.begin(i); } + local_iterator end(size_type i) const { return rep.end(i); } + + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } // tr1 name + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit dense_hash_set(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + } + + template <class InputIterator> + dense_hash_set(InputIterator f, InputIterator l, + const key_type& empty_key_val, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + set_empty_key(empty_key_val); + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + // This clears the hash set without resizing it down to the minimum + // bucket count, but rather keeps the number of buckets constant + void clear_no_resize() { rep.clear_no_resize(); } + void swap(dense_hash_set& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) const { return rep.find(key); } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair<iterator, iterator> equal_range(const key_type& key) const { + return rep.equal_range(key); + } + + + // Insertion routines + std::pair<iterator, bool> insert(const value_type& obj) { + std::pair<typename ht::iterator, bool> p = rep.insert(obj); + return std::pair<iterator, bool>(p.first, p.second); // const to non-const + } + template <class InputIterator> void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion and empty routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted and empty buckets. You can change the + // deleted key as time goes on, or get rid of it entirely to be insert-only. + void set_empty_key(const key_type& key) { rep.set_empty_key(key); } + key_type empty_key() const { return rep.empty_key(); } + + void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const dense_hash_set& hs) const { return rep == hs.rep; } + bool operator!=(const dense_hash_set& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing metainformation to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream<not_char> is not currently supported. + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream<not_char> is not currently supported. + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } +}; + +template <class Val, class HashFcn, class EqualKey, class Alloc> +inline void swap(dense_hash_set<Val, HashFcn, EqualKey, Alloc>& hs1, + dense_hash_set<Val, HashFcn, EqualKey, Alloc>& hs2) { + hs1.swap(hs2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _DENSE_HASH_SET_H_ */ diff --git a/contrib/libs/sparsehash/src/sparsehash/internal/densehashtable.h b/contrib/libs/sparsehash/src/sparsehash/internal/densehashtable.h index 7faa5913b5..cdf4ff624a 100644 --- a/contrib/libs/sparsehash/src/sparsehash/internal/densehashtable.h +++ b/contrib/libs/sparsehash/src/sparsehash/internal/densehashtable.h @@ -1,1327 +1,1327 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// A dense hashtable is a particular implementation of -// a hashtable: one that is meant to minimize memory allocation. -// It does this by using an array to store all the data. We -// steal a value from the key space to indicate "empty" array -// elements (ie indices where no item lives) and another to indicate -// "deleted" elements. -// -// (Note it is possible to change the value of the delete key -// on the fly; you can even remove it, though after that point -// the hashtable is insert_only until you set it again. The empty -// value however can't be changed.) -// -// To minimize allocation and pointer overhead, we use internal -// probing, in which the hashtable is a single table, and collisions -// are resolved by trying to insert again in another bucket. The -// most cache-efficient internal probing schemes are linear probing -// (which suffers, alas, from clumping) and quadratic probing, which -// is what we implement by default. -// -// Type requirements: value_type is required to be Copy Constructible -// and Default Constructible. It is not required to be (and commonly -// isn't) Assignable. -// -// You probably shouldn't use this code directly. Use dense_hash_map<> -// or dense_hash_set<> instead. - -// You can change the following below: -// HT_OCCUPANCY_PCT -- how full before we double size -// HT_EMPTY_PCT -- how empty before we halve size -// HT_MIN_BUCKETS -- default smallest bucket size -// -// You can also change enlarge_factor (which defaults to -// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to -// HT_EMPTY_PCT) with set_resizing_parameters(). -// -// How to decide what values to use? -// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good. -// HT_MIN_BUCKETS is probably unnecessary since you can specify -// (indirectly) the starting number of buckets at construct-time. -// For enlarge_factor, you can use this chart to try to trade-off -// expected lookup time to the space taken up. By default, this -// code uses quadratic probing, though you can change it to linear -// via JUMP_ below if you really want to. -// -// From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html -// NUMBER OF PROBES / LOOKUP Successful Unsuccessful -// Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L) -// Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2 -// -// -- enlarge_factor -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99 -// QUADRATIC COLLISION RES. -// probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11 -// probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6 -// LINEAR COLLISION RES. -// probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5 -// probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0 - -#ifndef _DENSEHASHTABLE_H_ -#define _DENSEHASHTABLE_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <assert.h> -#include <stdio.h> // for FILE, fwrite, fread -#include <algorithm> // For swap(), eg -#include <iterator> // For iterator tags -#include <limits> // for numeric_limits -#include <memory> // For uninitialized_fill -#include <utility> // for pair -#include <sparsehash/internal/hashtable-common.h> -#include <sparsehash/internal/libc_allocator_with_realloc.h> -#include <sparsehash/type_traits.h> -#include <stdexcept> // For length_error - -_START_GOOGLE_NAMESPACE_ - -namespace base { // just to make google->opensource transition easier -using GOOGLE_NAMESPACE::true_type; -using GOOGLE_NAMESPACE::false_type; -using GOOGLE_NAMESPACE::integral_constant; -using GOOGLE_NAMESPACE::is_same; -using GOOGLE_NAMESPACE::remove_const; -} - -// The probing method -// Linear probing -// #define JUMP_(key, num_probes) ( 1 ) -// Quadratic probing -#define JUMP_(key, num_probes) ( num_probes ) - -// Hashtable class, used to implement the hashed associative containers -// hash_set and hash_map. - -// Value: what is stored in the table (each bucket is a Value). -// Key: something in a 1-to-1 correspondence to a Value, that can be used -// to search for a Value in the table (find() takes a Key). -// HashFcn: Takes a Key and returns an integer, the more unique the better. -// ExtractKey: given a Value, returns the unique Key associated with it. -// Must inherit from unary_function, or at least have a -// result_type enum indicating the return type of operator(). -// SetKey: given a Value* and a Key, modifies the value such that -// ExtractKey(value) == key. We guarantee this is only called -// with key == deleted_key or key == empty_key. -// EqualKey: Given two Keys, says whether they are the same (that is, -// if they are both associated with the same Value). -// Alloc: STL allocator to use to allocate memory. - -template <class Value, class Key, class HashFcn, - class ExtractKey, class SetKey, class EqualKey, class Alloc> -class dense_hashtable; - -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct dense_hashtable_iterator; - -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct dense_hashtable_const_iterator; - -// We're just an array, but we need to skip over empty and deleted elements -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct dense_hashtable_iterator { - private: - typedef typename A::template rebind<V>::other value_alloc_type; - - public: - typedef dense_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; - typedef dense_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; - - typedef std::forward_iterator_tag iterator_category; // very little defined! - typedef V value_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::pointer pointer; - - // "Real" constructor and default constructor - dense_hashtable_iterator(const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, - pointer it, pointer it_end, bool advance) - : ht(h), pos(it), end(it_end) { - if (advance) advance_past_empty_and_deleted(); - } - dense_hashtable_iterator() { } - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // Happy dereferencer - reference operator*() const { return *pos; } - pointer operator->() const { return &(operator*()); } - - // Arithmetic. The only hard part is making sure that - // we're not on an empty or marked-deleted array element - void advance_past_empty_and_deleted() { - while ( pos != end && (ht->test_empty(*this) || ht->test_deleted(*this)) ) - ++pos; - } - iterator& operator++() { - assert(pos != end); ++pos; advance_past_empty_and_deleted(); return *this; - } - iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } - - // Comparison. - bool operator==(const iterator& it) const { return pos == it.pos; } - bool operator!=(const iterator& it) const { return pos != it.pos; } - - - // The actual data - const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; - pointer pos, end; -}; - - -// Now do it all again, but with const-ness! -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct dense_hashtable_const_iterator { - private: - typedef typename A::template rebind<V>::other value_alloc_type; - - public: - typedef dense_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; - typedef dense_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; - - typedef std::forward_iterator_tag iterator_category; // very little defined! - typedef V value_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::const_reference reference; - typedef typename value_alloc_type::const_pointer pointer; - - // "Real" constructor and default constructor - dense_hashtable_const_iterator( - const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, - pointer it, pointer it_end, bool advance) - : ht(h), pos(it), end(it_end) { - if (advance) advance_past_empty_and_deleted(); - } - dense_hashtable_const_iterator() - : ht(NULL), pos(pointer()), end(pointer()) { } - // This lets us convert regular iterators to const iterators - dense_hashtable_const_iterator(const iterator &it) - : ht(it.ht), pos(it.pos), end(it.end) { } - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // Happy dereferencer - reference operator*() const { return *pos; } - pointer operator->() const { return &(operator*()); } - - // Arithmetic. The only hard part is making sure that - // we're not on an empty or marked-deleted array element - void advance_past_empty_and_deleted() { - while ( pos != end && (ht->test_empty(*this) || ht->test_deleted(*this)) ) - ++pos; - } - const_iterator& operator++() { - assert(pos != end); ++pos; advance_past_empty_and_deleted(); return *this; - } - const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; } - - // Comparison. - bool operator==(const const_iterator& it) const { return pos == it.pos; } - bool operator!=(const const_iterator& it) const { return pos != it.pos; } - - - // The actual data - const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; - pointer pos, end; -}; - -template <class Value, class Key, class HashFcn, - class ExtractKey, class SetKey, class EqualKey, class Alloc> -class dense_hashtable { - private: - typedef typename Alloc::template rebind<Value>::other value_alloc_type; - - public: - typedef Key key_type; - typedef Value value_type; - typedef HashFcn hasher; - typedef EqualKey key_equal; - typedef Alloc allocator_type; - - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::const_reference const_reference; - typedef typename value_alloc_type::pointer pointer; - typedef typename value_alloc_type::const_pointer const_pointer; - typedef dense_hashtable_iterator<Value, Key, HashFcn, - ExtractKey, SetKey, EqualKey, Alloc> - iterator; - - typedef dense_hashtable_const_iterator<Value, Key, HashFcn, - ExtractKey, SetKey, EqualKey, Alloc> - const_iterator; - - // These come from tr1. For us they're the same as regular iterators. - typedef iterator local_iterator; - typedef const_iterator const_local_iterator; - - // How full we let the table get before we resize, by default. - // Knuth says .8 is good -- higher causes us to probe too much, - // though it saves memory. - static const int HT_OCCUPANCY_PCT; // defined at the bottom of this file - - // How empty we let the table get before we resize lower, by default. - // (0.0 means never resize lower.) - // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing - static const int HT_EMPTY_PCT; // defined at the bottom of this file - - // Minimum size we're willing to let hashtables be. - // Must be a power of two, and at least 4. - // Note, however, that for a given hashtable, the initial size is a - // function of the first constructor arg, and may be >HT_MIN_BUCKETS. - static const size_type HT_MIN_BUCKETS = 4; - - // By default, if you don't specify a hashtable size at - // construction-time, we use this size. Must be a power of two, and - // at least HT_MIN_BUCKETS. - static const size_type HT_DEFAULT_STARTING_BUCKETS = 32; - - // ITERATOR FUNCTIONS - iterator begin() { return iterator(this, table, - table + num_buckets, true); } - iterator end() { return iterator(this, table + num_buckets, - table + num_buckets, true); } - const_iterator begin() const { return const_iterator(this, table, - table+num_buckets,true);} - const_iterator end() const { return const_iterator(this, table + num_buckets, - table+num_buckets,true);} - - // These come from tr1 unordered_map. They iterate over 'bucket' n. - // We'll just consider bucket n to be the n-th element of the table. - local_iterator begin(size_type i) { - return local_iterator(this, table + i, table + i+1, false); - } - local_iterator end(size_type i) { - local_iterator it = begin(i); - if (!test_empty(i) && !test_deleted(i)) - ++it; - return it; - } - const_local_iterator begin(size_type i) const { - return const_local_iterator(this, table + i, table + i+1, false); - } - const_local_iterator end(size_type i) const { - const_local_iterator it = begin(i); - if (!test_empty(i) && !test_deleted(i)) - ++it; - return it; - } - - // ACCESSOR FUNCTIONS for the things we templatize on, basically - hasher hash_funct() const { return settings; } - key_equal key_eq() const { return key_info; } - allocator_type get_allocator() const { - return allocator_type(val_info); - } - - // Accessor function for statistics gathering. - int num_table_copies() const { return settings.num_ht_copies(); } - - private: - // Annoyingly, we can't copy values around, because they might have - // const components (they're probably pair<const X, Y>). We use - // explicit destructor invocation and placement new to get around - // this. Arg. - void set_value(pointer dst, const_reference src) { - dst->~value_type(); // delete the old value, if any - new(dst) value_type(src); - } - - void destroy_buckets(size_type first, size_type last) { - for ( ; first != last; ++first) - table[first].~value_type(); - } - - // DELETE HELPER FUNCTIONS - // This lets the user describe a key that will indicate deleted - // table entries. This key should be an "impossible" entry -- - // if you try to insert it for real, you won't be able to retrieve it! - // (NB: while you pass in an entire value, only the key part is looked - // at. This is just because I don't know how to assign just a key.) - private: - void squash_deleted() { // gets rid of any deleted entries we have - if ( num_deleted ) { // get rid of deleted before writing - dense_hashtable tmp(*this); // copying will get rid of deleted - swap(tmp); // now we are tmp - } - assert(num_deleted == 0); - } - - // Test if the given key is the deleted indicator. Requires - // num_deleted > 0, for correctness of read(), and because that - // guarantees that key_info.delkey is valid. - bool test_deleted_key(const key_type& key) const { - assert(num_deleted > 0); - return equals(key_info.delkey, key); - } - - public: - void set_deleted_key(const key_type &key) { - // the empty indicator (if specified) and the deleted indicator - // must be different - assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) - && "Passed the empty-key to set_deleted_key"); - // It's only safe to change what "deleted" means if we purge deleted guys - squash_deleted(); - settings.set_use_deleted(true); - key_info.delkey = key; - } - void clear_deleted_key() { - squash_deleted(); - settings.set_use_deleted(false); - } - key_type deleted_key() const { - assert(settings.use_deleted() - && "Must set deleted key before calling deleted_key"); - return key_info.delkey; - } - - // These are public so the iterators can use them - // True if the item at position bucknum is "deleted" marker - bool test_deleted(size_type bucknum) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && test_deleted_key(get_key(table[bucknum])); - } - bool test_deleted(const iterator &it) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && test_deleted_key(get_key(*it)); - } - bool test_deleted(const const_iterator &it) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && test_deleted_key(get_key(*it)); - } - - private: - void check_use_deleted(const char* caller) { - (void)caller; // could log it if the assert failed - assert(settings.use_deleted()); - } - - // Set it so test_deleted is true. true if object didn't used to be deleted. - bool set_deleted(iterator &it) { - check_use_deleted("set_deleted()"); - bool retval = !test_deleted(it); - // &* converts from iterator to value-type. - set_key(&(*it), key_info.delkey); - return retval; - } - // Set it so test_deleted is false. true if object used to be deleted. - bool clear_deleted(iterator &it) { - check_use_deleted("clear_deleted()"); - // Happens automatically when we assign something else in its place. - return test_deleted(it); - } - - // We also allow to set/clear the deleted bit on a const iterator. - // We allow a const_iterator for the same reason you can delete a - // const pointer: it's convenient, and semantically you can't use - // 'it' after it's been deleted anyway, so its const-ness doesn't - // really matter. - bool set_deleted(const_iterator &it) { - check_use_deleted("set_deleted()"); - bool retval = !test_deleted(it); - set_key(const_cast<pointer>(&(*it)), key_info.delkey); - return retval; - } - // Set it so test_deleted is false. true if object used to be deleted. - bool clear_deleted(const_iterator &it) { - check_use_deleted("clear_deleted()"); - return test_deleted(it); - } - - // EMPTY HELPER FUNCTIONS - // This lets the user describe a key that will indicate empty (unused) - // table entries. This key should be an "impossible" entry -- - // if you try to insert it for real, you won't be able to retrieve it! - // (NB: while you pass in an entire value, only the key part is looked - // at. This is just because I don't know how to assign just a key.) - public: - // These are public so the iterators can use them - // True if the item at position bucknum is "empty" marker - bool test_empty(size_type bucknum) const { - assert(settings.use_empty()); // we always need to know what's empty! - return equals(get_key(val_info.emptyval), get_key(table[bucknum])); - } - bool test_empty(const iterator &it) const { - assert(settings.use_empty()); // we always need to know what's empty! - return equals(get_key(val_info.emptyval), get_key(*it)); - } - bool test_empty(const const_iterator &it) const { - assert(settings.use_empty()); // we always need to know what's empty! - return equals(get_key(val_info.emptyval), get_key(*it)); - } - - private: - void fill_range_with_empty(pointer table_start, pointer table_end) { - std::uninitialized_fill(table_start, table_end, val_info.emptyval); - } - - public: - // TODO(csilvers): change all callers of this to pass in a key instead, - // and take a const key_type instead of const value_type. - void set_empty_key(const_reference val) { - // Once you set the empty key, you can't change it - assert(!settings.use_empty() && "Calling set_empty_key multiple times"); - // The deleted indicator (if specified) and the empty indicator - // must be different. - assert((!settings.use_deleted() || !equals(get_key(val), key_info.delkey)) - && "Setting the empty key the same as the deleted key"); - settings.set_use_empty(true); - set_value(&val_info.emptyval, val); - - assert(!table); // must set before first use - // num_buckets was set in constructor even though table was NULL - table = val_info.allocate(num_buckets); - assert(table); - fill_range_with_empty(table, table + num_buckets); - } - // TODO(user): return a key_type rather than a value_type - value_type empty_key() const { - assert(settings.use_empty()); - return val_info.emptyval; - } - - // FUNCTIONS CONCERNING SIZE - public: - size_type size() const { return num_elements - num_deleted; } - size_type max_size() const { return val_info.max_size(); } - bool empty() const { return size() == 0; } - size_type bucket_count() const { return num_buckets; } - size_type max_bucket_count() const { return max_size(); } - size_type nonempty_bucket_count() const { return num_elements; } - // These are tr1 methods. Their idea of 'bucket' doesn't map well to - // what we do. We just say every bucket has 0 or 1 items in it. - size_type bucket_size(size_type i) const { - return begin(i) == end(i) ? 0 : 1; - } - - private: - // Because of the above, size_type(-1) is never legal; use it for errors - static const size_type ILLEGAL_BUCKET = size_type(-1); - - // Used after a string of deletes. Returns true if we actually shrunk. - // TODO(csilvers): take a delta so we can take into account inserts - // done after shrinking. Maybe make part of the Settings class? - bool maybe_shrink() { - assert(num_elements >= num_deleted); - assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two - assert(bucket_count() >= HT_MIN_BUCKETS); - bool retval = false; - - // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS, - // we'll never shrink until you get relatively big, and we'll never - // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something - // like "dense_hash_set<int> x; x.insert(4); x.erase(4);" will - // shrink us down to HT_MIN_BUCKETS buckets, which is too small. - const size_type num_remain = num_elements - num_deleted; - const size_type shrink_threshold = settings.shrink_threshold(); - if (shrink_threshold > 0 && num_remain < shrink_threshold && - bucket_count() > HT_DEFAULT_STARTING_BUCKETS) { - const float shrink_factor = settings.shrink_factor(); - size_type sz = bucket_count() / 2; // find how much we should shrink - while (sz > HT_DEFAULT_STARTING_BUCKETS && - num_remain < sz * shrink_factor) { - sz /= 2; // stay a power of 2 - } - dense_hashtable tmp(*this, sz); // Do the actual resizing - swap(tmp); // now we are tmp - retval = true; - } - settings.set_consider_shrink(false); // because we just considered it - return retval; - } - - // We'll let you resize a hashtable -- though this makes us copy all! - // When you resize, you say, "make it big enough for this many more elements" - // Returns true if we actually resized, false if size was already ok. - bool resize_delta(size_type delta) { - bool did_resize = false; - if ( settings.consider_shrink() ) { // see if lots of deletes happened - if ( maybe_shrink() ) - did_resize = true; - } - if (num_elements >= - (std::numeric_limits<size_type>::max)() - delta) { - throw std::length_error("resize overflow"); - } - if ( bucket_count() >= HT_MIN_BUCKETS && - (num_elements + delta) <= settings.enlarge_threshold() ) - return did_resize; // we're ok as we are - - // Sometimes, we need to resize just to get rid of all the - // "deleted" buckets that are clogging up the hashtable. So when - // deciding whether to resize, count the deleted buckets (which - // are currently taking up room). But later, when we decide what - // size to resize to, *don't* count deleted buckets, since they - // get discarded during the resize. +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// A dense hashtable is a particular implementation of +// a hashtable: one that is meant to minimize memory allocation. +// It does this by using an array to store all the data. We +// steal a value from the key space to indicate "empty" array +// elements (ie indices where no item lives) and another to indicate +// "deleted" elements. +// +// (Note it is possible to change the value of the delete key +// on the fly; you can even remove it, though after that point +// the hashtable is insert_only until you set it again. The empty +// value however can't be changed.) +// +// To minimize allocation and pointer overhead, we use internal +// probing, in which the hashtable is a single table, and collisions +// are resolved by trying to insert again in another bucket. The +// most cache-efficient internal probing schemes are linear probing +// (which suffers, alas, from clumping) and quadratic probing, which +// is what we implement by default. +// +// Type requirements: value_type is required to be Copy Constructible +// and Default Constructible. It is not required to be (and commonly +// isn't) Assignable. +// +// You probably shouldn't use this code directly. Use dense_hash_map<> +// or dense_hash_set<> instead. + +// You can change the following below: +// HT_OCCUPANCY_PCT -- how full before we double size +// HT_EMPTY_PCT -- how empty before we halve size +// HT_MIN_BUCKETS -- default smallest bucket size +// +// You can also change enlarge_factor (which defaults to +// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to +// HT_EMPTY_PCT) with set_resizing_parameters(). +// +// How to decide what values to use? +// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good. +// HT_MIN_BUCKETS is probably unnecessary since you can specify +// (indirectly) the starting number of buckets at construct-time. +// For enlarge_factor, you can use this chart to try to trade-off +// expected lookup time to the space taken up. By default, this +// code uses quadratic probing, though you can change it to linear +// via JUMP_ below if you really want to. +// +// From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html +// NUMBER OF PROBES / LOOKUP Successful Unsuccessful +// Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L) +// Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2 +// +// -- enlarge_factor -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99 +// QUADRATIC COLLISION RES. +// probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11 +// probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6 +// LINEAR COLLISION RES. +// probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5 +// probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0 + +#ifndef _DENSEHASHTABLE_H_ +#define _DENSEHASHTABLE_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <assert.h> +#include <stdio.h> // for FILE, fwrite, fread +#include <algorithm> // For swap(), eg +#include <iterator> // For iterator tags +#include <limits> // for numeric_limits +#include <memory> // For uninitialized_fill +#include <utility> // for pair +#include <sparsehash/internal/hashtable-common.h> +#include <sparsehash/internal/libc_allocator_with_realloc.h> +#include <sparsehash/type_traits.h> +#include <stdexcept> // For length_error + +_START_GOOGLE_NAMESPACE_ + +namespace base { // just to make google->opensource transition easier +using GOOGLE_NAMESPACE::true_type; +using GOOGLE_NAMESPACE::false_type; +using GOOGLE_NAMESPACE::integral_constant; +using GOOGLE_NAMESPACE::is_same; +using GOOGLE_NAMESPACE::remove_const; +} + +// The probing method +// Linear probing +// #define JUMP_(key, num_probes) ( 1 ) +// Quadratic probing +#define JUMP_(key, num_probes) ( num_probes ) + +// Hashtable class, used to implement the hashed associative containers +// hash_set and hash_map. + +// Value: what is stored in the table (each bucket is a Value). +// Key: something in a 1-to-1 correspondence to a Value, that can be used +// to search for a Value in the table (find() takes a Key). +// HashFcn: Takes a Key and returns an integer, the more unique the better. +// ExtractKey: given a Value, returns the unique Key associated with it. +// Must inherit from unary_function, or at least have a +// result_type enum indicating the return type of operator(). +// SetKey: given a Value* and a Key, modifies the value such that +// ExtractKey(value) == key. We guarantee this is only called +// with key == deleted_key or key == empty_key. +// EqualKey: Given two Keys, says whether they are the same (that is, +// if they are both associated with the same Value). +// Alloc: STL allocator to use to allocate memory. + +template <class Value, class Key, class HashFcn, + class ExtractKey, class SetKey, class EqualKey, class Alloc> +class dense_hashtable; + +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct dense_hashtable_iterator; + +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct dense_hashtable_const_iterator; + +// We're just an array, but we need to skip over empty and deleted elements +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct dense_hashtable_iterator { + private: + typedef typename A::template rebind<V>::other value_alloc_type; + + public: + typedef dense_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; + typedef dense_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::pointer pointer; + + // "Real" constructor and default constructor + dense_hashtable_iterator(const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, + pointer it, pointer it_end, bool advance) + : ht(h), pos(it), end(it_end) { + if (advance) advance_past_empty_and_deleted(); + } + dense_hashtable_iterator() { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on an empty or marked-deleted array element + void advance_past_empty_and_deleted() { + while ( pos != end && (ht->test_empty(*this) || ht->test_deleted(*this)) ) + ++pos; + } + iterator& operator++() { + assert(pos != end); ++pos; advance_past_empty_and_deleted(); return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const iterator& it) const { return pos == it.pos; } + bool operator!=(const iterator& it) const { return pos != it.pos; } + + + // The actual data + const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; + pointer pos, end; +}; + + +// Now do it all again, but with const-ness! +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct dense_hashtable_const_iterator { + private: + typedef typename A::template rebind<V>::other value_alloc_type; + + public: + typedef dense_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; + typedef dense_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::const_reference reference; + typedef typename value_alloc_type::const_pointer pointer; + + // "Real" constructor and default constructor + dense_hashtable_const_iterator( + const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, + pointer it, pointer it_end, bool advance) + : ht(h), pos(it), end(it_end) { + if (advance) advance_past_empty_and_deleted(); + } + dense_hashtable_const_iterator() + : ht(NULL), pos(pointer()), end(pointer()) { } + // This lets us convert regular iterators to const iterators + dense_hashtable_const_iterator(const iterator &it) + : ht(it.ht), pos(it.pos), end(it.end) { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on an empty or marked-deleted array element + void advance_past_empty_and_deleted() { + while ( pos != end && (ht->test_empty(*this) || ht->test_deleted(*this)) ) + ++pos; + } + const_iterator& operator++() { + assert(pos != end); ++pos; advance_past_empty_and_deleted(); return *this; + } + const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const const_iterator& it) const { return pos == it.pos; } + bool operator!=(const const_iterator& it) const { return pos != it.pos; } + + + // The actual data + const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; + pointer pos, end; +}; + +template <class Value, class Key, class HashFcn, + class ExtractKey, class SetKey, class EqualKey, class Alloc> +class dense_hashtable { + private: + typedef typename Alloc::template rebind<Value>::other value_alloc_type; + + public: + typedef Key key_type; + typedef Value value_type; + typedef HashFcn hasher; + typedef EqualKey key_equal; + typedef Alloc allocator_type; + + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + typedef dense_hashtable_iterator<Value, Key, HashFcn, + ExtractKey, SetKey, EqualKey, Alloc> + iterator; + + typedef dense_hashtable_const_iterator<Value, Key, HashFcn, + ExtractKey, SetKey, EqualKey, Alloc> + const_iterator; + + // These come from tr1. For us they're the same as regular iterators. + typedef iterator local_iterator; + typedef const_iterator const_local_iterator; + + // How full we let the table get before we resize, by default. + // Knuth says .8 is good -- higher causes us to probe too much, + // though it saves memory. + static const int HT_OCCUPANCY_PCT; // defined at the bottom of this file + + // How empty we let the table get before we resize lower, by default. + // (0.0 means never resize lower.) + // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing + static const int HT_EMPTY_PCT; // defined at the bottom of this file + + // Minimum size we're willing to let hashtables be. + // Must be a power of two, and at least 4. + // Note, however, that for a given hashtable, the initial size is a + // function of the first constructor arg, and may be >HT_MIN_BUCKETS. + static const size_type HT_MIN_BUCKETS = 4; + + // By default, if you don't specify a hashtable size at + // construction-time, we use this size. Must be a power of two, and + // at least HT_MIN_BUCKETS. + static const size_type HT_DEFAULT_STARTING_BUCKETS = 32; + + // ITERATOR FUNCTIONS + iterator begin() { return iterator(this, table, + table + num_buckets, true); } + iterator end() { return iterator(this, table + num_buckets, + table + num_buckets, true); } + const_iterator begin() const { return const_iterator(this, table, + table+num_buckets,true);} + const_iterator end() const { return const_iterator(this, table + num_buckets, + table+num_buckets,true);} + + // These come from tr1 unordered_map. They iterate over 'bucket' n. + // We'll just consider bucket n to be the n-th element of the table. + local_iterator begin(size_type i) { + return local_iterator(this, table + i, table + i+1, false); + } + local_iterator end(size_type i) { + local_iterator it = begin(i); + if (!test_empty(i) && !test_deleted(i)) + ++it; + return it; + } + const_local_iterator begin(size_type i) const { + return const_local_iterator(this, table + i, table + i+1, false); + } + const_local_iterator end(size_type i) const { + const_local_iterator it = begin(i); + if (!test_empty(i) && !test_deleted(i)) + ++it; + return it; + } + + // ACCESSOR FUNCTIONS for the things we templatize on, basically + hasher hash_funct() const { return settings; } + key_equal key_eq() const { return key_info; } + allocator_type get_allocator() const { + return allocator_type(val_info); + } + + // Accessor function for statistics gathering. + int num_table_copies() const { return settings.num_ht_copies(); } + + private: + // Annoyingly, we can't copy values around, because they might have + // const components (they're probably pair<const X, Y>). We use + // explicit destructor invocation and placement new to get around + // this. Arg. + void set_value(pointer dst, const_reference src) { + dst->~value_type(); // delete the old value, if any + new(dst) value_type(src); + } + + void destroy_buckets(size_type first, size_type last) { + for ( ; first != last; ++first) + table[first].~value_type(); + } + + // DELETE HELPER FUNCTIONS + // This lets the user describe a key that will indicate deleted + // table entries. This key should be an "impossible" entry -- + // if you try to insert it for real, you won't be able to retrieve it! + // (NB: while you pass in an entire value, only the key part is looked + // at. This is just because I don't know how to assign just a key.) + private: + void squash_deleted() { // gets rid of any deleted entries we have + if ( num_deleted ) { // get rid of deleted before writing + dense_hashtable tmp(*this); // copying will get rid of deleted + swap(tmp); // now we are tmp + } + assert(num_deleted == 0); + } + + // Test if the given key is the deleted indicator. Requires + // num_deleted > 0, for correctness of read(), and because that + // guarantees that key_info.delkey is valid. + bool test_deleted_key(const key_type& key) const { + assert(num_deleted > 0); + return equals(key_info.delkey, key); + } + + public: + void set_deleted_key(const key_type &key) { + // the empty indicator (if specified) and the deleted indicator + // must be different + assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) + && "Passed the empty-key to set_deleted_key"); + // It's only safe to change what "deleted" means if we purge deleted guys + squash_deleted(); + settings.set_use_deleted(true); + key_info.delkey = key; + } + void clear_deleted_key() { + squash_deleted(); + settings.set_use_deleted(false); + } + key_type deleted_key() const { + assert(settings.use_deleted() + && "Must set deleted key before calling deleted_key"); + return key_info.delkey; + } + + // These are public so the iterators can use them + // True if the item at position bucknum is "deleted" marker + bool test_deleted(size_type bucknum) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(table[bucknum])); + } + bool test_deleted(const iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + bool test_deleted(const const_iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + + private: + void check_use_deleted(const char* caller) { + (void)caller; // could log it if the assert failed + assert(settings.use_deleted()); + } + + // Set it so test_deleted is true. true if object didn't used to be deleted. + bool set_deleted(iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + // &* converts from iterator to value-type. + set_key(&(*it), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(iterator &it) { + check_use_deleted("clear_deleted()"); + // Happens automatically when we assign something else in its place. + return test_deleted(it); + } + + // We also allow to set/clear the deleted bit on a const iterator. + // We allow a const_iterator for the same reason you can delete a + // const pointer: it's convenient, and semantically you can't use + // 'it' after it's been deleted anyway, so its const-ness doesn't + // really matter. + bool set_deleted(const_iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + set_key(const_cast<pointer>(&(*it)), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(const_iterator &it) { + check_use_deleted("clear_deleted()"); + return test_deleted(it); + } + + // EMPTY HELPER FUNCTIONS + // This lets the user describe a key that will indicate empty (unused) + // table entries. This key should be an "impossible" entry -- + // if you try to insert it for real, you won't be able to retrieve it! + // (NB: while you pass in an entire value, only the key part is looked + // at. This is just because I don't know how to assign just a key.) + public: + // These are public so the iterators can use them + // True if the item at position bucknum is "empty" marker + bool test_empty(size_type bucknum) const { + assert(settings.use_empty()); // we always need to know what's empty! + return equals(get_key(val_info.emptyval), get_key(table[bucknum])); + } + bool test_empty(const iterator &it) const { + assert(settings.use_empty()); // we always need to know what's empty! + return equals(get_key(val_info.emptyval), get_key(*it)); + } + bool test_empty(const const_iterator &it) const { + assert(settings.use_empty()); // we always need to know what's empty! + return equals(get_key(val_info.emptyval), get_key(*it)); + } + + private: + void fill_range_with_empty(pointer table_start, pointer table_end) { + std::uninitialized_fill(table_start, table_end, val_info.emptyval); + } + + public: + // TODO(csilvers): change all callers of this to pass in a key instead, + // and take a const key_type instead of const value_type. + void set_empty_key(const_reference val) { + // Once you set the empty key, you can't change it + assert(!settings.use_empty() && "Calling set_empty_key multiple times"); + // The deleted indicator (if specified) and the empty indicator + // must be different. + assert((!settings.use_deleted() || !equals(get_key(val), key_info.delkey)) + && "Setting the empty key the same as the deleted key"); + settings.set_use_empty(true); + set_value(&val_info.emptyval, val); + + assert(!table); // must set before first use + // num_buckets was set in constructor even though table was NULL + table = val_info.allocate(num_buckets); + assert(table); + fill_range_with_empty(table, table + num_buckets); + } + // TODO(user): return a key_type rather than a value_type + value_type empty_key() const { + assert(settings.use_empty()); + return val_info.emptyval; + } + + // FUNCTIONS CONCERNING SIZE + public: + size_type size() const { return num_elements - num_deleted; } + size_type max_size() const { return val_info.max_size(); } + bool empty() const { return size() == 0; } + size_type bucket_count() const { return num_buckets; } + size_type max_bucket_count() const { return max_size(); } + size_type nonempty_bucket_count() const { return num_elements; } + // These are tr1 methods. Their idea of 'bucket' doesn't map well to + // what we do. We just say every bucket has 0 or 1 items in it. + size_type bucket_size(size_type i) const { + return begin(i) == end(i) ? 0 : 1; + } + + private: + // Because of the above, size_type(-1) is never legal; use it for errors + static const size_type ILLEGAL_BUCKET = size_type(-1); + + // Used after a string of deletes. Returns true if we actually shrunk. + // TODO(csilvers): take a delta so we can take into account inserts + // done after shrinking. Maybe make part of the Settings class? + bool maybe_shrink() { + assert(num_elements >= num_deleted); + assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two + assert(bucket_count() >= HT_MIN_BUCKETS); + bool retval = false; + + // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS, + // we'll never shrink until you get relatively big, and we'll never + // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something + // like "dense_hash_set<int> x; x.insert(4); x.erase(4);" will + // shrink us down to HT_MIN_BUCKETS buckets, which is too small. + const size_type num_remain = num_elements - num_deleted; + const size_type shrink_threshold = settings.shrink_threshold(); + if (shrink_threshold > 0 && num_remain < shrink_threshold && + bucket_count() > HT_DEFAULT_STARTING_BUCKETS) { + const float shrink_factor = settings.shrink_factor(); + size_type sz = bucket_count() / 2; // find how much we should shrink + while (sz > HT_DEFAULT_STARTING_BUCKETS && + num_remain < sz * shrink_factor) { + sz /= 2; // stay a power of 2 + } + dense_hashtable tmp(*this, sz); // Do the actual resizing + swap(tmp); // now we are tmp + retval = true; + } + settings.set_consider_shrink(false); // because we just considered it + return retval; + } + + // We'll let you resize a hashtable -- though this makes us copy all! + // When you resize, you say, "make it big enough for this many more elements" + // Returns true if we actually resized, false if size was already ok. + bool resize_delta(size_type delta) { + bool did_resize = false; + if ( settings.consider_shrink() ) { // see if lots of deletes happened + if ( maybe_shrink() ) + did_resize = true; + } + if (num_elements >= + (std::numeric_limits<size_type>::max)() - delta) { + throw std::length_error("resize overflow"); + } + if ( bucket_count() >= HT_MIN_BUCKETS && + (num_elements + delta) <= settings.enlarge_threshold() ) + return did_resize; // we're ok as we are + + // Sometimes, we need to resize just to get rid of all the + // "deleted" buckets that are clogging up the hashtable. So when + // deciding whether to resize, count the deleted buckets (which + // are currently taking up room). But later, when we decide what + // size to resize to, *don't* count deleted buckets, since they + // get discarded during the resize. size_type needed_size = settings.min_buckets(num_elements + delta, 0); - if ( needed_size <= bucket_count() ) // we have enough buckets - return did_resize; - - size_type resize_to = - settings.min_buckets(num_elements - num_deleted + delta, bucket_count()); - + if ( needed_size <= bucket_count() ) // we have enough buckets + return did_resize; + + size_type resize_to = + settings.min_buckets(num_elements - num_deleted + delta, bucket_count()); + // When num_deleted is large, we may still grow but we do not want to // over expand. So we reduce needed_size by a portion of num_deleted // (the exact portion does not matter). This is especially helpful // when min_load_factor is zero (no shrink at all) to avoid doubling // the bucket count to infinity. See also test ResizeWithoutShrink. needed_size = settings.min_buckets(num_elements - num_deleted / 4 + delta, 0); - if (resize_to < needed_size && // may double resize_to - resize_to < (std::numeric_limits<size_type>::max)() / 2) { - // This situation means that we have enough deleted elements, - // that once we purge them, we won't actually have needed to - // grow. But we may want to grow anyway: if we just purge one - // element, say, we'll have to grow anyway next time we - // insert. Might as well grow now, since we're already going - // through the trouble of copying (in order to purge the - // deleted elements). - const size_type target = - static_cast<size_type>(settings.shrink_size(resize_to*2)); - if (num_elements - num_deleted + delta >= target) { - // Good, we won't be below the shrink threshhold even if we double. - resize_to *= 2; - } - } - dense_hashtable tmp(*this, resize_to); - swap(tmp); // now we are tmp - return true; - } - - // We require table be not-NULL and empty before calling this. - void resize_table(size_type /*old_size*/, size_type new_size, - base::true_type) { - table = val_info.realloc_or_die(table, new_size); - } - - void resize_table(size_type old_size, size_type new_size, base::false_type) { - val_info.deallocate(table, old_size); - table = val_info.allocate(new_size); - } - - // Used to actually do the rehashing when we grow/shrink a hashtable - void copy_from(const dense_hashtable &ht, size_type min_buckets_wanted) { - clear_to_size(settings.min_buckets(ht.size(), min_buckets_wanted)); - - // We use a normal iterator to get non-deleted bcks from ht - // We could use insert() here, but since we know there are - // no duplicates and no deleted items, we can be more efficient - assert((bucket_count() & (bucket_count()-1)) == 0); // a power of two - for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) { - size_type num_probes = 0; // how many times we've probed - size_type bucknum; - const size_type bucket_count_minus_one = bucket_count() - 1; - for (bucknum = hash(get_key(*it)) & bucket_count_minus_one; - !test_empty(bucknum); // not empty - bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) { - ++num_probes; - assert(num_probes < bucket_count() - && "Hashtable is full: an error in key_equal<> or hash<>"); - } - set_value(&table[bucknum], *it); // copies the value to here - num_elements++; - } - settings.inc_num_ht_copies(); - } - - // Required by the spec for hashed associative container - public: - // Though the docs say this should be num_buckets, I think it's much - // more useful as num_elements. As a special feature, calling with - // req_elements==0 will cause us to shrink if we can, saving space. - void resize(size_type req_elements) { // resize to this or larger - if ( settings.consider_shrink() || req_elements == 0 ) - maybe_shrink(); - if ( req_elements > num_elements ) - resize_delta(req_elements - num_elements); - } - - // Get and change the value of shrink_factor and enlarge_factor. The - // description at the beginning of this file explains how to choose - // the values. Setting the shrink parameter to 0.0 ensures that the - // table never shrinks. - void get_resizing_parameters(float* shrink, float* grow) const { - *shrink = settings.shrink_factor(); - *grow = settings.enlarge_factor(); - } - void set_resizing_parameters(float shrink, float grow) { - settings.set_resizing_parameters(shrink, grow); - settings.reset_thresholds(bucket_count()); - } - - // CONSTRUCTORS -- as required by the specs, we take a size, - // but also let you specify a hashfunction, key comparator, - // and key extractor. We also define a copy constructor and =. - // DESTRUCTOR -- needs to free the table - explicit dense_hashtable(size_type expected_max_items_in_table = 0, - const HashFcn& hf = HashFcn(), - const EqualKey& eql = EqualKey(), - const ExtractKey& ext = ExtractKey(), - const SetKey& set = SetKey(), - const Alloc& alloc = Alloc()) - : settings(hf), - key_info(ext, set, eql), - num_deleted(0), - num_elements(0), - num_buckets(expected_max_items_in_table == 0 - ? HT_DEFAULT_STARTING_BUCKETS - : settings.min_buckets(expected_max_items_in_table, 0)), - val_info(alloc_impl<value_alloc_type>(alloc)), - table(NULL) { - // table is NULL until emptyval is set. However, we set num_buckets - // here so we know how much space to allocate once emptyval is set - settings.reset_thresholds(bucket_count()); - } - - // As a convenience for resize(), we allow an optional second argument - // which lets you make this new hashtable a different size than ht - dense_hashtable(const dense_hashtable& ht, - size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) - : settings(ht.settings), - key_info(ht.key_info), - num_deleted(0), - num_elements(0), - num_buckets(0), - val_info(ht.val_info), - table(NULL) { - if (!ht.settings.use_empty()) { - // If use_empty isn't set, copy_from will crash, so we do our own copying. - assert(ht.empty()); - num_buckets = settings.min_buckets(ht.size(), min_buckets_wanted); - settings.reset_thresholds(bucket_count()); - return; - } - settings.reset_thresholds(bucket_count()); - copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries - } - - dense_hashtable& operator= (const dense_hashtable& ht) { - if (&ht == this) return *this; // don't copy onto ourselves - if (!ht.settings.use_empty()) { - assert(ht.empty()); - dense_hashtable empty_table(ht); // empty table with ht's thresholds - this->swap(empty_table); - return *this; - } - settings = ht.settings; - key_info = ht.key_info; - set_value(&val_info.emptyval, ht.val_info.emptyval); - // copy_from() calls clear and sets num_deleted to 0 too - copy_from(ht, HT_MIN_BUCKETS); - // we purposefully don't copy the allocator, which may not be copyable - return *this; - } - - ~dense_hashtable() { - if (table) { - destroy_buckets(0, num_buckets); - val_info.deallocate(table, num_buckets); - } - } - - // Many STL algorithms use swap instead of copy constructors - void swap(dense_hashtable& ht) { - std::swap(settings, ht.settings); - std::swap(key_info, ht.key_info); - std::swap(num_deleted, ht.num_deleted); - std::swap(num_elements, ht.num_elements); - std::swap(num_buckets, ht.num_buckets); - { value_type tmp; // for annoying reasons, swap() doesn't work - set_value(&tmp, val_info.emptyval); - set_value(&val_info.emptyval, ht.val_info.emptyval); - set_value(&ht.val_info.emptyval, tmp); - } - std::swap(table, ht.table); - settings.reset_thresholds(bucket_count()); // also resets consider_shrink - ht.settings.reset_thresholds(ht.bucket_count()); - // we purposefully don't swap the allocator, which may not be swap-able - } - - private: - void clear_to_size(size_type new_num_buckets) { - if (!table) { - table = val_info.allocate(new_num_buckets); - } else { - destroy_buckets(0, num_buckets); - if (new_num_buckets != num_buckets) { // resize, if necessary - typedef base::integral_constant<bool, - base::is_same<value_alloc_type, - libc_allocator_with_realloc<value_type> >::value> - realloc_ok; - resize_table(num_buckets, new_num_buckets, realloc_ok()); - } - } - assert(table); - fill_range_with_empty(table, table + new_num_buckets); - num_elements = 0; - num_deleted = 0; - num_buckets = new_num_buckets; // our new size - settings.reset_thresholds(bucket_count()); - } - - public: - // It's always nice to be able to clear a table without deallocating it - void clear() { - // If the table is already empty, and the number of buckets is - // already as we desire, there's nothing to do. - const size_type new_num_buckets = settings.min_buckets(0, 0); - if (num_elements == 0 && new_num_buckets == num_buckets) { - return; - } - clear_to_size(new_num_buckets); - } - - // Clear the table without resizing it. - // Mimicks the stl_hashtable's behaviour when clear()-ing in that it - // does not modify the bucket count - void clear_no_resize() { - if (num_elements > 0) { - assert(table); - destroy_buckets(0, num_buckets); - fill_range_with_empty(table, table + num_buckets); - } - // don't consider to shrink before another erase() - settings.reset_thresholds(bucket_count()); - num_elements = 0; - num_deleted = 0; - } - - // LOOKUP ROUTINES - private: - // Returns a pair of positions: 1st where the object is, 2nd where - // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET - // if object is not found; 2nd is ILLEGAL_BUCKET if it is. - // Note: because of deletions where-to-insert is not trivial: it's the - // first deleted bucket we see, as long as we don't find the key later - std::pair<size_type, size_type> find_position(const key_type &key) const { - size_type num_probes = 0; // how many times we've probed - const size_type bucket_count_minus_one = bucket_count() - 1; - size_type bucknum = hash(key) & bucket_count_minus_one; - size_type insert_pos = ILLEGAL_BUCKET; // where we would insert - while ( 1 ) { // probe until something happens - if ( test_empty(bucknum) ) { // bucket is empty - if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert - return std::pair<size_type,size_type>(ILLEGAL_BUCKET, bucknum); - else - return std::pair<size_type,size_type>(ILLEGAL_BUCKET, insert_pos); - - } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert - if ( insert_pos == ILLEGAL_BUCKET ) - insert_pos = bucknum; - - } else if ( equals(key, get_key(table[bucknum])) ) { - return std::pair<size_type,size_type>(bucknum, ILLEGAL_BUCKET); - } - ++num_probes; // we're doing another probe - bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one; - assert(num_probes < bucket_count() - && "Hashtable is full: an error in key_equal<> or hash<>"); - } - } - - public: - - iterator find(const key_type& key) { - if ( size() == 0 ) return end(); - std::pair<size_type, size_type> pos = find_position(key); - if ( pos.first == ILLEGAL_BUCKET ) // alas, not there - return end(); - else - return iterator(this, table + pos.first, table + num_buckets, false); - } - - const_iterator find(const key_type& key) const { - if ( size() == 0 ) return end(); - std::pair<size_type, size_type> pos = find_position(key); - if ( pos.first == ILLEGAL_BUCKET ) // alas, not there - return end(); - else - return const_iterator(this, table + pos.first, table+num_buckets, false); - } - - // This is a tr1 method: the bucket a given key is in, or what bucket - // it would be put in, if it were to be inserted. Shrug. - size_type bucket(const key_type& key) const { - std::pair<size_type, size_type> pos = find_position(key); - return pos.first == ILLEGAL_BUCKET ? pos.second : pos.first; - } - - // Counts how many elements have key key. For maps, it's either 0 or 1. - size_type count(const key_type &key) const { - std::pair<size_type, size_type> pos = find_position(key); - return pos.first == ILLEGAL_BUCKET ? 