47 files changed, 8058 insertions, 1172 deletions

diff --git a/.gitignore b/.gitignore
index 26fa702..a5b5926 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,22 +1,32 @@
-.dirstamp
+*.o
+*~
 .deps
+.dirstamp
 INSTALL
 Makefile
 Makefile.in
 aclocal.m4
 autom4te.cache
+build
 compile
 config.guess
 config.h
-config.h.in~
+config.h.in
 config.log
 config.status
 config.sub
+config.sub
 configure
 depcomp
 libtool
+libtool.m4
+ltmain.sh
+ltoptions.m4
+ltsugar.m4
+ltversion.m4
+lt~obsolete.m4
+missing
 stamp-h1
-*.o
-xdelta3
+xdelta3/xdelta3
 xdelta3decode
 xdelta3regtest
diff --git a/xdelta3/Makefile.all b/xdelta3/Makefile.all
new file mode 100644
index 0000000..c8e8cc0
--- /dev/null
+++ b/xdelta3/Makefile.all
@@ -0,0 +1,40 @@
+# -*- Mode: Makefile -*-
+all: 32_32_32 32_32_64 32_64_64 64_32_32 64_32_64 64_64_64
+
+32_32_32:
+        (cd ./build/m32/32size-32off && $(MAKE) all)
+
+32_32_64:
+        (cd ./build/m32/32size-64off && $(MAKE) all)
+
+32_64_64:
+        (cd ./build/m32/64size-64off && $(MAKE) all)
+
+64_32_32:
+        (cd ./build/m64/32size-32off && $(MAKE) all)
+
+64_32_64:
+        (cd ./build/m64/32size-64off && $(MAKE) all)
+
+64_64_64:
+        (cd ./build/m64/64size-64off && $(MAKE) all)
+
+clean: 32_32_32_clean 32_32_64_clean 32_64_64_clean 64_32_32_clean 64_32_64_clean 64_64_64_clean
+
+32_32_32_clean:
+        (cd ./build/m32/32size-32off && $(MAKE) clean)
+
+32_32_64_clean:
+        (cd ./build/m32/32size-64off && $(MAKE) clean)
+
+32_64_64_clean:
+        (cd ./build/m32/64size-64off && $(MAKE) clean)
+
+64_32_32_clean:
+        (cd ./build/m64/32size-32off && $(MAKE) clean)
+
+64_32_64_clean:
+        (cd ./build/m64/32size-64off && $(MAKE) clean)
+
+64_64_64_clean:
+        (cd ./build/m64/64size-64off && $(MAKE) clean)
diff --git a/xdelta3/Makefile.am b/xdelta3/Makefile.am
index 662ff02..f437764 100644
--- a/xdelta3/Makefile.am
+++ b/xdelta3/Makefile.am
@@ -1,7 +1,8 @@
 ACLOCAL_AMFLAGS = -I m4
+AUTOMAKE_OPTIONS = subdir-objects
 
 bin_PROGRAMS = xdelta3
-noinst_PROGRAMS = xdelta3regtest xdelta3decode
+noinst_PROGRAMS = xdelta3regtest xdelta3decode xdelta3checksum
 
 export AFL_HARDEN
 
@@ -37,14 +38,19 @@ xdelta3regtest_SOURCES = $(common_SOURCES) \
         testing/sizes.h \
         testing/test.h
 
+xdelta3checksum_SOURCES = $(common_SOURCES) \
+        testing/checksum_test.cc \
+        testing/checksum_test_c.c
+
 # Note: for extra sanity checks, enable -Wconversion. Note there
 # are a lot of false positives.
 WFLAGS = -Wall -Wshadow -fno-builtin -Wextra -Wsign-compare \
-         -Wextra -Wno-unused-parameter -Wno-unused-function
+         -Wno-unused-parameter -Wformat -Wno-unused-function # -Wconversion
 
 C_WFLAGS = $(WFLAGS) -pedantic -std=c99
 CXX_WFLAGS = $(WFLAGS)
 
+# TODO add -O3
 common_CFLAGS = \
               -DREGRESSION_TEST=1 \
               -DSECONDARY_DJW=1 \
@@ -80,6 +86,12 @@ xdelta3regtest_CFLAGS = \
         $(C_WFLAGS) $(common_CFLAGS) -DNOT_MAIN=1 -DXD3_DEBUG=1
 xdelta3regtest_LDADD = -lm
 
+xdelta3checksum_CXXFLAGS = \
+        $(CXX_WFLAGS) $(common_CFLAGS) -DNOT_MAIN=1 -DXD3_MAIN=1 -std=c++11
+xdelta3checksum_CFLAGS = \
+        $(C_WFLAGS) $(common_CFLAGS) -DNOT_MAIN=1 -DXD3_MAIN=1
+
+
 man1_MANS = xdelta3.1
 
 EXTRA_DIST = \
diff --git a/xdelta3/configure.ac b/xdelta3/configure.ac
index ce24045..dbcac4f 100644
--- a/xdelta3/configure.ac
+++ b/xdelta3/configure.ac
@@ -44,6 +44,7 @@ AC_ARG_ENABLE(debug-symbols,
    AS_HELP_STRING(--enable-debug-symbols,[Build with debug symbols (default is NO)]),,enableval=no)
 AM_CONDITIONAL([DEBUG_SYMBOLS], [test ${enableval} = "yes"])
 
+
 AC_CONFIG_HEADERS([config.h])
 AC_CONFIG_FILES([Makefile])
 AC_OUTPUT
diff --git a/xdelta3/cpp-btree/CMakeLists.txt b/xdelta3/cpp-btree/CMakeLists.txt
new file mode 100644
index 0000000..d005e15
--- /dev/null
+++ b/xdelta3/cpp-btree/CMakeLists.txt
@@ -0,0 +1,40 @@
+# Copyright 2013 Google Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+cmake_minimum_required(VERSION 2.6)
+
+project(cppbtree CXX)
+
+option(build_tests "Build B-tree tests" OFF)
+add_definitions(-std=c++11)
+set(CMAKE_CXX_FLAGS "-g -O2")
+
+# CMake doesn't have a way to pure template library, 
+# add_library(cppbtree btree.h btree_map.h btree_set.h 
+#             safe_