0 : 1; - } - - // Likewise, equal_range doesn't really make sense for us. Oh well. - std::pair<iterator,iterator> equal_range(const key_type& key) { - iterator pos = find(key); // either an iterator or end - if (pos == end()) { - return std::pair<iterator,iterator>(pos, pos); - } else { - const iterator startpos = pos++; - return std::pair<iterator,iterator>(startpos, pos); - } - } - std::pair<const_iterator,const_iterator> equal_range(const key_type& key) - const { - const_iterator pos = find(key); // either an iterator or end - if (pos == end()) { - return std::pair<const_iterator,const_iterator>(pos, pos); - } else { - const const_iterator startpos = pos++; - return std::pair<const_iterator,const_iterator>(startpos, pos); - } - } - - - // INSERTION ROUTINES - private: - // Private method used by insert_noresize and find_or_insert. - iterator insert_at(const_reference obj, size_type pos) { - if (size() >= max_size()) { - throw std::length_error("insert overflow"); - } - if ( test_deleted(pos) ) { // just replace if it's been del. - // shrug: shouldn't need to be const. - const_iterator delpos(this, table + pos, table + num_buckets, false); - clear_deleted(delpos); - assert( num_deleted > 0); - --num_deleted; // used to be, now it isn't - } else { - ++num_elements; // replacing an empty bucket - } - set_value(&table[pos], obj); - return iterator(this, table + pos, table + num_buckets, false); - } - - // If you know *this is big enough to hold obj, use this routine - std::pair<iterator, bool> insert_noresize(const_reference obj) { - // First, double-check we're not inserting delkey or emptyval - assert((!settings.use_empty() || !equals(get_key(obj), - get_key(val_info.emptyval))) - && "Inserting the empty key"); - assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey)) - && "Inserting the deleted key"); - const std::pair<size_type,size_type> pos = find_position(get_key(obj)); - if ( pos.first != ILLEGAL_BUCKET) { // object was already there - return std::pair<iterator,bool>(iterator(this, table + pos.first, - table + num_buckets, false), - false); // false: we didn't insert - } else { // pos.second says where to put it - return std::pair<iterator,bool>(insert_at(obj, pos.second), true); - } - } - - // Specializations of insert(it, it) depending on the power of the iterator: - // (1) Iterator supports operator-, resize before inserting - template <class ForwardIterator> - void insert(ForwardIterator f, ForwardIterator l, std::forward_iterator_tag) { - size_t dist = std::distance(f, l); - if (dist >= (std::numeric_limits<size_type>::max)()) { - throw std::length_error("insert-range overflow"); - } - resize_delta(static_cast<size_type>(dist)); - for ( ; dist > 0; --dist, ++f) { - insert_noresize(*f); - } - } - - // (2) Arbitrary iterator, can't tell how much to resize - template <class InputIterator> - void insert(InputIterator f, InputIterator l, std::input_iterator_tag) { - for ( ; f != l; ++f) - insert(*f); - } - - public: - // This is the normal insert routine, used by the outside world - std::pair<iterator, bool> insert(const_reference obj) { - resize_delta(1); // adding an object, grow if need be - return insert_noresize(obj); - } - - // When inserting a lot at a time, we specialize on the type of iterator - template <class InputIterator> - void insert(InputIterator f, InputIterator l) { - // specializes on iterator type - insert(f, l, - typename std::iterator_traits<InputIterator>::iterator_category()); - } - - // DefaultValue is a functor that takes a key and returns a value_type - // representing the default value to be inserted if none is found. - template <class DefaultValue> - value_type& find_or_insert(const key_type& key) { - // First, double-check we're not inserting emptykey or delkey - assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) - && "Inserting the empty key"); - assert((!settings.use_deleted() || !equals(key, key_info.delkey)) - && "Inserting the deleted key"); - const std::pair<size_type,size_type> pos = find_position(key); - DefaultValue default_value; - if ( pos.first != ILLEGAL_BUCKET) { // object was already there - return table[pos.first]; - } else if (resize_delta(1)) { // needed to rehash to make room - // Since we resized, we can't use pos, so recalculate where to insert. - return *insert_noresize(default_value(key)).first; - } else { // no need to rehash, insert right here - return *insert_at(default_value(key), pos.second); - } - } - - - // DELETION ROUTINES - size_type erase(const key_type& key) { - // First, double-check we're not trying to erase delkey or emptyval. - assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) - && "Erasing the empty key"); - assert((!settings.use_deleted() || !equals(key, key_info.delkey)) - && "Erasing the deleted key"); - const_iterator pos = find(key); // shrug: shouldn't need to be const - if ( pos != end() ) { - assert(!test_deleted(pos)); // or find() shouldn't have returned it - set_deleted(pos); - ++num_deleted; - settings.set_consider_shrink(true); // will think about shrink after next insert - return 1; // because we deleted one thing - } else { - return 0; // because we deleted nothing - } - } - - // We return the iterator past the deleted item. - void erase(iterator pos) { - if ( pos == end() ) return; // sanity check - if ( set_deleted(pos) ) { // true if object has been newly deleted - ++num_deleted; - settings.set_consider_shrink(true); // will think about shrink after next insert - } - } - - void erase(iterator f, iterator l) { - for ( ; f != l; ++f) { - if ( set_deleted(f) ) // should always be true - ++num_deleted; - } - settings.set_consider_shrink(true); // will think about shrink after next insert - } - - // We allow you to erase a const_iterator just like we allow you to - // erase an iterator. This is in parallel to 'delete': you can delete - // a const pointer just like a non-const pointer. The logic is that - // you can't use the object after it's erased anyway, so it doesn't matter - // if it's const or not. - void erase(const_iterator pos) { - if ( pos == end() ) return; // sanity check - if ( set_deleted(pos) ) { // true if object has been newly deleted - ++num_deleted; - settings.set_consider_shrink(true); // will think about shrink after next insert - } - } - void erase(const_iterator f, const_iterator l) { - for ( ; f != l; ++f) { - if ( set_deleted(f) ) // should always be true - ++num_deleted; - } - settings.set_consider_shrink(true); // will think about shrink after next insert - } - - - // COMPARISON - bool operator==(const dense_hashtable& ht) const { - if (size() != ht.size()) { - return false; - } else if (this == &ht) { - return true; - } else { - // Iterate through the elements in "this" and see if the - // corresponding element is in ht - for ( const_iterator it = begin(); it != end(); ++it ) { - const_iterator it2 = ht.find(get_key(*it)); - if ((it2 == ht.end()) || (*it != *it2)) { - return false; - } - } - return true; - } - } - bool operator!=(const dense_hashtable& ht) const { - return !(*this == ht); - } - - - // I/O - // We support reading and writing hashtables to disk. Alas, since - // I don't know how to write a hasher or key_equal, you have to make - // sure everything but the table is the same. We compact before writing. - private: - // Every time the disk format changes, this should probably change too - typedef unsigned long MagicNumberType; - static const MagicNumberType MAGIC_NUMBER = 0x13578642; - - public: - // I/O -- this is an add-on for writing hash table to disk - // - // INPUT and OUTPUT must be either a FILE, *or* a C++ stream - // (istream, ostream, etc) *or* a class providing - // Read(void*, size_t) and Write(const void*, size_t) - // (respectively), which writes a buffer into a stream - // (which the INPUT/OUTPUT instance presumably owns). - - typedef sparsehash_internal::pod_serializer<value_type> NopointerSerializer; - - // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT *fp) { - squash_deleted(); // so we don't have to worry about delkey - if ( !sparsehash_internal::write_bigendian_number(fp, MAGIC_NUMBER, 4) ) - return false; - if ( !sparsehash_internal::write_bigendian_number(fp, num_buckets, 8) ) - return false; - if ( !sparsehash_internal::write_bigendian_number(fp, num_elements, 8) ) - return false; - // Now write a bitmap of non-empty buckets. - for ( size_type i = 0; i < num_buckets; i += 8 ) { - unsigned char bits = 0; - for ( int bit = 0; bit < 8; ++bit ) { - if ( i + bit < num_buckets && !test_empty(i + bit) ) - bits |= (1 << bit); - } - if ( !sparsehash_internal::write_data(fp, &bits, sizeof(bits)) ) - return false; - for ( int bit = 0; bit < 8; ++bit ) { - if ( bits & (1 << bit) ) { - if ( !serializer(fp, table[i + bit]) ) return false; - } - } - } - return true; - } - - // INPUT: anything we've written an overload of read_data() for. - // ValueSerializer: a functor. operator()(INPUT*, value_type*) - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT *fp) { - assert(settings.use_empty() && "empty_key not set for read"); - - clear(); // just to be consistent - MagicNumberType magic_read; - if ( !sparsehash_internal::read_bigendian_number(fp, &magic_read, 4) ) - return false; - if ( magic_read != MAGIC_NUMBER ) { - return false; - } - size_type new_num_buckets; - if ( !sparsehash_internal::read_bigendian_number(fp, &new_num_buckets, 8) ) - return false; - clear_to_size(new_num_buckets); - if ( !sparsehash_internal::read_bigendian_number(fp, &num_elements, 8) ) - return false; - - // Read the bitmap of non-empty buckets. - for (size_type i = 0; i < num_buckets; i += 8) { - unsigned char bits; - if ( !sparsehash_internal::read_data(fp, &bits, sizeof(bits)) ) - return false; - for ( int bit = 0; bit < 8; ++bit ) { - if ( i + bit < num_buckets && (bits & (1 << bit)) ) { // not empty - if ( !serializer(fp, &table[i + bit]) ) return false; - } - } - } - return true; - } - - private: - template <class A> - class alloc_impl : public A { - public: - typedef typename A::pointer pointer; - typedef typename A::size_type size_type; - - // Convert a normal allocator to one that has realloc_or_die() - alloc_impl(const A& a) : A(a) { } - - // realloc_or_die should only be used when using the default - // allocator (libc_allocator_with_realloc). - pointer realloc_or_die(pointer /*ptr*/, size_type /*n*/) { - fprintf(stderr, "realloc_or_die is only supported for " - "libc_allocator_with_realloc\n"); - exit(1); - return NULL; - } - }; - - // A template specialization of alloc_impl for - // libc_allocator_with_realloc that can handle realloc_or_die. - template <class A> - class alloc_impl<libc_allocator_with_realloc<A> > - : public libc_allocator_with_realloc<A> { - public: - typedef typename libc_allocator_with_realloc<A>::pointer pointer; - typedef typename libc_allocator_with_realloc<A>::size_type size_type; - - alloc_impl(const libc_allocator_with_realloc<A>& a) - : libc_allocator_with_realloc<A>(a) { } - - pointer realloc_or_die(pointer ptr, size_type n) { - pointer retval = this->reallocate(ptr, n); - if (retval == NULL) { + if (resize_to < needed_size && // may double resize_to + resize_to < (std::numeric_limits<size_type>::max)() / 2) { + // This situation means that we have enough deleted elements, + // that once we purge them, we won't actually have needed to + // grow. But we may want to grow anyway: if we just purge one + // element, say, we'll have to grow anyway next time we + // insert. Might as well grow now, since we're already going + // through the trouble of copying (in order to purge the + // deleted elements). + const size_type target = + static_cast<size_type>(settings.shrink_size(resize_to*2)); + if (num_elements - num_deleted + delta >= target) { + // Good, we won't be below the shrink threshhold even if we double. + resize_to *= 2; + } + } + dense_hashtable tmp(*this, resize_to); + swap(tmp); // now we are tmp + return true; + } + + // We require table be not-NULL and empty before calling this. + void resize_table(size_type /*old_size*/, size_type new_size, + base::true_type) { + table = val_info.realloc_or_die(table, new_size); + } + + void resize_table(size_type old_size, size_type new_size, base::false_type) { + val_info.deallocate(table, old_size); + table = val_info.allocate(new_size); + } + + // Used to actually do the rehashing when we grow/shrink a hashtable + void copy_from(const dense_hashtable &ht, size_type min_buckets_wanted) { + clear_to_size(settings.min_buckets(ht.size(), min_buckets_wanted)); + + // We use a normal iterator to get non-deleted bcks from ht + // We could use insert() here, but since we know there are + // no duplicates and no deleted items, we can be more efficient + assert((bucket_count() & (bucket_count()-1)) == 0); // a power of two + for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) { + size_type num_probes = 0; // how many times we've probed + size_type bucknum; + const size_type bucket_count_minus_one = bucket_count() - 1; + for (bucknum = hash(get_key(*it)) & bucket_count_minus_one; + !test_empty(bucknum); // not empty + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) { + ++num_probes; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + set_value(&table[bucknum], *it); // copies the value to here + num_elements++; + } + settings.inc_num_ht_copies(); + } + + // Required by the spec for hashed associative container + public: + // Though the docs say this should be num_buckets, I think it's much + // more useful as num_elements. As a special feature, calling with + // req_elements==0 will cause us to shrink if we can, saving space. + void resize(size_type req_elements) { // resize to this or larger + if ( settings.consider_shrink() || req_elements == 0 ) + maybe_shrink(); + if ( req_elements > num_elements ) + resize_delta(req_elements - num_elements); + } + + // Get and change the value of shrink_factor and enlarge_factor. The + // description at the beginning of this file explains how to choose + // the values. Setting the shrink parameter to 0.0 ensures that the + // table never shrinks. + void get_resizing_parameters(float* shrink, float* grow) const { + *shrink = settings.shrink_factor(); + *grow = settings.enlarge_factor(); + } + void set_resizing_parameters(float shrink, float grow) { + settings.set_resizing_parameters(shrink, grow); + settings.reset_thresholds(bucket_count()); + } + + // CONSTRUCTORS -- as required by the specs, we take a size, + // but also let you specify a hashfunction, key comparator, + // and key extractor. We also define a copy constructor and =. + // DESTRUCTOR -- needs to free the table + explicit dense_hashtable(size_type expected_max_items_in_table = 0, + const HashFcn& hf = HashFcn(), + const EqualKey& eql = EqualKey(), + const ExtractKey& ext = ExtractKey(), + const SetKey& set = SetKey(), + const Alloc& alloc = Alloc()) + : settings(hf), + key_info(ext, set, eql), + num_deleted(0), + num_elements(0), + num_buckets(expected_max_items_in_table == 0 + ? HT_DEFAULT_STARTING_BUCKETS + : settings.min_buckets(expected_max_items_in_table, 0)), + val_info(alloc_impl<value_alloc_type>(alloc)), + table(NULL) { + // table is NULL until emptyval is set. However, we set num_buckets + // here so we know how much space to allocate once emptyval is set + settings.reset_thresholds(bucket_count()); + } + + // As a convenience for resize(), we allow an optional second argument + // which lets you make this new hashtable a different size than ht + dense_hashtable(const dense_hashtable& ht, + size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) + : settings(ht.settings), + key_info(ht.key_info), + num_deleted(0), + num_elements(0), + num_buckets(0), + val_info(ht.val_info), + table(NULL) { + if (!ht.settings.use_empty()) { + // If use_empty isn't set, copy_from will crash, so we do our own copying. + assert(ht.empty()); + num_buckets = settings.min_buckets(ht.size(), min_buckets_wanted); + settings.reset_thresholds(bucket_count()); + return; + } + settings.reset_thresholds(bucket_count()); + copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries + } + + dense_hashtable& operator= (const dense_hashtable& ht) { + if (&ht == this) return *this; // don't copy onto ourselves + if (!ht.settings.use_empty()) { + assert(ht.empty()); + dense_hashtable empty_table(ht); // empty table with ht's thresholds + this->swap(empty_table); + return *this; + } + settings = ht.settings; + key_info = ht.key_info; + set_value(&val_info.emptyval, ht.val_info.emptyval); + // copy_from() calls clear and sets num_deleted to 0 too + copy_from(ht, HT_MIN_BUCKETS); + // we purposefully don't copy the allocator, which may not be copyable + return *this; + } + + ~dense_hashtable() { + if (table) { + destroy_buckets(0, num_buckets); + val_info.deallocate(table, num_buckets); + } + } + + // Many STL algorithms use swap instead of copy constructors + void swap(dense_hashtable& ht) { + std::swap(settings, ht.settings); + std::swap(key_info, ht.key_info); + std::swap(num_deleted, ht.num_deleted); + std::swap(num_elements, ht.num_elements); + std::swap(num_buckets, ht.num_buckets); + { value_type tmp; // for annoying reasons, swap() doesn't work + set_value(&tmp, val_info.emptyval); + set_value(&val_info.emptyval, ht.val_info.emptyval); + set_value(&ht.val_info.emptyval, tmp); + } + std::swap(table, ht.table); + settings.reset_thresholds(bucket_count()); // also resets consider_shrink + ht.settings.reset_thresholds(ht.bucket_count()); + // we purposefully don't swap the allocator, which may not be swap-able + } + + private: + void clear_to_size(size_type new_num_buckets) { + if (!table) { + table = val_info.allocate(new_num_buckets); + } else { + destroy_buckets(0, num_buckets); + if (new_num_buckets != num_buckets) { // resize, if necessary + typedef base::integral_constant<bool, + base::is_same<value_alloc_type, + libc_allocator_with_realloc<value_type> >::value> + realloc_ok; + resize_table(num_buckets, new_num_buckets, realloc_ok()); + } + } + assert(table); + fill_range_with_empty(table, table + new_num_buckets); + num_elements = 0; + num_deleted = 0; + num_buckets = new_num_buckets; // our new size + settings.reset_thresholds(bucket_count()); + } + + public: + // It's always nice to be able to clear a table without deallocating it + void clear() { + // If the table is already empty, and the number of buckets is + // already as we desire, there's nothing to do. + const size_type new_num_buckets = settings.min_buckets(0, 0); + if (num_elements == 0 && new_num_buckets == num_buckets) { + return; + } + clear_to_size(new_num_buckets); + } + + // Clear the table without resizing it. + // Mimicks the stl_hashtable's behaviour when clear()-ing in that it + // does not modify the bucket count + void clear_no_resize() { + if (num_elements > 0) { + assert(table); + destroy_buckets(0, num_buckets); + fill_range_with_empty(table, table + num_buckets); + } + // don't consider to shrink before another erase() + settings.reset_thresholds(bucket_count()); + num_elements = 0; + num_deleted = 0; + } + + // LOOKUP ROUTINES + private: + // Returns a pair of positions: 1st where the object is, 2nd where + // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET + // if object is not found; 2nd is ILLEGAL_BUCKET if it is. + // Note: because of deletions where-to-insert is not trivial: it's the + // first deleted bucket we see, as long as we don't find the key later + std::pair<size_type, size_type> find_position(const key_type &key) const { + size_type num_probes = 0; // how many times we've probed + const size_type bucket_count_minus_one = bucket_count() - 1; + size_type bucknum = hash(key) & bucket_count_minus_one; + size_type insert_pos = ILLEGAL_BUCKET; // where we would insert + while ( 1 ) { // probe until something happens + if ( test_empty(bucknum) ) { // bucket is empty + if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert + return std::pair<size_type,size_type>(ILLEGAL_BUCKET, bucknum); + else + return std::pair<size_type,size_type>(ILLEGAL_BUCKET, insert_pos); + + } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert + if ( insert_pos == ILLEGAL_BUCKET ) + insert_pos = bucknum; + + } else if ( equals(key, get_key(table[bucknum])) ) { + return std::pair<size_type,size_type>(bucknum, ILLEGAL_BUCKET); + } + ++num_probes; // we're doing another probe + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + } + + public: + + iterator find(const key_type& key) { + if ( size() == 0 ) return end(); + std::pair<size_type, size_type> pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return iterator(this, table + pos.first, table + num_buckets, false); + } + + const_iterator find(const key_type& key) const { + if ( size() == 0 ) return end(); + std::pair<size_type, size_type> pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return const_iterator(this, table + pos.first, table+num_buckets, false); + } + + // This is a tr1 method: the bucket a given key is in, or what bucket + // it would be put in, if it were to be inserted. Shrug. + size_type bucket(const key_type& key) const { + std::pair<size_type, size_type> pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? pos.second : pos.first; + } + + // Counts how many elements have key key. For maps, it's either 0 or 1. + size_type count(const key_type &key) const { + std::pair<size_type, size_type> pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? 0 : 1; + } + + // Likewise, equal_range doesn't really make sense for us. Oh well. + std::pair<iterator,iterator> equal_range(const key_type& key) { + iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair<iterator,iterator>(pos, pos); + } else { + const iterator startpos = pos++; + return std::pair<iterator,iterator>(startpos, pos); + } + } + std::pair<const_iterator,const_iterator> equal_range(const key_type& key) + const { + const_iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair<const_iterator,const_iterator>(pos, pos); + } else { + const const_iterator startpos = pos++; + return std::pair<const_iterator,const_iterator>(startpos, pos); + } + } + + + // INSERTION ROUTINES + private: + // Private method used by insert_noresize and find_or_insert. + iterator insert_at(const_reference obj, size_type pos) { + if (size() >= max_size()) { + throw std::length_error("insert overflow"); + } + if ( test_deleted(pos) ) { // just replace if it's been del. + // shrug: shouldn't need to be const. + const_iterator delpos(this, table + pos, table + num_buckets, false); + clear_deleted(delpos); + assert( num_deleted > 0); + --num_deleted; // used to be, now it isn't + } else { + ++num_elements; // replacing an empty bucket + } + set_value(&table[pos], obj); + return iterator(this, table + pos, table + num_buckets, false); + } + + // If you know *this is big enough to hold obj, use this routine + std::pair<iterator, bool> insert_noresize(const_reference obj) { + // First, double-check we're not inserting delkey or emptyval + assert((!settings.use_empty() || !equals(get_key(obj), + get_key(val_info.emptyval))) + && "Inserting the empty key"); + assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey)) + && "Inserting the deleted key"); + const std::pair<size_type,size_type> pos = find_position(get_key(obj)); + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return std::pair<iterator,bool>(iterator(this, table + pos.first, + table + num_buckets, false), + false); // false: we didn't insert + } else { // pos.second says where to put it + return std::pair<iterator,bool>(insert_at(obj, pos.second), true); + } + } + + // Specializations of insert(it, it) depending on the power of the iterator: + // (1) Iterator supports operator-, resize before inserting + template <class ForwardIterator> + void insert(ForwardIterator f, ForwardIterator l, std::forward_iterator_tag) { + size_t dist = std::distance(f, l); + if (dist >= (std::numeric_limits<size_type>::max)()) { + throw std::length_error("insert-range overflow"); + } + resize_delta(static_cast<size_type>(dist)); + for ( ; dist > 0; --dist, ++f) { + insert_noresize(*f); + } + } + + // (2) Arbitrary iterator, can't tell how much to resize + template <class InputIterator> + void insert(InputIterator f, InputIterator l, std::input_iterator_tag) { + for ( ; f != l; ++f) + insert(*f); + } + + public: + // This is the normal insert routine, used by the outside world + std::pair<iterator, bool> insert(const_reference obj) { + resize_delta(1); // adding an object, grow if need be + return insert_noresize(obj); + } + + // When inserting a lot at a time, we specialize on the type of iterator + template <class InputIterator> + void insert(InputIterator f, InputIterator l) { + // specializes on iterator type + insert(f, l, + typename std::iterator_traits<InputIterator>::iterator_category()); + } + + // DefaultValue is a functor that takes a key and returns a value_type + // representing the default value to be inserted if none is found. + template <class DefaultValue> + value_type& find_or_insert(const key_type& key) { + // First, double-check we're not inserting emptykey or delkey + assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) + && "Inserting the empty key"); + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Inserting the deleted key"); + const std::pair<size_type,size_type> pos = find_position(key); + DefaultValue default_value; + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return table[pos.first]; + } else if (resize_delta(1)) { // needed to rehash to make room + // Since we resized, we can't use pos, so recalculate where to insert. + return *insert_noresize(default_value(key)).first; + } else { // no need to rehash, insert right here + return *insert_at(default_value(key), pos.second); + } + } + + + // DELETION ROUTINES + size_type erase(const key_type& key) { + // First, double-check we're not trying to erase delkey or emptyval. + assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) + && "Erasing the empty key"); + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Erasing the deleted key"); + const_iterator pos = find(key); // shrug: shouldn't need to be const + if ( pos != end() ) { + assert(!test_deleted(pos)); // or find() shouldn't have returned it + set_deleted(pos); + ++num_deleted; + settings.set_consider_shrink(true); // will think about shrink after next insert + return 1; // because we deleted one thing + } else { + return 0; // because we deleted nothing + } + } + + // We return the iterator past the deleted item. + void erase(iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + settings.set_consider_shrink(true); // will think about shrink after next insert + } + } + + void erase(iterator f, iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + settings.set_consider_shrink(true); // will think about shrink after next insert + } + + // We allow you to erase a const_iterator just like we allow you to + // erase an iterator. This is in parallel to 'delete': you can delete + // a const pointer just like a non-const pointer. The logic is that + // you can't use the object after it's erased anyway, so it doesn't matter + // if it's const or not. + void erase(const_iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + settings.set_consider_shrink(true); // will think about shrink after next insert + } + } + void erase(const_iterator f, const_iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + settings.set_consider_shrink(true); // will think about shrink after next insert + } + + + // COMPARISON + bool operator==(const dense_hashtable& ht) const { + if (size() != ht.size()) { + return false; + } else if (this == &ht) { + return true; + } else { + // Iterate through the elements in "this" and see if the + // corresponding element is in ht + for ( const_iterator it = begin(); it != end(); ++it ) { + const_iterator it2 = ht.find(get_key(*it)); + if ((it2 == ht.end()) || (*it != *it2)) { + return false; + } + } + return true; + } + } + bool operator!=(const dense_hashtable& ht) const { + return !(*this == ht); + } + + + // I/O + // We support reading and writing hashtables to disk. Alas, since + // I don't know how to write a hasher or key_equal, you have to make + // sure everything but the table is the same. We compact before writing. + private: + // Every time the disk format changes, this should probably change too + typedef unsigned long MagicNumberType; + static const MagicNumberType MAGIC_NUMBER = 0x13578642; + + public: + // I/O -- this is an add-on for writing hash table to disk + // + // INPUT and OUTPUT must be either a FILE, *or* a C++ stream + // (istream, ostream, etc) *or* a class providing + // Read(void*, size_t) and Write(const void*, size_t) + // (respectively), which writes a buffer into a stream + // (which the INPUT/OUTPUT instance presumably owns). + + typedef sparsehash_internal::pod_serializer<value_type> NopointerSerializer; + + // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT *fp) { + squash_deleted(); // so we don't have to worry about delkey + if ( !sparsehash_internal::write_bigendian_number(fp, MAGIC_NUMBER, 4) ) + return false; + if ( !sparsehash_internal::write_bigendian_number(fp, num_buckets, 8) ) + return false; + if ( !sparsehash_internal::write_bigendian_number(fp, num_elements, 8) ) + return false; + // Now write a bitmap of non-empty buckets. + for ( size_type i = 0; i < num_buckets; i += 8 ) { + unsigned char bits = 0; + for ( int bit = 0; bit < 8; ++bit ) { + if ( i + bit < num_buckets && !test_empty(i + bit) ) + bits |= (1 << bit); + } + if ( !sparsehash_internal::write_data(fp, &bits, sizeof(bits)) ) + return false; + for ( int bit = 0; bit < 8; ++bit ) { + if ( bits & (1 << bit) ) { + if ( !serializer(fp, table[i + bit]) ) return false; + } + } + } + return true; + } + + // INPUT: anything we've written an overload of read_data() for. + // ValueSerializer: a functor. operator()(INPUT*, value_type*) + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT *fp) { + assert(settings.use_empty() && "empty_key not set for read"); + + clear(); // just to be consistent + MagicNumberType magic_read; + if ( !sparsehash_internal::read_bigendian_number(fp, &magic_read, 4) ) + return false; + if ( magic_read != MAGIC_NUMBER ) { + return false; + } + size_type new_num_buckets; + if ( !sparsehash_internal::read_bigendian_number(fp, &new_num_buckets, 8) ) + return false; + clear_to_size(new_num_buckets); + if ( !sparsehash_internal::read_bigendian_number(fp, &num_elements, 8) ) + return false; + + // Read the bitmap of non-empty buckets. + for (size_type i = 0; i < num_buckets; i += 8) { + unsigned char bits; + if ( !sparsehash_internal::read_data(fp, &bits, sizeof(bits)) ) + return false; + for ( int bit = 0; bit < 8; ++bit ) { + if ( i + bit < num_buckets && (bits & (1 << bit)) ) { // not empty + if ( !serializer(fp, &table[i + bit]) ) return false; + } + } + } + return true; + } + + private: + template <class A> + class alloc_impl : public A { + public: + typedef typename A::pointer pointer; + typedef typename A::size_type size_type; + + // Convert a normal allocator to one that has realloc_or_die() + alloc_impl(const A& a) : A(a) { } + + // realloc_or_die should only be used when using the default + // allocator (libc_allocator_with_realloc). + pointer realloc_or_die(pointer /*ptr*/, size_type /*n*/) { + fprintf(stderr, "realloc_or_die is only supported for " + "libc_allocator_with_realloc\n"); + exit(1); + return NULL; + } + }; + + // A template specialization of alloc_impl for + // libc_allocator_with_realloc that can handle realloc_or_die. + template <class A> + class alloc_impl<libc_allocator_with_realloc<A> > + : public libc_allocator_with_realloc<A> { + public: + typedef typename libc_allocator_with_realloc<A>::pointer pointer; + typedef typename libc_allocator_with_realloc<A>::size_type size_type; + + alloc_impl(const libc_allocator_with_realloc<A>& a) + : libc_allocator_with_realloc<A>(a) { } + + pointer realloc_or_die(pointer ptr, size_type n) { + pointer retval = this->reallocate(ptr, n); + if (retval == NULL) { fprintf(stderr, "sparsehash: FATAL ERROR: failed to reallocate " "%lu elements for ptr %p", static_cast<unsigned long>(n), static_cast<void*>(ptr)); - exit(1); - } - return retval; - } - }; - - // Package allocator with emptyval to eliminate memory needed for - // the zero-size allocator. - // If new fields are added to this class, we should add them to - // operator= and swap. - class ValInfo : public alloc_impl<value_alloc_type> { - public: - typedef typename alloc_impl<value_alloc_type>::value_type value_type; - - ValInfo(const alloc_impl<value_alloc_type>& a) - : alloc_impl<value_alloc_type>(a), emptyval() { } - ValInfo(const ValInfo& v) - : alloc_impl<value_alloc_type>(v), emptyval(v.emptyval) { } - - value_type emptyval; // which key marks unused entries - }; - - - // Package functors with another class to eliminate memory needed for - // zero-size functors. Since ExtractKey and hasher's operator() might - // have the same function signature, they must be packaged in - // different classes. - struct Settings : - sparsehash_internal::sh_hashtable_settings<key_type, hasher, - size_type, HT_MIN_BUCKETS> { - explicit Settings(const hasher& hf) - : sparsehash_internal::sh_hashtable_settings<key_type, hasher, - size_type, HT_MIN_BUCKETS>( - hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {} - }; - - // Packages ExtractKey and SetKey functors. - class KeyInfo : public ExtractKey, public SetKey, public EqualKey { - public: - KeyInfo(const ExtractKey& ek, const SetKey& sk, const EqualKey& eq) - : ExtractKey(ek), - SetKey(sk), - EqualKey(eq) { - } - - // We want to return the exact same type as ExtractKey: Key or const Key& - typename ExtractKey::result_type get_key(const_reference v) const { - return ExtractKey::operator()(v); - } - void set_key(pointer v, const key_type& k) const { - SetKey::operator()(v, k); - } - bool equals(const key_type& a, const key_type& b) const { - return EqualKey::operator()(a, b); - } - - // Which key marks deleted entries. - // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!) - typename base::remove_const<key_type>::type delkey; - }; - - // Utility functions to access the templated operators - size_type hash(const key_type& v) const { - return settings.hash(v); - } - bool equals(const key_type& a, const key_type& b) const { - return key_info.equals(a, b); - } - typename ExtractKey::result_type get_key(const_reference v) const { - return key_info.get_key(v); - } - void set_key(pointer v, const key_type& k) const { - key_info.set_key(v, k); - } - - private: - // Actual data - Settings settings; - KeyInfo key_info; - - size_type num_deleted; // how many occupied buckets are marked deleted - size_type num_elements; - size_type num_buckets; - ValInfo val_info; // holds emptyval, and also the allocator - pointer table; -}; - - -// We need a global swap as well -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -inline void swap(dense_hashtable<V,K,HF,ExK,SetK,EqK,A> &x, - dense_hashtable<V,K,HF,ExK,SetK,EqK,A> &y) { - x.swap(y); -} - -#undef JUMP_ - -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -const typename dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::size_type - dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET; - -// How full we let the table get before we resize. Knuth says .8 is -// good -- higher causes us to probe too much, though saves memory. -// However, we go with .5, getting better performance at the cost of -// more space (a trade-off densehashtable explicitly chooses to make). -// Feel free to play around with different values, though, via -// max_load_factor() and/or set_resizing_parameters(). -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -const int dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT = 50; - -// How empty we let the table get before we resize lower. -// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing. -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -const int dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_PCT - = static_cast<int>(0.4 * - dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT); - -_END_GOOGLE_NAMESPACE_ - -#endif /* _DENSEHASHTABLE_H_ */ + exit(1); + } + return retval; + } + }; + + // Package allocator with emptyval to eliminate memory needed for + // the zero-size allocator. + // If new fields are added to this class, we should add them to + // operator= and swap. + class ValInfo : public alloc_impl<value_alloc_type> { + public: + typedef typename alloc_impl<value_alloc_type>::value_type value_type; + + ValInfo(const alloc_impl<value_alloc_type>& a) + : alloc_impl<value_alloc_type>(a), emptyval() { } + ValInfo(const ValInfo& v) + : alloc_impl<value_alloc_type>(v), emptyval(v.emptyval) { } + + value_type emptyval; // which key marks unused entries + }; + + + // Package functors with another class to eliminate memory needed for + // zero-size functors. Since ExtractKey and hasher's operator() might + // have the same function signature, they must be packaged in + // different classes. + struct Settings : + sparsehash_internal::sh_hashtable_settings<key_type, hasher, + size_type, HT_MIN_BUCKETS> { + explicit Settings(const hasher& hf) + : sparsehash_internal::sh_hashtable_settings<key_type, hasher, + size_type, HT_MIN_BUCKETS>( + hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {} + }; + + // Packages ExtractKey and SetKey functors. + class KeyInfo : public ExtractKey, public SetKey, public EqualKey { + public: + KeyInfo(const ExtractKey& ek, const SetKey& sk, const EqualKey& eq) + : ExtractKey(ek), + SetKey(sk), + EqualKey(eq) { + } + + // We want to return the exact same type as ExtractKey: Key or const Key& + typename ExtractKey::result_type get_key(const_reference v) const { + return ExtractKey::operator()(v); + } + void set_key(pointer v, const key_type& k) const { + SetKey::operator()(v, k); + } + bool equals(const key_type& a, const key_type& b) const { + return EqualKey::operator()(a, b); + } + + // Which key marks deleted entries. + // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!) + typename base::remove_const<key_type>::type delkey; + }; + + // Utility functions to access the templated operators + size_type hash(const key_type& v) const { + return settings.hash(v); + } + bool equals(const key_type& a, const key_type& b) const { + return key_info.equals(a, b); + } + typename ExtractKey::result_type get_key(const_reference v) const { + return key_info.get_key(v); + } + void set_key(pointer v, const key_type& k) const { + key_info.set_key(v, k); + } + + private: + // Actual data + Settings settings; + KeyInfo key_info; + + size_type num_deleted; // how many occupied buckets are marked deleted + size_type num_elements; + size_type num_buckets; + ValInfo val_info; // holds emptyval, and also the allocator + pointer table; +}; + + +// We need a global swap as well +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +inline void swap(dense_hashtable<V,K,HF,ExK,SetK,EqK,A> &x, + dense_hashtable<V,K,HF,ExK,SetK,EqK,A> &y) { + x.swap(y); +} + +#undef JUMP_ + +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +const typename dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::size_type + dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET; + +// How full we let the table get before we resize. Knuth says .8 is +// good -- higher causes us to probe too much, though saves memory. +// However, we go with .5, getting better performance at the cost of +// more space (a trade-off densehashtable explicitly chooses to make). +// Feel free to play around with different values, though, via +// max_load_factor() and/or set_resizing_parameters(). +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +const int dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT = 50; + +// How empty we let the table get before we resize lower. +// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing. +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +const int dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_PCT + = static_cast<int>(0.4 * + dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT); + +_END_GOOGLE_NAMESPACE_ + +#endif /* _DENSEHASHTABLE_H_ */ diff --git a/contrib/libs/sparsehash/src/sparsehash/internal/hashtable-common.h b/contrib/libs/sparsehash/src/sparsehash/internal/hashtable-common.h index b22e853830..9e49ec890f 100644 --- a/contrib/libs/sparsehash/src/sparsehash/internal/hashtable-common.h +++ b/contrib/libs/sparsehash/src/sparsehash/internal/hashtable-common.h @@ -1,381 +1,381 @@ -// Copyright (c) 2010, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// Provides classes shared by both sparse and dense hashtable. -// -// sh_hashtable_settings has parameters for growing and shrinking -// a hashtable. It also packages zero-size functor (ie. hasher). -// -// Other functions and classes provide common code for serializing -// and deserializing hashtables to a stream (such as a FILE*). - -#ifndef UTIL_GTL_HASHTABLE_COMMON_H_ -#define UTIL_GTL_HASHTABLE_COMMON_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <assert.h> -#include <stdio.h> -#include <stddef.h> // for size_t -#include <iosfwd> -#include <stdexcept> // For length_error - -_START_GOOGLE_NAMESPACE_ - -template <bool> struct SparsehashCompileAssert { }; -#define SPARSEHASH_COMPILE_ASSERT(expr, msg) \ +// Copyright (c) 2010, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// Provides classes shared by both sparse and dense hashtable. +// +// sh_hashtable_settings has parameters for growing and shrinking +// a hashtable. It also packages zero-size functor (ie. hasher). +// +// Other functions and classes provide common code for serializing +// and deserializing hashtables to a stream (such as a FILE*). + +#ifndef UTIL_GTL_HASHTABLE_COMMON_H_ +#define UTIL_GTL_HASHTABLE_COMMON_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <assert.h> +#include <stdio.h> +#include <stddef.h> // for size_t +#include <iosfwd> +#include <stdexcept> // For length_error + +_START_GOOGLE_NAMESPACE_ + +template <bool> struct SparsehashCompileAssert { }; +#define SPARSEHASH_COMPILE_ASSERT(expr, msg) \ [[maybe_unused]] typedef SparsehashCompileAssert<(bool(expr))> msg[bool(expr) ? 1 : -1] - -namespace sparsehash_internal { - -// Adaptor methods for reading/writing data from an INPUT or OUPTUT -// variable passed to serialize() or unserialize(). For now we -// have implemented INPUT/OUTPUT for FILE*, istream*/ostream* (note -// they are pointers, unlike typical use), or else a pointer to -// something that supports a Read()/Write() method. -// -// For technical reasons, we implement read_data/write_data in two -// stages. The actual work is done in *_data_internal, which takes -// the stream argument twice: once as a template type, and once with -// normal type information. (We only use the second version.) We do -// this because of how C++ picks what function overload to use. If we -// implemented this the naive way: -// bool read_data(istream* is, const void* data, size_t length); -// template<typename T> read_data(T* fp, const void* data, size_t length); -// C++ would prefer the second version for every stream type except -// istream. However, we want C++ to prefer the first version for -// streams that are *subclasses* of istream, such as istringstream. -// This is not possible given the way template types are resolved. So -// we split the stream argument in two, one of which is templated and -// one of which is not. The specialized functions (like the istream -// version above) ignore the template arg and use the second, 'type' -// arg, getting subclass matching as normal. The 'catch-all' -// functions (the second version above) use the template arg to deduce -// the type, and use a second, void* arg to achieve the desired -// 'catch-all' semantics. - -// ----- low-level I/O for FILE* ---- - -template<typename Ignored> -inline bool read_data_internal(Ignored*, FILE* fp, - void* data, size_t length) { - return fread(data, length, 1, fp) == 1; -} - -template<typename Ignored> -inline bool write_data_internal(Ignored*, FILE* fp, - const void* data, size_t length) { - return fwrite(data, length, 1, fp) == 1; -} - -// ----- low-level I/O for iostream ---- - -// We want the caller to be responsible for #including <iostream>, not -// us, because iostream is a big header! According to the standard, -// it's only legal to delay the instantiation the way we want to if -// the istream/ostream is a template type. So we jump through hoops. -template<typename ISTREAM> -inline bool read_data_internal_for_istream(ISTREAM* fp, - void* data, size_t length) { - return fp->read(reinterpret_cast<char*>(data), length).good(); -} -template<typename Ignored> -inline bool read_data_internal(Ignored*, std::istream* fp, - void* data, size_t length) { - return read_data_internal_for_istream(fp, data, length); -} - -template<typename OSTREAM> -inline bool write_data_internal_for_ostream(OSTREAM* fp, - const void* data, size_t length) { - return fp->write(reinterpret_cast<const char*>(data), length).good(); -} -template<typename Ignored> -inline bool write_data_internal(Ignored*, std::ostream* fp, - const void* data, size_t length) { - return write_data_internal_for_ostream(fp, data, length); -} - -// ----- low-level I/O for custom streams ---- - -// The INPUT type needs to support a Read() method that takes a -// buffer and a length and returns the number of bytes read. -template <typename INPUT> -inline bool read_data_internal(INPUT* fp, void*, - void* data, size_t length) { - return static_cast<size_t>(fp->Read(data, length)) == length; -} - -// The OUTPUT type needs to support a Write() operation that takes -// a buffer and a length and returns the number of bytes written. -template <typename OUTPUT> -inline bool write_data_internal(OUTPUT* fp, void*, - const void* data, size_t length) { - return static_cast<size_t>(fp->Write(data, length)) == length; -} - -// ----- low-level I/O: the public API ---- - -template <typename INPUT> -inline bool read_data(INPUT* fp, void* data, size_t length) { - return read_data_internal(fp, fp, data, length); -} - -template <typename OUTPUT> -inline bool write_data(OUTPUT* fp, const void* data, size_t length) { - return write_data_internal(fp, fp, data, length); -} - -// Uses read_data() and write_data() to read/write an integer. -// length is the number of bytes to read/write (which may differ -// from sizeof(IntType), allowing us to save on a 32-bit system -// and load on a 64-bit system). Excess bytes are taken to be 0. -// INPUT and OUTPUT must match legal inputs to read/write_data (above). -template <typename INPUT, typename IntType> -bool read_bigendian_number(INPUT* fp, IntType* value, size_t length) { - *value = 0; - unsigned char byte; - // We require IntType to be unsigned or else the shifting gets all screwy. - SPARSEHASH_COMPILE_ASSERT(static_cast<IntType>(-1) > static_cast<IntType>(0), - serializing_int_requires_an_unsigned_type); - for (size_t i = 0; i < length; ++i) { - if (!read_data(fp, &byte, sizeof(byte))) return false; - *value |= static_cast<IntType>(byte) << ((length - 1 - i) * 8); - } - return true; -} - -template <typename OUTPUT, typename IntType> -bool write_bigendian_number(OUTPUT* fp, IntType value, size_t length) { - unsigned char byte; - // We require IntType to be unsigned or else the shifting gets all screwy. - SPARSEHASH_COMPILE_ASSERT(static_cast<IntType>(-1) > static_cast<IntType>(0), - serializing_int_requires_an_unsigned_type); - for (size_t i = 0; i < length; ++i) { - byte = (sizeof(value) <= length-1 - i) - ? 0 : static_cast<unsigned char>((value >> ((length-1 - i) * 8)) & 255); - if (!write_data(fp, &byte, sizeof(byte))) return false; - } - return true; -} - -// If your keys and values are simple enough, you can pass this -// serializer to serialize()/unserialize(). "Simple enough" means -// value_type is a POD type that contains no pointers. Note, -// however, we don't try to normalize endianness. -// This is the type used for NopointerSerializer. -template <typename value_type> struct pod_serializer { - template <typename INPUT> - bool operator()(INPUT* fp, value_type* value) const { - return read_data(fp, value, sizeof(*value)); - } - - template <typename OUTPUT> - bool operator()(OUTPUT* fp, const value_type& value) const { - return write_data(fp, &value, sizeof(value)); - } -}; - - -// Settings contains parameters for growing and shrinking the table. -// It also packages zero-size functor (ie. hasher). -// -// It does some munging of the hash value in cases where we think -// (fear) the original hash function might not be very good. In -// particular, the default hash of pointers is the identity hash, -// so probably all the low bits are 0. We identify when we think -// we're hashing a pointer, and chop off the low bits. Note this -// isn't perfect: even when the key is a pointer, we can't tell -// for sure that the hash is the identity hash. If it's not, this -// is needless work (and possibly, though not likely, harmful). - -template<typename Key, typename HashFunc, - typename SizeType, int HT_MIN_BUCKETS> -class sh_hashtable_settings : public HashFunc { - public: - typedef Key key_type; - typedef HashFunc hasher; - typedef SizeType size_type; - - public: - sh_hashtable_settings(const hasher& hf, - const float ht_occupancy_flt, - const float ht_empty_flt) - : hasher(hf), - enlarge_threshold_(0), - shrink_threshold_(0), - consider_shrink_(false), - use_empty_(false), - use_deleted_(false), - num_ht_copies_(0) { - set_enlarge_factor(ht_occupancy_flt); - set_shrink_factor(ht_empty_flt); - } - - size_type hash(const key_type& v) const { - // We munge the hash value when we don't trust hasher::operator(). - return hash_munger<Key>::MungedHash(hasher::operator()(v)); - } - - float enlarge_factor() const { - return enlarge_factor_; - } - void set_enlarge_factor(float f) { - enlarge_factor_ = f; - } - float shrink_factor() const { - return shrink_factor_; - } - void set_shrink_factor(float f) { - shrink_factor_ = f; - } - - size_type enlarge_threshold() const { - return enlarge_threshold_; - } - void set_enlarge_threshold(size_type t) { - enlarge_threshold_ = t; - } - size_type shrink_threshold() const { - return shrink_threshold_; - } - void set_shrink_threshold(size_type t) { - shrink_threshold_ = t; - } - - size_type enlarge_size(size_type x) const { - return static_cast<size_type>(x * enlarge_factor_); - } - size_type shrink_size(size_type x) const { - return static_cast<size_type>(x * shrink_factor_); - } - - bool consider_shrink() const { - return consider_shrink_; - } - void set_consider_shrink(bool t) { - consider_shrink_ = t; - } - - bool use_empty() const { - return use_empty_; - } - void set_use_empty(bool t) { - use_empty_ = t; - } - - bool use_deleted() const { - return use_deleted_; - } - void set_use_deleted(bool t) { - use_deleted_ = t; - } - - size_type num_ht_copies() const { - return static_cast<size_type>(num_ht_copies_); - } - void inc_num_ht_copies() { - ++num_ht_copies_; - } - - // Reset the enlarge and shrink thresholds - void reset_thresholds(size_type num_buckets) { - set_enlarge_threshold(enlarge_size(num_buckets)); - set_shrink_threshold(shrink_size(num_buckets)); - // whatever caused us to reset already considered - set_consider_shrink(false); - } - - // Caller is resposible for calling reset_threshold right after - // set_resizing_parameters. - void set_resizing_parameters(float shrink, float grow) { - assert(shrink >= 0.0); - assert(grow <= 1.0); - if (shrink > grow/2.0f) - shrink = grow / 2.0f; // otherwise we thrash hashtable size - set_shrink_factor(shrink); - set_enlarge_factor(grow); - } - - // This is the smallest size a hashtable can be without being too crowded - // If you like, you can give a min #buckets as well as a min #elts - size_type min_buckets(size_type num_elts, size_type min_buckets_wanted) { - float enlarge = enlarge_factor(); - size_type sz = HT_MIN_BUCKETS; // min buckets allowed - while ( sz < min_buckets_wanted || - num_elts >= static_cast<size_type>(sz * enlarge) ) { - // This just prevents overflowing size_type, since sz can exceed - // max_size() here. - if (static_cast<size_type>(sz * 2) < sz) { - throw std::length_error("resize overflow"); // protect against overflow - } - sz *= 2; - } - return sz; - } - - private: - template<class HashKey> class hash_munger { - public: - static size_t MungedHash(size_t hash) { - return hash; - } - }; - // This matches when the hashtable key is a pointer. - template<class HashKey> class hash_munger<HashKey*> { - public: - static size_t MungedHash(size_t hash) { - // TODO(csilvers): consider rotating instead: - // static const int shift = (sizeof(void *) == 4) ? 2 : 3; - // return (hash << (sizeof(hash) * 8) - shift)) | (hash >> shift); - // This matters if we ever change sparse/dense_hash_* to compare - // hashes before comparing actual values. It's speedy on x86. - return hash / sizeof(void*); // get rid of known-0 bits - } - }; - - size_type enlarge_threshold_; // table.size() * enlarge_factor - size_type shrink_threshold_; // table.size() * shrink_factor - float enlarge_factor_; // how full before resize - float shrink_factor_; // how empty before resize - // consider_shrink=true if we should try to shrink before next insert - bool consider_shrink_; - bool use_empty_; // used only by densehashtable, not sparsehashtable - bool use_deleted_; // false until delkey has been set - // num_ht_copies is a counter incremented every Copy/Move - unsigned int num_ht_copies_; -}; - -} // namespace sparsehash_internal - -#undef SPARSEHASH_COMPILE_ASSERT -_END_GOOGLE_NAMESPACE_ - -#endif // UTIL_GTL_HASHTABLE_COMMON_H_ + +namespace sparsehash_internal { + +// Adaptor methods for reading/writing data from an INPUT or OUPTUT +// variable passed to serialize() or unserialize(). For now we +// have implemented INPUT/OUTPUT for FILE*, istream*/ostream* (note +// they are pointers, unlike typical use), or else a pointer to +// something that supports a Read()/Write() method. +// +// For technical reasons, we implement read_data/write_data in two +// stages. The actual work is done in *_data_internal, which takes +// the stream argument twice: once as a template type, and once with +// normal type information. (We only use the second version.) We do +// this because of how C++ picks what function overload to use. If we +// implemented this the naive way: +// bool read_data(istream* is, const void* data, size_t length); +// template<typename T> read_data(T* fp, const void* data, size_t length); +// C++ would prefer the second version for every stream type except +// istream. However, we want C++ to prefer the first version for +// streams that are *subclasses* of istream, such as istringstream. +// This is not possible given the way template types are resolved. So +// we split the stream argument in two, one of which is templated and +// one of which is not. The specialized functions (like the istream +// version above) ignore the template arg and use the second, 'type' +// arg, getting subclass matching as normal. The 'catch-all' +// functions (the second version above) use the template arg to deduce +// the type, and use a second, void* arg to achieve the desired +// 'catch-all' semantics. + +// ----- low-level I/O for FILE* ---- + +template<typename Ignored> +inline bool read_data_internal(Ignored*, FILE* fp, + void* data, size_t length) { + return fread(data, length, 1, fp) == 1; +} + +template<typename Ignored> +inline bool write_data_internal(Ignored*, FILE* fp, + const void* data, size_t length) { + return fwrite(data, length, 1, fp) == 1; +} + +// ----- low-level I/O for iostream ---- + +// We want the caller to be responsible for #including <iostream>, not +// us, because iostream is a big header! According to the standard, +// it's only legal to delay the instantiation the way we want to if +// the istream/ostream is a template type. So we jump through hoops. +template<typename ISTREAM> +inline bool read_data_internal_for_istream(ISTREAM* fp, + void* data, size_t length) { + return fp->read(reinterpret_cast<char*>(data), length).good(); +} +template<typename Ignored> +inline bool read_data_internal(Ignored*, std::istream* fp, + void* data, size_t length) { + return read_data_internal_for_istream(fp, data, length); +} + +template<typename OSTREAM> +inline bool write_data_internal_for_ostream(OSTREAM* fp, + const void* data, size_t length) { + return fp->write(reinterpret_cast<const char*>(data), length).good(); +} +template<typename Ignored> +inline bool write_data_internal(Ignored*, std::ostream* fp, + const void* data, size_t length) { + return write_data_internal_for_ostream(fp, data, length); +} + +// ----- low-level I/O for custom streams ---- + +// The INPUT type needs to support a Read() method that takes a +// buffer and a length and returns the number of bytes read. +template <typename INPUT> +inline bool read_data_internal(INPUT* fp, void*, + void* data, size_t length) { + return static_cast<size_t>(fp->Read(data, length)) == length; +} + +// The OUTPUT type needs to support a Write() operation that takes +// a buffer and a length and returns the number of bytes written. +template <typename OUTPUT> +inline bool write_data_internal(OUTPUT* fp, void*, + const void* data, size_t length) { + return static_cast<size_t>(fp->Write(data, length)) == length; +} + +// ----- low-level I/O: the public API ---- + +template <typename INPUT> +inline bool read_data(INPUT* fp, void* data, size_t length) { + return read_data_internal(fp, fp, data, length); +} + +template <typename OUTPUT> +inline bool write_data(OUTPUT* fp, const void* data, size_t length) { + return write_data_internal(fp, fp, data, length); +} + +// Uses read_data() and write_data() to read/write an integer. +// length is the number of bytes to read/write (which may differ +// from sizeof(IntType), allowing us to save on a 32-bit system +// and load on a 64-bit system). Excess bytes are taken to be 0. +// INPUT and OUTPUT must match legal inputs to read/write_data (above). +template <typename INPUT, typename IntType> +bool read_bigendian_number(INPUT* fp, IntType* value, size_t length) { + *value = 0; + unsigned char byte; + // We require IntType to be unsigned or else the shifting gets all screwy. + SPARSEHASH_COMPILE_ASSERT(static_cast<IntType>(-1) > static_cast<IntType>(0), + serializing_int_requires_an_unsigned_type); + for (size_t i = 0; i < length; ++i) { + if (!read_data(fp, &byte, sizeof(byte))) return false; + *value |= static_cast<IntType>(byte) << ((length - 1 - i) * 8); + } + return true; +} + +template <typename OUTPUT, typename IntType> +bool write_bigendian_number(OUTPUT* fp, IntType value, size_t length) { + unsigned char byte; + // We require IntType to be unsigned or else the shifting gets all screwy. + SPARSEHASH_COMPILE_ASSERT(static_cast<IntType>(-1) > static_cast<IntType>(0), + serializing_int_requires_an_unsigned_type); + for (size_t i = 0; i < length; ++i) { + byte = (sizeof(value) <= length-1 - i) + ? 0 : static_cast<unsigned char>((value >> ((length-1 - i) * 8)) & 255); + if (!write_data(fp, &byte, sizeof(byte))) return false; + } + return true; +} + +// If your keys and values are simple enough, you can pass this +// serializer to serialize()/unserialize(). "Simple enough" means +// value_type is a POD type that contains no pointers. Note, +// however, we don't try to normalize endianness. +// This is the type used for NopointerSerializer. +template <typename value_type> struct pod_serializer { + template <typename INPUT> + bool operator()(INPUT* fp, value_type* value) const { + return read_data(fp, value, sizeof(*value)); + } + + template <typename OUTPUT> + bool operator()(OUTPUT* fp, const value_type& value) const { + return write_data(fp, &value, sizeof(value)); + } +}; + + +// Settings contains parameters for growing and shrinking the table. +// It also packages zero-size functor (ie. hasher). +// +// It does some munging of the hash value in cases where we think +// (fear) the original hash function might not be very good. In +// particular, the default hash of pointers is the identity hash, +// so probably all the low bits are 0. We identify when we think +// we're hashing a pointer, and chop off the low bits. Note this +// isn't perfect: even when the key is a pointer, we can't tell +// for sure that the hash is the identity hash. If it's not, this +// is needless work (and possibly, though not likely, harmful). + +template<typename Key, typename HashFunc, + typename SizeType, int HT_MIN_BUCKETS> +class sh_hashtable_settings : public HashFunc { + public: + typedef Key key_type; + typedef HashFunc hasher; + typedef SizeType size_type; + + public: + sh_hashtable_settings(const hasher& hf, + const float ht_occupancy_flt, + const float ht_empty_flt) + : hasher(hf), + enlarge_threshold_(0), + shrink_threshold_(0), + consider_shrink_(false), + use_empty_(false), + use_deleted_(false), + num_ht_copies_(0) { + set_enlarge_factor(ht_occupancy_flt); + set_shrink_factor(ht_empty_flt); + } + + size_type hash(const key_type& v) const { + // We munge the hash value when we don't trust hasher::operator(). + return hash_munger<Key>::MungedHash(hasher::operator()(v)); + } + + float enlarge_factor() const { + return enlarge_factor_; + } + void set_enlarge_factor(float f) { + enlarge_factor_ = f; + } + float shrink_factor() const { + return shrink_factor_; + } + void set_shrink_factor(float f) { + shrink_factor_ = f; + } + + size_type enlarge_threshold() const { + return enlarge_threshold_; + } + void set_enlarge_threshold(size_type t) { + enlarge_threshold_ = t; + } + size_type shrink_threshold() const { + return shrink_threshold_; + } + void set_shrink_threshold(size_type t) { + shrink_threshold_ = t; + } + + size_type enlarge_size(size_type x) const { + return static_cast<size_type>(x * enlarge_factor_); + } + size_type shrink_size(size_type x) const { + return static_cast<size_type>(x * shrink_factor_); + } + + bool consider_shrink() const { + return consider_shrink_; + } + void set_consider_shrink(bool t) { + consider_shrink_ = t; + } + + bool use_empty() const { + return use_empty_; + } + void set_use_empty(bool t) { + use_empty_ = t; + } + + bool use_deleted() const { + return use_deleted_; + } + void set_use_deleted(bool t) { + use_deleted_ = t; + } + + size_type num_ht_copies() const { + return static_cast<size_type>(num_ht_copies_); + } + void inc_num_ht_copies() { + ++num_ht_copies_; + } + + // Reset the enlarge and shrink thresholds + void reset_thresholds(size_type num_buckets) { + set_enlarge_threshold(enlarge_size(num_buckets)); + set_shrink_threshold(shrink_size(num_buckets)); + // whatever caused us to reset already considered + set_consider_shrink(false); + } + + // Caller is resposible for calling reset_threshold right after + // set_resizing_parameters. + void set_resizing_parameters(float shrink, float grow) { + assert(shrink >= 0.0); + assert(grow <= 1.0); + if (shrink > grow/2.0f) + shrink = grow / 2.0f; // otherwise we thrash hashtable size + set_shrink_factor(shrink); + set_enlarge_factor(grow); + } + + // This is the smallest size a hashtable can be without being too crowded + // If you like, you can give a min #buckets as well as a min #elts + size_type min_buckets(size_type num_elts, size_type min_buckets_wanted) { + float enlarge = enlarge_factor(); + size_type sz = HT_MIN_BUCKETS; // min buckets allowed + while ( sz < min_buckets_wanted || + num_elts >= static_cast<size_type>(sz * enlarge) ) { + // This just prevents overflowing size_type, since sz can exceed + // max_size() here. + if (static_cast<size_type>(sz * 2) < sz) { + throw std::length_error("resize overflow"); // protect against overflow + } + sz *= 2; + } + return sz; + } + + private: + template<class HashKey> class hash_munger { + public: + static size_t MungedHash(size_t hash) { + return hash; + } + }; + // This matches when the hashtable key is a pointer. + template<class HashKey> class hash_munger<HashKey*> { + public: + static size_t MungedHash(size_t hash) { + // TODO(csilvers): consider rotating instead: + // static const int shift = (sizeof(void *) == 4) ? 2 : 3; + // return (hash << (sizeof(hash) * 8) - shift)) | (hash >> shift); + // This matters if we ever change sparse/dense_hash_* to compare + // hashes before comparing actual values. It's speedy on x86. + return hash / sizeof(void*); // get rid of known-0 bits + } + }; + + size_type enlarge_threshold_; // table.size() * enlarge_factor + size_type shrink_threshold_; // table.size() * shrink_factor + float enlarge_factor_; // how full before resize + float shrink_factor_; // how empty before resize + // consider_shrink=true if we should try to shrink before next insert + bool consider_shrink_; + bool use_empty_; // used only by densehashtable, not sparsehashtable + bool use_deleted_; // false until delkey has been set + // num_ht_copies is a counter incremented every Copy/Move + unsigned int num_ht_copies_; +}; + +} // namespace sparsehash_internal + +#undef SPARSEHASH_COMPILE_ASSERT +_END_GOOGLE_NAMESPACE_ + +#endif // UTIL_GTL_HASHTABLE_COMMON_H_ diff --git a/contrib/libs/sparsehash/src/sparsehash/internal/libc_allocator_with_realloc.h b/contrib/libs/sparsehash/src/sparsehash/internal/libc_allocator_with_realloc.h index 9429b84b7a..769329fa6c 100644 --- a/contrib/libs/sparsehash/src/sparsehash/internal/libc_allocator_with_realloc.h +++ b/contrib/libs/sparsehash/src/sparsehash/internal/libc_allocator_with_realloc.h @@ -1,122 +1,122 @@ -// Copyright (c) 2010, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- - -#ifndef UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ -#define UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <stdlib.h> // for malloc/realloc/free -#include <stddef.h> // for ptrdiff_t -#include <new> // for placement new - -_START_GOOGLE_NAMESPACE_ - -template<class T> -class libc_allocator_with_realloc { - public: - typedef T value_type; - typedef size_t size_type; - typedef ptrdiff_t difference_type; - - typedef T* pointer; - typedef const T* const_pointer; - typedef T& reference; - typedef const T& const_reference; - - libc_allocator_with_realloc() {} - libc_allocator_with_realloc(const libc_allocator_with_realloc&) {} - ~libc_allocator_with_realloc() {} - - pointer address(reference r) const { return &r; } - const_pointer address(const_reference r) const { return &r; } - - pointer allocate(size_type n, const_pointer = 0) { - return static_cast<pointer>(malloc(n * sizeof(value_type))); - } - void deallocate(pointer p, size_type) { - free(p); - } - pointer reallocate(pointer p, size_type n) { +// Copyright (c) 2010, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- + +#ifndef UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ +#define UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <stdlib.h> // for malloc/realloc/free +#include <stddef.h> // for ptrdiff_t +#include <new> // for placement new + +_START_GOOGLE_NAMESPACE_ + +template<class T> +class libc_allocator_with_realloc { + public: + typedef T value_type; + typedef size_t size_type; + typedef ptrdiff_t difference_type; + + typedef T* pointer; + typedef const T* const_pointer; + typedef T& reference; + typedef const T& const_reference; + + libc_allocator_with_realloc() {} + libc_allocator_with_realloc(const libc_allocator_with_realloc&) {} + ~libc_allocator_with_realloc() {} + + pointer address(reference r) const { return &r; } + const_pointer address(const_reference r) const { return &r; } + + pointer allocate(size_type n, const_pointer = 0) { + return static_cast<pointer>(malloc(n * sizeof(value_type))); + } + void deallocate(pointer p, size_type) { + free(p); + } + pointer reallocate(pointer p, size_type n) { // p points to a storage array whose objects have already been destroyed // cast to void* to prevent compiler warnings about calling realloc() on // an object which cannot be relocated in memory return static_cast<pointer>(realloc(static_cast<void*>(p), n * sizeof(value_type))); - } - - size_type max_size() const { - return static_cast<size_type>(-1) / sizeof(value_type); - } - - void construct(pointer p, const value_type& val) { - new(p) value_type(val); - } - void destroy(pointer p) { p->~value_type(); } - - template <class U> - libc_allocator_with_realloc(const libc_allocator_with_realloc<U>&) {} - - template<class U> - struct rebind { - typedef libc_allocator_with_realloc<U> other; - }; -}; - -// libc_allocator_with_realloc<void> specialization. -template<> -class libc_allocator_with_realloc<void> { - public: - typedef void value_type; - typedef size_t size_type; - typedef ptrdiff_t difference_type; - typedef void* pointer; - typedef const void* const_pointer; - - template<class U> - struct rebind { - typedef libc_allocator_with_realloc<U> other; - }; -}; - -template<class T> -inline bool operator==(const libc_allocator_with_realloc<T>&, - const libc_allocator_with_realloc<T>&) { - return true; -} - -template<class T> -inline bool operator!=(const libc_allocator_with_realloc<T>&, - const libc_allocator_with_realloc<T>&) { - return false; -} - -_END_GOOGLE_NAMESPACE_ - -#endif // UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ + } + + size_type max_size() const { + return static_cast<size_type>(-1) / sizeof(value_type); + } + + void construct(pointer p, const value_type& val) { + new(p) value_type(val); + } + void destroy(pointer p) { p->~value_type(); } + + template <class U> + libc_allocator_with_realloc(const libc_allocator_with_realloc<U>&) {} + + template<class U> + struct rebind { + typedef libc_allocator_with_realloc<U> other; + }; +}; + +// libc_allocator_with_realloc<void> specialization. +template<> +class libc_allocator_with_realloc<void> { + public: + typedef void value_type; + typedef size_t size_type; + typedef ptrdiff_t difference_type; + typedef void* pointer; + typedef const void* const_pointer; + + template<class U> + struct rebind { + typedef libc_allocator_with_realloc<U> other; + }; +}; + +template<class T> +inline bool operator==(const libc_allocator_with_realloc<T>&, + const libc_allocator_with_realloc<T>&) { + return true; +} + +template<class T> +inline bool operator!=(const libc_allocator_with_realloc<T>&, + const libc_allocator_with_realloc<T>&) { + return false; +} + +_END_GOOGLE_NAMESPACE_ + +#endif // UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ diff --git a/contrib/libs/sparsehash/src/sparsehash/internal/sparseconfig.h b/contrib/libs/sparsehash/src/sparsehash/internal/sparseconfig.h index 42d3dab03d..bb8c9a9428 100644 --- a/contrib/libs/sparsehash/src/sparsehash/internal/sparseconfig.h +++ b/contrib/libs/sparsehash/src/sparsehash/internal/sparseconfig.h @@ -1,48 +1,48 @@ -/* - * NOTE: This file is for internal use only. - * Do not use these #defines in your own program! - */ - -/* Namespace for Google classes */ +/* + * NOTE: This file is for internal use only. + * Do not use these #defines in your own program! + */ + +/* Namespace for Google classes */ #define GOOGLE_NAMESPACE ::google - + /* the location of the header defining hash functions */ #ifndef HASH_FUN_H -#define HASH_FUN_H <util/generic/hash.h> +#define HASH_FUN_H <util/generic/hash.h> #endif - -/* the namespace of the hash<> function */ -#define HASH_NAMESPACE - -/* Define to 1 if you have the <inttypes.h> header file. */ -#define HAVE_INTTYPES_H 1 - -/* Define to 1 if the system has the type `long long'. */ -#define HAVE_LONG_LONG 1 - -/* Define to 1 if you have the `memcpy' function. */ -#define HAVE_MEMCPY 1 - -/* Define to 1 if you have the <stdint.h> header file. */ -#define HAVE_STDINT_H 1 - -/* Define to 1 if you have the <sys/types.h> header file. */ -#define HAVE_SYS_TYPES_H 1 - -/* Define to 1 if the system has the type `uint16_t'. */ -#define HAVE_UINT16_T 1 - -/* Define to 1 if the system has the type `u_int16_t'. */ -#define HAVE_U_INT16_T 1 - -/* Define to 1 if the system has the type `__uint16'. */ -/* #undef HAVE___UINT16 */ - -/* The system-provided hash function including the namespace. */ + +/* the namespace of the hash<> function */ +#define HASH_NAMESPACE + +/* Define to 1 if you have the <inttypes.h> header file. */ +#define HAVE_INTTYPES_H 1 + +/* Define to 1 if the system has the type `long long'. */ +#define HAVE_LONG_LONG 1 + +/* Define to 1 if you have the `memcpy' function. */ +#define HAVE_MEMCPY 1 + +/* Define to 1 if you have the <stdint.h> header file. */ +#define HAVE_STDINT_H 1 + +/* Define to 1 if you have the <sys/types.h> header file. */ +#define HAVE_SYS_TYPES_H 1 + +/* Define to 1 if the system has the type `uint16_t'. */ +#define HAVE_UINT16_T 1 + +/* Define to 1 if the system has the type `u_int16_t'. */ +#define HAVE_U_INT16_T 1 + +/* Define to 1 if the system has the type `__uint16'. */ +/* #undef HAVE___UINT16 */ + +/* The system-provided hash function including the namespace. */ #define SPARSEHASH_HASH HASH_NAMESPACE::THash - -/* Stops putting the code inside the Google namespace */ + +/* Stops putting the code inside the Google namespace */ #define _END_GOOGLE_NAMESPACE_ } - -/* Puts following code inside the Google namespace */ + +/* Puts following code inside the Google namespace */ #define _START_GOOGLE_NAMESPACE_ namespace google { diff --git a/contrib/libs/sparsehash/src/sparsehash/internal/sparsehashtable.h b/contrib/libs/sparsehash/src/sparsehash/internal/sparsehashtable.h index 50eb01b767..f54ea51e9a 100644 --- a/contrib/libs/sparsehash/src/sparsehash/internal/sparsehashtable.h +++ b/contrib/libs/sparsehash/src/sparsehash/internal/sparsehashtable.h @@ -1,1247 +1,1247 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// A sparse hashtable is a particular implementation of -// a hashtable: one that is meant to minimize memory use. -// It does this by using a *sparse table* (cf sparsetable.h), -// which uses between 1 and 2 bits to store empty buckets -// (we may need another bit for hashtables that support deletion). -// -// When empty buckets are so cheap, an appealing hashtable -// implementation is internal probing, in which the hashtable -// is a single table, and collisions are resolved by trying -// to insert again in another bucket. The most cache-efficient -// internal probing schemes are linear probing (which suffers, -// alas, from clumping) and quadratic probing, which is what -// we implement by default. -// -// Deleted buckets are a bit of a pain. We have to somehow mark -// deleted buckets (the probing must distinguish them from empty -// buckets). The most principled way is to have another bitmap, -// but that's annoying and takes up space. Instead we let the -// user specify an "impossible" key. We set deleted buckets -// to have the impossible key. -// -// Note it is possible to change the value of the delete key -// on the fly; you can even remove it, though after that point -// the hashtable is insert_only until you set it again. -// -// You probably shouldn't use this code directly. Use -// sparse_hash_map<> or sparse_hash_set<> instead. -// -// You can modify the following, below: -// HT_OCCUPANCY_PCT -- how full before we double size -// HT_EMPTY_PCT -- how empty before we halve size -// HT_MIN_BUCKETS -- smallest bucket size -// HT_DEFAULT_STARTING_BUCKETS -- default bucket size at construct-time -// -// You can also change enlarge_factor (which defaults to -// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to -// HT_EMPTY_PCT) with set_resizing_parameters(). -// -// How to decide what values to use? -// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good. -// HT_MIN_BUCKETS is probably unnecessary since you can specify -// (indirectly) the starting number of buckets at construct-time. -// For enlarge_factor, you can use this chart to try to trade-off -// expected lookup time to the space taken up. By default, this -// code uses quadratic probing, though you can change it to linear -// via _JUMP below if you really want to. -// -// From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html -// NUMBER OF PROBES / LOOKUP Successful Unsuccessful -// Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L) -// Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2 -// -// -- enlarge_factor -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99 -// QUADRATIC COLLISION RES. -// probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11 -// probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6 -// LINEAR COLLISION RES. -// probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5 -// probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0 -// -// The value type is required to be copy constructible and default -// constructible, but it need not be (and commonly isn't) assignable. - -#ifndef _SPARSEHASHTABLE_H_ -#define _SPARSEHASHTABLE_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <assert.h> -#include <algorithm> // For swap(), eg -#include <iterator> // for iterator tags -#include <limits> // for numeric_limits -#include <utility> // for pair -#include <sparsehash/type_traits.h> // for remove_const -#include <sparsehash/internal/hashtable-common.h> -#include <sparsehash/sparsetable> // IWYU pragma: export -#include <stdexcept> // For length_error - -_START_GOOGLE_NAMESPACE_ - -namespace base { // just to make google->opensource transition easier -using GOOGLE_NAMESPACE::remove_const; -} - -#ifndef SPARSEHASH_STAT_UPDATE -#define SPARSEHASH_STAT_UPDATE(x) ((void) 0) -#endif - -// The probing method -// Linear probing -// #define JUMP_(key, num_probes) ( 1 ) -// Quadratic probing -#define JUMP_(key, num_probes) ( num_probes ) - -// The smaller this is, the faster lookup is (because the group bitmap is -// smaller) and the faster insert is, because there's less to move. -// On the other hand, there are more groups. Since group::size_type is -// a short, this number should be of the form 32*x + 16 to avoid waste. -static const u_int16_t DEFAULT_GROUP_SIZE = 48; // fits in 1.5 words - -// Hashtable class, used to implement the hashed associative containers -// hash_set and hash_map. -// -// Value: what is stored in the table (each bucket is a Value). -// Key: something in a 1-to-1 correspondence to a Value, that can be used -// to search for a Value in the table (find() takes a Key). -// HashFcn: Takes a Key and returns an integer, the more unique the better. -// ExtractKey: given a Value, returns the unique Key associated with it. -// Must inherit from unary_function, or at least have a -// result_type enum indicating the return type of operator(). -// SetKey: given a Value* and a Key, modifies the value such that -// ExtractKey(value) == key. We guarantee this is only called -// with key == deleted_key. -// EqualKey: Given two Keys, says whether they are the same (that is, -// if they are both associated with the same Value). -// Alloc: STL allocator to use to allocate memory. - -template <class Value, class Key, class HashFcn, - class ExtractKey, class SetKey, class EqualKey, class Alloc> -class sparse_hashtable; - -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct sparse_hashtable_iterator; - -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct sparse_hashtable_const_iterator; - -// As far as iterating, we're basically just a sparsetable -// that skips over deleted elements. -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct sparse_hashtable_iterator { - private: - typedef typename A::template rebind<V>::other value_alloc_type; - - public: - typedef sparse_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; - typedef sparse_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// A sparse hashtable is a particular implementation of +// a hashtable: one that is meant to minimize memory use. +// It does this by using a *sparse table* (cf sparsetable.h), +// which uses between 1 and 2 bits to store empty buckets +// (we may need another bit for hashtables that support deletion). +// +// When empty buckets are so cheap, an appealing hashtable +// implementation is internal probing, in which the hashtable +// is a single table, and collisions are resolved by trying +// to insert again in another bucket. The most cache-efficient +// internal probing schemes are linear probing (which suffers, +// alas, from clumping) and quadratic probing, which is what +// we implement by default. +// +// Deleted buckets are a bit of a pain. We have to somehow mark +// deleted buckets (the probing must distinguish them from empty +// buckets). The most principled way is to have another bitmap, +// but that's annoying and takes up space. Instead we let the +// user specify an "impossible" key. We set deleted buckets +// to have the impossible key. +// +// Note it is possible to change the value of the delete key +// on the fly; you can even remove it, though after that point +// the hashtable is insert_only until you set it again. +// +// You probably shouldn't use this code directly. Use +// sparse_hash_map<> or sparse_hash_set<> instead. +// +// You can modify the following, below: +// HT_OCCUPANCY_PCT -- how full before we double size +// HT_EMPTY_PCT -- how empty before we halve size +// HT_MIN_BUCKETS -- smallest bucket size +// HT_DEFAULT_STARTING_BUCKETS -- default bucket size at construct-time +// +// You can also change enlarge_factor (which defaults to +// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to +// HT_EMPTY_PCT) with set_resizing_parameters(). +// +// How to decide what values to use? +// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good. +// HT_MIN_BUCKETS is probably unnecessary since you can specify +// (indirectly) the starting number of buckets at construct-time. +// For enlarge_factor, you can use this chart to try to trade-off +// expected lookup time to the space taken up. By default, this +// code uses quadratic probing, though you can change it to linear +// via _JUMP below if you really want to. +// +// From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html +// NUMBER OF PROBES / LOOKUP Successful Unsuccessful +// Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L) +// Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2 +// +// -- enlarge_factor -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99 +// QUADRATIC COLLISION RES. +// probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11 +// probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6 +// LINEAR COLLISION RES. +// probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5 +// probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0 +// +// The value type is required to be copy constructible and default +// constructible, but it need not be (and commonly isn't) assignable. + +#ifndef _SPARSEHASHTABLE_H_ +#define _SPARSEHASHTABLE_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <assert.h> +#include <algorithm> // For swap(), eg +#include <iterator> // for iterator tags +#include <limits> // for numeric_limits +#include <utility> // for pair +#include <sparsehash/type_traits.h> // for remove_const +#include <sparsehash/internal/hashtable-common.h> +#include <sparsehash/sparsetable> // IWYU pragma: export +#include <stdexcept> // For length_error + +_START_GOOGLE_NAMESPACE_ + +namespace base { // just to make google->opensource transition easier +using GOOGLE_NAMESPACE::remove_const; +} + +#ifndef SPARSEHASH_STAT_UPDATE +#define SPARSEHASH_STAT_UPDATE(x) ((void) 0) +#endif + +// The probing method +// Linear probing +// #define JUMP_(key, num_probes) ( 1 ) +// Quadratic probing +#define JUMP_(key, num_probes) ( num_probes ) + +// The smaller this is, the faster lookup is (because the group bitmap is +// smaller) and the faster insert is, because there's less to move. +// On the other hand, there are more groups. Since group::size_type is +// a short, this number should be of the form 32*x + 16 to avoid waste. +static const u_int16_t DEFAULT_GROUP_SIZE = 48; // fits in 1.5 words + +// Hashtable class, used to implement the hashed associative containers +// hash_set and hash_map. +// +// Value: what is stored in the table (each bucket is a Value). +// Key: something in a 1-to-1 correspondence to a Value, that can be used +// to search for a Value in the table (find() takes a Key). +// HashFcn: Takes a Key and returns an integer, the more unique the better. +// ExtractKey: given a Value, returns the unique Key associated with it. +// Must inherit from unary_function, or at least have a +// result_type enum indicating the return type of operator(). +// SetKey: given a Value* and a Key, modifies the value such that +// ExtractKey(value) == key. We guarantee this is only called +// with key == deleted_key. +// EqualKey: Given two Keys, says whether they are the same (that is, +// if they are both associated with the same Value). +// Alloc: STL allocator to use to allocate memory. + +template <class Value, class Key, class HashFcn, + class ExtractKey, class SetKey, class EqualKey, class Alloc> +class sparse_hashtable; + +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct sparse_hashtable_iterator; + +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct sparse_hashtable_const_iterator; + +// As far as iterating, we're basically just a sparsetable +// that skips over deleted elements. +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct sparse_hashtable_iterator { + private: + typedef typename A::template rebind<V>::other value_alloc_type; + + public: + typedef sparse_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; + typedef sparse_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; typedef typename sparsetable<V,DEFAULT_GROUP_SIZE,value_alloc_type>::nonempty_iterator - st_iterator; - - typedef std::forward_iterator_tag iterator_category; // very little defined! - typedef V value_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::pointer pointer; - - // "Real" constructor and default constructor - sparse_hashtable_iterator(const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, - st_iterator it, st_iterator it_end) - : ht(h), pos(it), end(it_end) { advance_past_deleted(); } - sparse_hashtable_iterator() { } // not ever used internally - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // Happy dereferencer - reference operator*() const { return *pos; } - pointer operator->() const { return &(operator*()); } - - // Arithmetic. The only hard part is making sure that - // we're not on a marked-deleted array element - void advance_past_deleted() { - while ( pos != end && ht->test_deleted(*this) ) - ++pos; - } - iterator& operator++() { - assert(pos != end); ++pos; advance_past_deleted(); return *this; - } - iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } - - // Comparison. - bool operator==(const iterator& it) const { return pos == it.pos; } - bool operator!=(const iterator& it) const { return pos != it.pos; } - - - // The actual data - const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; - st_iterator pos, end; -}; - -// Now do it all again, but with const-ness! -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct sparse_hashtable_const_iterator { - private: - typedef typename A::template rebind<V>::other value_alloc_type; - - public: - typedef sparse_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; - typedef sparse_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; + st_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::pointer pointer; + + // "Real" constructor and default constructor + sparse_hashtable_iterator(const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, + st_iterator it, st_iterator it_end) + : ht(h), pos(it), end(it_end) { advance_past_deleted(); } + sparse_hashtable_iterator() { } // not ever used internally + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on a marked-deleted array element + void advance_past_deleted() { + while ( pos != end && ht->test_deleted(*this) ) + ++pos; + } + iterator& operator++() { + assert(pos != end); ++pos; advance_past_deleted(); return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const iterator& it) const { return pos == it.pos; } + bool operator!=(const iterator& it) const { return pos != it.pos; } + + + // The actual data + const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; + st_iterator pos, end; +}; + +// Now do it all again, but with const-ness! +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct sparse_hashtable_const_iterator { + private: + typedef typename A::template rebind<V>::other value_alloc_type; + + public: + typedef sparse_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; + typedef sparse_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator; typedef typename sparsetable<V,DEFAULT_GROUP_SIZE,value_alloc_type>::const_nonempty_iterator - st_iterator; - - typedef std::forward_iterator_tag iterator_category; // very little defined! - typedef V value_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::const_reference reference; - typedef typename value_alloc_type::const_pointer pointer; - - // "Real" constructor and default constructor - sparse_hashtable_const_iterator(const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, - st_iterator it, st_iterator it_end) - : ht(h), pos(it), end(it_end) { advance_past_deleted(); } - // This lets us convert regular iterators to const iterators - sparse_hashtable_const_iterator() { } // never used internally - sparse_hashtable_const_iterator(const iterator &it) - : ht(it.ht), pos(it.pos), end(it.end) { } - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // Happy dereferencer - reference operator*() const { return *pos; } - pointer operator->() const { return &(operator*()); } - - // Arithmetic. The only hard part is making sure that - // we're not on a marked-deleted array element - void advance_past_deleted() { - while ( pos != end && ht->test_deleted(*this) ) - ++pos; - } - const_iterator& operator++() { - assert(pos != end); ++pos; advance_past_deleted(); return *this; - } - const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; } - - // Comparison. - bool operator==(const const_iterator& it) const { return pos == it.pos; } - bool operator!=(const const_iterator& it) const { return pos != it.pos; } - - - // The actual data - const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; - st_iterator pos, end; -}; - -// And once again, but this time freeing up memory as we iterate -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -struct sparse_hashtable_destructive_iterator { - private: - typedef typename A::template rebind<V>::other value_alloc_type; - - public: - typedef sparse_hashtable_destructive_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; + st_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::const_reference reference; + typedef typename value_alloc_type::const_pointer pointer; + + // "Real" constructor and default constructor + sparse_hashtable_const_iterator(const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, + st_iterator it, st_iterator it_end) + : ht(h), pos(it), end(it_end) { advance_past_deleted(); } + // This lets us convert regular iterators to const iterators + sparse_hashtable_const_iterator() { } // never used internally + sparse_hashtable_const_iterator(const iterator &it) + : ht(it.ht), pos(it.pos), end(it.end) { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on a marked-deleted array element + void advance_past_deleted() { + while ( pos != end && ht->test_deleted(*this) ) + ++pos; + } + const_iterator& operator++() { + assert(pos != end); ++pos; advance_past_deleted(); return *this; + } + const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const const_iterator& it) const { return pos == it.pos; } + bool operator!=(const const_iterator& it) const { return pos != it.pos; } + + + // The actual data + const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; + st_iterator pos, end; +}; + +// And once again, but this time freeing up memory as we iterate +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +struct sparse_hashtable_destructive_iterator { + private: + typedef typename A::template rebind<V>::other value_alloc_type; + + public: + typedef sparse_hashtable_destructive_iterator<V,K,HF,ExK,SetK,EqK,A> iterator; typedef typename sparsetable<V,DEFAULT_GROUP_SIZE,value_alloc_type>::destructive_iterator - st_iterator; - - typedef std::forward_iterator_tag iterator_category; // very little defined! - typedef V value_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::pointer pointer; - - // "Real" constructor and default constructor - sparse_hashtable_destructive_iterator(const - sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, - st_iterator it, st_iterator it_end) - : ht(h), pos(it), end(it_end) { advance_past_deleted(); } - sparse_hashtable_destructive_iterator() { } // never used internally - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // Happy dereferencer - reference operator*() const { return *pos; } - pointer operator->() const { return &(operator*()); } - - // Arithmetic. The only hard part is making sure that - // we're not on a marked-deleted array element - void advance_past_deleted() { - while ( pos != end && ht->test_deleted(*this) ) - ++pos; - } - iterator& operator++() { - assert(pos != end); ++pos; advance_past_deleted(); return *this; - } - iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } - - // Comparison. - bool operator==(const iterator& it) const { return pos == it.pos; } - bool operator!=(const iterator& it) const { return pos != it.pos; } - - - // The actual data - const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; - st_iterator pos, end; -}; - - -template <class Value, class Key, class HashFcn, - class ExtractKey, class SetKey, class EqualKey, class Alloc> -class sparse_hashtable { - private: - typedef typename Alloc::template rebind<Value>::other value_alloc_type; - - public: - typedef Key key_type; - typedef Value value_type; - typedef HashFcn hasher; - typedef EqualKey key_equal; - typedef Alloc allocator_type; - - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::const_reference const_reference; - typedef typename value_alloc_type::pointer pointer; - typedef typename value_alloc_type::const_pointer const_pointer; - typedef sparse_hashtable_iterator<Value, Key, HashFcn, ExtractKey, - SetKey, EqualKey, Alloc> - iterator; - - typedef sparse_hashtable_const_iterator<Value, Key, HashFcn, ExtractKey, - SetKey, EqualKey, Alloc> - const_iterator; - - typedef sparse_hashtable_destructive_iterator<Value, Key, HashFcn, ExtractKey, - SetKey, EqualKey, Alloc> - destructive_iterator; - - // These come from tr1. For us they're the same as regular iterators. - typedef iterator local_iterator; - typedef const_iterator const_local_iterator; - - // How full we let the table get before we resize, by default. - // Knuth says .8 is good -- higher causes us to probe too much, - // though it saves memory. - static const int HT_OCCUPANCY_PCT; // = 80 (out of 100); - - // How empty we let the table get before we resize lower, by default. - // (0.0 means never resize lower.) - // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing - static const int HT_EMPTY_PCT; // = 0.4 * HT_OCCUPANCY_PCT; - - // Minimum size we're willing to let hashtables be. - // Must be a power of two, and at least 4. - // Note, however, that for a given hashtable, the initial size is a - // function of the first constructor arg, and may be >HT_MIN_BUCKETS. - static const size_type HT_MIN_BUCKETS = 4; - - // By default, if you don't specify a hashtable size at - // construction-time, we use this size. Must be a power of two, and - // at least HT_MIN_BUCKETS. - static const size_type HT_DEFAULT_STARTING_BUCKETS = 32; - - // ITERATOR FUNCTIONS - iterator begin() { return iterator(this, table.nonempty_begin(), - table.nonempty_end()); } - iterator end() { return iterator(this, table.nonempty_end(), - table.nonempty_end()); } - const_iterator begin() const { return const_iterator(this, - table.nonempty_begin(), - table.nonempty_end()); } - const_iterator end() const { return const_iterator(this, - table.nonempty_end(), - table.nonempty_end()); } - - // These come from tr1 unordered_map. They iterate over 'bucket' n. - // For sparsehashtable, we could consider each 'group' to be a bucket, - // I guess, but I don't really see the point. We'll just consider - // bucket n to be the n-th element of the sparsetable, if it's occupied, - // or some empty element, otherwise. - local_iterator begin(size_type i) { - if (table.test(i)) - return local_iterator(this, table.get_iter(i), table.nonempty_end()); - else - return local_iterator(this, table.nonempty_end(), table.nonempty_end()); - } - local_iterator end(size_type i) { - local_iterator it = begin(i); - if (table.test(i) && !test_deleted(i)) - ++it; - return it; - } - const_local_iterator begin(size_type i) const { - if (table.test(i)) - return const_local_iterator(this, table.get_iter(i), - table.nonempty_end()); - else - return const_local_iterator(this, table.nonempty_end(), - table.nonempty_end()); - } - const_local_iterator end(size_type i) const { - const_local_iterator it = begin(i); - if (table.test(i) && !test_deleted(i)) - ++it; - return it; - } - - // This is used when resizing - destructive_iterator destructive_begin() { - return destructive_iterator(this, table.destructive_begin(), - table.destructive_end()); - } - destructive_iterator destructive_end() { - return destructive_iterator(this, table.destructive_end(), - table.destructive_end()); - } - - - // ACCESSOR FUNCTIONS for the things we templatize on, basically - hasher hash_funct() const { return settings; } - key_equal key_eq() const { return key_info; } - allocator_type get_allocator() const { return table.get_allocator(); } - - // Accessor function for statistics gathering. - int num_table_copies() const { return settings.num_ht_copies(); } - - private: - // We need to copy values when we set the special marker for deleted - // elements, but, annoyingly, we can't just use the copy assignment - // operator because value_type might not be assignable (it's often - // pair<const X, Y>). We use explicit destructor invocation and - // placement new to get around this. Arg. - void set_value(pointer dst, const_reference src) { - dst->~value_type(); // delete the old value, if any - new(dst) value_type(src); - } - - // This is used as a tag for the copy constructor, saying to destroy its - // arg We have two ways of destructively copying: with potentially growing - // the hashtable as we copy, and without. To make sure the outside world - // can't do a destructive copy, we make the typename private. - enum MoveDontCopyT {MoveDontCopy, MoveDontGrow}; - - // DELETE HELPER FUNCTIONS - // This lets the user describe a key that will indicate deleted - // table entries. This key should be an "impossible" entry -- - // if you try to insert it for real, you won't be able to retrieve it! - // (NB: while you pass in an entire value, only the key part is looked - // at. This is just because I don't know how to assign just a key.) - private: - void squash_deleted() { // gets rid of any deleted entries we have - if ( num_deleted ) { // get rid of deleted before writing - sparse_hashtable tmp(MoveDontGrow, *this); - swap(tmp); // now we are tmp - } - assert(num_deleted == 0); - } - - // Test if the given key is the deleted indicator. Requires - // num_deleted > 0, for correctness of read(), and because that - // guarantees that key_info.delkey is valid. - bool test_deleted_key(const key_type& key) const { - assert(num_deleted > 0); - return equals(key_info.delkey, key); - } - - public: - void set_deleted_key(const key_type &key) { - // It's only safe to change what "deleted" means if we purge deleted guys - squash_deleted(); - settings.set_use_deleted(true); - key_info.delkey = key; - } - void clear_deleted_key() { - squash_deleted(); - settings.set_use_deleted(false); - } - key_type deleted_key() const { - assert(settings.use_deleted() - && "Must set deleted key before calling deleted_key"); - return key_info.delkey; - } - - // These are public so the iterators can use them - // True if the item at position bucknum is "deleted" marker - bool test_deleted(size_type bucknum) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && table.test(bucknum) && - test_deleted_key(get_key(table.unsafe_get(bucknum))); - } - bool test_deleted(const iterator &it) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && test_deleted_key(get_key(*it)); - } - bool test_deleted(const const_iterator &it) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && test_deleted_key(get_key(*it)); - } - bool test_deleted(const destructive_iterator &it) const { - // Invariant: !use_deleted() implies num_deleted is 0. - assert(settings.use_deleted() || num_deleted == 0); - return num_deleted > 0 && test_deleted_key(get_key(*it)); - } - - private: - void check_use_deleted(const char* caller) { - (void)caller; // could log it if the assert failed - assert(settings.use_deleted()); - } - - // Set it so test_deleted is true. true if object didn't used to be deleted. - // TODO(csilvers): make these private (also in densehashtable.h) - bool set_deleted(iterator &it) { - check_use_deleted("set_deleted()"); - bool retval = !test_deleted(it); - // &* converts from iterator to value-type. - set_key(&(*it), key_info.delkey); - return retval; - } - // Set it so test_deleted is false. true if object used to be deleted. - bool clear_deleted(iterator &it) { - check_use_deleted("clear_deleted()"); - // Happens automatically when we assign something else in its place. - return test_deleted(it); - } - - // We also allow to set/clear the deleted bit on a const iterator. - // We allow a const_iterator for the same reason you can delete a - // const pointer: it's convenient, and semantically you can't use - // 'it' after it's been deleted anyway, so its const-ness doesn't - // really matter. - bool set_deleted(const_iterator &it) { - check_use_deleted("set_deleted()"); - bool retval = !test_deleted(it); - set_key(const_cast<pointer>(&(*it)), key_info.delkey); - return retval; - } - // Set it so test_deleted is false. true if object used to be deleted. - bool clear_deleted(const_iterator &it) { - check_use_deleted("clear_deleted()"); - return test_deleted(it); - } - - // FUNCTIONS CONCERNING SIZE - public: - size_type size() const { return table.num_nonempty() - num_deleted; } - size_type max_size() const { return table.max_size(); } - bool empty() const { return size() == 0; } - size_type bucket_count() const { return table.size(); } - size_type max_bucket_count() const { return max_size(); } - // These are tr1 methods. Their idea of 'bucket' doesn't map well to - // what we do. We just say every bucket has 0 or 1 items in it. - size_type bucket_size(size_type i) const { - return begin(i) == end(i) ? 0 : 1; - } - - private: - // Because of the above, size_type(-1) is never legal; use it for errors - static const size_type ILLEGAL_BUCKET = size_type(-1); - - // Used after a string of deletes. Returns true if we actually shrunk. - // TODO(csilvers): take a delta so we can take into account inserts - // done after shrinking. Maybe make part of the Settings class? - bool maybe_shrink() { - assert(table.num_nonempty() >= num_deleted); - assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two - assert(bucket_count() >= HT_MIN_BUCKETS); - bool retval = false; - - // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS, - // we'll never shrink until you get relatively big, and we'll never - // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something - // like "dense_hash_set<int> x; x.insert(4); x.erase(4);" will - // shrink us down to HT_MIN_BUCKETS buckets, which is too small. - const size_type num_remain = table.num_nonempty() - num_deleted; - const size_type shrink_threshold = settings.shrink_threshold(); - if (shrink_threshold > 0 && num_remain < shrink_threshold && - bucket_count() > HT_DEFAULT_STARTING_BUCKETS) { - const float shrink_factor = settings.shrink_factor(); - size_type sz = bucket_count() / 2; // find how much we should shrink - while (sz > HT_DEFAULT_STARTING_BUCKETS && - num_remain < static_cast<size_type>(sz * shrink_factor)) { - sz /= 2; // stay a power of 2 - } - sparse_hashtable tmp(MoveDontCopy, *this, sz); - swap(tmp); // now we are tmp - retval = true; - } - settings.set_consider_shrink(false); // because we just considered it - return retval; - } - - // We'll let you resize a hashtable -- though this makes us copy all! - // When you resize, you say, "make it big enough for this many more elements" - // Returns true if we actually resized, false if size was already ok. - bool resize_delta(size_type delta) { - bool did_resize = false; - if ( settings.consider_shrink() ) { // see if lots of deletes happened - if ( maybe_shrink() ) - did_resize = true; - } - if (table.num_nonempty() >= - (std::numeric_limits<size_type>::max)() - delta) { - throw std::length_error("resize overflow"); - } - if ( bucket_count() >= HT_MIN_BUCKETS && - (table.num_nonempty() + delta) <= settings.enlarge_threshold() ) - return did_resize; // we're ok as we are - - // Sometimes, we need to resize just to get rid of all the - // "deleted" buckets that are clogging up the hashtable. So when - // deciding whether to resize, count the deleted buckets (which - // are currently taking up room). But later, when we decide what - // size to resize to, *don't* count deleted buckets, since they - // get discarded during the resize. - const size_type needed_size = - settings.min_buckets(table.num_nonempty() + delta, 0); - if ( needed_size <= bucket_count() ) // we have enough buckets - return did_resize; - - size_type resize_to = - settings.min_buckets(table.num_nonempty() - num_deleted + delta, - bucket_count()); - if (resize_to < needed_size && // may double resize_to - resize_to < (std::numeric_limits<size_type>::max)() / 2) { - // This situation means that we have enough deleted elements, - // that once we purge them, we won't actually have needed to - // grow. But we may want to grow anyway: if we just purge one - // element, say, we'll have to grow anyway next time we - // insert. Might as well grow now, since we're already going - // through the trouble of copying (in order to purge the - // deleted elements). - const size_type target = - static_cast<size_type>(settings.shrink_size(resize_to*2)); - if (table.num_nonempty() - num_deleted + delta >= target) { - // Good, we won't be below the shrink threshhold even if we double. - resize_to *= 2; - } - } - - sparse_hashtable tmp(MoveDontCopy, *this, resize_to); - swap(tmp); // now we are tmp - return true; - } - - // Used to actually do the rehashing when we grow/shrink a hashtable - void copy_from(const sparse_hashtable &ht, size_type min_buckets_wanted) { - clear(); // clear table, set num_deleted to 0 - - // If we need to change the size of our table, do it now - const size_type resize_to = - settings.min_buckets(ht.size(), min_buckets_wanted); - if ( resize_to > bucket_count() ) { // we don't have enough buckets - table.resize(resize_to); // sets the number of buckets - settings.reset_thresholds(bucket_count()); - } - - // We use a normal iterator to get non-deleted bcks from ht - // We could use insert() here, but since we know there are - // no duplicates and no deleted items, we can be more efficient - assert((bucket_count() & (bucket_count()-1)) == 0); // a power of two - for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) { - size_type num_probes = 0; // how many times we've probed - size_type bucknum; - const size_type bucket_count_minus_one = bucket_count() - 1; - for (bucknum = hash(get_key(*it)) & bucket_count_minus_one; - table.test(bucknum); // not empty - bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) { - ++num_probes; - assert(num_probes < bucket_count() - && "Hashtable is full: an error in key_equal<> or hash<>"); - } - table.set(bucknum, *it); // copies the value to here - } - settings.inc_num_ht_copies(); - } - - // Implementation is like copy_from, but it destroys the table of the - // "from" guy by freeing sparsetable memory as we iterate. This is - // useful in resizing, since we're throwing away the "from" guy anyway. - void move_from(MoveDontCopyT mover, sparse_hashtable &ht, - size_type min_buckets_wanted) { - clear(); // clear table, set num_deleted to 0 - - // If we need to change the size of our table, do it now - size_type resize_to; - if ( mover == MoveDontGrow ) - resize_to = ht.bucket_count(); // keep same size as old ht - else // MoveDontCopy - resize_to = settings.min_buckets(ht.size(), min_buckets_wanted); - if ( resize_to > bucket_count() ) { // we don't have enough buckets - table.resize(resize_to); // sets the number of buckets - settings.reset_thresholds(bucket_count()); - } - - // We use a normal iterator to get non-deleted bcks from ht - // We could use insert() here, but since we know there are - // no duplicates and no deleted items, we can be more efficient - assert( (bucket_count() & (bucket_count()-1)) == 0); // a power of two - // THIS IS THE MAJOR LINE THAT DIFFERS FROM COPY_FROM(): - for ( destructive_iterator it = ht.destructive_begin(); - it != ht.destructive_end(); ++it ) { - size_type num_probes = 0; // how many times we've probed - size_type bucknum; - for ( bucknum = hash(get_key(*it)) & (bucket_count()-1); // h % buck_cnt - table.test(bucknum); // not empty - bucknum = (bucknum + JUMP_(key, num_probes)) & (bucket_count()-1) ) { - ++num_probes; - assert(num_probes < bucket_count() - && "Hashtable is full: an error in key_equal<> or hash<>"); - } - table.set(bucknum, *it); // copies the value to here - } - settings.inc_num_ht_copies(); - } - - - // Required by the spec for hashed associative container - public: - // Though the docs say this should be num_buckets, I think it's much - // more useful as num_elements. As a special feature, calling with - // req_elements==0 will cause us to shrink if we can, saving space. - void resize(size_type req_elements) { // resize to this or larger - if ( settings.consider_shrink() || req_elements == 0 ) - maybe_shrink(); - if ( req_elements > table.num_nonempty() ) // we only grow - resize_delta(req_elements - table.num_nonempty()); - } - - // Get and change the value of shrink_factor and enlarge_factor. The - // description at the beginning of this file explains how to choose - // the values. Setting the shrink parameter to 0.0 ensures that the - // table never shrinks. - void get_resizing_parameters(float* shrink, float* grow) const { - *shrink = settings.shrink_factor(); - *grow = settings.enlarge_factor(); - } - void set_resizing_parameters(float shrink, float grow) { - settings.set_resizing_parameters(shrink, grow); - settings.reset_thresholds(bucket_count()); - } - - // CONSTRUCTORS -- as required by the specs, we take a size, - // but also let you specify a hashfunction, key comparator, - // and key extractor. We also define a copy constructor and =. - // DESTRUCTOR -- the default is fine, surprisingly. - explicit sparse_hashtable(size_type expected_max_items_in_table = 0, - const HashFcn& hf = HashFcn(), - const EqualKey& eql = EqualKey(), - const ExtractKey& ext = ExtractKey(), - const SetKey& set = SetKey(), - const Alloc& alloc = Alloc()) - : settings(hf), - key_info(ext, set, eql), - num_deleted(0), - table((expected_max_items_in_table == 0 - ? HT_DEFAULT_STARTING_BUCKETS - : settings.min_buckets(expected_max_items_in_table, 0)), - alloc) { - settings.reset_thresholds(bucket_count()); - } - - // As a convenience for resize(), we allow an optional second argument - // which lets you make this new hashtable a different size than ht. - // We also provide a mechanism of saying you want to "move" the ht argument - // into us instead of copying. - sparse_hashtable(const sparse_hashtable& ht, - size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) - : settings(ht.settings), - key_info(ht.key_info), - num_deleted(0), - table(0, ht.get_allocator()) { - settings.reset_thresholds(bucket_count()); - copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries - } - sparse_hashtable(MoveDontCopyT mover, sparse_hashtable& ht, - size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) - : settings(ht.settings), - key_info(ht.key_info), - num_deleted(0), - table(0, ht.get_allocator()) { - settings.reset_thresholds(bucket_count()); - move_from(mover, ht, min_buckets_wanted); // ignores deleted entries - } - - sparse_hashtable& operator= (const sparse_hashtable& ht) { - if (&ht == this) return *this; // don't copy onto ourselves - settings = ht.settings; - key_info = ht.key_info; - num_deleted = ht.num_deleted; - // copy_from() calls clear and sets num_deleted to 0 too - copy_from(ht, HT_MIN_BUCKETS); - // we purposefully don't copy the allocator, which may not be copyable - return *this; - } - - // Many STL algorithms use swap instead of copy constructors - void swap(sparse_hashtable& ht) { - std::swap(settings, ht.settings); - std::swap(key_info, ht.key_info); - std::swap(num_deleted, ht.num_deleted); - table.swap(ht.table); - settings.reset_thresholds(bucket_count()); // also resets consider_shrink - ht.settings.reset_thresholds(ht.bucket_count()); - // we purposefully don't swap the allocator, which may not be swap-able - } - - // It's always nice to be able to clear a table without deallocating it - void clear() { - if (!empty() || (num_deleted != 0)) { - table.clear(); - } - settings.reset_thresholds(bucket_count()); - num_deleted = 0; - } - - // LOOKUP ROUTINES - private: - // Returns a pair of positions: 1st where the object is, 2nd where - // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET - // if object is not found; 2nd is ILLEGAL_BUCKET if it is. - // Note: because of deletions where-to-insert is not trivial: it's the - // first deleted bucket we see, as long as we don't find the key later - std::pair<size_type, size_type> find_position(const key_type &key) const { - size_type num_probes = 0; // how many times we've probed - const size_type bucket_count_minus_one = bucket_count() - 1; - size_type bucknum = hash(key) & bucket_count_minus_one; - size_type insert_pos = ILLEGAL_BUCKET; // where we would insert - SPARSEHASH_STAT_UPDATE(total_lookups += 1); - while ( 1 ) { // probe until something happens - if ( !table.test(bucknum) ) { // bucket is empty - SPARSEHASH_STAT_UPDATE(total_probes += num_probes); - if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert - return std::pair<size_type,size_type>(ILLEGAL_BUCKET, bucknum); - else - return std::pair<size_type,size_type>(ILLEGAL_BUCKET, insert_pos); - - } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert - if ( insert_pos == ILLEGAL_BUCKET ) - insert_pos = bucknum; - - } else if ( equals(key, get_key(table.unsafe_get(bucknum))) ) { - SPARSEHASH_STAT_UPDATE(total_probes += num_probes); - return std::pair<size_type,size_type>(bucknum, ILLEGAL_BUCKET); - } - ++num_probes; // we're doing another probe - bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one; - assert(num_probes < bucket_count() - && "Hashtable is full: an error in key_equal<> or hash<>"); - } - } - - public: - - iterator find(const key_type& key) { - if ( size() == 0 ) return end(); - std::pair<size_type, size_type> pos = find_position(key); - if ( pos.first == ILLEGAL_BUCKET ) // alas, not there - return end(); - else - return iterator(this, table.get_iter(pos.first), table.nonempty_end()); - } - - const_iterator find(const key_type& key) const { - if ( size() == 0 ) return end(); - std::pair<size_type, size_type> pos = find_position(key); - if ( pos.first == ILLEGAL_BUCKET ) // alas, not there - return end(); - else - return const_iterator(this, - table.get_iter(pos.first), table.nonempty_end()); - } - - // This is a tr1 method: the bucket a given key is in, or what bucket - // it would be put in, if it were to be inserted. Shrug. - size_type bucket(const key_type& key) const { - std::pair<size_type, size_type> pos = find_position(key); - return pos.first == ILLEGAL_BUCKET ? pos.second : pos.first; - } - - // Counts how many elements have key key. For maps, it's either 0 or 1. - size_type count(const key_type &key) const { - std::pair<size_type, size_type> pos = find_position(key); - return pos.first == ILLEGAL_BUCKET ? 0 : 1; - } - - // Likewise, equal_range doesn't really make sense for us. Oh well. - std::pair<iterator,iterator> equal_range(const key_type& key) { - iterator pos = find(key); // either an iterator or end - if (pos == end()) { - return std::pair<iterator,iterator>(pos, pos); - } else { - const iterator startpos = pos++; - return std::pair<iterator,iterator>(startpos, pos); - } - } - std::pair<const_iterator,const_iterator> equal_range(const key_type& key) - const { - const_iterator pos = find(key); // either an iterator or end - if (pos == end()) { - return std::pair<const_iterator,const_iterator>(pos, pos); - } else { - const const_iterator startpos = pos++; - return std::pair<const_iterator,const_iterator>(startpos, pos); - } - } - - - // INSERTION ROUTINES - private: - // Private method used by insert_noresize and find_or_insert. - iterator insert_at(const_reference obj, size_type pos) { - if (size() >= max_size()) { - throw std::length_error("insert overflow"); - } - if ( test_deleted(pos) ) { // just replace if it's been deleted - // The set() below will undelete this object. We just worry about stats - assert(num_deleted > 0); - --num_deleted; // used to be, now it isn't - } - table.set(pos, obj); - return iterator(this, table.get_iter(pos), table.nonempty_end()); - } - - // If you know *this is big enough to hold obj, use this routine - std::pair<iterator, bool> insert_noresize(const_reference obj) { - // First, double-check we're not inserting delkey - assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey)) - && "Inserting the deleted key"); - const std::pair<size_type,size_type> pos = find_position(get_key(obj)); - if ( pos.first != ILLEGAL_BUCKET) { // object was already there - return std::pair<iterator,bool>(iterator(this, table.get_iter(pos.first), - table.nonempty_end()), - false); // false: we didn't insert - } else { // pos.second says where to put it - return std::pair<iterator,bool>(insert_at(obj, pos.second), true); - } - } - - // Specializations of insert(it, it) depending on the power of the iterator: - // (1) Iterator supports operator-, resize before inserting - template <class ForwardIterator> - void insert(ForwardIterator f, ForwardIterator l, std::forward_iterator_tag) { - size_t dist = std::distance(f, l); - if (dist >= (std::numeric_limits<size_type>::max)()) { - throw std::length_error("insert-range overflow"); - } - resize_delta(static_cast<size_type>(dist)); - for ( ; dist > 0; --dist, ++f) { - insert_noresize(*f); - } - } - - // (2) Arbitrary iterator, can't tell how much to resize - template <class InputIterator> - void insert(InputIterator f, InputIterator l, std::input_iterator_tag) { - for ( ; f != l; ++f) - insert(*f); - } - - public: - // This is the normal insert routine, used by the outside world - std::pair<iterator, bool> insert(const_reference obj) { - resize_delta(1); // adding an object, grow if need be - return insert_noresize(obj); - } - - // When inserting a lot at a time, we specialize on the type of iterator - template <class InputIterator> - void insert(InputIterator f, InputIterator l) { - // specializes on iterator type - insert(f, l, - typename std::iterator_traits<InputIterator>::iterator_category()); - } - - // DefaultValue is a functor that takes a key and returns a value_type - // representing the default value to be inserted if none is found. - template <class DefaultValue> - value_type& find_or_insert(const key_type& key) { - // First, double-check we're not inserting delkey - assert((!settings.use_deleted() || !equals(key, key_info.delkey)) - && "Inserting the deleted key"); - const std::pair<size_type,size_type> pos = find_position(key); - DefaultValue default_value; - if ( pos.first != ILLEGAL_BUCKET) { // object was already there - return *table.get_iter(pos.first); - } else if (resize_delta(1)) { // needed to rehash to make room - // Since we resized, we can't use pos, so recalculate where to insert. - return *insert_noresize(default_value(key)).first; - } else { // no need to rehash, insert right here - return *insert_at(default_value(key), pos.second); - } - } - - // DELETION ROUTINES - size_type erase(const key_type& key) { - // First, double-check we're not erasing delkey. - assert((!settings.use_deleted() || !equals(key, key_info.delkey)) - && "Erasing the deleted key"); - assert(!settings.use_deleted() || !equals(key, key_info.delkey)); - const_iterator pos = find(key); // shrug: shouldn't need to be const - if ( pos != end() ) { - assert(!test_deleted(pos)); // or find() shouldn't have returned it - set_deleted(pos); - ++num_deleted; - // will think about shrink after next insert - settings.set_consider_shrink(true); - return 1; // because we deleted one thing - } else { - return 0; // because we deleted nothing - } - } - - // We return the iterator past the deleted item. - void erase(iterator pos) { - if ( pos == end() ) return; // sanity check - if ( set_deleted(pos) ) { // true if object has been newly deleted - ++num_deleted; - // will think about shrink after next insert - settings.set_consider_shrink(true); - } - } - - void erase(iterator f, iterator l) { - for ( ; f != l; ++f) { - if ( set_deleted(f) ) // should always be true - ++num_deleted; - } - // will think about shrink after next insert - settings.set_consider_shrink(true); - } - - // We allow you to erase a const_iterator just like we allow you to - // erase an iterator. This is in parallel to 'delete': you can delete - // a const pointer just like a non-const pointer. The logic is that - // you can't use the object after it's erased anyway, so it doesn't matter - // if it's const or not. - void erase(const_iterator pos) { - if ( pos == end() ) return; // sanity check - if ( set_deleted(pos) ) { // true if object has been newly deleted - ++num_deleted; - // will think about shrink after next insert - settings.set_consider_shrink(true); - } - } - void erase(const_iterator f, const_iterator l) { - for ( ; f != l; ++f) { - if ( set_deleted(f) ) // should always be true - ++num_deleted; - } - // will think about shrink after next insert - settings.set_consider_shrink(true); - } - - - // COMPARISON - bool operator==(const sparse_hashtable& ht) const { - if (size() != ht.size()) { - return false; - } else if (this == &ht) { - return true; - } else { - // Iterate through the elements in "this" and see if the - // corresponding element is in ht - for ( const_iterator it = begin(); it != end(); ++it ) { - const_iterator it2 = ht.find(get_key(*it)); - if ((it2 == ht.end()) || (*it != *it2)) { - return false; - } - } - return true; - } - } - bool operator!=(const sparse_hashtable& ht) const { - return !(*this == ht); - } - - - // I/O - // We support reading and writing hashtables to disk. NOTE that - // this only stores the hashtable metadata, not the stuff you've - // actually put in the hashtable! Alas, since I don't know how to - // write a hasher or key_equal, you have to make sure everything - // but the table is the same. We compact before writing. - // - // The OUTPUT type needs to support a Write() operation. File and - // OutputBuffer are appropriate types to pass in. - // - // The INPUT type needs to support a Read() operation. File and - // InputBuffer are appropriate types to pass in. - template <typename OUTPUT> - bool write_metadata(OUTPUT *fp) { - squash_deleted(); // so we don't have to worry about delkey - return table.write_metadata(fp); - } - - template <typename INPUT> - bool read_metadata(INPUT *fp) { - num_deleted = 0; // since we got rid before writing - const bool result = table.read_metadata(fp); - settings.reset_thresholds(bucket_count()); - return result; - } - - // Only meaningful if value_type is a POD. - template <typename OUTPUT> - bool write_nopointer_data(OUTPUT *fp) { - return table.write_nopointer_data(fp); - } - - // Only meaningful if value_type is a POD. - template <typename INPUT> - bool read_nopointer_data(INPUT *fp) { - return table.read_nopointer_data(fp); - } - - // INPUT and OUTPUT must be either a FILE, *or* a C++ stream - // (istream, ostream, etc) *or* a class providing - // Read(void*, size_t) and Write(const void*, size_t) - // (respectively), which writes a buffer into a stream - // (which the INPUT/OUTPUT instance presumably owns). - - typedef sparsehash_internal::pod_serializer<value_type> NopointerSerializer; - - // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT *fp) { - squash_deleted(); // so we don't have to worry about delkey - return table.serialize(serializer, fp); - } - - // ValueSerializer: a functor. operator()(INPUT*, value_type*) - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT *fp) { - num_deleted = 0; // since we got rid before writing - const bool result = table.unserialize(serializer, fp); - settings.reset_thresholds(bucket_count()); - return result; - } - - private: - // Table is the main storage class. - typedef sparsetable<value_type, DEFAULT_GROUP_SIZE, value_alloc_type> Table; - - // Package templated functors with the other types to eliminate memory - // needed for storing these zero-size operators. Since ExtractKey and - // hasher's operator() might have the same function signature, they - // must be packaged in different classes. - struct Settings : - sparsehash_internal::sh_hashtable_settings<key_type, hasher, - size_type, HT_MIN_BUCKETS> { - explicit Settings(const hasher& hf) - : sparsehash_internal::sh_hashtable_settings<key_type, hasher, - size_type, HT_MIN_BUCKETS>( - hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {} - }; - - // KeyInfo stores delete key and packages zero-size functors: - // ExtractKey and SetKey. - class KeyInfo : public ExtractKey, public SetKey, public EqualKey { - public: - KeyInfo(const ExtractKey& ek, const SetKey& sk, const EqualKey& eq) - : ExtractKey(ek), - SetKey(sk), - EqualKey(eq) { - } - // We want to return the exact same type as ExtractKey: Key or const Key& - typename ExtractKey::result_type get_key(const_reference v) const { - return ExtractKey::operator()(v); - } - void set_key(pointer v, const key_type& k) const { - SetKey::operator()(v, k); - } - bool equals(const key_type& a, const key_type& b) const { - return EqualKey::operator()(a, b); - } - - // Which key marks deleted entries. - // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!) - typename base::remove_const<key_type>::type delkey; - }; - - // Utility functions to access the templated operators - size_type hash(const key_type& v) const { - return settings.hash(v); - } - bool equals(const key_type& a, const key_type& b) const { - return key_info.equals(a, b); - } - typename ExtractKey::result_type get_key(const_reference v) const { - return key_info.get_key(v); - } - void set_key(pointer v, const key_type& k) const { - key_info.set_key(v, k); - } - - private: - // Actual data - Settings settings; - KeyInfo key_info; - size_type num_deleted; // how many occupied buckets are marked deleted - Table table; // holds num_buckets and num_elements too -}; - - -// We need a global swap as well -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -inline void swap(sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> &x, - sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> &y) { - x.swap(y); -} - -#undef JUMP_ - -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -const typename sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::size_type - sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET; - -// How full we let the table get before we resize. Knuth says .8 is -// good -- higher causes us to probe too much, though saves memory -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -const int sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT = 80; - -// How empty we let the table get before we resize lower. -// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing -template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> -const int sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_PCT - = static_cast<int>(0.4 * - sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT); - -_END_GOOGLE_NAMESPACE_ - -#endif /* _SPARSEHASHTABLE_H_ */ + st_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::pointer pointer; + + // "Real" constructor and default constructor + sparse_hashtable_destructive_iterator(const + sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h, + st_iterator it, st_iterator it_end) + : ht(h), pos(it), end(it_end) { advance_past_deleted(); } + sparse_hashtable_destructive_iterator() { } // never used internally + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on a marked-deleted array element + void advance_past_deleted() { + while ( pos != end && ht->test_deleted(*this) ) + ++pos; + } + iterator& operator++() { + assert(pos != end); ++pos; advance_past_deleted(); return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const iterator& it) const { return pos == it.pos; } + bool operator!=(const iterator& it) const { return pos != it.pos; } + + + // The actual data + const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *ht; + st_iterator pos, end; +}; + + +template <class Value, class Key, class HashFcn, + class ExtractKey, class SetKey, class EqualKey, class Alloc> +class sparse_hashtable { + private: + typedef typename Alloc::template rebind<Value>::other value_alloc_type; + + public: + typedef Key key_type; + typedef Value value_type; + typedef HashFcn hasher; + typedef EqualKey key_equal; + typedef Alloc allocator_type; + + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + typedef sparse_hashtable_iterator<Value, Key, HashFcn, ExtractKey, + SetKey, EqualKey, Alloc> + iterator; + + typedef sparse_hashtable_const_iterator<Value, Key, HashFcn, ExtractKey, + SetKey, EqualKey, Alloc> + const_iterator; + + typedef sparse_hashtable_destructive_iterator<Value, Key, HashFcn, ExtractKey, + SetKey, EqualKey, Alloc> + destructive_iterator; + + // These come from tr1. For us they're the same as regular iterators. + typedef iterator local_iterator; + typedef const_iterator const_local_iterator; + + // How full we let the table get before we resize, by default. + // Knuth says .8 is good -- higher causes us to probe too much, + // though it saves memory. + static const int HT_OCCUPANCY_PCT; // = 80 (out of 100); + + // How empty we let the table get before we resize lower, by default. + // (0.0 means never resize lower.) + // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing + static const int HT_EMPTY_PCT; // = 0.4 * HT_OCCUPANCY_PCT; + + // Minimum size we're willing to let hashtables be. + // Must be a power of two, and at least 4. + // Note, however, that for a given hashtable, the initial size is a + // function of the first constructor arg, and may be >HT_MIN_BUCKETS. + static const size_type HT_MIN_BUCKETS = 4; + + // By default, if you don't specify a hashtable size at + // construction-time, we use this size. Must be a power of two, and + // at least HT_MIN_BUCKETS. + static const size_type HT_DEFAULT_STARTING_BUCKETS = 32; + + // ITERATOR FUNCTIONS + iterator begin() { return iterator(this, table.nonempty_begin(), + table.nonempty_end()); } + iterator end() { return iterator(this, table.nonempty_end(), + table.nonempty_end()); } + const_iterator begin() const { return const_iterator(this, + table.nonempty_begin(), + table.nonempty_end()); } + const_iterator end() const { return const_iterator(this, + table.nonempty_end(), + table.nonempty_end()); } + + // These come from tr1 unordered_map. They iterate over 'bucket' n. + // For sparsehashtable, we could consider each 'group' to be a bucket, + // I guess, but I don't really see the point. We'll just consider + // bucket n to be the n-th element of the sparsetable, if it's occupied, + // or some empty element, otherwise. + local_iterator begin(size_type i) { + if (table.test(i)) + return local_iterator(this, table.get_iter(i), table.nonempty_end()); + else + return local_iterator(this, table.nonempty_end(), table.nonempty_end()); + } + local_iterator end(size_type i) { + local_iterator it = begin(i); + if (table.test(i) && !test_deleted(i)) + ++it; + return it; + } + const_local_iterator begin(size_type i) const { + if (table.test(i)) + return const_local_iterator(this, table.get_iter(i), + table.nonempty_end()); + else + return const_local_iterator(this, table.nonempty_end(), + table.nonempty_end()); + } + const_local_iterator end(size_type i) const { + const_local_iterator it = begin(i); + if (table.test(i) && !test_deleted(i)) + ++it; + return it; + } + + // This is used when resizing + destructive_iterator destructive_begin() { + return destructive_iterator(this, table.destructive_begin(), + table.destructive_end()); + } + destructive_iterator destructive_end() { + return destructive_iterator(this, table.destructive_end(), + table.destructive_end()); + } + + + // ACCESSOR FUNCTIONS for the things we templatize on, basically + hasher hash_funct() const { return settings; } + key_equal key_eq() const { return key_info; } + allocator_type get_allocator() const { return table.get_allocator(); } + + // Accessor function for statistics gathering. + int num_table_copies() const { return settings.num_ht_copies(); } + + private: + // We need to copy values when we set the special marker for deleted + // elements, but, annoyingly, we can't just use the copy assignment + // operator because value_type might not be assignable (it's often + // pair<const X, Y>). We use explicit destructor invocation and + // placement new to get around this. Arg. + void set_value(pointer dst, const_reference src) { + dst->~value_type(); // delete the old value, if any + new(dst) value_type(src); + } + + // This is used as a tag for the copy constructor, saying to destroy its + // arg We have two ways of destructively copying: with potentially growing + // the hashtable as we copy, and without. To make sure the outside world + // can't do a destructive copy, we make the typename private. + enum MoveDontCopyT {MoveDontCopy, MoveDontGrow}; + + // DELETE HELPER FUNCTIONS + // This lets the user describe a key that will indicate deleted + // table entries. This key should be an "impossible" entry -- + // if you try to insert it for real, you won't be able to retrieve it! + // (NB: while you pass in an entire value, only the key part is looked + // at. This is just because I don't know how to assign just a key.) + private: + void squash_deleted() { // gets rid of any deleted entries we have + if ( num_deleted ) { // get rid of deleted before writing + sparse_hashtable tmp(MoveDontGrow, *this); + swap(tmp); // now we are tmp + } + assert(num_deleted == 0); + } + + // Test if the given key is the deleted indicator. Requires + // num_deleted > 0, for correctness of read(), and because that + // guarantees that key_info.delkey is valid. + bool test_deleted_key(const key_type& key) const { + assert(num_deleted > 0); + return equals(key_info.delkey, key); + } + + public: + void set_deleted_key(const key_type &key) { + // It's only safe to change what "deleted" means if we purge deleted guys + squash_deleted(); + settings.set_use_deleted(true); + key_info.delkey = key; + } + void clear_deleted_key() { + squash_deleted(); + settings.set_use_deleted(false); + } + key_type deleted_key() const { + assert(settings.use_deleted() + && "Must set deleted key before calling deleted_key"); + return key_info.delkey; + } + + // These are public so the iterators can use them + // True if the item at position bucknum is "deleted" marker + bool test_deleted(size_type bucknum) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && table.test(bucknum) && + test_deleted_key(get_key(table.unsafe_get(bucknum))); + } + bool test_deleted(const iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + bool test_deleted(const const_iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + bool test_deleted(const destructive_iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + + private: + void check_use_deleted(const char* caller) { + (void)caller; // could log it if the assert failed + assert(settings.use_deleted()); + } + + // Set it so test_deleted is true. true if object didn't used to be deleted. + // TODO(csilvers): make these private (also in densehashtable.h) + bool set_deleted(iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + // &* converts from iterator to value-type. + set_key(&(*it), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(iterator &it) { + check_use_deleted("clear_deleted()"); + // Happens automatically when we assign something else in its place. + return test_deleted(it); + } + + // We also allow to set/clear the deleted bit on a const iterator. + // We allow a const_iterator for the same reason you can delete a + // const pointer: it's convenient, and semantically you can't use + // 'it' after it's been deleted anyway, so its const-ness doesn't + // really matter. + bool set_deleted(const_iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + set_key(const_cast<pointer>(&(*it)), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(const_iterator &it) { + check_use_deleted("clear_deleted()"); + return test_deleted(it); + } + + // FUNCTIONS CONCERNING SIZE + public: + size_type size() const { return table.num_nonempty() - num_deleted; } + size_type max_size() const { return table.max_size(); } + bool empty() const { return size() == 0; } + size_type bucket_count() const { return table.size(); } + size_type max_bucket_count() const { return max_size(); } + // These are tr1 methods. Their idea of 'bucket' doesn't map well to + // what we do. We just say every bucket has 0 or 1 items in it. + size_type bucket_size(size_type i) const { + return begin(i) == end(i) ? 0 : 1; + } + + private: + // Because of the above, size_type(-1) is never legal; use it for errors + static const size_type ILLEGAL_BUCKET = size_type(-1); + + // Used after a string of deletes. Returns true if we actually shrunk. + // TODO(csilvers): take a delta so we can take into account inserts + // done after shrinking. Maybe make part of the Settings class? + bool maybe_shrink() { + assert(table.num_nonempty() >= num_deleted); + assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two + assert(bucket_count() >= HT_MIN_BUCKETS); + bool retval = false; + + // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS, + // we'll never shrink until you get relatively big, and we'll never + // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something + // like "dense_hash_set<int> x; x.insert(4); x.erase(4);" will + // shrink us down to HT_MIN_BUCKETS buckets, which is too small. + const size_type num_remain = table.num_nonempty() - num_deleted; + const size_type shrink_threshold = settings.shrink_threshold(); + if (shrink_threshold > 0 && num_remain < shrink_threshold && + bucket_count() > HT_DEFAULT_STARTING_BUCKETS) { + const float shrink_factor = settings.shrink_factor(); + size_type sz = bucket_count() / 2; // find how much we should shrink + while (sz > HT_DEFAULT_STARTING_BUCKETS && + num_remain < static_cast<size_type>(sz * shrink_factor)) { + sz /= 2; // stay a power of 2 + } + sparse_hashtable tmp(MoveDontCopy, *this, sz); + swap(tmp); // now we are tmp + retval = true; + } + settings.set_consider_shrink(false); // because we just considered it + return retval; + } + + // We'll let you resize a hashtable -- though this makes us copy all! + // When you resize, you say, "make it big enough for this many more elements" + // Returns true if we actually resized, false if size was already ok. + bool resize_delta(size_type delta) { + bool did_resize = false; + if ( settings.consider_shrink() ) { // see if lots of deletes happened + if ( maybe_shrink() ) + did_resize = true; + } + if (table.num_nonempty() >= + (std::numeric_limits<size_type>::max)() - delta) { + throw std::length_error("resize overflow"); + } + if ( bucket_count() >= HT_MIN_BUCKETS && + (table.num_nonempty() + delta) <= settings.enlarge_threshold() ) + return did_resize; // we're ok as we are + + // Sometimes, we need to resize just to get rid of all the + // "deleted" buckets that are clogging up the hashtable. So when + // deciding whether to resize, count the deleted buckets (which + // are currently taking up room). But later, when we decide what + // size to resize to, *don't* count deleted buckets, since they + // get discarded during the resize. + const size_type needed_size = + settings.min_buckets(table.num_nonempty() + delta, 0); + if ( needed_size <= bucket_count() ) // we have enough buckets + return did_resize; + + size_type resize_to = + settings.min_buckets(table.num_nonempty() - num_deleted + delta, + bucket_count()); + if (resize_to < needed_size && // may double resize_to + resize_to < (std::numeric_limits<size_type>::max)() / 2) { + // This situation means that we have enough deleted elements, + // that once we purge them, we won't actually have needed to + // grow. But we may want to grow anyway: if we just purge one + // element, say, we'll have to grow anyway next time we + // insert. Might as well grow now, since we're already going + // through the trouble of copying (in order to purge the + // deleted elements). + const size_type target = + static_cast<size_type>(settings.shrink_size(resize_to*2)); + if (table.num_nonempty() - num_deleted + delta >= target) { + // Good, we won't be below the shrink threshhold even if we double. + resize_to *= 2; + } + } + + sparse_hashtable tmp(MoveDontCopy, *this, resize_to); + swap(tmp); // now we are tmp + return true; + } + + // Used to actually do the rehashing when we grow/shrink a hashtable + void copy_from(const sparse_hashtable &ht, size_type min_buckets_wanted) { + clear(); // clear table, set num_deleted to 0 + + // If we need to change the size of our table, do it now + const size_type resize_to = + settings.min_buckets(ht.size(), min_buckets_wanted); + if ( resize_to > bucket_count() ) { // we don't have enough buckets + table.resize(resize_to); // sets the number of buckets + settings.reset_thresholds(bucket_count()); + } + + // We use a normal iterator to get non-deleted bcks from ht + // We could use insert() here, but since we know there are + // no duplicates and no deleted items, we can be more efficient + assert((bucket_count() & (bucket_count()-1)) == 0); // a power of two + for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) { + size_type num_probes = 0; // how many times we've probed + size_type bucknum; + const size_type bucket_count_minus_one = bucket_count() - 1; + for (bucknum = hash(get_key(*it)) & bucket_count_minus_one; + table.test(bucknum); // not empty + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) { + ++num_probes; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + table.set(bucknum, *it); // copies the value to here + } + settings.inc_num_ht_copies(); + } + + // Implementation is like copy_from, but it destroys the table of the + // "from" guy by freeing sparsetable memory as we iterate. This is + // useful in resizing, since we're throwing away the "from" guy anyway. + void move_from(MoveDontCopyT mover, sparse_hashtable &ht, + size_type min_buckets_wanted) { + clear(); // clear table, set num_deleted to 0 + + // If we need to change the size of our table, do it now + size_type resize_to; + if ( mover == MoveDontGrow ) + resize_to = ht.bucket_count(); // keep same size as old ht + else // MoveDontCopy + resize_to = settings.min_buckets(ht.size(), min_buckets_wanted); + if ( resize_to > bucket_count() ) { // we don't have enough buckets + table.resize(resize_to); // sets the number of buckets + settings.reset_thresholds(bucket_count()); + } + + // We use a normal iterator to get non-deleted bcks from ht + // We could use insert() here, but since we know there are + // no duplicates and no deleted items, we can be more efficient + assert( (bucket_count() & (bucket_count()-1)) == 0); // a power of two + // THIS IS THE MAJOR LINE THAT DIFFERS FROM COPY_FROM(): + for ( destructive_iterator it = ht.destructive_begin(); + it != ht.destructive_end(); ++it ) { + size_type num_probes = 0; // how many times we've probed + size_type bucknum; + for ( bucknum = hash(get_key(*it)) & (bucket_count()-1); // h % buck_cnt + table.test(bucknum); // not empty + bucknum = (bucknum + JUMP_(key, num_probes)) & (bucket_count()-1) ) { + ++num_probes; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + table.set(bucknum, *it); // copies the value to here + } + settings.inc_num_ht_copies(); + } + + + // Required by the spec for hashed associative container + public: + // Though the docs say this should be num_buckets, I think it's much + // more useful as num_elements. As a special feature, calling with + // req_elements==0 will cause us to shrink if we can, saving space. + void resize(size_type req_elements) { // resize to this or larger + if ( settings.consider_shrink() || req_elements == 0 ) + maybe_shrink(); + if ( req_elements > table.num_nonempty() ) // we only grow + resize_delta(req_elements - table.num_nonempty()); + } + + // Get and change the value of shrink_factor and enlarge_factor. The + // description at the beginning of this file explains how to choose + // the values. Setting the shrink parameter to 0.0 ensures that the + // table never shrinks. + void get_resizing_parameters(float* shrink, float* grow) const { + *shrink = settings.shrink_factor(); + *grow = settings.enlarge_factor(); + } + void set_resizing_parameters(float shrink, float grow) { + settings.set_resizing_parameters(shrink, grow); + settings.reset_thresholds(bucket_count()); + } + + // CONSTRUCTORS -- as required by the specs, we take a size, + // but also let you specify a hashfunction, key comparator, + // and key extractor. We also define a copy constructor and =. + // DESTRUCTOR -- the default is fine, surprisingly. + explicit sparse_hashtable(size_type expected_max_items_in_table = 0, + const HashFcn& hf = HashFcn(), + const EqualKey& eql = EqualKey(), + const ExtractKey& ext = ExtractKey(), + const SetKey& set = SetKey(), + const Alloc& alloc = Alloc()) + : settings(hf), + key_info(ext, set, eql), + num_deleted(0), + table((expected_max_items_in_table == 0 + ? HT_DEFAULT_STARTING_BUCKETS + : settings.min_buckets(expected_max_items_in_table, 0)), + alloc) { + settings.reset_thresholds(bucket_count()); + } + + // As a convenience for resize(), we allow an optional second argument + // which lets you make this new hashtable a different size than ht. + // We also provide a mechanism of saying you want to "move" the ht argument + // into us instead of copying. + sparse_hashtable(const sparse_hashtable& ht, + size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) + : settings(ht.settings), + key_info(ht.key_info), + num_deleted(0), + table(0, ht.get_allocator()) { + settings.reset_thresholds(bucket_count()); + copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries + } + sparse_hashtable(MoveDontCopyT mover, sparse_hashtable& ht, + size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) + : settings(ht.settings), + key_info(ht.key_info), + num_deleted(0), + table(0, ht.get_allocator()) { + settings.reset_thresholds(bucket_count()); + move_from(mover, ht, min_buckets_wanted); // ignores deleted entries + } + + sparse_hashtable& operator= (const sparse_hashtable& ht) { + if (&ht == this) return *this; // don't copy onto ourselves + settings = ht.settings; + key_info = ht.key_info; + num_deleted = ht.num_deleted; + // copy_from() calls clear and sets num_deleted to 0 too + copy_from(ht, HT_MIN_BUCKETS); + // we purposefully don't copy the allocator, which may not be copyable + return *this; + } + + // Many STL algorithms use swap instead of copy constructors + void swap(sparse_hashtable& ht) { + std::swap(settings, ht.settings); + std::swap(key_info, ht.key_info); + std::swap(num_deleted, ht.num_deleted); + table.swap(ht.table); + settings.reset_thresholds(bucket_count()); // also resets consider_shrink + ht.settings.reset_thresholds(ht.bucket_count()); + // we purposefully don't swap the allocator, which may not be swap-able + } + + // It's always nice to be able to clear a table without deallocating it + void clear() { + if (!empty() || (num_deleted != 0)) { + table.clear(); + } + settings.reset_thresholds(bucket_count()); + num_deleted = 0; + } + + // LOOKUP ROUTINES + private: + // Returns a pair of positions: 1st where the object is, 2nd where + // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET + // if object is not found; 2nd is ILLEGAL_BUCKET if it is. + // Note: because of deletions where-to-insert is not trivial: it's the + // first deleted bucket we see, as long as we don't find the key later + std::pair<size_type, size_type> find_position(const key_type &key) const { + size_type num_probes = 0; // how many times we've probed + const size_type bucket_count_minus_one = bucket_count() - 1; + size_type bucknum = hash(key) & bucket_count_minus_one; + size_type insert_pos = ILLEGAL_BUCKET; // where we would insert + SPARSEHASH_STAT_UPDATE(total_lookups += 1); + while ( 1 ) { // probe until something happens + if ( !table.test(bucknum) ) { // bucket is empty + SPARSEHASH_STAT_UPDATE(total_probes += num_probes); + if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert + return std::pair<size_type,size_type>(ILLEGAL_BUCKET, bucknum); + else + return std::pair<size_type,size_type>(ILLEGAL_BUCKET, insert_pos); + + } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert + if ( insert_pos == ILLEGAL_BUCKET ) + insert_pos = bucknum; + + } else if ( equals(key, get_key(table.unsafe_get(bucknum))) ) { + SPARSEHASH_STAT_UPDATE(total_probes += num_probes); + return std::pair<size_type,size_type>(bucknum, ILLEGAL_BUCKET); + } + ++num_probes; // we're doing another probe + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + } + + public: + + iterator find(const key_type& key) { + if ( size() == 0 ) return end(); + std::pair<size_type, size_type> pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return iterator(this, table.get_iter(pos.first), table.nonempty_end()); + } + + const_iterator find(const key_type& key) const { + if ( size() == 0 ) return end(); + std::pair<size_type, size_type> pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return const_iterator(this, + table.get_iter(pos.first), table.nonempty_end()); + } + + // This is a tr1 method: the bucket a given key is in, or what bucket + // it would be put in, if it were to be inserted. Shrug. + size_type bucket(const key_type& key) const { + std::pair<size_type, size_type> pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? pos.second : pos.first; + } + + // Counts how many elements have key key. For maps, it's either 0 or 1. + size_type count(const key_type &key) const { + std::pair<size_type, size_type> pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? 0 : 1; + } + + // Likewise, equal_range doesn't really make sense for us. Oh well. + std::pair<iterator,iterator> equal_range(const key_type& key) { + iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair<iterator,iterator>(pos, pos); + } else { + const iterator startpos = pos++; + return std::pair<iterator,iterator>(startpos, pos); + } + } + std::pair<const_iterator,const_iterator> equal_range(const key_type& key) + const { + const_iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair<const_iterator,const_iterator>(pos, pos); + } else { + const const_iterator startpos = pos++; + return std::pair<const_iterator,const_iterator>(startpos, pos); + } + } + + + // INSERTION ROUTINES + private: + // Private method used by insert_noresize and find_or_insert. + iterator insert_at(const_reference obj, size_type pos) { + if (size() >= max_size()) { + throw std::length_error("insert overflow"); + } + if ( test_deleted(pos) ) { // just replace if it's been deleted + // The set() below will undelete this object. We just worry about stats + assert(num_deleted > 0); + --num_deleted; // used to be, now it isn't + } + table.set(pos, obj); + return iterator(this, table.get_iter(pos), table.nonempty_end()); + } + + // If you know *this is big enough to hold obj, use this routine + std::pair<iterator, bool> insert_noresize(const_reference obj) { + // First, double-check we're not inserting delkey + assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey)) + && "Inserting the deleted key"); + const std::pair<size_type,size_type> pos = find_position(get_key(obj)); + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return std::pair<iterator,bool>(iterator(this, table.get_iter(pos.first), + table.nonempty_end()), + false); // false: we didn't insert + } else { // pos.second says where to put it + return std::pair<iterator,bool>(insert_at(obj, pos.second), true); + } + } + + // Specializations of insert(it, it) depending on the power of the iterator: + // (1) Iterator supports operator-, resize before inserting + template <class ForwardIterator> + void insert(ForwardIterator f, ForwardIterator l, std::forward_iterator_tag) { + size_t dist = std::distance(f, l); + if (dist >= (std::numeric_limits<size_type>::max)()) { + throw std::length_error("insert-range overflow"); + } + resize_delta(static_cast<size_type>(dist)); + for ( ; dist > 0; --dist, ++f) { + insert_noresize(*f); + } + } + + // (2) Arbitrary iterator, can't tell how much to resize + template <class InputIterator> + void insert(InputIterator f, InputIterator l, std::input_iterator_tag) { + for ( ; f != l; ++f) + insert(*f); + } + + public: + // This is the normal insert routine, used by the outside world + std::pair<iterator, bool> insert(const_reference obj) { + resize_delta(1); // adding an object, grow if need be + return insert_noresize(obj); + } + + // When inserting a lot at a time, we specialize on the type of iterator + template <class InputIterator> + void insert(InputIterator f, InputIterator l) { + // specializes on iterator type + insert(f, l, + typename std::iterator_traits<InputIterator>::iterator_category()); + } + + // DefaultValue is a functor that takes a key and returns a value_type + // representing the default value to be inserted if none is found. + template <class DefaultValue> + value_type& find_or_insert(const key_type& key) { + // First, double-check we're not inserting delkey + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Inserting the deleted key"); + const std::pair<size_type,size_type> pos = find_position(key); + DefaultValue default_value; + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return *table.get_iter(pos.first); + } else if (resize_delta(1)) { // needed to rehash to make room + // Since we resized, we can't use pos, so recalculate where to insert. + return *insert_noresize(default_value(key)).first; + } else { // no need to rehash, insert right here + return *insert_at(default_value(key), pos.second); + } + } + + // DELETION ROUTINES + size_type erase(const key_type& key) { + // First, double-check we're not erasing delkey. + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Erasing the deleted key"); + assert(!settings.use_deleted() || !equals(key, key_info.delkey)); + const_iterator pos = find(key); // shrug: shouldn't need to be const + if ( pos != end() ) { + assert(!test_deleted(pos)); // or find() shouldn't have returned it + set_deleted(pos); + ++num_deleted; + // will think about shrink after next insert + settings.set_consider_shrink(true); + return 1; // because we deleted one thing + } else { + return 0; // because we deleted nothing + } + } + + // We return the iterator past the deleted item. + void erase(iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + } + + void erase(iterator f, iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + + // We allow you to erase a const_iterator just like we allow you to + // erase an iterator. This is in parallel to 'delete': you can delete + // a const pointer just like a non-const pointer. The logic is that + // you can't use the object after it's erased anyway, so it doesn't matter + // if it's const or not. + void erase(const_iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + } + void erase(const_iterator f, const_iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + + + // COMPARISON + bool operator==(const sparse_hashtable& ht) const { + if (size() != ht.size()) { + return false; + } else if (this == &ht) { + return true; + } else { + // Iterate through the elements in "this" and see if the + // corresponding element is in ht + for ( const_iterator it = begin(); it != end(); ++it ) { + const_iterator it2 = ht.find(get_key(*it)); + if ((it2 == ht.end()) || (*it != *it2)) { + return false; + } + } + return true; + } + } + bool operator!=(const sparse_hashtable& ht) const { + return !(*this == ht); + } + + + // I/O + // We support reading and writing hashtables to disk. NOTE that + // this only stores the hashtable metadata, not the stuff you've + // actually put in the hashtable! Alas, since I don't know how to + // write a hasher or key_equal, you have to make sure everything + // but the table is the same. We compact before writing. + // + // The OUTPUT type needs to support a Write() operation. File and + // OutputBuffer are appropriate types to pass in. + // + // The INPUT type needs to support a Read() operation. File and + // InputBuffer are appropriate types to pass in. + template <typename OUTPUT> + bool write_metadata(OUTPUT *fp) { + squash_deleted(); // so we don't have to worry about delkey + return table.write_metadata(fp); + } + + template <typename INPUT> + bool read_metadata(INPUT *fp) { + num_deleted = 0; // since we got rid before writing + const bool result = table.read_metadata(fp); + settings.reset_thresholds(bucket_count()); + return result; + } + + // Only meaningful if value_type is a POD. + template <typename OUTPUT> + bool write_nopointer_data(OUTPUT *fp) { + return table.write_nopointer_data(fp); + } + + // Only meaningful if value_type is a POD. + template <typename INPUT> + bool read_nopointer_data(INPUT *fp) { + return table.read_nopointer_data(fp); + } + + // INPUT and OUTPUT must be either a FILE, *or* a C++ stream + // (istream, ostream, etc) *or* a class providing + // Read(void*, size_t) and Write(const void*, size_t) + // (respectively), which writes a buffer into a stream + // (which the INPUT/OUTPUT instance presumably owns). + + typedef sparsehash_internal::pod_serializer<value_type> NopointerSerializer; + + // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT *fp) { + squash_deleted(); // so we don't have to worry about delkey + return table.serialize(serializer, fp); + } + + // ValueSerializer: a functor. operator()(INPUT*, value_type*) + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT *fp) { + num_deleted = 0; // since we got rid before writing + const bool result = table.unserialize(serializer, fp); + settings.reset_thresholds(bucket_count()); + return result; + } + + private: + // Table is the main storage class. + typedef sparsetable<value_type, DEFAULT_GROUP_SIZE, value_alloc_type> Table; + + // Package templated functors with the other types to eliminate memory + // needed for storing these zero-size operators. Since ExtractKey and + // hasher's operator() might have the same function signature, they + // must be packaged in different classes. + struct Settings : + sparsehash_internal::sh_hashtable_settings<key_type, hasher, + size_type, HT_MIN_BUCKETS> { + explicit Settings(const hasher& hf) + : sparsehash_internal::sh_hashtable_settings<key_type, hasher, + size_type, HT_MIN_BUCKETS>( + hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {} + }; + + // KeyInfo stores delete key and packages zero-size functors: + // ExtractKey and SetKey. + class KeyInfo : public ExtractKey, public SetKey, public EqualKey { + public: + KeyInfo(const ExtractKey& ek, const SetKey& sk, const EqualKey& eq) + : ExtractKey(ek), + SetKey(sk), + EqualKey(eq) { + } + // We want to return the exact same type as ExtractKey: Key or const Key& + typename ExtractKey::result_type get_key(const_reference v) const { + return ExtractKey::operator()(v); + } + void set_key(pointer v, const key_type& k) const { + SetKey::operator()(v, k); + } + bool equals(const key_type& a, const key_type& b) const { + return EqualKey::operator()(a, b); + } + + // Which key marks deleted entries. + // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!) + typename base::remove_const<key_type>::type delkey; + }; + + // Utility functions to access the templated operators + size_type hash(const key_type& v) const { + return settings.hash(v); + } + bool equals(const key_type& a, const key_type& b) const { + return key_info.equals(a, b); + } + typename ExtractKey::result_type get_key(const_reference v) const { + return key_info.get_key(v); + } + void set_key(pointer v, const key_type& k) const { + key_info.set_key(v, k); + } + + private: + // Actual data + Settings settings; + KeyInfo key_info; + size_type num_deleted; // how many occupied buckets are marked deleted + Table table; // holds num_buckets and num_elements too +}; + + +// We need a global swap as well +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +inline void swap(sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> &x, + sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> &y) { + x.swap(y); +} + +#undef JUMP_ + +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +const typename sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::size_type + sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET; + +// How full we let the table get before we resize. Knuth says .8 is +// good -- higher causes us to probe too much, though saves memory +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +const int sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT = 80; + +// How empty we let the table get before we resize lower. +// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing +template <class V, class K, class HF, class ExK, class SetK, class EqK, class A> +const int sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_PCT + = static_cast<int>(0.4 * + sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT); + +_END_GOOGLE_NAMESPACE_ + +#endif /* _SPARSEHASHTABLE_H_ */ diff --git a/contrib/libs/sparsehash/src/sparsehash/sparse_hash_map b/contrib/libs/sparsehash/src/sparsehash/sparse_hash_map index 2520deb505..1687a8b11c 100644 --- a/contrib/libs/sparsehash/src/sparsehash/sparse_hash_map +++ b/contrib/libs/sparsehash/src/sparsehash/sparse_hash_map @@ -1,363 +1,363 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// This is just a very thin wrapper over sparsehashtable.h, just -// like sgi stl's stl_hash_map is a very thin wrapper over -// stl_hashtable. The major thing we define is operator[], because -// we have a concept of a data_type which stl_hashtable doesn't -// (it only has a key and a value). -// -// We adhere mostly to the STL semantics for hash-map. One important -// exception is that insert() may invalidate iterators entirely -- STL -// semantics are that insert() may reorder iterators, but they all -// still refer to something valid in the hashtable. Not so for us. -// Likewise, insert() may invalidate pointers into the hashtable. -// (Whether insert invalidates iterators and pointers depends on -// whether it results in a hashtable resize). On the plus side, -// delete() doesn't invalidate iterators or pointers at all, or even -// change the ordering of elements. -// -// Here are a few "power user" tips: -// -// 1) set_deleted_key(): -// Unlike STL's hash_map, if you want to use erase() you -// *must* call set_deleted_key() after construction. -// -// 2) resize(0): -// When an item is deleted, its memory isn't freed right -// away. This is what allows you to iterate over a hashtable -// and call erase() without invalidating the iterator. -// To force the memory to be freed, call resize(0). -// For tr1 compatibility, this can also be called as rehash(0). -// -// 3) min_load_factor(0.0) -// Setting the minimum load factor to 0.0 guarantees that -// the hash table will never shrink. -// -// Roughly speaking: -// (1) dense_hash_map: fastest, uses the most memory unless entries are small -// (2) sparse_hash_map: slowest, uses the least memory -// (3) hash_map / unordered_map (STL): in the middle -// -// Typically I use sparse_hash_map when I care about space and/or when -// I need to save the hashtable on disk. I use hash_map otherwise. I -// don't personally use dense_hash_map ever; some people use it for -// small maps with lots of lookups. -// -// - dense_hash_map has, typically, about 78% memory overhead (if your -// data takes up X bytes, the hash_map uses .78X more bytes in overhead). -// - sparse_hash_map has about 4 bits overhead per entry. -// - sparse_hash_map can be 3-7 times slower than the others for lookup and, -// especially, inserts. See time_hash_map.cc for details. -// -// See /usr/(local/)?doc/sparsehash-*/sparse_hash_map.html -// for information about how to use this class. - -#ifndef _SPARSE_HASH_MAP_H_ -#define _SPARSE_HASH_MAP_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <algorithm> // needed by stl_alloc -#include <functional> // for equal_to<>, select1st<>, etc -#include <memory> // for alloc -#include <utility> // for pair<> -#include <sparsehash/internal/libc_allocator_with_realloc.h> -#include <sparsehash/internal/sparsehashtable.h> // IWYU pragma: export -#include HASH_FUN_H // for hash<> -_START_GOOGLE_NAMESPACE_ - -template <class Key, class T, - class HashFcn = SPARSEHASH_HASH<Key>, // defined in sparseconfig.h - class EqualKey = std::equal_to<Key>, - class Alloc = libc_allocator_with_realloc<std::pair<const Key, T> > > -class sparse_hash_map { - private: - // Apparently select1st is not stl-standard, so we define our own - struct SelectKey { - typedef const Key& result_type; - const Key& operator()(const std::pair<const Key, T>& p) const { - return p.first; - } - }; - struct SetKey { - void operator()(std::pair<const Key, T>* value, const Key& new_key) const { - *const_cast<Key*>(&value->first) = new_key; - // It would be nice to clear the rest of value here as well, in - // case it's taking up a lot of memory. We do this by clearing - // the value. This assumes T has a zero-arg constructor! - value->second = T(); - } - }; - // For operator[]. - struct DefaultValue { - std::pair<const Key, T> operator()(const Key& key) { - return std::make_pair(key, T()); - } - }; - - // The actual data - typedef sparse_hashtable<std::pair<const Key, T>, Key, HashFcn, SelectKey, - SetKey, EqualKey, Alloc> ht; - ht rep; - - public: - typedef typename ht::key_type key_type; - typedef T data_type; - typedef T mapped_type; - typedef typename ht::value_type value_type; - typedef typename ht::hasher hasher; - typedef typename ht::key_equal key_equal; - typedef Alloc allocator_type; - - typedef typename ht::size_type size_type; - typedef typename ht::difference_type difference_type; - typedef typename ht::pointer pointer; - typedef typename ht::const_pointer const_pointer; - typedef typename ht::reference reference; - typedef typename ht::const_reference const_reference; - - typedef typename ht::iterator iterator; - typedef typename ht::const_iterator const_iterator; - typedef typename ht::local_iterator local_iterator; - typedef typename ht::const_local_iterator const_local_iterator; - - // Iterator functions - iterator begin() { return rep.begin(); } - iterator end() { return rep.end(); } - const_iterator begin() const { return rep.begin(); } - const_iterator end() const { return rep.end(); } - - // These come from tr1's unordered_map. For us, a bucket has 0 or 1 elements. - local_iterator begin(size_type i) { return rep.begin(i); } - local_iterator end(size_type i) { return rep.end(i); } - const_local_iterator begin(size_type i) const { return rep.begin(i); } - const_local_iterator end(size_type i) const { return rep.end(i); } - - // Accessor functions - allocator_type get_allocator() const { return rep.get_allocator(); } - hasher hash_funct() const { return rep.hash_funct(); } - hasher hash_function() const { return hash_funct(); } - key_equal key_eq() const { return rep.key_eq(); } - - - // Constructors - explicit sparse_hash_map(size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { - } - - template <class InputIterator> - sparse_hash_map(InputIterator f, InputIterator l, - size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { - rep.insert(f, l); - } - // We use the default copy constructor - // We use the default operator=() - // We use the default destructor - - void clear() { rep.clear(); } - void swap(sparse_hash_map& hs) { rep.swap(hs.rep); } - - - // Functions concerning size - size_type size() const { return rep.size(); } - size_type max_size() const { return rep.max_size(); } - bool empty() const { return rep.empty(); } - size_type bucket_count() const { return rep.bucket_count(); } - size_type max_bucket_count() const { return rep.max_bucket_count(); } - - // These are tr1 methods. bucket() is the bucket the key is or would be in. - size_type bucket_size(size_type i) const { return rep.bucket_size(i); } - size_type bucket(const key_type& key) const { return rep.bucket(key); } - float load_factor() const { - return size() * 1.0f / bucket_count(); - } - float max_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return grow; - } - void max_load_factor(float new_grow) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(shrink, new_grow); - } - // These aren't tr1 methods but perhaps ought to be. - float min_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return shrink; - } - void min_load_factor(float new_shrink) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(new_shrink, grow); - } - // Deprecated; use min_load_factor() or max_load_factor() instead. - void set_resizing_parameters(float shrink, float grow) { - rep.set_resizing_parameters(shrink, grow); - } - - void resize(size_type hint) { rep.resize(hint); } - void rehash(size_type hint) { resize(hint); } // the tr1 name - - // Lookup routines - iterator find(const key_type& key) { return rep.find(key); } - const_iterator find(const key_type& key) const { return rep.find(key); } - - data_type& operator[](const key_type& key) { // This is our value-add! - // If key is in the hashtable, returns find(key)->second, - // otherwise returns insert(value_type(key, T()).first->second. - // Note it does not create an empty T unless the find fails. - return rep.template find_or_insert<DefaultValue>(key).second; - } - - size_type count(const key_type& key) const { return rep.count(key); } - - std::pair<iterator, iterator> equal_range(const key_type& key) { - return rep.equal_range(key); - } - std::pair<const_iterator, const_iterator> equal_range(const key_type& key) - const { - return rep.equal_range(key); - } - - // Insertion routines - std::pair<iterator, bool> insert(const value_type& obj) { - return rep.insert(obj); - } - template <class InputIterator> void insert(InputIterator f, InputIterator l) { - rep.insert(f, l); - } - void insert(const_iterator f, const_iterator l) { - rep.insert(f, l); - } - // Required for std::insert_iterator; the passed-in iterator is ignored. - iterator insert(iterator, const value_type& obj) { - return insert(obj).first; - } - - // Deletion routines - // THESE ARE NON-STANDARD! I make you specify an "impossible" key - // value to identify deleted buckets. You can change the key as - // time goes on, or get rid of it entirely to be insert-only. - void set_deleted_key(const key_type& key) { - rep.set_deleted_key(key); - } - void clear_deleted_key() { rep.clear_deleted_key(); } - key_type deleted_key() const { return rep.deleted_key(); } - - // These are standard - size_type erase(const key_type& key) { return rep.erase(key); } - void erase(iterator it) { rep.erase(it); } - void erase(iterator f, iterator l) { rep.erase(f, l); } - - - // Comparison - bool operator==(const sparse_hash_map& hs) const { return rep == hs.rep; } - bool operator!=(const sparse_hash_map& hs) const { return rep != hs.rep; } - - - // I/O -- this is an add-on for writing metainformation to disk - // - // For maximum flexibility, this does not assume a particular - // file type (though it will probably be a FILE *). We just pass - // the fp through to rep. - - // If your keys and values are simple enough, you can pass this - // serializer to serialize()/unserialize(). "Simple enough" means - // value_type is a POD type that contains no pointers. Note, - // however, we don't try to normalize endianness. - typedef typename ht::NopointerSerializer NopointerSerializer; - - // serializer: a class providing operator()(OUTPUT*, const value_type&) - // (writing value_type to OUTPUT). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a - // pointer to a class providing size_t Write(const void*, size_t), - // which writes a buffer into a stream (which fp presumably - // owns) and returns the number of bytes successfully written. - // Note basic_ostream<not_char> is not currently supported. - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT* fp) { - return rep.serialize(serializer, fp); - } - - // serializer: a functor providing operator()(INPUT*, value_type*) - // (reading from INPUT and into value_type). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a - // pointer to a class providing size_t Read(void*, size_t), - // which reads into a buffer from a stream (which fp presumably - // owns) and returns the number of bytes successfully read. - // Note basic_istream<not_char> is not currently supported. - // NOTE: Since value_type is std::pair<const Key, T>, ValueSerializer - // may need to do a const cast in order to fill in the key. - // NOTE: if Key or T are not POD types, the serializer MUST use - // placement-new to initialize their values, rather than a normal - // equals-assignment or similar. (The value_type* passed into the - // serializer points to garbage memory.) - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT* fp) { - return rep.unserialize(serializer, fp); - } - - // The four methods below are DEPRECATED. - // Use serialize() and unserialize() for new code. - template <typename OUTPUT> - bool write_metadata(OUTPUT *fp) { return rep.write_metadata(fp); } - - template <typename INPUT> - bool read_metadata(INPUT *fp) { return rep.read_metadata(fp); } - - template <typename OUTPUT> - bool write_nopointer_data(OUTPUT *fp) { return rep.write_nopointer_data(fp); } - - template <typename INPUT> - bool read_nopointer_data(INPUT *fp) { return rep.read_nopointer_data(fp); } -}; - -// We need a global swap as well -template <class Key, class T, class HashFcn, class EqualKey, class Alloc> -inline void swap(sparse_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm1, - sparse_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm2) { - hm1.swap(hm2); -} - -_END_GOOGLE_NAMESPACE_ - -#endif /* _SPARSE_HASH_MAP_H_ */ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// This is just a very thin wrapper over sparsehashtable.h, just +// like sgi stl's stl_hash_map is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// We adhere mostly to the STL semantics for hash-map. One important +// exception is that insert() may invalidate iterators entirely -- STL +// semantics are that insert() may reorder iterators, but they all +// still refer to something valid in the hashtable. Not so for us. +// Likewise, insert() may invalidate pointers into the hashtable. +// (Whether insert invalidates iterators and pointers depends on +// whether it results in a hashtable resize). On the plus side, +// delete() doesn't invalidate iterators or pointers at all, or even +// change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// Unlike STL's hash_map, if you want to use erase() you +// *must* call set_deleted_key() after construction. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This is what allows you to iterate over a hashtable +// and call erase() without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_map: fastest, uses the most memory unless entries are small +// (2) sparse_hash_map: slowest, uses the least memory +// (3) hash_map / unordered_map (STL): in the middle +// +// Typically I use sparse_hash_map when I care about space and/or when +// I need to save the hashtable on disk. I use hash_map otherwise. I +// don't personally use dense_hash_map ever; some people use it for +// small maps with lots of lookups. +// +// - dense_hash_map has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_map uses .78X more bytes in overhead). +// - sparse_hash_map has about 4 bits overhead per entry. +// - sparse_hash_map can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/sparse_hash_map.html +// for information about how to use this class. + +#ifndef _SPARSE_HASH_MAP_H_ +#define _SPARSE_HASH_MAP_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <algorithm> // needed by stl_alloc +#include <functional> // for equal_to<>, select1st<>, etc +#include <memory> // for alloc +#include <utility> // for pair<> +#include <sparsehash/internal/libc_allocator_with_realloc.h> +#include <sparsehash/internal/sparsehashtable.h> // IWYU pragma: export +#include HASH_FUN_H // for hash<> +_START_GOOGLE_NAMESPACE_ + +template <class Key, class T, + class HashFcn = SPARSEHASH_HASH<Key>, // defined in sparseconfig.h + class EqualKey = std::equal_to<Key>, + class Alloc = libc_allocator_with_realloc<std::pair<const Key, T> > > +class sparse_hash_map { + private: + // Apparently select1st is not stl-standard, so we define our own + struct SelectKey { + typedef const Key& result_type; + const Key& operator()(const std::pair<const Key, T>& p) const { + return p.first; + } + }; + struct SetKey { + void operator()(std::pair<const Key, T>* value, const Key& new_key) const { + *const_cast<Key*>(&value->first) = new_key; + // It would be nice to clear the rest of value here as well, in + // case it's taking up a lot of memory. We do this by clearing + // the value. This assumes T has a zero-arg constructor! + value->second = T(); + } + }; + // For operator[]. + struct DefaultValue { + std::pair<const Key, T> operator()(const Key& key) { + return std::make_pair(key, T()); + } + }; + + // The actual data + typedef sparse_hashtable<std::pair<const Key, T>, Key, HashFcn, SelectKey, + SetKey, EqualKey, Alloc> ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef T data_type; + typedef T mapped_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + // Iterator functions + iterator begin() { return rep.begin(); } + iterator end() { return rep.end(); } + const_iterator begin() const { return rep.begin(); } + const_iterator end() const { return rep.end(); } + + // These come from tr1's unordered_map. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) { return rep.begin(i); } + local_iterator end(size_type i) { return rep.end(i); } + const_local_iterator begin(size_type i) const { return rep.begin(i); } + const_local_iterator end(size_type i) const { return rep.end(i); } + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit sparse_hash_map(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + } + + template <class InputIterator> + sparse_hash_map(InputIterator f, InputIterator l, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + void swap(sparse_hash_map& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) { return rep.find(key); } + const_iterator find(const key_type& key) const { return rep.find(key); } + + data_type& operator[](const key_type& key) { // This is our value-add! + // If key is in the hashtable, returns find(key)->second, + // otherwise returns insert(value_type(key, T()).first->second. + // Note it does not create an empty T unless the find fails. + return rep.template find_or_insert<DefaultValue>(key).second; + } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair<iterator, iterator> equal_range(const key_type& key) { + return rep.equal_range(key); + } + std::pair<const_iterator, const_iterator> equal_range(const key_type& key) + const { + return rep.equal_range(key); + } + + // Insertion routines + std::pair<iterator, bool> insert(const value_type& obj) { + return rep.insert(obj); + } + template <class InputIterator> void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted buckets. You can change the key as + // time goes on, or get rid of it entirely to be insert-only. + void set_deleted_key(const key_type& key) { + rep.set_deleted_key(key); + } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const sparse_hash_map& hs) const { return rep == hs.rep; } + bool operator!=(const sparse_hash_map& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing metainformation to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream<not_char> is not currently supported. + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream<not_char> is not currently supported. + // NOTE: Since value_type is std::pair<const Key, T>, ValueSerializer + // may need to do a const cast in order to fill in the key. + // NOTE: if Key or T are not POD types, the serializer MUST use + // placement-new to initialize their values, rather than a normal + // equals-assignment or similar. (The value_type* passed into the + // serializer points to garbage memory.) + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } + + // The four methods below are DEPRECATED. + // Use serialize() and unserialize() for new code. + template <typename OUTPUT> + bool write_metadata(OUTPUT *fp) { return rep.write_metadata(fp); } + + template <typename INPUT> + bool read_metadata(INPUT *fp) { return rep.read_metadata(fp); } + + template <typename OUTPUT> + bool write_nopointer_data(OUTPUT *fp) { return rep.write_nopointer_data(fp); } + + template <typename INPUT> + bool read_nopointer_data(INPUT *fp) { return rep.read_nopointer_data(fp); } +}; + +// We need a global swap as well +template <class Key, class T, class HashFcn, class EqualKey, class Alloc> +inline void swap(sparse_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm1, + sparse_hash_map<Key, T, HashFcn, EqualKey, Alloc>& hm2) { + hm1.swap(hm2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _SPARSE_HASH_MAP_H_ */ diff --git a/contrib/libs/sparsehash/src/sparsehash/sparse_hash_set b/contrib/libs/sparsehash/src/sparsehash/sparse_hash_set index 555c425514..ae4a97a62c 100644 --- a/contrib/libs/sparsehash/src/sparsehash/sparse_hash_set +++ b/contrib/libs/sparsehash/src/sparsehash/sparse_hash_set @@ -1,338 +1,338 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// This is just a very thin wrapper over sparsehashtable.h, just -// like sgi stl's stl_hash_set is a very thin wrapper over -// stl_hashtable. The major thing we define is operator[], because -// we have a concept of a data_type which stl_hashtable doesn't -// (it only has a key and a value). -// -// This is more different from sparse_hash_map than you might think, -// because all iterators for sets are const (you obviously can't -// change the key, and for sets there is no value). -// -// We adhere mostly to the STL semantics for hash-map. One important -// exception is that insert() may invalidate iterators entirely -- STL -// semantics are that insert() may reorder iterators, but they all -// still refer to something valid in the hashtable. Not so for us. -// Likewise, insert() may invalidate pointers into the hashtable. -// (Whether insert invalidates iterators and pointers depends on -// whether it results in a hashtable resize). On the plus side, -// delete() doesn't invalidate iterators or pointers at all, or even -// change the ordering of elements. -// -// Here are a few "power user" tips: -// -// 1) set_deleted_key(): -// Unlike STL's hash_map, if you want to use erase() you -// *must* call set_deleted_key() after construction. -// -// 2) resize(0): -// When an item is deleted, its memory isn't freed right -// away. This allows you to iterate over a hashtable, -// and call erase(), without invalidating the iterator. -// To force the memory to be freed, call resize(0). -// For tr1 compatibility, this can also be called as rehash(0). -// -// 3) min_load_factor(0.0) -// Setting the minimum load factor to 0.0 guarantees that -// the hash table will never shrink. -// -// Roughly speaking: -// (1) dense_hash_set: fastest, uses the most memory unless entries are small -// (2) sparse_hash_set: slowest, uses the least memory -// (3) hash_set / unordered_set (STL): in the middle -// -// Typically I use sparse_hash_set when I care about space and/or when -// I need to save the hashtable on disk. I use hash_set otherwise. I -// don't personally use dense_hash_set ever; some people use it for -// small sets with lots of lookups. -// -// - dense_hash_set has, typically, about 78% memory overhead (if your -// data takes up X bytes, the hash_set uses .78X more bytes in overhead). -// - sparse_hash_set has about 4 bits overhead per entry. -// - sparse_hash_set can be 3-7 times slower than the others for lookup and, -// especially, inserts. See time_hash_map.cc for details. -// -// See /usr/(local/)?doc/sparsehash-*/sparse_hash_set.html -// for information about how to use this class. - -#ifndef _SPARSE_HASH_SET_H_ -#define _SPARSE_HASH_SET_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <algorithm> // needed by stl_alloc -#include <functional> // for equal_to<> -#include <memory> // for alloc (which we don't use) -#include <utility> // for pair<> -#include <sparsehash/internal/libc_allocator_with_realloc.h> -#include <sparsehash/internal/sparsehashtable.h> // IWYU pragma: export -#include HASH_FUN_H // for hash<> - -_START_GOOGLE_NAMESPACE_ - -template <class Value, - class HashFcn = SPARSEHASH_HASH<Value>, // defined in sparseconfig.h - class EqualKey = std::equal_to<Value>, - class Alloc = libc_allocator_with_realloc<Value> > -class sparse_hash_set { - private: - // Apparently identity is not stl-standard, so we define our own - struct Identity { - typedef const Value& result_type; - const Value& operator()(const Value& v) const { return v; } - }; - struct SetKey { - void operator()(Value* value, const Value& new_key) const { - *value = new_key; - } - }; - - typedef sparse_hashtable<Value, Value, HashFcn, Identity, SetKey, - EqualKey, Alloc> ht; - ht rep; - - public: - typedef typename ht::key_type key_type; - typedef typename ht::value_type value_type; - typedef typename ht::hasher hasher; - typedef typename ht::key_equal key_equal; - typedef Alloc allocator_type; - - typedef typename ht::size_type size_type; - typedef typename ht::difference_type difference_type; - typedef typename ht::const_pointer pointer; - typedef typename ht::const_pointer const_pointer; - typedef typename ht::const_reference reference; - typedef typename ht::const_reference const_reference; - - typedef typename ht::const_iterator iterator; - typedef typename ht::const_iterator const_iterator; - typedef typename ht::const_local_iterator local_iterator; - typedef typename ht::const_local_iterator const_local_iterator; - - - // Iterator functions -- recall all iterators are const - iterator begin() const { return rep.begin(); } - iterator end() const { return rep.end(); } - - // These come from tr1's unordered_set. For us, a bucket has 0 or 1 elements. - local_iterator begin(size_type i) const { return rep.begin(i); } - local_iterator end(size_type i) const { return rep.end(i); } - - - // Accessor functions - allocator_type get_allocator() const { return rep.get_allocator(); } - hasher hash_funct() const { return rep.hash_funct(); } - hasher hash_function() const { return hash_funct(); } // tr1 name - key_equal key_eq() const { return rep.key_eq(); } - - - // Constructors - explicit sparse_hash_set(size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { - } - - template <class InputIterator> - sparse_hash_set(InputIterator f, InputIterator l, - size_type expected_max_items_in_table = 0, - const hasher& hf = hasher(), - const key_equal& eql = key_equal(), - const allocator_type& alloc = allocator_type()) - : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { - rep.insert(f, l); - } - // We use the default copy constructor - // We use the default operator=() - // We use the default destructor - - void clear() { rep.clear(); } - void swap(sparse_hash_set& hs) { rep.swap(hs.rep); } - - - // Functions concerning size - size_type size() const { return rep.size(); } - size_type max_size() const { return rep.max_size(); } - bool empty() const { return rep.empty(); } - size_type bucket_count() const { return rep.bucket_count(); } - size_type max_bucket_count() const { return rep.max_bucket_count(); } - - // These are tr1 methods. bucket() is the bucket the key is or would be in. - size_type bucket_size(size_type i) const { return rep.bucket_size(i); } - size_type bucket(const key_type& key) const { return rep.bucket(key); } - float load_factor() const { - return size() * 1.0f / bucket_count(); - } - float max_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return grow; - } - void max_load_factor(float new_grow) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(shrink, new_grow); - } - // These aren't tr1 methods but perhaps ought to be. - float min_load_factor() const { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - return shrink; - } - void min_load_factor(float new_shrink) { - float shrink, grow; - rep.get_resizing_parameters(&shrink, &grow); - rep.set_resizing_parameters(new_shrink, grow); - } - // Deprecated; use min_load_factor() or max_load_factor() instead. - void set_resizing_parameters(float shrink, float grow) { - rep.set_resizing_parameters(shrink, grow); - } - - void resize(size_type hint) { rep.resize(hint); } - void rehash(size_type hint) { resize(hint); } // the tr1 name - - // Lookup routines - iterator find(const key_type& key) const { return rep.find(key); } - - size_type count(const key_type& key) const { return rep.count(key); } - - std::pair<iterator, iterator> equal_range(const key_type& key) const { - return rep.equal_range(key); - } - - - // Insertion routines - std::pair<iterator, bool> insert(const value_type& obj) { - std::pair<typename ht::iterator, bool> p = rep.insert(obj); - return std::pair<iterator, bool>(p.first, p.second); // const to non-const - } - template <class InputIterator> void insert(InputIterator f, InputIterator l) { - rep.insert(f, l); - } - void insert(const_iterator f, const_iterator l) { - rep.insert(f, l); - } - // Required for std::insert_iterator; the passed-in iterator is ignored. - iterator insert(iterator, const value_type& obj) { - return insert(obj).first; - } - - // Deletion routines - // THESE ARE NON-STANDARD! I make you specify an "impossible" key - // value to identify deleted buckets. You can change the key as - // time goes on, or get rid of it entirely to be insert-only. - void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } - void clear_deleted_key() { rep.clear_deleted_key(); } - key_type deleted_key() const { return rep.deleted_key(); } - - // These are standard - size_type erase(const key_type& key) { return rep.erase(key); } - void erase(iterator it) { rep.erase(it); } - void erase(iterator f, iterator l) { rep.erase(f, l); } - - - // Comparison - bool operator==(const sparse_hash_set& hs) const { return rep == hs.rep; } - bool operator!=(const sparse_hash_set& hs) const { return rep != hs.rep; } - - - // I/O -- this is an add-on for writing metainformation to disk - // - // For maximum flexibility, this does not assume a particular - // file type (though it will probably be a FILE *). We just pass - // the fp through to rep. - - // If your keys and values are simple enough, you can pass this - // serializer to serialize()/unserialize(). "Simple enough" means - // value_type is a POD type that contains no pointers. Note, - // however, we don't try to normalize endianness. - typedef typename ht::NopointerSerializer NopointerSerializer; - - // serializer: a class providing operator()(OUTPUT*, const value_type&) - // (writing value_type to OUTPUT). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a - // pointer to a class providing size_t Write(const void*, size_t), - // which writes a buffer into a stream (which fp presumably - // owns) and returns the number of bytes successfully written. - // Note basic_ostream<not_char> is not currently supported. - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT* fp) { - return rep.serialize(serializer, fp); - } - - // serializer: a functor providing operator()(INPUT*, value_type*) - // (reading from INPUT and into value_type). You can specify a - // NopointerSerializer object if appropriate (see above). - // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a - // pointer to a class providing size_t Read(void*, size_t), - // which reads into a buffer from a stream (which fp presumably - // owns) and returns the number of bytes successfully read. - // Note basic_istream<not_char> is not currently supported. - // NOTE: Since value_type is const Key, ValueSerializer - // may need to do a const cast in order to fill in the key. - // NOTE: if Key is not a POD type, the serializer MUST use - // placement-new to initialize its value, rather than a normal - // equals-assignment or similar. (The value_type* passed into - // the serializer points to garbage memory.) - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT* fp) { - return rep.unserialize(serializer, fp); - } - - // The four methods below are DEPRECATED. - // Use serialize() and unserialize() for new code. - template <typename OUTPUT> - bool write_metadata(OUTPUT *fp) { return rep.write_metadata(fp); } - - template <typename INPUT> - bool read_metadata(INPUT *fp) { return rep.read_metadata(fp); } - - template <typename OUTPUT> - bool write_nopointer_data(OUTPUT *fp) { return rep.write_nopointer_data(fp); } - - template <typename INPUT> - bool read_nopointer_data(INPUT *fp) { return rep.read_nopointer_data(fp); } -}; - -template <class Val, class HashFcn, class EqualKey, class Alloc> -inline void swap(sparse_hash_set<Val, HashFcn, EqualKey, Alloc>& hs1, - sparse_hash_set<Val, HashFcn, EqualKey, Alloc>& hs2) { - hs1.swap(hs2); -} - -_END_GOOGLE_NAMESPACE_ - -#endif /* _SPARSE_HASH_SET_H_ */ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// This is just a very thin wrapper over sparsehashtable.h, just +// like sgi stl's stl_hash_set is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// This is more different from sparse_hash_map than you might think, +// because all iterators for sets are const (you obviously can't +// change the key, and for sets there is no value). +// +// We adhere mostly to the STL semantics for hash-map. One important +// exception is that insert() may invalidate iterators entirely -- STL +// semantics are that insert() may reorder iterators, but they all +// still refer to something valid in the hashtable. Not so for us. +// Likewise, insert() may invalidate pointers into the hashtable. +// (Whether insert invalidates iterators and pointers depends on +// whether it results in a hashtable resize). On the plus side, +// delete() doesn't invalidate iterators or pointers at all, or even +// change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// Unlike STL's hash_map, if you want to use erase() you +// *must* call set_deleted_key() after construction. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This allows you to iterate over a hashtable, +// and call erase(), without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_set: fastest, uses the most memory unless entries are small +// (2) sparse_hash_set: slowest, uses the least memory +// (3) hash_set / unordered_set (STL): in the middle +// +// Typically I use sparse_hash_set when I care about space and/or when +// I need to save the hashtable on disk. I use hash_set otherwise. I +// don't personally use dense_hash_set ever; some people use it for +// small sets with lots of lookups. +// +// - dense_hash_set has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_set uses .78X more bytes in overhead). +// - sparse_hash_set has about 4 bits overhead per entry. +// - sparse_hash_set can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/sparse_hash_set.html +// for information about how to use this class. + +#ifndef _SPARSE_HASH_SET_H_ +#define _SPARSE_HASH_SET_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <algorithm> // needed by stl_alloc +#include <functional> // for equal_to<> +#include <memory> // for alloc (which we don't use) +#include <utility> // for pair<> +#include <sparsehash/internal/libc_allocator_with_realloc.h> +#include <sparsehash/internal/sparsehashtable.h> // IWYU pragma: export +#include HASH_FUN_H // for hash<> + +_START_GOOGLE_NAMESPACE_ + +template <class Value, + class HashFcn = SPARSEHASH_HASH<Value>, // defined in sparseconfig.h + class EqualKey = std::equal_to<Value>, + class Alloc = libc_allocator_with_realloc<Value> > +class sparse_hash_set { + private: + // Apparently identity is not stl-standard, so we define our own + struct Identity { + typedef const Value& result_type; + const Value& operator()(const Value& v) const { return v; } + }; + struct SetKey { + void operator()(Value* value, const Value& new_key) const { + *value = new_key; + } + }; + + typedef sparse_hashtable<Value, Value, HashFcn, Identity, SetKey, + EqualKey, Alloc> ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::const_pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::const_reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::const_iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::const_local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + + // Iterator functions -- recall all iterators are const + iterator begin() const { return rep.begin(); } + iterator end() const { return rep.end(); } + + // These come from tr1's unordered_set. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) const { return rep.begin(i); } + local_iterator end(size_type i) const { return rep.end(i); } + + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } // tr1 name + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit sparse_hash_set(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + } + + template <class InputIterator> + sparse_hash_set(InputIterator f, InputIterator l, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + void swap(sparse_hash_set& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) const { return rep.find(key); } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair<iterator, iterator> equal_range(const key_type& key) const { + return rep.equal_range(key); + } + + + // Insertion routines + std::pair<iterator, bool> insert(const value_type& obj) { + std::pair<typename ht::iterator, bool> p = rep.insert(obj); + return std::pair<iterator, bool>(p.first, p.second); // const to non-const + } + template <class InputIterator> void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted buckets. You can change the key as + // time goes on, or get rid of it entirely to be insert-only. + void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const sparse_hash_set& hs) const { return rep == hs.rep; } + bool operator!=(const sparse_hash_set& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing metainformation to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream<not_char> is not currently supported. + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream<not_char> is not currently supported. + // NOTE: Since value_type is const Key, ValueSerializer + // may need to do a const cast in order to fill in the key. + // NOTE: if Key is not a POD type, the serializer MUST use + // placement-new to initialize its value, rather than a normal + // equals-assignment or similar. (The value_type* passed into + // the serializer points to garbage memory.) + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } + + // The four methods below are DEPRECATED. + // Use serialize() and unserialize() for new code. + template <typename OUTPUT> + bool write_metadata(OUTPUT *fp) { return rep.write_metadata(fp); } + + template <typename INPUT> + bool read_metadata(INPUT *fp) { return rep.read_metadata(fp); } + + template <typename OUTPUT> + bool write_nopointer_data(OUTPUT *fp) { return rep.write_nopointer_data(fp); } + + template <typename INPUT> + bool read_nopointer_data(INPUT *fp) { return rep.read_nopointer_data(fp); } +}; + +template <class Val, class HashFcn, class EqualKey, class Alloc> +inline void swap(sparse_hash_set<Val, HashFcn, EqualKey, Alloc>& hs1, + sparse_hash_set<Val, HashFcn, EqualKey, Alloc>& hs2) { + hs1.swap(hs2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _SPARSE_HASH_SET_H_ */ diff --git a/contrib/libs/sparsehash/src/sparsehash/sparsetable b/contrib/libs/sparsehash/src/sparsehash/sparsetable index 11fd3990ae..6259ebdb04 100644 --- a/contrib/libs/sparsehash/src/sparsehash/sparsetable +++ b/contrib/libs/sparsehash/src/sparsehash/sparsetable @@ -1,1825 +1,1825 @@ -// Copyright (c) 2005, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// --- -// -// -// A sparsetable is a random container that implements a sparse array, -// that is, an array that uses very little memory to store unassigned -// indices (in this case, between 1-2 bits per unassigned index). For -// instance, if you allocate an array of size 5 and assign a[2] = <big -// struct>, then a[2] will take up a lot of memory but a[0], a[1], -// a[3], and a[4] will not. Array elements that have a value are -// called "assigned". Array elements that have no value yet, or have -// had their value cleared using erase() or clear(), are called -// "unassigned". -// -// Unassigned values seem to have the default value of T (see below). -// Nevertheless, there is a difference between an unassigned index and -// one explicitly assigned the value of T(). The latter is considered -// assigned. -// -// Access to an array element is constant time, as is insertion and -// deletion. Insertion and deletion may be fairly slow, however: -// because of this container's memory economy, each insert and delete -// causes a memory reallocation. -// -// NOTE: You should not test(), get(), or set() any index that is -// greater than sparsetable.size(). If you need to do that, call -// resize() first. -// -// --- Template parameters -// PARAMETER DESCRIPTION DEFAULT -// T The value of the array: the type of -- -// object that is stored in the array. -// -// GROUP_SIZE How large each "group" in the table 48 -// is (see below). Larger values use -// a little less memory but cause most -// operations to be a little slower -// -// Alloc: Allocator to use to allocate memory. libc_allocator_with_realloc -// -// --- Model of -// Random Access Container -// -// --- Type requirements -// T must be Copy Constructible. It need not be Assignable. -// -// --- Public base classes -// None. -// -// --- Members -// Type members -// -// MEMBER WHERE DEFINED DESCRIPTION -// value_type container The type of object, T, stored in the array -// allocator_type container Allocator to use -// pointer container Pointer to p -// const_pointer container Const pointer to p -// reference container Reference to t -// const_reference container Const reference to t -// size_type container An unsigned integral type -// difference_type container A signed integral type -// iterator [*] container Iterator used to iterate over a sparsetable -// const_iterator container Const iterator used to iterate over a table -// reverse_iterator reversible Iterator used to iterate backwards over -// container a sparsetable -// const_reverse_iterator reversible container Guess -// nonempty_iterator [+] sparsetable Iterates over assigned -// array elements only -// const_nonempty_iterator sparsetable Iterates over assigned -// array elements only -// reverse_nonempty_iterator sparsetable Iterates backwards over -// assigned array elements only -// const_reverse_nonempty_iterator sparsetable Iterates backwards over -// assigned array elements only -// -// [*] All iterators are const in a sparsetable (though nonempty_iterators -// may not be). Use get() and set() to assign values, not iterators. -// -// [+] iterators are random-access iterators. nonempty_iterators are -// bidirectional iterators. - -// Iterator members -// MEMBER WHERE DEFINED DESCRIPTION -// -// iterator begin() container An iterator to the beginning of the table -// iterator end() container An iterator to the end of the table -// const_iterator container A const_iterator pointing to the -// begin() const beginning of a sparsetable -// const_iterator container A const_iterator pointing to the -// end() const end of a sparsetable -// -// reverse_iterator reversable Points to beginning of a reversed -// rbegin() container sparsetable -// reverse_iterator reversable Points to end of a reversed table -// rend() container -// const_reverse_iterator reversable Points to beginning of a -// rbegin() const container reversed sparsetable -// const_reverse_iterator reversable Points to end of a reversed table -// rend() const container -// -// nonempty_iterator sparsetable Points to first assigned element -// begin() of a sparsetable -// nonempty_iterator sparsetable Points past last assigned element -// end() of a sparsetable -// const_nonempty_iterator sparsetable Points to first assigned element -// begin() const of a sparsetable -// const_nonempty_iterator sparsetable Points past last assigned element -// end() const of a sparsetable -// -// reverse_nonempty_iterator sparsetable Points to first assigned element -// begin() of a reversed sparsetable -// reverse_nonempty_iterator sparsetable Points past last assigned element -// end() of a reversed sparsetable -// const_reverse_nonempty_iterator sparsetable Points to first assigned -// begin() const elt of a reversed sparsetable -// const_reverse_nonempty_iterator sparsetable Points past last assigned -// end() const elt of a reversed sparsetable -// -// -// Other members -// MEMBER WHERE DEFINED DESCRIPTION -// sparsetable() sparsetable A table of size 0; must resize() -// before using. -// sparsetable(size_type size) sparsetable A table of size size. All -// indices are unassigned. -// sparsetable( -// const sparsetable &tbl) sparsetable Copy constructor -// ~sparsetable() sparsetable The destructor -// sparsetable &operator=( sparsetable The assignment operator -// const sparsetable &tbl) -// -// void resize(size_type size) sparsetable Grow or shrink a table to -// have size indices [*] -// -// void swap(sparsetable &x) sparsetable Swap two sparsetables -// void swap(sparsetable &x, sparsetable Swap two sparsetables -// sparsetable &y) (global, not member, function) -// -// size_type size() const sparsetable Number of "buckets" in the table -// size_type max_size() const sparsetable Max allowed size of a sparsetable -// bool empty() const sparsetable true if size() == 0 -// size_type num_nonempty() const sparsetable Number of assigned "buckets" -// -// const_reference get( sparsetable Value at index i, or default -// size_type i) const value if i is unassigned -// const_reference operator[]( sparsetable Identical to get(i) [+] -// difference_type i) const -// reference set(size_type i, sparsetable Set element at index i to -// const_reference val) be a copy of val -// bool test(size_type i) sparsetable True if element at index i -// const has been assigned to -// bool test(iterator pos) sparsetable True if element pointed to -// const by pos has been assigned to -// void erase(iterator pos) sparsetable Set element pointed to by -// pos to be unassigned [!] -// void erase(size_type i) sparsetable Set element i to be unassigned -// void erase(iterator start, sparsetable Erases all elements between -// iterator end) start and end -// void clear() sparsetable Erases all elements in the table -// -// I/O versions exist for both FILE* and for File* (Google2-style files): -// bool write_metadata(FILE *fp) sparsetable Writes a sparsetable to the -// bool write_metadata(File *fp) given file. true if write -// completes successfully -// bool read_metadata(FILE *fp) sparsetable Replaces sparsetable with -// bool read_metadata(File *fp) version read from fp. true -// if read completes sucessfully -// bool write_nopointer_data(FILE *fp) Read/write the data stored in -// bool read_nopointer_data(FILE*fp) the table, if it's simple -// -// bool operator==( forward Tests two tables for equality. -// const sparsetable &t1, container This is a global function, -// const sparsetable &t2) not a member function. -// bool operator<( forward Lexicographical comparison. -// const sparsetable &t1, container This is a global function, -// const sparsetable &t2) not a member function. -// -// [*] If you shrink a sparsetable using resize(), assigned elements -// past the end of the table are removed using erase(). If you grow -// a sparsetable, new unassigned indices are created. -// -// [+] Note that operator[] returns a const reference. You must use -// set() to change the value of a table element. -// -// [!] Unassignment also calls the destructor. -// -// Iterators are invalidated whenever an item is inserted or -// deleted (ie set() or erase() is used) or when the size of -// the table changes (ie resize() or clear() is used). -// -// See doc/sparsetable.html for more information about how to use this class. - -// Note: this uses STL style for naming, rather than Google naming. -// That's because this is an STL-y container - -#ifndef UTIL_GTL_SPARSETABLE_H_ -#define UTIL_GTL_SPARSETABLE_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <stdlib.h> // for malloc/free -#include <stdio.h> // to read/write tables -#include <string.h> // for memcpy -#ifdef HAVE_STDINT_H -#include <stdint.h> // the normal place uint16_t is defined -#endif -#ifdef HAVE_SYS_TYPES_H -#include <sys/types.h> // the normal place u_int16_t is defined -#endif -#ifdef HAVE_INTTYPES_H -#include <inttypes.h> // a third place for uint16_t or u_int16_t -#endif -#include <assert.h> // for bounds checking -#include <iterator> // to define reverse_iterator for me -#include <algorithm> // equal, lexicographical_compare, swap,... -#include <memory> // uninitialized_copy, uninitialized_fill -#include <vector> // a sparsetable is a vector of groups -#include <sparsehash/type_traits.h> -#include <sparsehash/internal/hashtable-common.h> -#include <sparsehash/internal/libc_allocator_with_realloc.h> - -// A lot of work to get a type that's guaranteed to be 16 bits... -#ifndef HAVE_U_INT16_T -# if defined HAVE_UINT16_T - typedef uint16_t u_int16_t; // true on solaris, possibly other C99 libc's -# elif defined HAVE___UINT16 - typedef __int16 int16_t; // true on vc++7 - typedef unsigned __int16 u_int16_t; -# else - // Cannot find a 16-bit integer type. Hoping for the best with "short"... - typedef short int int16_t; - typedef unsigned short int u_int16_t; -# endif -#endif - -_START_GOOGLE_NAMESPACE_ - -namespace base { // just to make google->opensource transition easier -using GOOGLE_NAMESPACE::true_type; -using GOOGLE_NAMESPACE::false_type; -using GOOGLE_NAMESPACE::integral_constant; -using GOOGLE_NAMESPACE::has_trivial_copy; -using GOOGLE_NAMESPACE::has_trivial_destructor; -using GOOGLE_NAMESPACE::is_same; -} - - -// The smaller this is, the faster lookup is (because the group bitmap is -// smaller) and the faster insert is, because there's less to move. -// On the other hand, there are more groups. Since group::size_type is -// a short, this number should be of the form 32*x + 16 to avoid waste. -static const u_int16_t DEFAULT_SPARSEGROUP_SIZE = 48; // fits in 1.5 words - - -// Our iterator as simple as iterators can be: basically it's just -// the index into our table. Dereference, the only complicated -// thing, we punt to the table class. This just goes to show how -// much machinery STL requires to do even the most trivial tasks. -// -// A NOTE ON ASSIGNING: -// A sparse table does not actually allocate memory for entries -// that are not filled. Because of this, it becomes complicated -// to have a non-const iterator: we don't know, if the iterator points -// to a not-filled bucket, whether you plan to fill it with something -// or whether you plan to read its value (in which case you'll get -// the default bucket value). Therefore, while we can define const -// operations in a pretty 'normal' way, for non-const operations, we -// define something that returns a helper object with operator= and -// operator& that allocate a bucket lazily. We use this for table[] -// and also for regular table iterators. - -template <class tabletype> -class table_element_adaptor { - public: - typedef typename tabletype::value_type value_type; - typedef typename tabletype::size_type size_type; - typedef typename tabletype::reference reference; - typedef typename tabletype::pointer pointer; - - table_element_adaptor(tabletype *tbl, size_type p) - : table(tbl), pos(p) { } - table_element_adaptor& operator= (const value_type &val) { - table->set(pos, val); - return *this; - } - operator value_type() { return table->get(pos); } // we look like a value - pointer operator& () { return &table->mutating_get(pos); } - - private: - tabletype* table; - size_type pos; -}; - -// Our iterator as simple as iterators can be: basically it's just -// the index into our table. Dereference, the only complicated -// thing, we punt to the table class. This just goes to show how -// much machinery STL requires to do even the most trivial tasks. -// -// By templatizing over tabletype, we have one iterator type which -// we can use for both sparsetables and sparsebins. In fact it -// works on any class that allows size() and operator[] (eg vector), -// as long as it does the standard STL typedefs too (eg value_type). - -template <class tabletype> -class table_iterator { - public: - typedef table_iterator iterator; - - typedef std::random_access_iterator_tag iterator_category; - typedef typename tabletype::value_type value_type; - typedef typename tabletype::difference_type difference_type; - typedef typename tabletype::size_type size_type; - typedef table_element_adaptor<tabletype> reference; - typedef table_element_adaptor<tabletype>* pointer; - - // The "real" constructor - table_iterator(tabletype *tbl, size_type p) - : table(tbl), pos(p) { } - // The default constructor, used when I define vars of type table::iterator - table_iterator() : table(NULL), pos(0) { } - // The copy constructor, for when I say table::iterator foo = tbl.begin() - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // The main thing our iterator does is dereference. If the table entry - // we point to is empty, we return the default value type. - // This is the big different function from the const iterator. - reference operator*() { - return table_element_adaptor<tabletype>(table, pos); - } - pointer operator->() { return &(operator*()); } - - // Helper function to assert things are ok; eg pos is still in range - void check() const { - assert(table); - assert(pos <= table->size()); - } - - // Arithmetic: we just do arithmetic on pos. We don't even need to - // do bounds checking, since STL doesn't consider that its job. :-) - iterator& operator+=(size_type t) { pos += t; check(); return *this; } - iterator& operator-=(size_type t) { pos -= t; check(); return *this; } - iterator& operator++() { ++pos; check(); return *this; } - iterator& operator--() { --pos; check(); return *this; } - iterator operator++(int) { iterator tmp(*this); // for x++ - ++pos; check(); return tmp; } - iterator operator--(int) { iterator tmp(*this); // for x-- - --pos; check(); return tmp; } - iterator operator+(difference_type i) const { iterator tmp(*this); - tmp += i; return tmp; } - iterator operator-(difference_type i) const { iterator tmp(*this); - tmp -= i; return tmp; } - difference_type operator-(iterator it) const { // for "x = it2 - it" - assert(table == it.table); - return pos - it.pos; - } - reference operator[](difference_type n) const { - return *(*this + n); // simple though not totally efficient - } - - // Comparisons. - bool operator==(const iterator& it) const { - return table == it.table && pos == it.pos; - } - bool operator<(const iterator& it) const { - assert(table == it.table); // life is bad bad bad otherwise - return pos < it.pos; - } - bool operator!=(const iterator& it) const { return !(*this == it); } - bool operator<=(const iterator& it) const { return !(it < *this); } - bool operator>(const iterator& it) const { return it < *this; } - bool operator>=(const iterator& it) const { return !(*this < it); } - - // Here's the info we actually need to be an iterator - tabletype *table; // so we can dereference and bounds-check - size_type pos; // index into the table -}; - -// support for "3 + iterator" has to be defined outside the class, alas -template<class T> -table_iterator<T> operator+(typename table_iterator<T>::difference_type i, - table_iterator<T> it) { - return it + i; // so people can say it2 = 3 + it -} - -template <class tabletype> -class const_table_iterator { - public: - typedef table_iterator<tabletype> iterator; - typedef const_table_iterator const_iterator; - - typedef std::random_access_iterator_tag iterator_category; - typedef typename tabletype::value_type value_type; - typedef typename tabletype::difference_type difference_type; - typedef typename tabletype::size_type size_type; - typedef typename tabletype::const_reference reference; // we're const-only - typedef typename tabletype::const_pointer pointer; - - // The "real" constructor - const_table_iterator(const tabletype *tbl, size_type p) - : table(tbl), pos(p) { } - // The default constructor, used when I define vars of type table::iterator - const_table_iterator() : table(NULL), pos(0) { } - // The copy constructor, for when I say table::iterator foo = tbl.begin() - // Also converts normal iterators to const iterators - const_table_iterator(const iterator &from) - : table(from.table), pos(from.pos) { } - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // The main thing our iterator does is dereference. If the table entry - // we point to is empty, we return the default value type. - reference operator*() const { return (*table)[pos]; } - pointer operator->() const { return &(operator*()); } - - // Helper function to assert things are ok; eg pos is still in range - void check() const { - assert(table); - assert(pos <= table->size()); - } - - // Arithmetic: we just do arithmetic on pos. We don't even need to - // do bounds checking, since STL doesn't consider that its job. :-) - const_iterator& operator+=(size_type t) { pos += t; check(); return *this; } - const_iterator& operator-=(size_type t) { pos -= t; check(); return *this; } - const_iterator& operator++() { ++pos; check(); return *this; } - const_iterator& operator--() { --pos; check(); return *this; } - const_iterator operator++(int) { const_iterator tmp(*this); // for x++ - ++pos; check(); return tmp; } - const_iterator operator--(int) { const_iterator tmp(*this); // for x-- - --pos; check(); return tmp; } - const_iterator operator+(difference_type i) const { const_iterator tmp(*this); - tmp += i; return tmp; } - const_iterator operator-(difference_type i) const { const_iterator tmp(*this); - tmp -= i; return tmp; } - difference_type operator-(const_iterator it) const { // for "x = it2 - it" - assert(table == it.table); - return pos - it.pos; - } - reference operator[](difference_type n) const { - return *(*this + n); // simple though not totally efficient - } - - // Comparisons. - bool operator==(const const_iterator& it) const { - return table == it.table && pos == it.pos; - } - bool operator<(const const_iterator& it) const { - assert(table == it.table); // life is bad bad bad otherwise - return pos < it.pos; - } - bool operator!=(const const_iterator& it) const { return !(*this == it); } - bool operator<=(const const_iterator& it) const { return !(it < *this); } - bool operator>(const const_iterator& it) const { return it < *this; } - bool operator>=(const const_iterator& it) const { return !(*this < it); } - - // Here's the info we actually need to be an iterator - const tabletype *table; // so we can dereference and bounds-check - size_type pos; // index into the table -}; - -// support for "3 + iterator" has to be defined outside the class, alas -template<class T> -const_table_iterator<T> operator+(typename - const_table_iterator<T>::difference_type i, - const_table_iterator<T> it) { - return it + i; // so people can say it2 = 3 + it -} - - -// --------------------------------------------------------------------------- - - -/* -// This is a 2-D iterator. You specify a begin and end over a list -// of *containers*. We iterate over each container by iterating over -// it. It's actually simple: -// VECTOR.begin() VECTOR[0].begin() --------> VECTOR[0].end() ---, -// | ________________________________________________/ -// | \_> VECTOR[1].begin() --------> VECTOR[1].end() -, -// | ___________________________________________________/ -// v \_> ...... -// VECTOR.end() -// -// It's impossible to do random access on one of these things in constant -// time, so it's just a bidirectional iterator. -// -// Unfortunately, because we need to use this for a non-empty iterator, -// we use nonempty_begin() and nonempty_end() instead of begin() and end() -// (though only going across, not down). -*/ - -#define TWOD_BEGIN_ nonempty_begin -#define TWOD_END_ nonempty_end -#define TWOD_ITER_ nonempty_iterator -#define TWOD_CONST_ITER_ const_nonempty_iterator - -template <class containertype> -class two_d_iterator { - public: - typedef two_d_iterator iterator; - - typedef std::bidirectional_iterator_tag iterator_category; - // apparently some versions of VC++ have trouble with two ::'s in a typename - typedef typename containertype::value_type _tmp_vt; - typedef typename _tmp_vt::value_type value_type; - typedef typename _tmp_vt::difference_type difference_type; - typedef typename _tmp_vt::reference reference; - typedef typename _tmp_vt::pointer pointer; - - // The "real" constructor. begin and end specify how many rows we have - // (in the diagram above); we always iterate over each row completely. - two_d_iterator(typename containertype::iterator begin, - typename containertype::iterator end, - typename containertype::iterator curr) - : row_begin(begin), row_end(end), row_current(curr), col_current() { - if ( row_current != row_end ) { - col_current = row_current->TWOD_BEGIN_(); - advance_past_end(); // in case cur->begin() == cur->end() - } - } - // If you want to start at an arbitrary place, you can, I guess - two_d_iterator(typename containertype::iterator begin, - typename containertype::iterator end, - typename containertype::iterator curr, - typename containertype::value_type::TWOD_ITER_ col) - : row_begin(begin), row_end(end), row_current(curr), col_current(col) { - advance_past_end(); // in case cur->begin() == cur->end() - } - // The default constructor, used when I define vars of type table::iterator - two_d_iterator() : row_begin(), row_end(), row_current(), col_current() { } - // The default destructor is fine; we don't define one - // The default operator= is fine; we don't define one - - // Happy dereferencer - reference operator*() const { return *col_current; } - pointer operator->() const { return &(operator*()); } - - // Arithmetic: we just do arithmetic on pos. We don't even need to - // do bounds checking, since STL doesn't consider that its job. :-) - // NOTE: this is not amortized constant time! What do we do about it? - void advance_past_end() { // used when col_current points to end() - while ( col_current == row_current->TWOD_END_() ) { // end of current row - ++row_current; // go to beginning of next - if ( row_current != row_end ) // col is irrelevant at end - col_current = row_current->TWOD_BEGIN_(); - else - break; // don't go past row_end - } - } - - iterator& operator++() { - assert(row_current != row_end); // how to ++ from there? - ++col_current; - advance_past_end(); // in case col_current is at end() - return *this; - } - iterator& operator--() { - while ( row_current == row_end || - col_current == row_current->TWOD_BEGIN_() ) { - assert(row_current != row_begin); - --row_current; - col_current = row_current->TWOD_END_(); // this is 1 too far - } - --col_current; - return *this; - } - iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } - iterator operator--(int) { iterator tmp(*this); --*this; return tmp; } - - - // Comparisons. - bool operator==(const iterator& it) const { - return ( row_begin == it.row_begin && - row_end == it.row_end && - row_current == it.row_current && - (row_current == row_end || col_current == it.col_current) ); - } - bool operator!=(const iterator& it) const { return !(*this == it); } - - - // Here's the info we actually need to be an iterator - // These need to be public so we convert from iterator to const_iterator - typename containertype::iterator row_begin, row_end, row_current; - typename containertype::value_type::TWOD_ITER_ col_current; -}; - -// The same thing again, but this time const. :-( -template <class containertype> -class const_two_d_iterator { - public: - typedef const_two_d_iterator iterator; - - typedef std::bidirectional_iterator_tag iterator_category; - // apparently some versions of VC++ have trouble with two ::'s in a typename - typedef typename containertype::value_type _tmp_vt; - typedef typename _tmp_vt::value_type value_type; - typedef typename _tmp_vt::difference_type difference_type; - typedef typename _tmp_vt::const_reference reference; - typedef typename _tmp_vt::const_pointer pointer; - - const_two_d_iterator(typename containertype::const_iterator begin, - typename containertype::const_iterator end, - typename containertype::const_iterator curr) - : row_begin(begin), row_end(end), row_current(curr), col_current() { - if ( curr != end ) { - col_current = curr->TWOD_BEGIN_(); - advance_past_end(); // in case cur->begin() == cur->end() - } - } - const_two_d_iterator(typename containertype::const_iterator begin, - typename containertype::const_iterator end, - typename containertype::const_iterator curr, - typename containertype::value_type::TWOD_CONST_ITER_ col) - : row_begin(begin), row_end(end), row_current(curr), col_current(col) { - advance_past_end(); // in case cur->begin() == cur->end() - } - const_two_d_iterator() - : row_begin(), row_end(), row_current(), col_current() { - } - // Need this explicitly so we can convert normal iterators to const iterators - const_two_d_iterator(const two_d_iterator<containertype>& it) : - row_begin(it.row_begin), row_end(it.row_end), row_current(it.row_current), - col_current(it.col_current) { } - - typename containertype::const_iterator row_begin, row_end, row_current; - typename containertype::value_type::TWOD_CONST_ITER_ col_current; - - - // EVERYTHING FROM HERE DOWN IS THE SAME AS THE NON-CONST ITERATOR - reference operator*() const { return *col_current; } - pointer operator->() const { return &(operator*()); } - - void advance_past_end() { // used when col_current points to end() - while ( col_current == row_current->TWOD_END_() ) { // end of current row - ++row_current; // go to beginning of next - if ( row_current != row_end ) // col is irrelevant at end - col_current = row_current->TWOD_BEGIN_(); - else - break; // don't go past row_end - } - } - iterator& operator++() { - assert(row_current != row_end); // how to ++ from there? - ++col_current; - advance_past_end(); // in case col_current is at end() - return *this; - } - iterator& operator--() { - while ( row_current == row_end || - col_current == row_current->TWOD_BEGIN_() ) { - assert(row_current != row_begin); - --row_current; - col_current = row_current->TWOD_END_(); // this is 1 too far - } - --col_current; - return *this; - } - iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } - iterator operator--(int) { iterator tmp(*this); --*this; return tmp; } - - bool operator==(const iterator& it) const { - return ( row_begin == it.row_begin && - row_end == it.row_end && - row_current == it.row_current && - (row_current == row_end || col_current == it.col_current) ); - } - bool operator!=(const iterator& it) const { return !(*this == it); } -}; - -// We provide yet another version, to be as frugal with memory as -// possible. This one frees each block of memory as it finishes -// iterating over it. By the end, the entire table is freed. -// For understandable reasons, you can only iterate over it once, -// which is why it's an input iterator -template <class containertype> -class destructive_two_d_iterator { - public: - typedef destructive_two_d_iterator iterator; - - typedef std::input_iterator_tag iterator_category; - // apparently some versions of VC++ have trouble with two ::'s in a typename - typedef typename containertype::value_type _tmp_vt; - typedef typename _tmp_vt::value_type value_type; - typedef typename _tmp_vt::difference_type difference_type; - typedef typename _tmp_vt::reference reference; - typedef typename _tmp_vt::pointer pointer; - - destructive_two_d_iterator(typename containertype::iterator begin, - typename containertype::iterator end, - typename containertype::iterator curr) - : row_begin(begin), row_end(end), row_current(curr), col_current() { - if ( curr != end ) { - col_current = curr->TWOD_BEGIN_(); - advance_past_end(); // in case cur->begin() == cur->end() - } - } - destructive_two_d_iterator(typename containertype::iterator begin, - typename containertype::iterator end, - typename containertype::iterator curr, - typename containertype::value_type::TWOD_ITER_ col) - : row_begin(begin), row_end(end), row_current(curr), col_current(col) { - advance_past_end(); // in case cur->begin() == cur->end() - } - destructive_two_d_iterator() - : row_begin(), row_end(), row_current(), col_current() { - } - - typename containertype::iterator row_begin, row_end, row_current; - typename containertype::value_type::TWOD_ITER_ col_current; - - // This is the part that destroys - void advance_past_end() { // used when col_current points to end() - while ( col_current == row_current->TWOD_END_() ) { // end of current row - row_current->clear(); // the destructive part - // It would be nice if we could decrement sparsetable->num_buckets here - ++row_current; // go to beginning of next - if ( row_current != row_end ) // col is irrelevant at end - col_current = row_current->TWOD_BEGIN_(); - else - break; // don't go past row_end - } - } - - // EVERYTHING FROM HERE DOWN IS THE SAME AS THE REGULAR ITERATOR - reference operator*() const { return *col_current; } - pointer operator->() const { return &(operator*()); } - - iterator& operator++() { - assert(row_current != row_end); // how to ++ from there? - ++col_current; - advance_past_end(); // in case col_current is at end() - return *this; - } - iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } - - bool operator==(const iterator& it) const { - return ( row_begin == it.row_begin && - row_end == it.row_end && - row_current == it.row_current && - (row_current == row_end || col_current == it.col_current) ); - } - bool operator!=(const iterator& it) const { return !(*this == it); } -}; - -#undef TWOD_BEGIN_ -#undef TWOD_END_ -#undef TWOD_ITER_ -#undef TWOD_CONST_ITER_ - - - - -// SPARSE-TABLE -// ------------ -// The idea is that a table with (logically) t buckets is divided -// into t/M *groups* of M buckets each. (M is a constant set in -// GROUP_SIZE for efficiency.) Each group is stored sparsely. -// Thus, inserting into the table causes some array to grow, which is -// slow but still constant time. Lookup involves doing a -// logical-position-to-sparse-position lookup, which is also slow but -// constant time. The larger M is, the slower these operations are -// but the less overhead (slightly). -// -// To store the sparse array, we store a bitmap B, where B[i] = 1 iff -// bucket i is non-empty. Then to look up bucket i we really look up -// array[# of 1s before i in B]. This is constant time for fixed M. -// -// Terminology: the position of an item in the overall table (from -// 1 .. t) is called its "location." The logical position in a group -// (from 1 .. M ) is called its "position." The actual location in -// the array (from 1 .. # of non-empty buckets in the group) is -// called its "offset." - -template <class T, u_int16_t GROUP_SIZE, class Alloc> -class sparsegroup { - private: - typedef typename Alloc::template rebind<T>::other value_alloc_type; - - public: - // Basic types - typedef T value_type; - typedef Alloc allocator_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::const_reference const_reference; - typedef typename value_alloc_type::pointer pointer; - typedef typename value_alloc_type::const_pointer const_pointer; - - typedef table_iterator<sparsegroup<T, GROUP_SIZE, Alloc> > iterator; - typedef const_table_iterator<sparsegroup<T, GROUP_SIZE, Alloc> > - const_iterator; - typedef table_element_adaptor<sparsegroup<T, GROUP_SIZE, Alloc> > - element_adaptor; - typedef u_int16_t size_type; // max # of buckets - typedef int16_t difference_type; - typedef std::reverse_iterator<const_iterator> const_reverse_iterator; - typedef std::reverse_iterator<iterator> reverse_iterator; // from iterator.h - - // These are our special iterators, that go over non-empty buckets in a - // group. These aren't const-only because you can change non-empty bcks. - typedef pointer nonempty_iterator; - typedef const_pointer const_nonempty_iterator; - typedef std::reverse_iterator<nonempty_iterator> reverse_nonempty_iterator; - typedef std::reverse_iterator<const_nonempty_iterator> const_reverse_nonempty_iterator; - - // Iterator functions - iterator begin() { return iterator(this, 0); } - const_iterator begin() const { return const_iterator(this, 0); } - iterator end() { return iterator(this, size()); } - const_iterator end() const { return const_iterator(this, size()); } - reverse_iterator rbegin() { return reverse_iterator(end()); } - const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } - reverse_iterator rend() { return reverse_iterator(begin()); } - const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } - - // We'll have versions for our special non-empty iterator too - nonempty_iterator nonempty_begin() { return group; } - const_nonempty_iterator nonempty_begin() const { return group; } - nonempty_iterator nonempty_end() { - return group + settings.num_buckets; - } - const_nonempty_iterator nonempty_end() const { - return group + settings.num_buckets; - } - reverse_nonempty_iterator nonempty_rbegin() { - return reverse_nonempty_iterator(nonempty_end()); - } - const_reverse_nonempty_iterator nonempty_rbegin() const { - return const_reverse_nonempty_iterator(nonempty_end()); - } - reverse_nonempty_iterator nonempty_rend() { - return reverse_nonempty_iterator(nonempty_begin()); - } - const_reverse_nonempty_iterator nonempty_rend() const { - return const_reverse_nonempty_iterator(nonempty_begin()); - } - - - // This gives us the "default" value to return for an empty bucket. - // We just use the default constructor on T, the template type - const_reference default_value() const { - static value_type defaultval = value_type(); - return defaultval; - } - - - private: - // We need to do all this bit manipulation, of course. ick - static size_type charbit(size_type i) { return i >> 3; } - static size_type modbit(size_type i) { return 1 << (i&7); } - int bmtest(size_type i) const { return bitmap[charbit(i)] & modbit(i); } - void bmset(size_type i) { bitmap[charbit(i)] |= modbit(i); } - void bmclear(size_type i) { bitmap[charbit(i)] &= ~modbit(i); } - - pointer allocate_group(size_type n) { - pointer retval = settings.allocate(n); - if (retval == NULL) { - // We really should use PRIuS here, but I don't want to have to add - // a whole new configure option, with concomitant macro namespace - // pollution, just to print this (unlikely) error message. So I cast. - fprintf(stderr, "sparsehash FATAL ERROR: failed to allocate %lu groups\n", - static_cast<unsigned long>(n)); - exit(1); - } - return retval; - } - - void free_group() { - if (!group) return; - pointer end_it = group + settings.num_buckets; - for (pointer p = group; p != end_it; ++p) - p->~value_type(); - settings.deallocate(group, settings.num_buckets); - group = NULL; - } - - static size_type bits_in_char(unsigned char c) { - // We could make these ints. The tradeoff is size (eg does it overwhelm - // the cache?) vs efficiency in referencing sub-word-sized array elements. - static const char bits_in[256] = { - 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, - 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, - 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, - 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, - 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, - 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, - 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, - 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, - 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, - 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, - 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, - 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, - 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, - 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, - 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, - 4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8, - }; - return bits_in[c]; - } - - public: // get_iter() in sparsetable needs it - // We need a small function that tells us how many set bits there are - // in positions 0..i-1 of the bitmap. It uses a big table. - // We make it static so templates don't allocate lots of these tables. - // There are lots of ways to do this calculation (called 'popcount'). - // The 8-bit table lookup is one of the fastest, though this - // implementation suffers from not doing any loop unrolling. See, eg, - // http://www.dalkescientific.com/writings/diary/archive/2008/07/03/hakmem_and_other_popcounts.html - // http://gurmeetsingh.wordpress.com/2008/08/05/fast-bit-counting-routines/ - static size_type pos_to_offset(const unsigned char *bm, size_type pos) { - size_type retval = 0; - - // [Note: condition pos > 8 is an optimization; convince yourself we - // give exactly the same result as if we had pos >= 8 here instead.] - for ( ; pos > 8; pos -= 8 ) // bm[0..pos/8-1] - retval += bits_in_char(*bm++); // chars we want *all* bits in - return retval + bits_in_char(*bm & ((1 << pos)-1)); // char including pos - } - - size_type pos_to_offset(size_type pos) const { // not static but still const - return pos_to_offset(bitmap, pos); - } - - // Returns the (logical) position in the bm[] array, i, such that - // bm[i] is the offset-th set bit in the array. It is the inverse - // of pos_to_offset. get_pos() uses this function to find the index - // of an nonempty_iterator in the table. Bit-twiddling from - // http://hackersdelight.org/basics.pdf - static size_type offset_to_pos(const unsigned char *bm, size_type offset) { - size_type retval = 0; - // This is sizeof(this->bitmap). - const size_type group_size = (GROUP_SIZE-1) / 8 + 1; - for (size_type i = 0; i < group_size; i++) { // forward scan - const size_type pop_count = bits_in_char(*bm); - if (pop_count > offset) { - unsigned char last_bm = *bm; - for (; offset > 0; offset--) { - last_bm &= (last_bm-1); // remove right-most set bit - } - // Clear all bits to the left of the rightmost bit (the &), - // and then clear the rightmost bit but set all bits to the - // right of it (the -1). - last_bm = (last_bm & -last_bm) - 1; - retval += bits_in_char(last_bm); - return retval; - } - offset -= pop_count; - retval += 8; - bm++; - } - return retval; - } - - size_type offset_to_pos(size_type offset) const { - return offset_to_pos(bitmap, offset); - } - - - public: - // Constructors -- default and copy -- and destructor - explicit sparsegroup(allocator_type& a) : - group(0), settings(alloc_impl<value_alloc_type>(a)) { - memset(bitmap, 0, sizeof(bitmap)); - } - sparsegroup(const sparsegroup& x) : group(0), settings(x.settings) { - if ( settings.num_buckets ) { - group = allocate_group(x.settings.num_buckets); - std::uninitialized_copy(x.group, x.group + x.settings.num_buckets, group); - } - memcpy(bitmap, x.bitmap, sizeof(bitmap)); - } - ~sparsegroup() { free_group(); } - - // Operator= is just like the copy constructor, I guess - // TODO(austern): Make this exception safe. Handle exceptions in value_type's - // copy constructor. - sparsegroup &operator=(const sparsegroup& x) { - if ( &x == this ) return *this; // x = x - if ( x.settings.num_buckets == 0 ) { - free_group(); - } else { - pointer p = allocate_group(x.settings.num_buckets); - std::uninitialized_copy(x.group, x.group + x.settings.num_buckets, p); - free_group(); - group = p; - } - memcpy(bitmap, x.bitmap, sizeof(bitmap)); - settings.num_buckets = x.settings.num_buckets; - return *this; - } - - // Many STL algorithms use swap instead of copy constructors - void swap(sparsegroup& x) { - std::swap(group, x.group); // defined in <algorithm> - for ( int i = 0; i < sizeof(bitmap) / sizeof(*bitmap); ++i ) - std::swap(bitmap[i], x.bitmap[i]); // swap not defined on arrays - std::swap(settings.num_buckets, x.settings.num_buckets); - // we purposefully don't swap the allocator, which may not be swap-able - } - - // It's always nice to be able to clear a table without deallocating it - void clear() { - free_group(); - memset(bitmap, 0, sizeof(bitmap)); - settings.num_buckets = 0; - } - - // Functions that tell you about size. Alas, these aren't so useful - // because our table is always fixed size. - size_type size() const { return GROUP_SIZE; } - size_type max_size() const { return GROUP_SIZE; } - bool empty() const { return false; } - // We also may want to know how many *used* buckets there are - size_type num_nonempty() const { return settings.num_buckets; } - - - // get()/set() are explicitly const/non-const. You can use [] if - // you want something that can be either (potentially more expensive). - const_reference get(size_type i) const { - if ( bmtest(i) ) // bucket i is occupied - return group[pos_to_offset(bitmap, i)]; - else - return default_value(); // return the default reference - } - - // TODO(csilvers): make protected + friend - // This is used by sparse_hashtable to get an element from the table - // when we know it exists. - const_reference unsafe_get(size_type i) const { - assert(bmtest(i)); - return group[pos_to_offset(bitmap, i)]; - } - - // TODO(csilvers): make protected + friend - reference mutating_get(size_type i) { // fills bucket i before getting - if ( !bmtest(i) ) - set(i, default_value()); - return group[pos_to_offset(bitmap, i)]; - } - - // Syntactic sugar. It's easy to return a const reference. To - // return a non-const reference, we need to use the assigner adaptor. - const_reference operator[](size_type i) const { - return get(i); - } - - element_adaptor operator[](size_type i) { - return element_adaptor(this, i); - } - - private: - // Create space at group[offset], assuming value_type has trivial - // copy constructor and destructor, and the allocator_type is - // the default libc_allocator_with_alloc. (Really, we want it to have - // "trivial move", because that's what realloc and memmove both do. - // But there's no way to capture that using type_traits, so we - // pretend that move(x, y) is equivalent to "x.~T(); new(x) T(y);" - // which is pretty much correct, if a bit conservative.) - void set_aux(size_type offset, base::true_type) { - group = settings.realloc_or_die(group, settings.num_buckets+1); - // This is equivalent to memmove(), but faster on my Intel P4, - // at least with gcc4.1 -O2 / glibc 2.3.6. - for (size_type i = settings.num_buckets; i > offset; --i) +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// +// A sparsetable is a random container that implements a sparse array, +// that is, an array that uses very little memory to store unassigned +// indices (in this case, between 1-2 bits per unassigned index). For +// instance, if you allocate an array of size 5 and assign a[2] = <big +// struct>, then a[2] will take up a lot of memory but a[0], a[1], +// a[3], and a[4] will not. Array elements that have a value are +// called "assigned". Array elements that have no value yet, or have +// had their value cleared using erase() or clear(), are called +// "unassigned". +// +// Unassigned values seem to have the default value of T (see below). +// Nevertheless, there is a difference between an unassigned index and +// one explicitly assigned the value of T(). The latter is considered +// assigned. +// +// Access to an array element is constant time, as is insertion and +// deletion. Insertion and deletion may be fairly slow, however: +// because of this container's memory economy, each insert and delete +// causes a memory reallocation. +// +// NOTE: You should not test(), get(), or set() any index that is +// greater than sparsetable.size(). If you need to do that, call +// resize() first. +// +// --- Template parameters +// PARAMETER DESCRIPTION DEFAULT +// T The value of the array: the type of -- +// object that is stored in the array. +// +// GROUP_SIZE How large each "group" in the table 48 +// is (see below). Larger values use +// a little less memory but cause most +// operations to be a little slower +// +// Alloc: Allocator to use to allocate memory. libc_allocator_with_realloc +// +// --- Model of +// Random Access Container +// +// --- Type requirements +// T must be Copy Constructible. It need not be Assignable. +// +// --- Public base classes +// None. +// +// --- Members +// Type members +// +// MEMBER WHERE DEFINED DESCRIPTION +// value_type container The type of object, T, stored in the array +// allocator_type container Allocator to use +// pointer container Pointer to p +// const_pointer container Const pointer to p +// reference container Reference to t +// const_reference container Const reference to t +// size_type container An unsigned integral type +// difference_type container A signed integral type +// iterator [*] container Iterator used to iterate over a sparsetable +// const_iterator container Const iterator used to iterate over a table +// reverse_iterator reversible Iterator used to iterate backwards over +// container a sparsetable +// const_reverse_iterator reversible container Guess +// nonempty_iterator [+] sparsetable Iterates over assigned +// array elements only +// const_nonempty_iterator sparsetable Iterates over assigned +// array elements only +// reverse_nonempty_iterator sparsetable Iterates backwards over +// assigned array elements only +// const_reverse_nonempty_iterator sparsetable Iterates backwards over +// assigned array elements only +// +// [*] All iterators are const in a sparsetable (though nonempty_iterators +// may not be). Use get() and set() to assign values, not iterators. +// +// [+] iterators are random-access iterators. nonempty_iterators are +// bidirectional iterators. + +// Iterator members +// MEMBER WHERE DEFINED DESCRIPTION +// +// iterator begin() container An iterator to the beginning of the table +// iterator end() container An iterator to the end of the table +// const_iterator container A const_iterator pointing to the +// begin() const beginning of a sparsetable +// const_iterator container A const_iterator pointing to the +// end() const end of a sparsetable +// +// reverse_iterator reversable Points to beginning of a reversed +// rbegin() container sparsetable +// reverse_iterator reversable Points to end of a reversed table +// rend() container +// const_reverse_iterator reversable Points to beginning of a +// rbegin() const container reversed sparsetable +// const_reverse_iterator reversable Points to end of a reversed table +// rend() const container +// +// nonempty_iterator sparsetable Points to first assigned element +// begin() of a sparsetable +// nonempty_iterator sparsetable Points past last assigned element +// end() of a sparsetable +// const_nonempty_iterator sparsetable Points to first assigned element +// begin() const of a sparsetable +// const_nonempty_iterator sparsetable Points past last assigned element +// end() const of a sparsetable +// +// reverse_nonempty_iterator sparsetable Points to first assigned element +// begin() of a reversed sparsetable +// reverse_nonempty_iterator sparsetable Points past last assigned element +// end() of a reversed sparsetable +// const_reverse_nonempty_iterator sparsetable Points to first assigned +// begin() const elt of a reversed sparsetable +// const_reverse_nonempty_iterator sparsetable Points past last assigned +// end() const elt of a reversed sparsetable +// +// +// Other members +// MEMBER WHERE DEFINED DESCRIPTION +// sparsetable() sparsetable A table of size 0; must resize() +// before using. +// sparsetable(size_type size) sparsetable A table of size size. All +// indices are unassigned. +// sparsetable( +// const sparsetable &tbl) sparsetable Copy constructor +// ~sparsetable() sparsetable The destructor +// sparsetable &operator=( sparsetable The assignment operator +// const sparsetable &tbl) +// +// void resize(size_type size) sparsetable Grow or shrink a table to +// have size indices [*] +// +// void swap(sparsetable &x) sparsetable Swap two sparsetables +// void swap(sparsetable &x, sparsetable Swap two sparsetables +// sparsetable &y) (global, not member, function) +// +// size_type size() const sparsetable Number of "buckets" in the table +// size_type max_size() const sparsetable Max allowed size of a sparsetable +// bool empty() const sparsetable true if size() == 0 +// size_type num_nonempty() const sparsetable Number of assigned "buckets" +// +// const_reference get( sparsetable Value at index i, or default +// size_type i) const value if i is unassigned +// const_reference operator[]( sparsetable Identical to get(i) [+] +// difference_type i) const +// reference set(size_type i, sparsetable Set element at index i to +// const_reference val) be a copy of val +// bool test(size_type i) sparsetable True if element at index i +// const has been assigned to +// bool test(iterator pos) sparsetable True if element pointed to +// const by pos has been assigned to +// void erase(iterator pos) sparsetable Set element pointed to by +// pos to be unassigned [!] +// void erase(size_type i) sparsetable Set element i to be unassigned +// void erase(iterator start, sparsetable Erases all elements between +// iterator end) start and end +// void clear() sparsetable Erases all elements in the table +// +// I/O versions exist for both FILE* and for File* (Google2-style files): +// bool write_metadata(FILE *fp) sparsetable Writes a sparsetable to the +// bool write_metadata(File *fp) given file. true if write +// completes successfully +// bool read_metadata(FILE *fp) sparsetable Replaces sparsetable with +// bool read_metadata(File *fp) version read from fp. true +// if read completes sucessfully +// bool write_nopointer_data(FILE *fp) Read/write the data stored in +// bool read_nopointer_data(FILE*fp) the table, if it's simple +// +// bool operator==( forward Tests two tables for equality. +// const sparsetable &t1, container This is a global function, +// const sparsetable &t2) not a member function. +// bool operator<( forward Lexicographical comparison. +// const sparsetable &t1, container This is a global function, +// const sparsetable &t2) not a member function. +// +// [*] If you shrink a sparsetable using resize(), assigned elements +// past the end of the table are removed using erase(). If you grow +// a sparsetable, new unassigned indices are created. +// +// [+] Note that operator[] returns a const reference. You must use +// set() to change the value of a table element. +// +// [!] Unassignment also calls the destructor. +// +// Iterators are invalidated whenever an item is inserted or +// deleted (ie set() or erase() is used) or when the size of +// the table changes (ie resize() or clear() is used). +// +// See doc/sparsetable.html for more information about how to use this class. + +// Note: this uses STL style for naming, rather than Google naming. +// That's because this is an STL-y container + +#ifndef UTIL_GTL_SPARSETABLE_H_ +#define UTIL_GTL_SPARSETABLE_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <stdlib.h> // for malloc/free +#include <stdio.h> // to read/write tables +#include <string.h> // for memcpy +#ifdef HAVE_STDINT_H +#include <stdint.h> // the normal place uint16_t is defined +#endif +#ifdef HAVE_SYS_TYPES_H +#include <sys/types.h> // the normal place u_int16_t is defined +#endif +#ifdef HAVE_INTTYPES_H +#include <inttypes.h> // a third place for uint16_t or u_int16_t +#endif +#include <assert.h> // for bounds checking +#include <iterator> // to define reverse_iterator for me +#include <algorithm> // equal, lexicographical_compare, swap,... +#include <memory> // uninitialized_copy, uninitialized_fill +#include <vector> // a sparsetable is a vector of groups +#include <sparsehash/type_traits.h> +#include <sparsehash/internal/hashtable-common.h> +#include <sparsehash/internal/libc_allocator_with_realloc.h> + +// A lot of work to get a type that's guaranteed to be 16 bits... +#ifndef HAVE_U_INT16_T +# if defined HAVE_UINT16_T + typedef uint16_t u_int16_t; // true on solaris, possibly other C99 libc's +# elif defined HAVE___UINT16 + typedef __int16 int16_t; // true on vc++7 + typedef unsigned __int16 u_int16_t; +# else + // Cannot find a 16-bit integer type. Hoping for the best with "short"... + typedef short int int16_t; + typedef unsigned short int u_int16_t; +# endif +#endif + +_START_GOOGLE_NAMESPACE_ + +namespace base { // just to make google->opensource transition easier +using GOOGLE_NAMESPACE::true_type; +using GOOGLE_NAMESPACE::false_type; +using GOOGLE_NAMESPACE::integral_constant; +using GOOGLE_NAMESPACE::has_trivial_copy; +using GOOGLE_NAMESPACE::has_trivial_destructor; +using GOOGLE_NAMESPACE::is_same; +} + + +// The smaller this is, the faster lookup is (because the group bitmap is +// smaller) and the faster insert is, because there's less to move. +// On the other hand, there are more groups. Since group::size_type is +// a short, this number should be of the form 32*x + 16 to avoid waste. +static const u_int16_t DEFAULT_SPARSEGROUP_SIZE = 48; // fits in 1.5 words + + +// Our iterator as simple as iterators can be: basically it's just +// the index into our table. Dereference, the only complicated +// thing, we punt to the table class. This just goes to show how +// much machinery STL requires to do even the most trivial tasks. +// +// A NOTE ON ASSIGNING: +// A sparse table does not actually allocate memory for entries +// that are not filled. Because of this, it becomes complicated +// to have a non-const iterator: we don't know, if the iterator points +// to a not-filled bucket, whether you plan to fill it with something +// or whether you plan to read its value (in which case you'll get +// the default bucket value). Therefore, while we can define const +// operations in a pretty 'normal' way, for non-const operations, we +// define something that returns a helper object with operator= and +// operator& that allocate a bucket lazily. We use this for table[] +// and also for regular table iterators. + +template <class tabletype> +class table_element_adaptor { + public: + typedef typename tabletype::value_type value_type; + typedef typename tabletype::size_type size_type; + typedef typename tabletype::reference reference; + typedef typename tabletype::pointer pointer; + + table_element_adaptor(tabletype *tbl, size_type p) + : table(tbl), pos(p) { } + table_element_adaptor& operator= (const value_type &val) { + table->set(pos, val); + return *this; + } + operator value_type() { return table->get(pos); } // we look like a value + pointer operator& () { return &table->mutating_get(pos); } + + private: + tabletype* table; + size_type pos; +}; + +// Our iterator as simple as iterators can be: basically it's just +// the index into our table. Dereference, the only complicated +// thing, we punt to the table class. This just goes to show how +// much machinery STL requires to do even the most trivial tasks. +// +// By templatizing over tabletype, we have one iterator type which +// we can use for both sparsetables and sparsebins. In fact it +// works on any class that allows size() and operator[] (eg vector), +// as long as it does the standard STL typedefs too (eg value_type). + +template <class tabletype> +class table_iterator { + public: + typedef table_iterator iterator; + + typedef std::random_access_iterator_tag iterator_category; + typedef typename tabletype::value_type value_type; + typedef typename tabletype::difference_type difference_type; + typedef typename tabletype::size_type size_type; + typedef table_element_adaptor<tabletype> reference; + typedef table_element_adaptor<tabletype>* pointer; + + // The "real" constructor + table_iterator(tabletype *tbl, size_type p) + : table(tbl), pos(p) { } + // The default constructor, used when I define vars of type table::iterator + table_iterator() : table(NULL), pos(0) { } + // The copy constructor, for when I say table::iterator foo = tbl.begin() + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // The main thing our iterator does is dereference. If the table entry + // we point to is empty, we return the default value type. + // This is the big different function from the const iterator. + reference operator*() { + return table_element_adaptor<tabletype>(table, pos); + } + pointer operator->() { return &(operator*()); } + + // Helper function to assert things are ok; eg pos is still in range + void check() const { + assert(table); + assert(pos <= table->size()); + } + + // Arithmetic: we just do arithmetic on pos. We don't even need to + // do bounds checking, since STL doesn't consider that its job. :-) + iterator& operator+=(size_type t) { pos += t; check(); return *this; } + iterator& operator-=(size_type t) { pos -= t; check(); return *this; } + iterator& operator++() { ++pos; check(); return *this; } + iterator& operator--() { --pos; check(); return *this; } + iterator operator++(int) { iterator tmp(*this); // for x++ + ++pos; check(); return tmp; } + iterator operator--(int) { iterator tmp(*this); // for x-- + --pos; check(); return tmp; } + iterator operator+(difference_type i) const { iterator tmp(*this); + tmp += i; return tmp; } + iterator operator-(difference_type i) const { iterator tmp(*this); + tmp -= i; return tmp; } + difference_type operator-(iterator it) const { // for "x = it2 - it" + assert(table == it.table); + return pos - it.pos; + } + reference operator[](difference_type n) const { + return *(*this + n); // simple though not totally efficient + } + + // Comparisons. + bool operator==(const iterator& it) const { + return table == it.table && pos == it.pos; + } + bool operator<(const iterator& it) const { + assert(table == it.table); // life is bad bad bad otherwise + return pos < it.pos; + } + bool operator!=(const iterator& it) const { return !(*this == it); } + bool operator<=(const iterator& it) const { return !(it < *this); } + bool operator>(const iterator& it) const { return it < *this; } + bool operator>=(const iterator& it) const { return !(*this < it); } + + // Here's the info we actually need to be an iterator + tabletype *table; // so we can dereference and bounds-check + size_type pos; // index into the table +}; + +// support for "3 + iterator" has to be defined outside the class, alas +template<class T> +table_iterator<T> operator+(typename table_iterator<T>::difference_type i, + table_iterator<T> it) { + return it + i; // so people can say it2 = 3 + it +} + +template <class tabletype> +class const_table_iterator { + public: + typedef table_iterator<tabletype> iterator; + typedef const_table_iterator const_iterator; + + typedef std::random_access_iterator_tag iterator_category; + typedef typename tabletype::value_type value_type; + typedef typename tabletype::difference_type difference_type; + typedef typename tabletype::size_type size_type; + typedef typename tabletype::const_reference reference; // we're const-only + typedef typename tabletype::const_pointer pointer; + + // The "real" constructor + const_table_iterator(const tabletype *tbl, size_type p) + : table(tbl), pos(p) { } + // The default constructor, used when I define vars of type table::iterator + const_table_iterator() : table(NULL), pos(0) { } + // The copy constructor, for when I say table::iterator foo = tbl.begin() + // Also converts normal iterators to const iterators + const_table_iterator(const iterator &from) + : table(from.table), pos(from.pos) { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // The main thing our iterator does is dereference. If the table entry + // we point to is empty, we return the default value type. + reference operator*() const { return (*table)[pos]; } + pointer operator->() const { return &(operator*()); } + + // Helper function to assert things are ok; eg pos is still in range + void check() const { + assert(table); + assert(pos <= table->size()); + } + + // Arithmetic: we just do arithmetic on pos. We don't even need to + // do bounds checking, since STL doesn't consider that its job. :-) + const_iterator& operator+=(size_type t) { pos += t; check(); return *this; } + const_iterator& operator-=(size_type t) { pos -= t; check(); return *this; } + const_iterator& operator++() { ++pos; check(); return *this; } + const_iterator& operator--() { --pos; check(); return *this; } + const_iterator operator++(int) { const_iterator tmp(*this); // for x++ + ++pos; check(); return tmp; } + const_iterator operator--(int) { const_iterator tmp(*this); // for x-- + --pos; check(); return tmp; } + const_iterator operator+(difference_type i) const { const_iterator tmp(*this); + tmp += i; return tmp; } + const_iterator operator-(difference_type i) const { const_iterator tmp(*this); + tmp -= i; return tmp; } + difference_type operator-(const_iterator it) const { // for "x = it2 - it" + assert(table == it.table); + return pos - it.pos; + } + reference operator[](difference_type n) const { + return *(*this + n); // simple though not totally efficient + } + + // Comparisons. + bool operator==(const const_iterator& it) const { + return table == it.table && pos == it.pos; + } + bool operator<(const const_iterator& it) const { + assert(table == it.table); // life is bad bad bad otherwise + return pos < it.pos; + } + bool operator!=(const const_iterator& it) const { return !(*this == it); } + bool operator<=(const const_iterator& it) const { return !(it < *this); } + bool operator>(const const_iterator& it) const { return it < *this; } + bool operator>=(const const_iterator& it) const { return !(*this < it); } + + // Here's the info we actually need to be an iterator + const tabletype *table; // so we can dereference and bounds-check + size_type pos; // index into the table +}; + +// support for "3 + iterator" has to be defined outside the class, alas +template<class T> +const_table_iterator<T> operator+(typename + const_table_iterator<T>::difference_type i, + const_table_iterator<T> it) { + return it + i; // so people can say it2 = 3 + it +} + + +// --------------------------------------------------------------------------- + + +/* +// This is a 2-D iterator. You specify a begin and end over a list +// of *containers*. We iterate over each container by iterating over +// it. It's actually simple: +// VECTOR.begin() VECTOR[0].begin() --------> VECTOR[0].end() ---, +// | ________________________________________________/ +// | \_> VECTOR[1].begin() --------> VECTOR[1].end() -, +// | ___________________________________________________/ +// v \_> ...... +// VECTOR.end() +// +// It's impossible to do random access on one of these things in constant +// time, so it's just a bidirectional iterator. +// +// Unfortunately, because we need to use this for a non-empty iterator, +// we use nonempty_begin() and nonempty_end() instead of begin() and end() +// (though only going across, not down). +*/ + +#define TWOD_BEGIN_ nonempty_begin +#define TWOD_END_ nonempty_end +#define TWOD_ITER_ nonempty_iterator +#define TWOD_CONST_ITER_ const_nonempty_iterator + +template <class containertype> +class two_d_iterator { + public: + typedef two_d_iterator iterator; + + typedef std::bidirectional_iterator_tag iterator_category; + // apparently some versions of VC++ have trouble with two ::'s in a typename + typedef typename containertype::value_type _tmp_vt; + typedef typename _tmp_vt::value_type value_type; + typedef typename _tmp_vt::difference_type difference_type; + typedef typename _tmp_vt::reference reference; + typedef typename _tmp_vt::pointer pointer; + + // The "real" constructor. begin and end specify how many rows we have + // (in the diagram above); we always iterate over each row completely. + two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr) + : row_begin(begin), row_end(end), row_current(curr), col_current() { + if ( row_current != row_end ) { + col_current = row_current->TWOD_BEGIN_(); + advance_past_end(); // in case cur->begin() == cur->end() + } + } + // If you want to start at an arbitrary place, you can, I guess + two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr, + typename containertype::value_type::TWOD_ITER_ col) + : row_begin(begin), row_end(end), row_current(curr), col_current(col) { + advance_past_end(); // in case cur->begin() == cur->end() + } + // The default constructor, used when I define vars of type table::iterator + two_d_iterator() : row_begin(), row_end(), row_current(), col_current() { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *col_current; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic: we just do arithmetic on pos. We don't even need to + // do bounds checking, since STL doesn't consider that its job. :-) + // NOTE: this is not amortized constant time! What do we do about it? + void advance_past_end() { // used when col_current points to end() + while ( col_current == row_current->TWOD_END_() ) { // end of current row + ++row_current; // go to beginning of next + if ( row_current != row_end ) // col is irrelevant at end + col_current = row_current->TWOD_BEGIN_(); + else + break; // don't go past row_end + } + } + + iterator& operator++() { + assert(row_current != row_end); // how to ++ from there? + ++col_current; + advance_past_end(); // in case col_current is at end() + return *this; + } + iterator& operator--() { + while ( row_current == row_end || + col_current == row_current->TWOD_BEGIN_() ) { + assert(row_current != row_begin); + --row_current; + col_current = row_current->TWOD_END_(); // this is 1 too far + } + --col_current; + return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + iterator operator--(int) { iterator tmp(*this); --*this; return tmp; } + + + // Comparisons. + bool operator==(const iterator& it) const { + return ( row_begin == it.row_begin && + row_end == it.row_end && + row_current == it.row_current && + (row_current == row_end || col_current == it.col_current) ); + } + bool operator!=(const iterator& it) const { return !(*this == it); } + + + // Here's the info we actually need to be an iterator + // These need to be public so we convert from iterator to const_iterator + typename containertype::iterator row_begin, row_end, row_current; + typename containertype::value_type::TWOD_ITER_ col_current; +}; + +// The same thing again, but this time const. :-( +template <class containertype> +class const_two_d_iterator { + public: + typedef const_two_d_iterator iterator; + + typedef std::bidirectional_iterator_tag iterator_category; + // apparently some versions of VC++ have trouble with two ::'s in a typename + typedef typename containertype::value_type _tmp_vt; + typedef typename _tmp_vt::value_type value_type; + typedef typename _tmp_vt::difference_type difference_type; + typedef typename _tmp_vt::const_reference reference; + typedef typename _tmp_vt::const_pointer pointer; + + const_two_d_iterator(typename containertype::const_iterator begin, + typename containertype::const_iterator end, + typename containertype::const_iterator curr) + : row_begin(begin), row_end(end), row_current(curr), col_current() { + if ( curr != end ) { + col_current = curr->TWOD_BEGIN_(); + advance_past_end(); // in case cur->begin() == cur->end() + } + } + const_two_d_iterator(typename containertype::const_iterator begin, + typename containertype::const_iterator end, + typename containertype::const_iterator curr, + typename containertype::value_type::TWOD_CONST_ITER_ col) + : row_begin(begin), row_end(end), row_current(curr), col_current(col) { + advance_past_end(); // in case cur->begin() == cur->end() + } + const_two_d_iterator() + : row_begin(), row_end(), row_current(), col_current() { + } + // Need this explicitly so we can convert normal iterators to const iterators + const_two_d_iterator(const two_d_iterator<containertype>& it) : + row_begin(it.row_begin), row_end(it.row_end), row_current(it.row_current), + col_current(it.col_current) { } + + typename containertype::const_iterator row_begin, row_end, row_current; + typename containertype::value_type::TWOD_CONST_ITER_ col_current; + + + // EVERYTHING FROM HERE DOWN IS THE SAME AS THE NON-CONST ITERATOR + reference operator*() const { return *col_current; } + pointer operator->() const { return &(operator*()); } + + void advance_past_end() { // used when col_current points to end() + while ( col_current == row_current->TWOD_END_() ) { // end of current row + ++row_current; // go to beginning of next + if ( row_current != row_end ) // col is irrelevant at end + col_current = row_current->TWOD_BEGIN_(); + else + break; // don't go past row_end + } + } + iterator& operator++() { + assert(row_current != row_end); // how to ++ from there? + ++col_current; + advance_past_end(); // in case col_current is at end() + return *this; + } + iterator& operator--() { + while ( row_current == row_end || + col_current == row_current->TWOD_BEGIN_() ) { + assert(row_current != row_begin); + --row_current; + col_current = row_current->TWOD_END_(); // this is 1 too far + } + --col_current; + return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + iterator operator--(int) { iterator tmp(*this); --*this; return tmp; } + + bool operator==(const iterator& it) const { + return ( row_begin == it.row_begin && + row_end == it.row_end && + row_current == it.row_current && + (row_current == row_end || col_current == it.col_current) ); + } + bool operator!=(const iterator& it) const { return !(*this == it); } +}; + +// We provide yet another version, to be as frugal with memory as +// possible. This one frees each block of memory as it finishes +// iterating over it. By the end, the entire table is freed. +// For understandable reasons, you can only iterate over it once, +// which is why it's an input iterator +template <class containertype> +class destructive_two_d_iterator { + public: + typedef destructive_two_d_iterator iterator; + + typedef std::input_iterator_tag iterator_category; + // apparently some versions of VC++ have trouble with two ::'s in a typename + typedef typename containertype::value_type _tmp_vt; + typedef typename _tmp_vt::value_type value_type; + typedef typename _tmp_vt::difference_type difference_type; + typedef typename _tmp_vt::reference reference; + typedef typename _tmp_vt::pointer pointer; + + destructive_two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr) + : row_begin(begin), row_end(end), row_current(curr), col_current() { + if ( curr != end ) { + col_current = curr->TWOD_BEGIN_(); + advance_past_end(); // in case cur->begin() == cur->end() + } + } + destructive_two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr, + typename containertype::value_type::TWOD_ITER_ col) + : row_begin(begin), row_end(end), row_current(curr), col_current(col) { + advance_past_end(); // in case cur->begin() == cur->end() + } + destructive_two_d_iterator() + : row_begin(), row_end(), row_current(), col_current() { + } + + typename containertype::iterator row_begin, row_end, row_current; + typename containertype::value_type::TWOD_ITER_ col_current; + + // This is the part that destroys + void advance_past_end() { // used when col_current points to end() + while ( col_current == row_current->TWOD_END_() ) { // end of current row + row_current->clear(); // the destructive part + // It would be nice if we could decrement sparsetable->num_buckets here + ++row_current; // go to beginning of next + if ( row_current != row_end ) // col is irrelevant at end + col_current = row_current->TWOD_BEGIN_(); + else + break; // don't go past row_end + } + } + + // EVERYTHING FROM HERE DOWN IS THE SAME AS THE REGULAR ITERATOR + reference operator*() const { return *col_current; } + pointer operator->() const { return &(operator*()); } + + iterator& operator++() { + assert(row_current != row_end); // how to ++ from there? + ++col_current; + advance_past_end(); // in case col_current is at end() + return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + bool operator==(const iterator& it) const { + return ( row_begin == it.row_begin && + row_end == it.row_end && + row_current == it.row_current && + (row_current == row_end || col_current == it.col_current) ); + } + bool operator!=(const iterator& it) const { return !(*this == it); } +}; + +#undef TWOD_BEGIN_ +#undef TWOD_END_ +#undef TWOD_ITER_ +#undef TWOD_CONST_ITER_ + + + + +// SPARSE-TABLE +// ------------ +// The idea is that a table with (logically) t buckets is divided +// into t/M *groups* of M buckets each. (M is a constant set in +// GROUP_SIZE for efficiency.) Each group is stored sparsely. +// Thus, inserting into the table causes some array to grow, which is +// slow but still constant time. Lookup involves doing a +// logical-position-to-sparse-position lookup, which is also slow but +// constant time. The larger M is, the slower these operations are +// but the less overhead (slightly). +// +// To store the sparse array, we store a bitmap B, where B[i] = 1 iff +// bucket i is non-empty. Then to look up bucket i we really look up +// array[# of 1s before i in B]. This is constant time for fixed M. +// +// Terminology: the position of an item in the overall table (from +// 1 .. t) is called its "location." The logical position in a group +// (from 1 .. M ) is called its "position." The actual location in +// the array (from 1 .. # of non-empty buckets in the group) is +// called its "offset." + +template <class T, u_int16_t GROUP_SIZE, class Alloc> +class sparsegroup { + private: + typedef typename Alloc::template rebind<T>::other value_alloc_type; + + public: + // Basic types + typedef T value_type; + typedef Alloc allocator_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + + typedef table_iterator<sparsegroup<T, GROUP_SIZE, Alloc> > iterator; + typedef const_table_iterator<sparsegroup<T, GROUP_SIZE, Alloc> > + const_iterator; + typedef table_element_adaptor<sparsegroup<T, GROUP_SIZE, Alloc> > + element_adaptor; + typedef u_int16_t size_type; // max # of buckets + typedef int16_t difference_type; + typedef std::reverse_iterator<const_iterator> const_reverse_iterator; + typedef std::reverse_iterator<iterator> reverse_iterator; // from iterator.h + + // These are our special iterators, that go over non-empty buckets in a + // group. These aren't const-only because you can change non-empty bcks. + typedef pointer nonempty_iterator; + typedef const_pointer const_nonempty_iterator; + typedef std::reverse_iterator<nonempty_iterator> reverse_nonempty_iterator; + typedef std::reverse_iterator<const_nonempty_iterator> const_reverse_nonempty_iterator; + + // Iterator functions + iterator begin() { return iterator(this, 0); } + const_iterator begin() const { return const_iterator(this, 0); } + iterator end() { return iterator(this, size()); } + const_iterator end() const { return const_iterator(this, size()); } + reverse_iterator rbegin() { return reverse_iterator(end()); } + const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } + reverse_iterator rend() { return reverse_iterator(begin()); } + const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } + + // We'll have versions for our special non-empty iterator too + nonempty_iterator nonempty_begin() { return group; } + const_nonempty_iterator nonempty_begin() const { return group; } + nonempty_iterator nonempty_end() { + return group + settings.num_buckets; + } + const_nonempty_iterator nonempty_end() const { + return group + settings.num_buckets; + } + reverse_nonempty_iterator nonempty_rbegin() { + return reverse_nonempty_iterator(nonempty_end()); + } + const_reverse_nonempty_iterator nonempty_rbegin() const { + return const_reverse_nonempty_iterator(nonempty_end()); + } + reverse_nonempty_iterator nonempty_rend() { + return reverse_nonempty_iterator(nonempty_begin()); + } + const_reverse_nonempty_iterator nonempty_rend() const { + return const_reverse_nonempty_iterator(nonempty_begin()); + } + + + // This gives us the "default" value to return for an empty bucket. + // We just use the default constructor on T, the template type + const_reference default_value() const { + static value_type defaultval = value_type(); + return defaultval; + } + + + private: + // We need to do all this bit manipulation, of course. ick + static size_type charbit(size_type i) { return i >> 3; } + static size_type modbit(size_type i) { return 1 << (i&7); } + int bmtest(size_type i) const { return bitmap[charbit(i)] & modbit(i); } + void bmset(size_type i) { bitmap[charbit(i)] |= modbit(i); } + void bmclear(size_type i) { bitmap[charbit(i)] &= ~modbit(i); } + + pointer allocate_group(size_type n) { + pointer retval = settings.allocate(n); + if (retval == NULL) { + // We really should use PRIuS here, but I don't want to have to add + // a whole new configure option, with concomitant macro namespace + // pollution, just to print this (unlikely) error message. So I cast. + fprintf(stderr, "sparsehash FATAL ERROR: failed to allocate %lu groups\n", + static_cast<unsigned long>(n)); + exit(1); + } + return retval; + } + + void free_group() { + if (!group) return; + pointer end_it = group + settings.num_buckets; + for (pointer p = group; p != end_it; ++p) + p->~value_type(); + settings.deallocate(group, settings.num_buckets); + group = NULL; + } + + static size_type bits_in_char(unsigned char c) { + // We could make these ints. The tradeoff is size (eg does it overwhelm + // the cache?) vs efficiency in referencing sub-word-sized array elements. + static const char bits_in[256] = { + 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8, + }; + return bits_in[c]; + } + + public: // get_iter() in sparsetable needs it + // We need a small function that tells us how many set bits there are + // in positions 0..i-1 of the bitmap. It uses a big table. + // We make it static so templates don't allocate lots of these tables. + // There are lots of ways to do this calculation (called 'popcount'). + // The 8-bit table lookup is one of the fastest, though this + // implementation suffers from not doing any loop unrolling. See, eg, + // http://www.dalkescientific.com/writings/diary/archive/2008/07/03/hakmem_and_other_popcounts.html + // http://gurmeetsingh.wordpress.com/2008/08/05/fast-bit-counting-routines/ + static size_type pos_to_offset(const unsigned char *bm, size_type pos) { + size_type retval = 0; + + // [Note: condition pos > 8 is an optimization; convince yourself we + // give exactly the same result as if we had pos >= 8 here instead.] + for ( ; pos > 8; pos -= 8 ) // bm[0..pos/8-1] + retval += bits_in_char(*bm++); // chars we want *all* bits in + return retval + bits_in_char(*bm & ((1 << pos)-1)); // char including pos + } + + size_type pos_to_offset(size_type pos) const { // not static but still const + return pos_to_offset(bitmap, pos); + } + + // Returns the (logical) position in the bm[] array, i, such that + // bm[i] is the offset-th set bit in the array. It is the inverse + // of pos_to_offset. get_pos() uses this function to find the index + // of an nonempty_iterator in the table. Bit-twiddling from + // http://hackersdelight.org/basics.pdf + static size_type offset_to_pos(const unsigned char *bm, size_type offset) { + size_type retval = 0; + // This is sizeof(this->bitmap). + const size_type group_size = (GROUP_SIZE-1) / 8 + 1; + for (size_type i = 0; i < group_size; i++) { // forward scan + const size_type pop_count = bits_in_char(*bm); + if (pop_count > offset) { + unsigned char last_bm = *bm; + for (; offset > 0; offset--) { + last_bm &= (last_bm-1); // remove right-most set bit + } + // Clear all bits to the left of the rightmost bit (the &), + // and then clear the rightmost bit but set all bits to the + // right of it (the -1). + last_bm = (last_bm & -last_bm) - 1; + retval += bits_in_char(last_bm); + return retval; + } + offset -= pop_count; + retval += 8; + bm++; + } + return retval; + } + + size_type offset_to_pos(size_type offset) const { + return offset_to_pos(bitmap, offset); + } + + + public: + // Constructors -- default and copy -- and destructor + explicit sparsegroup(allocator_type& a) : + group(0), settings(alloc_impl<value_alloc_type>(a)) { + memset(bitmap, 0, sizeof(bitmap)); + } + sparsegroup(const sparsegroup& x) : group(0), settings(x.settings) { + if ( settings.num_buckets ) { + group = allocate_group(x.settings.num_buckets); + std::uninitialized_copy(x.group, x.group + x.settings.num_buckets, group); + } + memcpy(bitmap, x.bitmap, sizeof(bitmap)); + } + ~sparsegroup() { free_group(); } + + // Operator= is just like the copy constructor, I guess + // TODO(austern): Make this exception safe. Handle exceptions in value_type's + // copy constructor. + sparsegroup &operator=(const sparsegroup& x) { + if ( &x == this ) return *this; // x = x + if ( x.settings.num_buckets == 0 ) { + free_group(); + } else { + pointer p = allocate_group(x.settings.num_buckets); + std::uninitialized_copy(x.group, x.group + x.settings.num_buckets, p); + free_group(); + group = p; + } + memcpy(bitmap, x.bitmap, sizeof(bitmap)); + settings.num_buckets = x.settings.num_buckets; + return *this; + } + + // Many STL algorithms use swap instead of copy constructors + void swap(sparsegroup& x) { + std::swap(group, x.group); // defined in <algorithm> + for ( int i = 0; i < sizeof(bitmap) / sizeof(*bitmap); ++i ) + std::swap(bitmap[i], x.bitmap[i]); // swap not defined on arrays + std::swap(settings.num_buckets, x.settings.num_buckets); + // we purposefully don't swap the allocator, which may not be swap-able + } + + // It's always nice to be able to clear a table without deallocating it + void clear() { + free_group(); + memset(bitmap, 0, sizeof(bitmap)); + settings.num_buckets = 0; + } + + // Functions that tell you about size. Alas, these aren't so useful + // because our table is always fixed size. + size_type size() const { return GROUP_SIZE; } + size_type max_size() const { return GROUP_SIZE; } + bool empty() const { return false; } + // We also may want to know how many *used* buckets there are + size_type num_nonempty() const { return settings.num_buckets; } + + + // get()/set() are explicitly const/non-const. You can use [] if + // you want something that can be either (potentially more expensive). + const_reference get(size_type i) const { + if ( bmtest(i) ) // bucket i is occupied + return group[pos_to_offset(bitmap, i)]; + else + return default_value(); // return the default reference + } + + // TODO(csilvers): make protected + friend + // This is used by sparse_hashtable to get an element from the table + // when we know it exists. + const_reference unsafe_get(size_type i) const { + assert(bmtest(i)); + return group[pos_to_offset(bitmap, i)]; + } + + // TODO(csilvers): make protected + friend + reference mutating_get(size_type i) { // fills bucket i before getting + if ( !bmtest(i) ) + set(i, default_value()); + return group[pos_to_offset(bitmap, i)]; + } + + // Syntactic sugar. It's easy to return a const reference. To + // return a non-const reference, we need to use the assigner adaptor. + const_reference operator[](size_type i) const { + return get(i); + } + + element_adaptor operator[](size_type i) { + return element_adaptor(this, i); + } + + private: + // Create space at group[offset], assuming value_type has trivial + // copy constructor and destructor, and the allocator_type is + // the default libc_allocator_with_alloc. (Really, we want it to have + // "trivial move", because that's what realloc and memmove both do. + // But there's no way to capture that using type_traits, so we + // pretend that move(x, y) is equivalent to "x.~T(); new(x) T(y);" + // which is pretty much correct, if a bit conservative.) + void set_aux(size_type offset, base::true_type) { + group = settings.realloc_or_die(group, settings.num_buckets+1); + // This is equivalent to memmove(), but faster on my Intel P4, + // at least with gcc4.1 -O2 / glibc 2.3.6. + for (size_type i = settings.num_buckets; i > offset; --i) // cast to void* to prevent compiler warnings about writing to an object // with no trivial copy-assignment memcpy(static_cast<void*>(group + i), group + i-1, sizeof(*group)); - } - - // Create space at group[offset], without special assumptions about value_type - // and allocator_type. - void set_aux(size_type offset, base::false_type) { - // This is valid because 0 <= offset <= num_buckets - pointer p = allocate_group(settings.num_buckets + 1); - std::uninitialized_copy(group, group + offset, p); - std::uninitialized_copy(group + offset, group + settings.num_buckets, - p + offset + 1); - free_group(); - group = p; - } - - public: - // This returns a reference to the inserted item (which is a copy of val). - // TODO(austern): Make this exception safe: handle exceptions from - // value_type's copy constructor. - reference set(size_type i, const_reference val) { - size_type offset = pos_to_offset(bitmap, i); // where we'll find (or insert) - if ( bmtest(i) ) { - // Delete the old value, which we're replacing with the new one - group[offset].~value_type(); - } else { - typedef base::integral_constant<bool, - (base::has_trivial_copy<value_type>::value && - base::has_trivial_destructor<value_type>::value && - base::is_same< - allocator_type, - libc_allocator_with_realloc<value_type> >::value)> - realloc_and_memmove_ok; // we pretend mv(x,y) == "x.~T(); new(x) T(y)" - set_aux(offset, realloc_and_memmove_ok()); - ++settings.num_buckets; - bmset(i); - } - // This does the actual inserting. Since we made the array using - // malloc, we use "placement new" to just call the constructor. - new(&group[offset]) value_type(val); - return group[offset]; - } - - // We let you see if a bucket is non-empty without retrieving it - bool test(size_type i) const { - return bmtest(i) != 0; - } - bool test(iterator pos) const { - return bmtest(pos.pos) != 0; - } - - private: - // Shrink the array, assuming value_type has trivial copy - // constructor and destructor, and the allocator_type is the default - // libc_allocator_with_alloc. (Really, we want it to have "trivial - // move", because that's what realloc and memmove both do. But - // there's no way to capture that using type_traits, so we pretend - // that move(x, y) is equivalent to ""x.~T(); new(x) T(y);" - // which is pretty much correct, if a bit conservative.) - void erase_aux(size_type offset, base::true_type) { - // This isn't technically necessary, since we know we have a - // trivial destructor, but is a cheap way to get a bit more safety. - group[offset].~value_type(); - // This is equivalent to memmove(), but faster on my Intel P4, - // at lesat with gcc4.1 -O2 / glibc 2.3.6. - assert(settings.num_buckets > 0); - for (size_type i = offset; i < settings.num_buckets-1; ++i) + } + + // Create space at group[offset], without special assumptions about value_type + // and allocator_type. + void set_aux(size_type offset, base::false_type) { + // This is valid because 0 <= offset <= num_buckets + pointer p = allocate_group(settings.num_buckets + 1); + std::uninitialized_copy(group, group + offset, p); + std::uninitialized_copy(group + offset, group + settings.num_buckets, + p + offset + 1); + free_group(); + group = p; + } + + public: + // This returns a reference to the inserted item (which is a copy of val). + // TODO(austern): Make this exception safe: handle exceptions from + // value_type's copy constructor. + reference set(size_type i, const_reference val) { + size_type offset = pos_to_offset(bitmap, i); // where we'll find (or insert) + if ( bmtest(i) ) { + // Delete the old value, which we're replacing with the new one + group[offset].~value_type(); + } else { + typedef base::integral_constant<bool, + (base::has_trivial_copy<value_type>::value && + base::has_trivial_destructor<value_type>::value && + base::is_same< + allocator_type, + libc_allocator_with_realloc<value_type> >::value)> + realloc_and_memmove_ok; // we pretend mv(x,y) == "x.~T(); new(x) T(y)" + set_aux(offset, realloc_and_memmove_ok()); + ++settings.num_buckets; + bmset(i); + } + // This does the actual inserting. Since we made the array using + // malloc, we use "placement new" to just call the constructor. + new(&group[offset]) value_type(val); + return group[offset]; + } + + // We let you see if a bucket is non-empty without retrieving it + bool test(size_type i) const { + return bmtest(i) != 0; + } + bool test(iterator pos) const { + return bmtest(pos.pos) != 0; + } + + private: + // Shrink the array, assuming value_type has trivial copy + // constructor and destructor, and the allocator_type is the default + // libc_allocator_with_alloc. (Really, we want it to have "trivial + // move", because that's what realloc and memmove both do. But + // there's no way to capture that using type_traits, so we pretend + // that move(x, y) is equivalent to ""x.~T(); new(x) T(y);" + // which is pretty much correct, if a bit conservative.) + void erase_aux(size_type offset, base::true_type) { + // This isn't technically necessary, since we know we have a + // trivial destructor, but is a cheap way to get a bit more safety. + group[offset].~value_type(); + // This is equivalent to memmove(), but faster on my Intel P4, + // at lesat with gcc4.1 -O2 / glibc 2.3.6. + assert(settings.num_buckets > 0); + for (size_type i = offset; i < settings.num_buckets-1; ++i) // cast to void* to prevent compiler warnings about writing to an object // with no trivial copy-assignment // hopefully inlined! memcpy(static_cast<void*>(group + i), group + i+1, sizeof(*group)); - group = settings.realloc_or_die(group, settings.num_buckets-1); - } - - // Shrink the array, without any special assumptions about value_type and - // allocator_type. - void erase_aux(size_type offset, base::false_type) { - // This is valid because 0 <= offset < num_buckets. Note the inequality. - pointer p = allocate_group(settings.num_buckets - 1); - std::uninitialized_copy(group, group + offset, p); - std::uninitialized_copy(group + offset + 1, group + settings.num_buckets, - p + offset); - free_group(); - group = p; - } - - public: - // This takes the specified elements out of the group. This is - // "undefining", rather than "clearing". - // TODO(austern): Make this exception safe: handle exceptions from - // value_type's copy constructor. - void erase(size_type i) { - if ( bmtest(i) ) { // trivial to erase empty bucket - size_type offset = pos_to_offset(bitmap,i); // where we'll find (or insert) - if ( settings.num_buckets == 1 ) { - free_group(); - group = NULL; - } else { - typedef base::integral_constant<bool, - (base::has_trivial_copy<value_type>::value && - base::has_trivial_destructor<value_type>::value && - base::is_same< - allocator_type, - libc_allocator_with_realloc<value_type> >::value)> - realloc_and_memmove_ok; // pretend mv(x,y) == "x.~T(); new(x) T(y)" - erase_aux(offset, realloc_and_memmove_ok()); - } - --settings.num_buckets; - bmclear(i); - } - } - - void erase(iterator pos) { - erase(pos.pos); - } - - void erase(iterator start_it, iterator end_it) { - // This could be more efficient, but to do so we'd need to make - // bmclear() clear a range of indices. Doesn't seem worth it. - for ( ; start_it != end_it; ++start_it ) - erase(start_it); - } - - - // I/O - // We support reading and writing groups to disk. We don't store - // the actual array contents (which we don't know how to store), - // just the bitmap and size. Meant to be used with table I/O. - - template <typename OUTPUT> bool write_metadata(OUTPUT *fp) const { - // we explicitly set to u_int16_t - assert(sizeof(settings.num_buckets) == 2); - if ( !sparsehash_internal::write_bigendian_number(fp, settings.num_buckets, - 2) ) - return false; - if ( !sparsehash_internal::write_data(fp, bitmap, sizeof(bitmap)) ) - return false; - return true; - } - - // Reading destroys the old group contents! Returns true if all was ok. - template <typename INPUT> bool read_metadata(INPUT *fp) { - clear(); - if ( !sparsehash_internal::read_bigendian_number(fp, &settings.num_buckets, - 2) ) - return false; - if ( !sparsehash_internal::read_data(fp, bitmap, sizeof(bitmap)) ) - return false; - // We'll allocate the space, but we won't fill it: it will be - // left as uninitialized raw memory. - group = allocate_group(settings.num_buckets); - return true; - } - - // Again, only meaningful if value_type is a POD. - template <typename INPUT> bool read_nopointer_data(INPUT *fp) { - for ( nonempty_iterator it = nonempty_begin(); - it != nonempty_end(); ++it ) { - if ( !sparsehash_internal::read_data(fp, &(*it), sizeof(*it)) ) - return false; - } - return true; - } - - // If your keys and values are simple enough, we can write them - // to disk for you. "simple enough" means POD and no pointers. - // However, we don't try to normalize endianness. - template <typename OUTPUT> bool write_nopointer_data(OUTPUT *fp) const { - for ( const_nonempty_iterator it = nonempty_begin(); - it != nonempty_end(); ++it ) { - if ( !sparsehash_internal::write_data(fp, &(*it), sizeof(*it)) ) - return false; - } - return true; - } - - - // Comparisons. We only need to define == and < -- we get - // != > <= >= via relops.h (which we happily included above). - // Note the comparisons are pretty arbitrary: we compare - // values of the first index that isn't equal (using default - // value for empty buckets). - bool operator==(const sparsegroup& x) const { - return ( settings.num_buckets == x.settings.num_buckets && - memcmp(bitmap, x.bitmap, sizeof(bitmap)) == 0 && - std::equal(begin(), end(), x.begin()) ); // from <algorithm> - } - - bool operator<(const sparsegroup& x) const { // also from <algorithm> - return std::lexicographical_compare(begin(), end(), x.begin(), x.end()); - } - bool operator!=(const sparsegroup& x) const { return !(*this == x); } - bool operator<=(const sparsegroup& x) const { return !(x < *this); } - bool operator>(const sparsegroup& x) const { return x < *this; } - bool operator>=(const sparsegroup& x) const { return !(*this < x); } - - private: - template <class A> - class alloc_impl : public A { - public: - typedef typename A::pointer pointer; - typedef typename A::size_type size_type; - - // Convert a normal allocator to one that has realloc_or_die() - alloc_impl(const A& a) : A(a) { } - - // realloc_or_die should only be used when using the default - // allocator (libc_allocator_with_realloc). - pointer realloc_or_die(pointer /*ptr*/, size_type /*n*/) { - fprintf(stderr, "realloc_or_die is only supported for " - "libc_allocator_with_realloc\n"); - exit(1); - return NULL; - } - }; - - // A template specialization of alloc_impl for - // libc_allocator_with_realloc that can handle realloc_or_die. - template <class A> - class alloc_impl<libc_allocator_with_realloc<A> > - : public libc_allocator_with_realloc<A> { - public: - typedef typename libc_allocator_with_realloc<A>::pointer pointer; - typedef typename libc_allocator_with_realloc<A>::size_type size_type; - - alloc_impl(const libc_allocator_with_realloc<A>& a) - : libc_allocator_with_realloc<A>(a) { } - - pointer realloc_or_die(pointer ptr, size_type n) { - pointer retval = this->reallocate(ptr, n); - if (retval == NULL) { - fprintf(stderr, "sparsehash: FATAL ERROR: failed to reallocate " - "%lu elements for ptr %p", static_cast<unsigned long>(n), ptr); - exit(1); - } - return retval; - } - }; - - // Package allocator with num_buckets to eliminate memory needed for the - // zero-size allocator. - // If new fields are added to this class, we should add them to - // operator= and swap. - class Settings : public alloc_impl<value_alloc_type> { - public: - Settings(const alloc_impl<value_alloc_type>& a, u_int16_t n = 0) - : alloc_impl<value_alloc_type>(a), num_buckets(n) { } - Settings(const Settings& s) - : alloc_impl<value_alloc_type>(s), num_buckets(s.num_buckets) { } - - u_int16_t num_buckets; // limits GROUP_SIZE to 64K - }; - - // The actual data - pointer group; // (small) array of T's - Settings settings; // allocator and num_buckets - unsigned char bitmap[(GROUP_SIZE-1)/8 + 1]; // fancy math is so we round up -}; - -// We need a global swap as well -template <class T, u_int16_t GROUP_SIZE, class Alloc> -inline void swap(sparsegroup<T,GROUP_SIZE,Alloc> &x, - sparsegroup<T,GROUP_SIZE,Alloc> &y) { - x.swap(y); -} - -// --------------------------------------------------------------------------- - - -template <class T, u_int16_t GROUP_SIZE = DEFAULT_SPARSEGROUP_SIZE, - class Alloc = libc_allocator_with_realloc<T> > -class sparsetable { - private: - typedef typename Alloc::template rebind<T>::other value_alloc_type; - typedef typename Alloc::template rebind< - sparsegroup<T, GROUP_SIZE, value_alloc_type> >::other vector_alloc; - - public: - // Basic types - typedef T value_type; // stolen from stl_vector.h - typedef Alloc allocator_type; - typedef typename value_alloc_type::size_type size_type; - typedef typename value_alloc_type::difference_type difference_type; - typedef typename value_alloc_type::reference reference; - typedef typename value_alloc_type::const_reference const_reference; - typedef typename value_alloc_type::pointer pointer; - typedef typename value_alloc_type::const_pointer const_pointer; - typedef table_iterator<sparsetable<T, GROUP_SIZE, Alloc> > iterator; - typedef const_table_iterator<sparsetable<T, GROUP_SIZE, Alloc> > - const_iterator; - typedef table_element_adaptor<sparsetable<T, GROUP_SIZE, Alloc> > - element_adaptor; - typedef std::reverse_iterator<const_iterator> const_reverse_iterator; - typedef std::reverse_iterator<iterator> reverse_iterator; // from iterator.h - - // These are our special iterators, that go over non-empty buckets in a - // table. These aren't const only because you can change non-empty bcks. - typedef two_d_iterator< std::vector< sparsegroup<value_type, GROUP_SIZE, - value_alloc_type>, - vector_alloc> > - nonempty_iterator; - typedef const_two_d_iterator< std::vector< sparsegroup<value_type, - GROUP_SIZE, - value_alloc_type>, - vector_alloc> > - const_nonempty_iterator; - typedef std::reverse_iterator<nonempty_iterator> reverse_nonempty_iterator; - typedef std::reverse_iterator<const_nonempty_iterator> const_reverse_nonempty_iterator; - // Another special iterator: it frees memory as it iterates (used to resize) - typedef destructive_two_d_iterator< std::vector< sparsegroup<value_type, - GROUP_SIZE, - value_alloc_type>, - vector_alloc> > - destructive_iterator; - - // Iterator functions - iterator begin() { return iterator(this, 0); } - const_iterator begin() const { return const_iterator(this, 0); } - iterator end() { return iterator(this, size()); } - const_iterator end() const { return const_iterator(this, size()); } - reverse_iterator rbegin() { return reverse_iterator(end()); } - const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } - reverse_iterator rend() { return reverse_iterator(begin()); } - const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } - - // Versions for our special non-empty iterator - nonempty_iterator nonempty_begin() { - return nonempty_iterator(groups.begin(), groups.end(), groups.begin()); - } - const_nonempty_iterator nonempty_begin() const { - return const_nonempty_iterator(groups.begin(),groups.end(), groups.begin()); - } - nonempty_iterator nonempty_end() { - return nonempty_iterator(groups.begin(), groups.end(), groups.end()); - } - const_nonempty_iterator nonempty_end() const { - return const_nonempty_iterator(groups.begin(), groups.end(), groups.end()); - } - reverse_nonempty_iterator nonempty_rbegin() { - return reverse_nonempty_iterator(nonempty_end()); - } - const_reverse_nonempty_iterator nonempty_rbegin() const { - return const_reverse_nonempty_iterator(nonempty_end()); - } - reverse_nonempty_iterator nonempty_rend() { - return reverse_nonempty_iterator(nonempty_begin()); - } - const_reverse_nonempty_iterator nonempty_rend() const { - return const_reverse_nonempty_iterator(nonempty_begin()); - } - destructive_iterator destructive_begin() { - return destructive_iterator(groups.begin(), groups.end(), groups.begin()); - } - destructive_iterator destructive_end() { - return destructive_iterator(groups.begin(), groups.end(), groups.end()); - } - - typedef sparsegroup<value_type, GROUP_SIZE, allocator_type> group_type; - typedef std::vector<group_type, vector_alloc > group_vector_type; - - typedef typename group_vector_type::reference GroupsReference; - typedef typename group_vector_type::const_reference GroupsConstReference; - typedef typename group_vector_type::iterator GroupsIterator; - typedef typename group_vector_type::const_iterator GroupsConstIterator; - - // How to deal with the proper group - static size_type num_groups(size_type num) { // how many to hold num buckets - return num == 0 ? 0 : ((num-1) / GROUP_SIZE) + 1; - } - - u_int16_t pos_in_group(size_type i) const { - return static_cast<u_int16_t>(i % GROUP_SIZE); - } - size_type group_num(size_type i) const { - return i / GROUP_SIZE; - } - GroupsReference which_group(size_type i) { - return groups[group_num(i)]; - } - GroupsConstReference which_group(size_type i) const { - return groups[group_num(i)]; - } - - public: - // Constructors -- default, normal (when you specify size), and copy - explicit sparsetable(size_type sz = 0, Alloc alloc = Alloc()) - : groups(vector_alloc(alloc)), settings(alloc, sz) { - groups.resize(num_groups(sz), group_type(settings)); - } - // We can get away with using the default copy constructor, - // and default destructor, and hence the default operator=. Huzzah! - - // Many STL algorithms use swap instead of copy constructors - void swap(sparsetable& x) { - std::swap(groups, x.groups); // defined in stl_algobase.h - std::swap(settings.table_size, x.settings.table_size); - std::swap(settings.num_buckets, x.settings.num_buckets); - } - - // It's always nice to be able to clear a table without deallocating it - void clear() { - GroupsIterator group; - for ( group = groups.begin(); group != groups.end(); ++group ) { - group->clear(); - } - settings.num_buckets = 0; - } - - // ACCESSOR FUNCTIONS for the things we templatize on, basically - allocator_type get_allocator() const { - return allocator_type(settings); - } - - - // Functions that tell you about size. - // NOTE: empty() is non-intuitive! It does not tell you the number - // of not-empty buckets (use num_nonempty() for that). Instead - // it says whether you've allocated any buckets or not. - size_type size() const { return settings.table_size; } - size_type max_size() const { return settings.max_size(); } - bool empty() const { return settings.table_size == 0; } - // We also may want to know how many *used* buckets there are - size_type num_nonempty() const { return settings.num_buckets; } - - // OK, we'll let you resize one of these puppies - void resize(size_type new_size) { - groups.resize(num_groups(new_size), group_type(settings)); - if ( new_size < settings.table_size) { - // lower num_buckets, clear last group - if ( pos_in_group(new_size) > 0 ) // need to clear inside last group - groups.back().erase(groups.back().begin() + pos_in_group(new_size), - groups.back().end()); - settings.num_buckets = 0; // refigure # of used buckets - GroupsConstIterator group; - for ( group = groups.begin(); group != groups.end(); ++group ) - settings.num_buckets += group->num_nonempty(); - } - settings.table_size = new_size; - } - - - // We let you see if a bucket is non-empty without retrieving it - bool test(size_type i) const { - assert(i < settings.table_size); - return which_group(i).test(pos_in_group(i)); - } - bool test(iterator pos) const { - return which_group(pos.pos).test(pos_in_group(pos.pos)); - } - bool test(const_iterator pos) const { - return which_group(pos.pos).test(pos_in_group(pos.pos)); - } - - // We only return const_references because it's really hard to - // return something settable for empty buckets. Use set() instead. - const_reference get(size_type i) const { - assert(i < settings.table_size); - return which_group(i).get(pos_in_group(i)); - } - - // TODO(csilvers): make protected + friend - // This is used by sparse_hashtable to get an element from the table - // when we know it exists (because the caller has called test(i)). - const_reference unsafe_get(size_type i) const { - assert(i < settings.table_size); - assert(test(i)); - return which_group(i).unsafe_get(pos_in_group(i)); - } - - // TODO(csilvers): make protected + friend element_adaptor - reference mutating_get(size_type i) { // fills bucket i before getting - assert(i < settings.table_size); - typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); - reference retval = which_group(i).mutating_get(pos_in_group(i)); - settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; - return retval; - } - - // Syntactic sugar. As in sparsegroup, the non-const version is harder - const_reference operator[](size_type i) const { - return get(i); - } - - element_adaptor operator[](size_type i) { - return element_adaptor(this, i); - } - - // Needed for hashtables, gets as a nonempty_iterator. Crashes for empty bcks - const_nonempty_iterator get_iter(size_type i) const { - assert(test(i)); // how can a nonempty_iterator point to an empty bucket? - return const_nonempty_iterator( - groups.begin(), groups.end(), - groups.begin() + group_num(i), - (groups[group_num(i)].nonempty_begin() + - groups[group_num(i)].pos_to_offset(pos_in_group(i)))); - } - // For nonempty we can return a non-const version - nonempty_iterator get_iter(size_type i) { - assert(test(i)); // how can a nonempty_iterator point to an empty bucket? - return nonempty_iterator( - groups.begin(), groups.end(), - groups.begin() + group_num(i), - (groups[group_num(i)].nonempty_begin() + - groups[group_num(i)].pos_to_offset(pos_in_group(i)))); - } - - // And the reverse transformation. + group = settings.realloc_or_die(group, settings.num_buckets-1); + } + + // Shrink the array, without any special assumptions about value_type and + // allocator_type. + void erase_aux(size_type offset, base::false_type) { + // This is valid because 0 <= offset < num_buckets. Note the inequality. + pointer p = allocate_group(settings.num_buckets - 1); + std::uninitialized_copy(group, group + offset, p); + std::uninitialized_copy(group + offset + 1, group + settings.num_buckets, + p + offset); + free_group(); + group = p; + } + + public: + // This takes the specified elements out of the group. This is + // "undefining", rather than "clearing". + // TODO(austern): Make this exception safe: handle exceptions from + // value_type's copy constructor. + void erase(size_type i) { + if ( bmtest(i) ) { // trivial to erase empty bucket + size_type offset = pos_to_offset(bitmap,i); // where we'll find (or insert) + if ( settings.num_buckets == 1 ) { + free_group(); + group = NULL; + } else { + typedef base::integral_constant<bool, + (base::has_trivial_copy<value_type>::value && + base::has_trivial_destructor<value_type>::value && + base::is_same< + allocator_type, + libc_allocator_with_realloc<value_type> >::value)> + realloc_and_memmove_ok; // pretend mv(x,y) == "x.~T(); new(x) T(y)" + erase_aux(offset, realloc_and_memmove_ok()); + } + --settings.num_buckets; + bmclear(i); + } + } + + void erase(iterator pos) { + erase(pos.pos); + } + + void erase(iterator start_it, iterator end_it) { + // This could be more efficient, but to do so we'd need to make + // bmclear() clear a range of indices. Doesn't seem worth it. + for ( ; start_it != end_it; ++start_it ) + erase(start_it); + } + + + // I/O + // We support reading and writing groups to disk. We don't store + // the actual array contents (which we don't know how to store), + // just the bitmap and size. Meant to be used with table I/O. + + template <typename OUTPUT> bool write_metadata(OUTPUT *fp) const { + // we explicitly set to u_int16_t + assert(sizeof(settings.num_buckets) == 2); + if ( !sparsehash_internal::write_bigendian_number(fp, settings.num_buckets, + 2) ) + return false; + if ( !sparsehash_internal::write_data(fp, bitmap, sizeof(bitmap)) ) + return false; + return true; + } + + // Reading destroys the old group contents! Returns true if all was ok. + template <typename INPUT> bool read_metadata(INPUT *fp) { + clear(); + if ( !sparsehash_internal::read_bigendian_number(fp, &settings.num_buckets, + 2) ) + return false; + if ( !sparsehash_internal::read_data(fp, bitmap, sizeof(bitmap)) ) + return false; + // We'll allocate the space, but we won't fill it: it will be + // left as uninitialized raw memory. + group = allocate_group(settings.num_buckets); + return true; + } + + // Again, only meaningful if value_type is a POD. + template <typename INPUT> bool read_nopointer_data(INPUT *fp) { + for ( nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !sparsehash_internal::read_data(fp, &(*it), sizeof(*it)) ) + return false; + } + return true; + } + + // If your keys and values are simple enough, we can write them + // to disk for you. "simple enough" means POD and no pointers. + // However, we don't try to normalize endianness. + template <typename OUTPUT> bool write_nopointer_data(OUTPUT *fp) const { + for ( const_nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !sparsehash_internal::write_data(fp, &(*it), sizeof(*it)) ) + return false; + } + return true; + } + + + // Comparisons. We only need to define == and < -- we get + // != > <= >= via relops.h (which we happily included above). + // Note the comparisons are pretty arbitrary: we compare + // values of the first index that isn't equal (using default + // value for empty buckets). + bool operator==(const sparsegroup& x) const { + return ( settings.num_buckets == x.settings.num_buckets && + memcmp(bitmap, x.bitmap, sizeof(bitmap)) == 0 && + std::equal(begin(), end(), x.begin()) ); // from <algorithm> + } + + bool operator<(const sparsegroup& x) const { // also from <algorithm> + return std::lexicographical_compare(begin(), end(), x.begin(), x.end()); + } + bool operator!=(const sparsegroup& x) const { return !(*this == x); } + bool operator<=(const sparsegroup& x) const { return !(x < *this); } + bool operator>(const sparsegroup& x) const { return x < *this; } + bool operator>=(const sparsegroup& x) const { return !(*this < x); } + + private: + template <class A> + class alloc_impl : public A { + public: + typedef typename A::pointer pointer; + typedef typename A::size_type size_type; + + // Convert a normal allocator to one that has realloc_or_die() + alloc_impl(const A& a) : A(a) { } + + // realloc_or_die should only be used when using the default + // allocator (libc_allocator_with_realloc). + pointer realloc_or_die(pointer /*ptr*/, size_type /*n*/) { + fprintf(stderr, "realloc_or_die is only supported for " + "libc_allocator_with_realloc\n"); + exit(1); + return NULL; + } + }; + + // A template specialization of alloc_impl for + // libc_allocator_with_realloc that can handle realloc_or_die. + template <class A> + class alloc_impl<libc_allocator_with_realloc<A> > + : public libc_allocator_with_realloc<A> { + public: + typedef typename libc_allocator_with_realloc<A>::pointer pointer; + typedef typename libc_allocator_with_realloc<A>::size_type size_type; + + alloc_impl(const libc_allocator_with_realloc<A>& a) + : libc_allocator_with_realloc<A>(a) { } + + pointer realloc_or_die(pointer ptr, size_type n) { + pointer retval = this->reallocate(ptr, n); + if (retval == NULL) { + fprintf(stderr, "sparsehash: FATAL ERROR: failed to reallocate " + "%lu elements for ptr %p", static_cast<unsigned long>(n), ptr); + exit(1); + } + return retval; + } + }; + + // Package allocator with num_buckets to eliminate memory needed for the + // zero-size allocator. + // If new fields are added to this class, we should add them to + // operator= and swap. + class Settings : public alloc_impl<value_alloc_type> { + public: + Settings(const alloc_impl<value_alloc_type>& a, u_int16_t n = 0) + : alloc_impl<value_alloc_type>(a), num_buckets(n) { } + Settings(const Settings& s) + : alloc_impl<value_alloc_type>(s), num_buckets(s.num_buckets) { } + + u_int16_t num_buckets; // limits GROUP_SIZE to 64K + }; + + // The actual data + pointer group; // (small) array of T's + Settings settings; // allocator and num_buckets + unsigned char bitmap[(GROUP_SIZE-1)/8 + 1]; // fancy math is so we round up +}; + +// We need a global swap as well +template <class T, u_int16_t GROUP_SIZE, class Alloc> +inline void swap(sparsegroup<T,GROUP_SIZE,Alloc> &x, + sparsegroup<T,GROUP_SIZE,Alloc> &y) { + x.swap(y); +} + +// --------------------------------------------------------------------------- + + +template <class T, u_int16_t GROUP_SIZE = DEFAULT_SPARSEGROUP_SIZE, + class Alloc = libc_allocator_with_realloc<T> > +class sparsetable { + private: + typedef typename Alloc::template rebind<T>::other value_alloc_type; + typedef typename Alloc::template rebind< + sparsegroup<T, GROUP_SIZE, value_alloc_type> >::other vector_alloc; + + public: + // Basic types + typedef T value_type; // stolen from stl_vector.h + typedef Alloc allocator_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + typedef table_iterator<sparsetable<T, GROUP_SIZE, Alloc> > iterator; + typedef const_table_iterator<sparsetable<T, GROUP_SIZE, Alloc> > + const_iterator; + typedef table_element_adaptor<sparsetable<T, GROUP_SIZE, Alloc> > + element_adaptor; + typedef std::reverse_iterator<const_iterator> const_reverse_iterator; + typedef std::reverse_iterator<iterator> reverse_iterator; // from iterator.h + + // These are our special iterators, that go over non-empty buckets in a + // table. These aren't const only because you can change non-empty bcks. + typedef two_d_iterator< std::vector< sparsegroup<value_type, GROUP_SIZE, + value_alloc_type>, + vector_alloc> > + nonempty_iterator; + typedef const_two_d_iterator< std::vector< sparsegroup<value_type, + GROUP_SIZE, + value_alloc_type>, + vector_alloc> > + const_nonempty_iterator; + typedef std::reverse_iterator<nonempty_iterator> reverse_nonempty_iterator; + typedef std::reverse_iterator<const_nonempty_iterator> const_reverse_nonempty_iterator; + // Another special iterator: it frees memory as it iterates (used to resize) + typedef destructive_two_d_iterator< std::vector< sparsegroup<value_type, + GROUP_SIZE, + value_alloc_type>, + vector_alloc> > + destructive_iterator; + + // Iterator functions + iterator begin() { return iterator(this, 0); } + const_iterator begin() const { return const_iterator(this, 0); } + iterator end() { return iterator(this, size()); } + const_iterator end() const { return const_iterator(this, size()); } + reverse_iterator rbegin() { return reverse_iterator(end()); } + const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } + reverse_iterator rend() { return reverse_iterator(begin()); } + const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } + + // Versions for our special non-empty iterator + nonempty_iterator nonempty_begin() { + return nonempty_iterator(groups.begin(), groups.end(), groups.begin()); + } + const_nonempty_iterator nonempty_begin() const { + return const_nonempty_iterator(groups.begin(),groups.end(), groups.begin()); + } + nonempty_iterator nonempty_end() { + return nonempty_iterator(groups.begin(), groups.end(), groups.end()); + } + const_nonempty_iterator nonempty_end() const { + return const_nonempty_iterator(groups.begin(), groups.end(), groups.end()); + } + reverse_nonempty_iterator nonempty_rbegin() { + return reverse_nonempty_iterator(nonempty_end()); + } + const_reverse_nonempty_iterator nonempty_rbegin() const { + return const_reverse_nonempty_iterator(nonempty_end()); + } + reverse_nonempty_iterator nonempty_rend() { + return reverse_nonempty_iterator(nonempty_begin()); + } + const_reverse_nonempty_iterator nonempty_rend() const { + return const_reverse_nonempty_iterator(nonempty_begin()); + } + destructive_iterator destructive_begin() { + return destructive_iterator(groups.begin(), groups.end(), groups.begin()); + } + destructive_iterator destructive_end() { + return destructive_iterator(groups.begin(), groups.end(), groups.end()); + } + + typedef sparsegroup<value_type, GROUP_SIZE, allocator_type> group_type; + typedef std::vector<group_type, vector_alloc > group_vector_type; + + typedef typename group_vector_type::reference GroupsReference; + typedef typename group_vector_type::const_reference GroupsConstReference; + typedef typename group_vector_type::iterator GroupsIterator; + typedef typename group_vector_type::const_iterator GroupsConstIterator; + + // How to deal with the proper group + static size_type num_groups(size_type num) { // how many to hold num buckets + return num == 0 ? 0 : ((num-1) / GROUP_SIZE) + 1; + } + + u_int16_t pos_in_group(size_type i) const { + return static_cast<u_int16_t>(i % GROUP_SIZE); + } + size_type group_num(size_type i) const { + return i / GROUP_SIZE; + } + GroupsReference which_group(size_type i) { + return groups[group_num(i)]; + } + GroupsConstReference which_group(size_type i) const { + return groups[group_num(i)]; + } + + public: + // Constructors -- default, normal (when you specify size), and copy + explicit sparsetable(size_type sz = 0, Alloc alloc = Alloc()) + : groups(vector_alloc(alloc)), settings(alloc, sz) { + groups.resize(num_groups(sz), group_type(settings)); + } + // We can get away with using the default copy constructor, + // and default destructor, and hence the default operator=. Huzzah! + + // Many STL algorithms use swap instead of copy constructors + void swap(sparsetable& x) { + std::swap(groups, x.groups); // defined in stl_algobase.h + std::swap(settings.table_size, x.settings.table_size); + std::swap(settings.num_buckets, x.settings.num_buckets); + } + + // It's always nice to be able to clear a table without deallocating it + void clear() { + GroupsIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) { + group->clear(); + } + settings.num_buckets = 0; + } + + // ACCESSOR FUNCTIONS for the things we templatize on, basically + allocator_type get_allocator() const { + return allocator_type(settings); + } + + + // Functions that tell you about size. + // NOTE: empty() is non-intuitive! It does not tell you the number + // of not-empty buckets (use num_nonempty() for that). Instead + // it says whether you've allocated any buckets or not. + size_type size() const { return settings.table_size; } + size_type max_size() const { return settings.max_size(); } + bool empty() const { return settings.table_size == 0; } + // We also may want to know how many *used* buckets there are + size_type num_nonempty() const { return settings.num_buckets; } + + // OK, we'll let you resize one of these puppies + void resize(size_type new_size) { + groups.resize(num_groups(new_size), group_type(settings)); + if ( new_size < settings.table_size) { + // lower num_buckets, clear last group + if ( pos_in_group(new_size) > 0 ) // need to clear inside last group + groups.back().erase(groups.back().begin() + pos_in_group(new_size), + groups.back().end()); + settings.num_buckets = 0; // refigure # of used buckets + GroupsConstIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) + settings.num_buckets += group->num_nonempty(); + } + settings.table_size = new_size; + } + + + // We let you see if a bucket is non-empty without retrieving it + bool test(size_type i) const { + assert(i < settings.table_size); + return which_group(i).test(pos_in_group(i)); + } + bool test(iterator pos) const { + return which_group(pos.pos).test(pos_in_group(pos.pos)); + } + bool test(const_iterator pos) const { + return which_group(pos.pos).test(pos_in_group(pos.pos)); + } + + // We only return const_references because it's really hard to + // return something settable for empty buckets. Use set() instead. + const_reference get(size_type i) const { + assert(i < settings.table_size); + return which_group(i).get(pos_in_group(i)); + } + + // TODO(csilvers): make protected + friend + // This is used by sparse_hashtable to get an element from the table + // when we know it exists (because the caller has called test(i)). + const_reference unsafe_get(size_type i) const { + assert(i < settings.table_size); + assert(test(i)); + return which_group(i).unsafe_get(pos_in_group(i)); + } + + // TODO(csilvers): make protected + friend element_adaptor + reference mutating_get(size_type i) { // fills bucket i before getting + assert(i < settings.table_size); + typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); + reference retval = which_group(i).mutating_get(pos_in_group(i)); + settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; + return retval; + } + + // Syntactic sugar. As in sparsegroup, the non-const version is harder + const_reference operator[](size_type i) const { + return get(i); + } + + element_adaptor operator[](size_type i) { + return element_adaptor(this, i); + } + + // Needed for hashtables, gets as a nonempty_iterator. Crashes for empty bcks + const_nonempty_iterator get_iter(size_type i) const { + assert(test(i)); // how can a nonempty_iterator point to an empty bucket? + return const_nonempty_iterator( + groups.begin(), groups.end(), + groups.begin() + group_num(i), + (groups[group_num(i)].nonempty_begin() + + groups[group_num(i)].pos_to_offset(pos_in_group(i)))); + } + // For nonempty we can return a non-const version + nonempty_iterator get_iter(size_type i) { + assert(test(i)); // how can a nonempty_iterator point to an empty bucket? + return nonempty_iterator( + groups.begin(), groups.end(), + groups.begin() + group_num(i), + (groups[group_num(i)].nonempty_begin() + + groups[group_num(i)].pos_to_offset(pos_in_group(i)))); + } + + // And the reverse transformation. size_type get_pos(const const_nonempty_iterator& it) const { - difference_type current_row = it.row_current - it.row_begin; - difference_type current_col = (it.col_current - - groups[current_row].nonempty_begin()); - return ((current_row * GROUP_SIZE) + - groups[current_row].offset_to_pos(current_col)); - } - - - // This returns a reference to the inserted item (which is a copy of val) - // The trick is to figure out whether we're replacing or inserting anew - reference set(size_type i, const_reference val) { - assert(i < settings.table_size); - typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); - reference retval = which_group(i).set(pos_in_group(i), val); - settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; - return retval; - } - - // This takes the specified elements out of the table. This is - // "undefining", rather than "clearing". - void erase(size_type i) { - assert(i < settings.table_size); - typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); - which_group(i).erase(pos_in_group(i)); - settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; - } - - void erase(iterator pos) { - erase(pos.pos); - } - - void erase(iterator start_it, iterator end_it) { - // This could be more efficient, but then we'd need to figure - // out if we spanned groups or not. Doesn't seem worth it. - for ( ; start_it != end_it; ++start_it ) - erase(start_it); - } - - - // We support reading and writing tables to disk. We don't store - // the actual array contents (which we don't know how to store), - // just the groups and sizes. Returns true if all went ok. - - private: - // Every time the disk format changes, this should probably change too - typedef unsigned long MagicNumberType; - static const MagicNumberType MAGIC_NUMBER = 0x24687531; - - // Old versions of this code write all data in 32 bits. We need to - // support these files as well as having support for 64-bit systems. - // So we use the following encoding scheme: for values < 2^32-1, we - // store in 4 bytes in big-endian order. For values > 2^32, we - // store 0xFFFFFFF followed by 8 bytes in big-endian order. This - // causes us to mis-read old-version code that stores exactly - // 0xFFFFFFF, but I don't think that is likely to have happened for - // these particular values. - template <typename OUTPUT, typename IntType> - static bool write_32_or_64(OUTPUT* fp, IntType value) { - if ( value < 0xFFFFFFFFULL ) { // fits in 4 bytes - if ( !sparsehash_internal::write_bigendian_number(fp, value, 4) ) - return false; - } else { - if ( !sparsehash_internal::write_bigendian_number(fp, 0xFFFFFFFFUL, 4) ) - return false; - if ( !sparsehash_internal::write_bigendian_number(fp, value, 8) ) - return false; - } - return true; - } - - template <typename INPUT, typename IntType> - static bool read_32_or_64(INPUT* fp, IntType *value) { // reads into value - MagicNumberType first4 = 0; // a convenient 32-bit unsigned type - if ( !sparsehash_internal::read_bigendian_number(fp, &first4, 4) ) - return false; - if ( first4 < 0xFFFFFFFFULL ) { - *value = first4; - } else { - if ( !sparsehash_internal::read_bigendian_number(fp, value, 8) ) - return false; - } - return true; - } - - public: - // read/write_metadata() and read_write/nopointer_data() are DEPRECATED. - // Use serialize() and unserialize(), below, for new code. - - template <typename OUTPUT> bool write_metadata(OUTPUT *fp) const { - if ( !write_32_or_64(fp, MAGIC_NUMBER) ) return false; - if ( !write_32_or_64(fp, settings.table_size) ) return false; - if ( !write_32_or_64(fp, settings.num_buckets) ) return false; - - GroupsConstIterator group; - for ( group = groups.begin(); group != groups.end(); ++group ) - if ( group->write_metadata(fp) == false ) return false; - return true; - } - - // Reading destroys the old table contents! Returns true if read ok. - template <typename INPUT> bool read_metadata(INPUT *fp) { - size_type magic_read = 0; - if ( !read_32_or_64(fp, &magic_read) ) return false; - if ( magic_read != MAGIC_NUMBER ) { - clear(); // just to be consistent - return false; - } - - if ( !read_32_or_64(fp, &settings.table_size) ) return false; - if ( !read_32_or_64(fp, &settings.num_buckets) ) return false; - - resize(settings.table_size); // so the vector's sized ok - GroupsIterator group; - for ( group = groups.begin(); group != groups.end(); ++group ) - if ( group->read_metadata(fp) == false ) return false; - return true; - } - - // This code is identical to that for SparseGroup - // If your keys and values are simple enough, we can write them - // to disk for you. "simple enough" means no pointers. - // However, we don't try to normalize endianness - bool write_nopointer_data(FILE *fp) const { - for ( const_nonempty_iterator it = nonempty_begin(); - it != nonempty_end(); ++it ) { - if ( !fwrite(&*it, sizeof(*it), 1, fp) ) return false; - } - return true; - } - - // When reading, we have to override the potential const-ness of *it - bool read_nopointer_data(FILE *fp) { - for ( nonempty_iterator it = nonempty_begin(); - it != nonempty_end(); ++it ) { - if ( !fread(reinterpret_cast<void*>(&(*it)), sizeof(*it), 1, fp) ) - return false; - } - return true; - } - - // INPUT and OUTPUT must be either a FILE, *or* a C++ stream - // (istream, ostream, etc) *or* a class providing - // Read(void*, size_t) and Write(const void*, size_t) - // (respectively), which writes a buffer into a stream - // (which the INPUT/OUTPUT instance presumably owns). - - typedef sparsehash_internal::pod_serializer<value_type> NopointerSerializer; - - // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) - template <typename ValueSerializer, typename OUTPUT> - bool serialize(ValueSerializer serializer, OUTPUT *fp) { - if ( !write_metadata(fp) ) - return false; - for ( const_nonempty_iterator it = nonempty_begin(); - it != nonempty_end(); ++it ) { - if ( !serializer(fp, *it) ) return false; - } - return true; - } - - // ValueSerializer: a functor. operator()(INPUT*, value_type*) - template <typename ValueSerializer, typename INPUT> - bool unserialize(ValueSerializer serializer, INPUT *fp) { - clear(); - if ( !read_metadata(fp) ) - return false; - for ( nonempty_iterator it = nonempty_begin(); - it != nonempty_end(); ++it ) { - if ( !serializer(fp, &*it) ) return false; - } - return true; - } - - // Comparisons. Note the comparisons are pretty arbitrary: we - // compare values of the first index that isn't equal (using default - // value for empty buckets). - bool operator==(const sparsetable& x) const { - return ( settings.table_size == x.settings.table_size && - settings.num_buckets == x.settings.num_buckets && - groups == x.groups ); - } - - bool operator<(const sparsetable& x) const { - return std::lexicographical_compare(begin(), end(), x.begin(), x.end()); - } - bool operator!=(const sparsetable& x) const { return !(*this == x); } - bool operator<=(const sparsetable& x) const { return !(x < *this); } - bool operator>(const sparsetable& x) const { return x < *this; } - bool operator>=(const sparsetable& x) const { return !(*this < x); } - - - private: - // Package allocator with table_size and num_buckets to eliminate memory - // needed for the zero-size allocator. - // If new fields are added to this class, we should add them to - // operator= and swap. - class Settings : public allocator_type { - public: - typedef typename allocator_type::size_type size_type; - - Settings(const allocator_type& a, size_type sz = 0, size_type n = 0) - : allocator_type(a), table_size(sz), num_buckets(n) { } - - Settings(const Settings& s) - : allocator_type(s), - table_size(s.table_size), num_buckets(s.num_buckets) { } - - size_type table_size; // how many buckets they want - size_type num_buckets; // number of non-empty buckets - }; - - // The actual data - group_vector_type groups; // our list of groups - Settings settings; // allocator, table size, buckets -}; - -// We need a global swap as well -template <class T, u_int16_t GROUP_SIZE, class Alloc> -inline void swap(sparsetable<T,GROUP_SIZE,Alloc> &x, - sparsetable<T,GROUP_SIZE,Alloc> &y) { - x.swap(y); -} - -_END_GOOGLE_NAMESPACE_ - -#endif // UTIL_GTL_SPARSETABLE_H_ + difference_type current_row = it.row_current - it.row_begin; + difference_type current_col = (it.col_current - + groups[current_row].nonempty_begin()); + return ((current_row * GROUP_SIZE) + + groups[current_row].offset_to_pos(current_col)); + } + + + // This returns a reference to the inserted item (which is a copy of val) + // The trick is to figure out whether we're replacing or inserting anew + reference set(size_type i, const_reference val) { + assert(i < settings.table_size); + typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); + reference retval = which_group(i).set(pos_in_group(i), val); + settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; + return retval; + } + + // This takes the specified elements out of the table. This is + // "undefining", rather than "clearing". + void erase(size_type i) { + assert(i < settings.table_size); + typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); + which_group(i).erase(pos_in_group(i)); + settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; + } + + void erase(iterator pos) { + erase(pos.pos); + } + + void erase(iterator start_it, iterator end_it) { + // This could be more efficient, but then we'd need to figure + // out if we spanned groups or not. Doesn't seem worth it. + for ( ; start_it != end_it; ++start_it ) + erase(start_it); + } + + + // We support reading and writing tables to disk. We don't store + // the actual array contents (which we don't know how to store), + // just the groups and sizes. Returns true if all went ok. + + private: + // Every time the disk format changes, this should probably change too + typedef unsigned long MagicNumberType; + static const MagicNumberType MAGIC_NUMBER = 0x24687531; + + // Old versions of this code write all data in 32 bits. We need to + // support these files as well as having support for 64-bit systems. + // So we use the following encoding scheme: for values < 2^32-1, we + // store in 4 bytes in big-endian order. For values > 2^32, we + // store 0xFFFFFFF followed by 8 bytes in big-endian order. This + // causes us to mis-read old-version code that stores exactly + // 0xFFFFFFF, but I don't think that is likely to have happened for + // these particular values. + template <typename OUTPUT, typename IntType> + static bool write_32_or_64(OUTPUT* fp, IntType value) { + if ( value < 0xFFFFFFFFULL ) { // fits in 4 bytes + if ( !sparsehash_internal::write_bigendian_number(fp, value, 4) ) + return false; + } else { + if ( !sparsehash_internal::write_bigendian_number(fp, 0xFFFFFFFFUL, 4) ) + return false; + if ( !sparsehash_internal::write_bigendian_number(fp, value, 8) ) + return false; + } + return true; + } + + template <typename INPUT, typename IntType> + static bool read_32_or_64(INPUT* fp, IntType *value) { // reads into value + MagicNumberType first4 = 0; // a convenient 32-bit unsigned type + if ( !sparsehash_internal::read_bigendian_number(fp, &first4, 4) ) + return false; + if ( first4 < 0xFFFFFFFFULL ) { + *value = first4; + } else { + if ( !sparsehash_internal::read_bigendian_number(fp, value, 8) ) + return false; + } + return true; + } + + public: + // read/write_metadata() and read_write/nopointer_data() are DEPRECATED. + // Use serialize() and unserialize(), below, for new code. + + template <typename OUTPUT> bool write_metadata(OUTPUT *fp) const { + if ( !write_32_or_64(fp, MAGIC_NUMBER) ) return false; + if ( !write_32_or_64(fp, settings.table_size) ) return false; + if ( !write_32_or_64(fp, settings.num_buckets) ) return false; + + GroupsConstIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) + if ( group->write_metadata(fp) == false ) return false; + return true; + } + + // Reading destroys the old table contents! Returns true if read ok. + template <typename INPUT> bool read_metadata(INPUT *fp) { + size_type magic_read = 0; + if ( !read_32_or_64(fp, &magic_read) ) return false; + if ( magic_read != MAGIC_NUMBER ) { + clear(); // just to be consistent + return false; + } + + if ( !read_32_or_64(fp, &settings.table_size) ) return false; + if ( !read_32_or_64(fp, &settings.num_buckets) ) return false; + + resize(settings.table_size); // so the vector's sized ok + GroupsIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) + if ( group->read_metadata(fp) == false ) return false; + return true; + } + + // This code is identical to that for SparseGroup + // If your keys and values are simple enough, we can write them + // to disk for you. "simple enough" means no pointers. + // However, we don't try to normalize endianness + bool write_nopointer_data(FILE *fp) const { + for ( const_nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !fwrite(&*it, sizeof(*it), 1, fp) ) return false; + } + return true; + } + + // When reading, we have to override the potential const-ness of *it + bool read_nopointer_data(FILE *fp) { + for ( nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !fread(reinterpret_cast<void*>(&(*it)), sizeof(*it), 1, fp) ) + return false; + } + return true; + } + + // INPUT and OUTPUT must be either a FILE, *or* a C++ stream + // (istream, ostream, etc) *or* a class providing + // Read(void*, size_t) and Write(const void*, size_t) + // (respectively), which writes a buffer into a stream + // (which the INPUT/OUTPUT instance presumably owns). + + typedef sparsehash_internal::pod_serializer<value_type> NopointerSerializer; + + // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) + template <typename ValueSerializer, typename OUTPUT> + bool serialize(ValueSerializer serializer, OUTPUT *fp) { + if ( !write_metadata(fp) ) + return false; + for ( const_nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !serializer(fp, *it) ) return false; + } + return true; + } + + // ValueSerializer: a functor. operator()(INPUT*, value_type*) + template <typename ValueSerializer, typename INPUT> + bool unserialize(ValueSerializer serializer, INPUT *fp) { + clear(); + if ( !read_metadata(fp) ) + return false; + for ( nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !serializer(fp, &*it) ) return false; + } + return true; + } + + // Comparisons. Note the comparisons are pretty arbitrary: we + // compare values of the first index that isn't equal (using default + // value for empty buckets). + bool operator==(const sparsetable& x) const { + return ( settings.table_size == x.settings.table_size && + settings.num_buckets == x.settings.num_buckets && + groups == x.groups ); + } + + bool operator<(const sparsetable& x) const { + return std::lexicographical_compare(begin(), end(), x.begin(), x.end()); + } + bool operator!=(const sparsetable& x) const { return !(*this == x); } + bool operator<=(const sparsetable& x) const { return !(x < *this); } + bool operator>(const sparsetable& x) const { return x < *this; } + bool operator>=(const sparsetable& x) const { return !(*this < x); } + + + private: + // Package allocator with table_size and num_buckets to eliminate memory + // needed for the zero-size allocator. + // If new fields are added to this class, we should add them to + // operator= and swap. + class Settings : public allocator_type { + public: + typedef typename allocator_type::size_type size_type; + + Settings(const allocator_type& a, size_type sz = 0, size_type n = 0) + : allocator_type(a), table_size(sz), num_buckets(n) { } + + Settings(const Settings& s) + : allocator_type(s), + table_size(s.table_size), num_buckets(s.num_buckets) { } + + size_type table_size; // how many buckets they want + size_type num_buckets; // number of non-empty buckets + }; + + // The actual data + group_vector_type groups; // our list of groups + Settings settings; // allocator, table size, buckets +}; + +// We need a global swap as well +template <class T, u_int16_t GROUP_SIZE, class Alloc> +inline void swap(sparsetable<T,GROUP_SIZE,Alloc> &x, + sparsetable<T,GROUP_SIZE,Alloc> &y) { + x.swap(y); +} + +_END_GOOGLE_NAMESPACE_ + +#endif // UTIL_GTL_SPARSETABLE_H_ diff --git a/contrib/libs/sparsehash/src/sparsehash/template_util.h b/contrib/libs/sparsehash/src/sparsehash/template_util.h index 182cef24a1..6fec3d0924 100644 --- a/contrib/libs/sparsehash/src/sparsehash/template_util.h +++ b/contrib/libs/sparsehash/src/sparsehash/template_util.h @@ -1,134 +1,134 @@ -// Copyright 2005 Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// ---- -// -// Template metaprogramming utility functions. -// -// This code is compiled directly on many platforms, including client -// platforms like Windows, Mac, and embedded systems. Before making -// any changes here, make sure that you're not breaking any platforms. -// -// -// The names choosen here reflect those used in tr1 and the boost::mpl -// library, there are similar operations used in the Loki library as -// well. I prefer the boost names for 2 reasons: -// 1. I think that portions of the Boost libraries are more likely to -// be included in the c++ standard. -// 2. It is not impossible that some of the boost libraries will be -// included in our own build in the future. -// Both of these outcomes means that we may be able to directly replace -// some of these with boost equivalents. -// -#ifndef BASE_TEMPLATE_UTIL_H_ -#define BASE_TEMPLATE_UTIL_H_ - -#include <sparsehash/internal/sparseconfig.h> -_START_GOOGLE_NAMESPACE_ - -// Types small_ and big_ are guaranteed such that sizeof(small_) < -// sizeof(big_) -typedef char small_; - -struct big_ { - char dummy[2]; -}; - -// Identity metafunction. -template <class T> -struct identity_ { - typedef T type; -}; - -// integral_constant, defined in tr1, is a wrapper for an integer -// value. We don't really need this generality; we could get away -// with hardcoding the integer type to bool. We use the fully -// general integer_constant for compatibility with tr1. - -template<class T, T v> -struct integral_constant { - static const T value = v; - typedef T value_type; - typedef integral_constant<T, v> type; -}; - -template <class T, T v> const T integral_constant<T, v>::value; - - -// Abbreviations: true_type and false_type are structs that represent boolean -// true and false values. Also define the boost::mpl versions of those names, -// true_ and false_. -typedef integral_constant<bool, true> true_type; -typedef integral_constant<bool, false> false_type; -typedef true_type true_; -typedef false_type false_; - -// if_ is a templatized conditional statement. -// if_<cond, A, B> is a compile time evaluation of cond. -// if_<>::type contains A if cond is true, B otherwise. -template<bool cond, typename A, typename B> -struct if_{ - typedef A type; -}; - -template<typename A, typename B> -struct if_<false, A, B> { - typedef B type; -}; - - -// type_equals_ is a template type comparator, similar to Loki IsSameType. -// type_equals_<A, B>::value is true iff "A" is the same type as "B". -// -// New code should prefer base::is_same, defined in base/type_traits.h. -// It is functionally identical, but is_same is the standard spelling. -template<typename A, typename B> -struct type_equals_ : public false_ { -}; - -template<typename A> -struct type_equals_<A, A> : public true_ { -}; - -// and_ is a template && operator. -// and_<A, B>::value evaluates "A::value && B::value". -template<typename A, typename B> -struct and_ : public integral_constant<bool, (A::value && B::value)> { -}; - -// or_ is a template || operator. -// or_<A, B>::value evaluates "A::value || B::value". -template<typename A, typename B> -struct or_ : public integral_constant<bool, (A::value || B::value)> { -}; - - -_END_GOOGLE_NAMESPACE_ - -#endif // BASE_TEMPLATE_UTIL_H_ +// Copyright 2005 Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// ---- +// +// Template metaprogramming utility functions. +// +// This code is compiled directly on many platforms, including client +// platforms like Windows, Mac, and embedded systems. Before making +// any changes here, make sure that you're not breaking any platforms. +// +// +// The names choosen here reflect those used in tr1 and the boost::mpl +// library, there are similar operations used in the Loki library as +// well. I prefer the boost names for 2 reasons: +// 1. I think that portions of the Boost libraries are more likely to +// be included in the c++ standard. +// 2. It is not impossible that some of the boost libraries will be +// included in our own build in the future. +// Both of these outcomes means that we may be able to directly replace +// some of these with boost equivalents. +// +#ifndef BASE_TEMPLATE_UTIL_H_ +#define BASE_TEMPLATE_UTIL_H_ + +#include <sparsehash/internal/sparseconfig.h> +_START_GOOGLE_NAMESPACE_ + +// Types small_ and big_ are guaranteed such that sizeof(small_) < +// sizeof(big_) +typedef char small_; + +struct big_ { + char dummy[2]; +}; + +// Identity metafunction. +template <class T> +struct identity_ { + typedef T type; +}; + +// integral_constant, defined in tr1, is a wrapper for an integer +// value. We don't really need this generality; we could get away +// with hardcoding the integer type to bool. We use the fully +// general integer_constant for compatibility with tr1. + +template<class T, T v> +struct integral_constant { + static const T value = v; + typedef T value_type; + typedef integral_constant<T, v> type; +}; + +template <class T, T v> const T integral_constant<T, v>::value; + + +// Abbreviations: true_type and false_type are structs that represent boolean +// true and false values. Also define the boost::mpl versions of those names, +// true_ and false_. +typedef integral_constant<bool, true> true_type; +typedef integral_constant<bool, false> false_type; +typedef true_type true_; +typedef false_type false_; + +// if_ is a templatized conditional statement. +// if_<cond, A, B> is a compile time evaluation of cond. +// if_<>::type contains A if cond is true, B otherwise. +template<bool cond, typename A, typename B> +struct if_{ + typedef A type; +}; + +template<typename A, typename B> +struct if_<false, A, B> { + typedef B type; +}; + + +// type_equals_ is a template type comparator, similar to Loki IsSameType. +// type_equals_<A, B>::value is true iff "A" is the same type as "B". +// +// New code should prefer base::is_same, defined in base/type_traits.h. +// It is functionally identical, but is_same is the standard spelling. +template<typename A, typename B> +struct type_equals_ : public false_ { +}; + +template<typename A> +struct type_equals_<A, A> : public true_ { +}; + +// and_ is a template && operator. +// and_<A, B>::value evaluates "A::value && B::value". +template<typename A, typename B> +struct and_ : public integral_constant<bool, (A::value && B::value)> { +}; + +// or_ is a template || operator. +// or_<A, B>::value evaluates "A::value || B::value". +template<typename A, typename B> +struct or_ : public integral_constant<bool, (A::value || B::value)> { +}; + + +_END_GOOGLE_NAMESPACE_ + +#endif // BASE_TEMPLATE_UTIL_H_ diff --git a/contrib/libs/sparsehash/src/sparsehash/type_traits.h b/contrib/libs/sparsehash/src/sparsehash/type_traits.h index 478d9068a7..f909cf9a37 100644 --- a/contrib/libs/sparsehash/src/sparsehash/type_traits.h +++ b/contrib/libs/sparsehash/src/sparsehash/type_traits.h @@ -1,342 +1,342 @@ -// Copyright (c) 2006, Google Inc. -// All rights reserved. -// -// Redistribution and use in source and binary forms, with or without -// modification, are permitted provided that the following conditions are -// met: -// -// * Redistributions of source code must retain the above copyright -// notice, this list of conditions and the following disclaimer. -// * Redistributions in binary form must reproduce the above -// copyright notice, this list of conditions and the following disclaimer -// in the documentation and/or other materials provided with the -// distribution. -// * Neither the name of Google Inc. nor the names of its -// contributors may be used to endorse or promote products derived from -// this software without specific prior written permission. -// -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -// ---- -// -// This code is compiled directly on many platforms, including client -// platforms like Windows, Mac, and embedded systems. Before making -// any changes here, make sure that you're not breaking any platforms. -// -// Define a small subset of tr1 type traits. The traits we define are: -// is_integral -// is_floating_point -// is_pointer -// is_enum -// is_reference -// is_pod -// has_trivial_constructor -// has_trivial_copy -// has_trivial_assign -// has_trivial_destructor -// remove_const -// remove_volatile -// remove_cv -// remove_reference -// add_reference -// remove_pointer -// is_same -// is_convertible -// We can add more type traits as required. - -#ifndef BASE_TYPE_TRAITS_H_ -#define BASE_TYPE_TRAITS_H_ - -#include <sparsehash/internal/sparseconfig.h> -#include <utility> // For pair - -#include <sparsehash/template_util.h> // For true_type and false_type - -_START_GOOGLE_NAMESPACE_ - -template <class T> struct is_integral; -template <class T> struct is_floating_point; -template <class T> struct is_pointer; -// MSVC can't compile this correctly, and neither can gcc 3.3.5 (at least) -#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) -// is_enum uses is_convertible, which is not available on MSVC. -template <class T> struct is_enum; -#endif -template <class T> struct is_reference; -template <class T> struct is_pod; -template <class T> struct has_trivial_constructor; -template <class T> struct has_trivial_copy; -template <class T> struct has_trivial_assign; -template <class T> struct has_trivial_destructor; -template <class T> struct remove_const; -template <class T> struct remove_volatile; -template <class T> struct remove_cv; -template <class T> struct remove_reference; -template <class T> struct add_reference; -template <class T> struct remove_pointer; -template <class T, class U> struct is_same; -#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) -template <class From, class To> struct is_convertible; -#endif - -// is_integral is false except for the built-in integer types. A -// cv-qualified type is integral if and only if the underlying type is. -template <class T> struct is_integral : false_type { }; -template<> struct is_integral<bool> : true_type { }; -template<> struct is_integral<char> : true_type { }; -template<> struct is_integral<unsigned char> : true_type { }; -template<> struct is_integral<signed char> : true_type { }; -#if defined(_MSC_VER) -// wchar_t is not by default a distinct type from unsigned short in -// Microsoft C. -// See http://msdn2.microsoft.com/en-us/library/dh8che7s(VS.80).aspx -template<> struct is_integral<__wchar_t> : true_type { }; -#else -template<> struct is_integral<wchar_t> : true_type { }; -#endif -template<> struct is_integral<short> : true_type { }; -template<> struct is_integral<unsigned short> : true_type { }; -template<> struct is_integral<int> : true_type { }; -template<> struct is_integral<unsigned int> : true_type { }; -template<> struct is_integral<long> : true_type { }; -template<> struct is_integral<unsigned long> : true_type { }; -#ifdef HAVE_LONG_LONG -template<> struct is_integral<long long> : true_type { }; -template<> struct is_integral<unsigned long long> : true_type { }; -#endif -template <class T> struct is_integral<const T> : is_integral<T> { }; -template <class T> struct is_integral<volatile T> : is_integral<T> { }; -template <class T> struct is_integral<const volatile T> : is_integral<T> { }; - -// is_floating_point is false except for the built-in floating-point types. -// A cv-qualified type is integral if and only if the underlying type is. -template <class T> struct is_floating_point : false_type { }; -template<> struct is_floating_point<float> : true_type { }; -template<> struct is_floating_point<double> : true_type { }; -template<> struct is_floating_point<long double> : true_type { }; -template <class T> struct is_floating_point<const T> - : is_floating_point<T> { }; -template <class T> struct is_floating_point<volatile T> - : is_floating_point<T> { }; -template <class T> struct is_floating_point<const volatile T> - : is_floating_point<T> { }; - -// is_pointer is false except for pointer types. A cv-qualified type (e.g. -// "int* const", as opposed to "int const*") is cv-qualified if and only if -// the underlying type is. -template <class T> struct is_pointer : false_type { }; -template <class T> struct is_pointer<T*> : true_type { }; -template <class T> struct is_pointer<const T> : is_pointer<T> { }; -template <class T> struct is_pointer<volatile T> : is_pointer<T> { }; -template <class T> struct is_pointer<const volatile T> : is_pointer<T> { }; - -#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) - -namespace internal { - -template <class T> struct is_class_or_union { - template <class U> static small_ tester(void (U::*)()); - template <class U> static big_ tester(...); - static const bool value = sizeof(tester<T>(0)) == sizeof(small_); -}; - -// is_convertible chokes if the first argument is an array. That's why -// we use add_reference here. -template <bool NotUnum, class T> struct is_enum_impl - : is_convertible<typename add_reference<T>::type, int> { }; - -template <class T> struct is_enum_impl<true, T> : false_type { }; - -} // namespace internal - -// Specified by TR1 [4.5.1] primary type categories. - -// Implementation note: -// -// Each type is either void, integral, floating point, array, pointer, -// reference, member object pointer, member function pointer, enum, -// union or class. Out of these, only integral, floating point, reference, -// class and enum types are potentially convertible to int. Therefore, -// if a type is not a reference, integral, floating point or class and -// is convertible to int, it's a enum. Adding cv-qualification to a type -// does not change whether it's an enum. -// -// Is-convertible-to-int check is done only if all other checks pass, -// because it can't be used with some types (e.g. void or classes with -// inaccessible conversion operators). -template <class T> struct is_enum - : internal::is_enum_impl< - is_same<T, void>::value || - is_integral<T>::value || - is_floating_point<T>::value || - is_reference<T>::value || - internal::is_class_or_union<T>::value, - T> { }; - -template <class T> struct is_enum<const T> : is_enum<T> { }; -template <class T> struct is_enum<volatile T> : is_enum<T> { }; -template <class T> struct is_enum<const volatile T> : is_enum<T> { }; - -#endif - -// is_reference is false except for reference types. -template<typename T> struct is_reference : false_type {}; -template<typename T> struct is_reference<T&> : true_type {}; - - -// We can't get is_pod right without compiler help, so fail conservatively. -// We will assume it's false except for arithmetic types, enumerations, -// pointers and cv-qualified versions thereof. Note that std::pair<T,U> -// is not a POD even if T and U are PODs. -template <class T> struct is_pod - : integral_constant<bool, (is_integral<T>::value || - is_floating_point<T>::value || -#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) - // is_enum is not available on MSVC. - is_enum<T>::value || -#endif - is_pointer<T>::value)> { }; -template <class T> struct is_pod<const T> : is_pod<T> { }; -template <class T> struct is_pod<volatile T> : is_pod<T> { }; -template <class T> struct is_pod<const volatile T> : is_pod<T> { }; - - -// We can't get has_trivial_constructor right without compiler help, so -// fail conservatively. We will assume it's false except for: (1) types -// for which is_pod is true. (2) std::pair of types with trivial -// constructors. (3) array of a type with a trivial constructor. -// (4) const versions thereof. -template <class T> struct has_trivial_constructor : is_pod<T> { }; -template <class T, class U> struct has_trivial_constructor<std::pair<T, U> > - : integral_constant<bool, - (has_trivial_constructor<T>::value && - has_trivial_constructor<U>::value)> { }; -template <class A, int N> struct has_trivial_constructor<A[N]> - : has_trivial_constructor<A> { }; -template <class T> struct has_trivial_constructor<const T> - : has_trivial_constructor<T> { }; - -// We can't get has_trivial_copy right without compiler help, so fail -// conservatively. We will assume it's false except for: (1) types -// for which is_pod is true. (2) std::pair of types with trivial copy -// constructors. (3) array of a type with a trivial copy constructor. -// (4) const versions thereof. -template <class T> struct has_trivial_copy : is_pod<T> { }; -template <class T, class U> struct has_trivial_copy<std::pair<T, U> > - : integral_constant<bool, - (has_trivial_copy<T>::value && - has_trivial_copy<U>::value)> { }; -template <class A, int N> struct has_trivial_copy<A[N]> - : has_trivial_copy<A> { }; -template <class T> struct has_trivial_copy<const T> : has_trivial_copy<T> { }; - -// We can't get has_trivial_assign right without compiler help, so fail -// conservatively. We will assume it's false except for: (1) types -// for which is_pod is true. (2) std::pair of types with trivial copy -// constructors. (3) array of a type with a trivial assign constructor. -template <class T> struct has_trivial_assign : is_pod<T> { }; -template <class T, class U> struct has_trivial_assign<std::pair<T, U> > - : integral_constant<bool, - (has_trivial_assign<T>::value && - has_trivial_assign<U>::value)> { }; -template <class A, int N> struct has_trivial_assign<A[N]> - : has_trivial_assign<A> { }; - -// We can't get has_trivial_destructor right without compiler help, so -// fail conservatively. We will assume it's false except for: (1) types -// for which is_pod is true. (2) std::pair of types with trivial -// destructors. (3) array of a type with a trivial destructor. -// (4) const versions thereof. -template <class T> struct has_trivial_destructor : is_pod<T> { }; -template <class T, class U> struct has_trivial_destructor<std::pair<T, U> > - : integral_constant<bool, - (has_trivial_destructor<T>::value && - has_trivial_destructor<U>::value)> { }; -template <class A, int N> struct has_trivial_destructor<A[N]> - : has_trivial_destructor<A> { }; -template <class T> struct has_trivial_destructor<const T> - : has_trivial_destructor<T> { }; - -// Specified by TR1 [4.7.1] -template<typename T> struct remove_const { typedef T type; }; -template<typename T> struct remove_const<T const> { typedef T type; }; -template<typename T> struct remove_volatile { typedef T type; }; -template<typename T> struct remove_volatile<T volatile> { typedef T type; }; -template<typename T> struct remove_cv { - typedef typename remove_const<typename remove_volatile<T>::type>::type type; -}; - - -// Specified by TR1 [4.7.2] Reference modifications. -template<typename T> struct remove_reference { typedef T type; }; -template<typename T> struct remove_reference<T&> { typedef T type; }; - -template <typename T> struct add_reference { typedef T& type; }; -template <typename T> struct add_reference<T&> { typedef T& type; }; - -// Specified by TR1 [4.7.4] Pointer modifications. -template<typename T> struct remove_pointer { typedef T type; }; -template<typename T> struct remove_pointer<T*> { typedef T type; }; -template<typename T> struct remove_pointer<T* const> { typedef T type; }; -template<typename T> struct remove_pointer<T* volatile> { typedef T type; }; -template<typename T> struct remove_pointer<T* const volatile> { - typedef T type; }; - -// Specified by TR1 [4.6] Relationships between types -template<typename T, typename U> struct is_same : public false_type { }; -template<typename T> struct is_same<T, T> : public true_type { }; - -// Specified by TR1 [4.6] Relationships between types -#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) -namespace internal { - -// This class is an implementation detail for is_convertible, and you -// don't need to know how it works to use is_convertible. For those -// who care: we declare two different functions, one whose argument is -// of type To and one with a variadic argument list. We give them -// return types of different size, so we can use sizeof to trick the -// compiler into telling us which function it would have chosen if we -// had called it with an argument of type From. See Alexandrescu's -// _Modern C++ Design_ for more details on this sort of trick. - -template <typename From, typename To> -struct ConvertHelper { - static small_ Test(To); - static big_ Test(...); - static From Create(); -}; -} // namespace internal - -// Inherits from true_type if From is convertible to To, false_type otherwise. -template <typename From, typename To> -struct is_convertible - : integral_constant<bool, - sizeof(internal::ConvertHelper<From, To>::Test( - internal::ConvertHelper<From, To>::Create())) - == sizeof(small_)> { -}; -#endif - -_END_GOOGLE_NAMESPACE_ - -// Right now these macros are no-ops, and mostly just document the fact -// these types are PODs, for human use. They may be made more contentful -// later. The typedef is just to make it legal to put a semicolon after -// these macros. -#define DECLARE_POD(TypeName) typedef int Dummy_Type_For_DECLARE_POD -#define DECLARE_NESTED_POD(TypeName) DECLARE_POD(TypeName) -#define PROPAGATE_POD_FROM_TEMPLATE_ARGUMENT(TemplateName) \ - typedef int Dummy_Type_For_PROPAGATE_POD_FROM_TEMPLATE_ARGUMENT -#define ENFORCE_POD(TypeName) typedef int Dummy_Type_For_ENFORCE_POD - -#endif // BASE_TYPE_TRAITS_H_ +// Copyright (c) 2006, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// ---- +// +// This code is compiled directly on many platforms, including client +// platforms like Windows, Mac, and embedded systems. Before making +// any changes here, make sure that you're not breaking any platforms. +// +// Define a small subset of tr1 type traits. The traits we define are: +// is_integral +// is_floating_point +// is_pointer +// is_enum +// is_reference +// is_pod +// has_trivial_constructor +// has_trivial_copy +// has_trivial_assign +// has_trivial_destructor +// remove_const +// remove_volatile +// remove_cv +// remove_reference +// add_reference +// remove_pointer +// is_same +// is_convertible +// We can add more type traits as required. + +#ifndef BASE_TYPE_TRAITS_H_ +#define BASE_TYPE_TRAITS_H_ + +#include <sparsehash/internal/sparseconfig.h> +#include <utility> // For pair + +#include <sparsehash/template_util.h> // For true_type and false_type + +_START_GOOGLE_NAMESPACE_ + +template <class T> struct is_integral; +template <class T> struct is_floating_point; +template <class T> struct is_pointer; +// MSVC can't compile this correctly, and neither can gcc 3.3.5 (at least) +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) +// is_enum uses is_convertible, which is not available on MSVC. +template <class T> struct is_enum; +#endif +template <class T> struct is_reference; +template <class T> struct is_pod; +template <class T> struct has_trivial_constructor; +template <class T> struct has_trivial_copy; +template <class T> struct has_trivial_assign; +template <class T> struct has_trivial_destructor; +template <class T> struct remove_const; +template <class T> struct remove_volatile; +template <class T> struct remove_cv; +template <class T> struct remove_reference; +template <class T> struct add_reference; +template <class T> struct remove_pointer; +template <class T, class U> struct is_same; +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) +template <class From, class To> struct is_convertible; +#endif + +// is_integral is false except for the built-in integer types. A +// cv-qualified type is integral if and only if the underlying type is. +template <class T> struct is_integral : false_type { }; +template<> struct is_integral<bool> : true_type { }; +template<> struct is_integral<char> : true_type { }; +template<> struct is_integral<unsigned char> : true_type { }; +template<> struct is_integral<signed char> : true_type { }; +#if defined(_MSC_VER) +// wchar_t is not by default a distinct type from unsigned short in +// Microsoft C. +// See http://msdn2.microsoft.com/en-us/library/dh8che7s(VS.80).aspx +template<> struct is_integral<__wchar_t> : true_type { }; +#else +template<> struct is_integral<wchar_t> : true_type { }; +#endif +template<> struct is_integral<short> : true_type { }; +template<> struct is_integral<unsigned short> : true_type { }; +template<> struct is_integral<int> : true_type { }; +template<> struct is_integral<unsigned int> : true_type { }; +template<> struct is_integral<long> : true_type { }; +template<> struct is_integral<unsigned long> : true_type { }; +#ifdef HAVE_LONG_LONG +template<> struct is_integral<long long> : true_type { }; +template<> struct is_integral<unsigned long long> : true_type { }; +#endif +template <class T> struct is_integral<const T> : is_integral<T> { }; +template <class T> struct is_integral<volatile T> : is_integral<T> { }; +template <class T> struct is_integral<const volatile T> : is_integral<T> { }; + +// is_floating_point is false except for the built-in floating-point types. +// A cv-qualified type is integral if and only if the underlying type is. +template <class T> struct is_floating_point : false_type { }; +template<> struct is_floating_point<float> : true_type { }; +template<> struct is_floating_point<double> : true_type { }; +template<> struct is_floating_point<long double> : true_type { }; +template <class T> struct is_floating_point<const T> + : is_floating_point<T> { }; +template <class T> struct is_floating_point<volatile T> + : is_floating_point<T> { }; +template <class T> struct is_floating_point<const volatile T> + : is_floating_point<T> { }; + +// is_pointer is false except for pointer types. A cv-qualified type (e.g. +// "int* const", as opposed to "int const*") is cv-qualified if and only if +// the underlying type is. +template <class T> struct is_pointer : false_type { }; +template <class T> struct is_pointer<T*> : true_type { }; +template <class T> struct is_pointer<const T> : is_pointer<T> { }; +template <class T> struct is_pointer<volatile T> : is_pointer<T> { }; +template <class T> struct is_pointer<const volatile T> : is_pointer<T> { }; + +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) + +namespace internal { + +template <class T> struct is_class_or_union { + template <class U> static small_ tester(void (U::*)()); + template <class U> static big_ tester(...); + static const bool value = sizeof(tester<T>(0)) == sizeof(small_); +}; + +// is_convertible chokes if the first argument is an array. That's why +// we use add_reference here. +template <bool NotUnum, class T> struct is_enum_impl + : is_convertible<typename add_reference<T>::type, int> { }; + +template <class T> struct is_enum_impl<true, T> : false_type { }; + +} // namespace internal + +// Specified by TR1 [4.5.1] primary type categories. + +// Implementation note: +// +// Each type is either void, integral, floating point, array, pointer, +// reference, member object pointer, member function pointer, enum, +// union or class. Out of these, only integral, floating point, reference, +// class and enum types are potentially convertible to int. Therefore, +// if a type is not a reference, integral, floating point or class and +// is convertible to int, it's a enum. Adding cv-qualification to a type +// does not change whether it's an enum. +// +// Is-convertible-to-int check is done only if all other checks pass, +// because it can't be used with some types (e.g. void or classes with +// inaccessible conversion operators). +template <class T> struct is_enum + : internal::is_enum_impl< + is_same<T, void>::value || + is_integral<T>::value || + is_floating_point<T>::value || + is_reference<T>::value || + internal::is_class_or_union<T>::value, + T> { }; + +template <class T> struct is_enum<const T> : is_enum<T> { }; +template <class T> struct is_enum<volatile T> : is_enum<T> { }; +template <class T> struct is_enum<const volatile T> : is_enum<T> { }; + +#endif + +// is_reference is false except for reference types. +template<typename T> struct is_reference : false_type {}; +template<typename T> struct is_reference<T&> : true_type {}; + + +// We can't get is_pod right without compiler help, so fail conservatively. +// We will assume it's false except for arithmetic types, enumerations, +// pointers and cv-qualified versions thereof. Note that std::pair<T,U> +// is not a POD even if T and U are PODs. +template <class T> struct is_pod + : integral_constant<bool, (is_integral<T>::value || + is_floating_point<T>::value || +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) + // is_enum is not available on MSVC. + is_enum<T>::value || +#endif + is_pointer<T>::value)> { }; +template <class T> struct is_pod<const T> : is_pod<T> { }; +template <class T> struct is_pod<volatile T> : is_pod<T> { }; +template <class T> struct is_pod<const volatile T> : is_pod<T> { }; + + +// We can't get has_trivial_constructor right without compiler help, so +// fail conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial +// constructors. (3) array of a type with a trivial constructor. +// (4) const versions thereof. +template <class T> struct has_trivial_constructor : is_pod<T> { }; +template <class T, class U> struct has_trivial_constructor<std::pair<T, U> > + : integral_constant<bool, + (has_trivial_constructor<T>::value && + has_trivial_constructor<U>::value)> { }; +template <class A, int N> struct has_trivial_constructor<A[N]> + : has_trivial_constructor<A> { }; +template <class T> struct has_trivial_constructor<const T> + : has_trivial_constructor<T> { }; + +// We can't get has_trivial_copy right without compiler help, so fail +// conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial copy +// constructors. (3) array of a type with a trivial copy constructor. +// (4) const versions thereof. +template <class T> struct has_trivial_copy : is_pod<T> { }; +template <class T, class U> struct has_trivial_copy<std::pair<T, U> > + : integral_constant<bool, + (has_trivial_copy<T>::value && + has_trivial_copy<U>::value)> { }; +template <class A, int N> struct has_trivial_copy<A[N]> + : has_trivial_copy<A> { }; +template <class T> struct has_trivial_copy<const T> : has_trivial_copy<T> { }; + +// We can't get has_trivial_assign right without compiler help, so fail +// conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial copy +// constructors. (3) array of a type with a trivial assign constructor. +template <class T> struct has_trivial_assign : is_pod<T> { }; +template <class T, class U> struct has_trivial_assign<std::pair<T, U> > + : integral_constant<bool, + (has_trivial_assign<T>::value && + has_trivial_assign<U>::value)> { }; +template <class A, int N> struct has_trivial_assign<A[N]> + : has_trivial_assign<A> { }; + +// We can't get has_trivial_destructor right without compiler help, so +// fail conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial +// destructors. (3) array of a type with a trivial destructor. +// (4) const versions thereof. +template <class T> struct has_trivial_destructor : is_pod<T> { }; +template <class T, class U> struct has_trivial_destructor<std::pair<T, U> > + : integral_constant<bool, + (has_trivial_destructor<T>::value && + has_trivial_destructor<U>::value)> { }; +template <class A, int N> struct has_trivial_destructor<A[N]> + : has_trivial_destructor<A> { }; +template <class T> struct has_trivial_destructor<const T> + : has_trivial_destructor<T> { }; + +// Specified by TR1 [4.7.1] +template<typename T> struct remove_const { typedef T type; }; +template<typename T> struct remove_const<T const> { typedef T type; }; +template<typename T> struct remove_volatile { typedef T type; }; +template<typename T> struct remove_volatile<T volatile> { typedef T type; }; +template<typename T> struct remove_cv { + typedef typename remove_const<typename remove_volatile<T>::type>::type type; +}; + + +// Specified by TR1 [4.7.2] Reference modifications. +template<typename T> struct remove_reference { typedef T type; }; +template<typename T> struct remove_reference<T&> { typedef T type; }; + +template <typename T> struct add_reference { typedef T& type; }; +template <typename T> struct add_reference<T&> { typedef T& type; }; + +// Specified by TR1 [4.7.4] Pointer modifications. +template<typename T> struct remove_pointer { typedef T type; }; +template<typename T> struct remove_pointer<T*> { typedef T type; }; +template<typename T> struct remove_pointer<T* const> { typedef T type; }; +template<typename T> struct remove_pointer<T* volatile> { typedef T type; }; +template<typename T> struct remove_pointer<T* const volatile> { + typedef T type; }; + +// Specified by TR1 [4.6] Relationships between types +template<typename T, typename U> struct is_same : public false_type { }; +template<typename T> struct is_same<T, T> : public true_type { }; + +// Specified by TR1 [4.6] Relationships between types +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) +namespace internal { + +// This class is an implementation detail for is_convertible, and you +// don't need to know how it works to use is_convertible. For those +// who care: we declare two different functions, one whose argument is +// of type To and one with a variadic argument list. We give them +// return types of different size, so we can use sizeof to trick the +// compiler into telling us which function it would have chosen if we +// had called it with an argument of type From. See Alexandrescu's +// _Modern C++ Design_ for more details on this sort of trick. + +template <typename From, typename To> +struct ConvertHelper { + static small_ Test(To); + static big_ Test(...); + static From Create(); +}; +} // namespace internal + +// Inherits from true_type if From is convertible to To, false_type otherwise. +template <typename From, typename To> +struct is_convertible + : integral_constant<bool, + sizeof(internal::ConvertHelper<From, To>::Test( + internal::ConvertHelper<From, To>::Create())) + == sizeof(small_)> { +}; +#endif + +_END_GOOGLE_NAMESPACE_ + +// Right now these macros are no-ops, and mostly just document the fact +// these types are PODs, for human use. They may be made more contentful +// later. The typedef is just to make it legal to put a semicolon after +// these macros. +#define DECLARE_POD(TypeName) typedef int Dummy_Type_For_DECLARE_POD +#define DECLARE_NESTED_POD(TypeName) DECLARE_POD(TypeName) +#define PROPAGATE_POD_FROM_TEMPLATE_ARGUMENT(TemplateName) \ + typedef int Dummy_Type_For_PROPAGATE_POD_FROM_TEMPLATE_ARGUMENT +#define ENFORCE_POD(TypeName) typedef int Dummy_Type_For_ENFORCE_POD + +#endif // BASE_TYPE_TRAITS_H_ diff --git a/contrib/libs/sparsehash/ya.make b/contrib/libs/sparsehash/ya.make index 15ba31eb9d..d12785158a 100644 --- a/contrib/libs/sparsehash/ya.make +++ b/contrib/libs/sparsehash/ya.make @@ -1,7 +1,7 @@ # Generated by devtools/yamaker from nixpkgs 930da485d9af8100f8858bd6fe8f28e3eca26933. -LIBRARY() - +LIBRARY() + LICENSE(BSD-3-Clause) LICENSE_TEXTS(.yandex_meta/licenses.list.txt) @@ -13,5 +13,5 @@ VERSION(2.0.4) ADDINCL( GLOBAL contrib/libs/sparsehash/src ) - -END() + +END() |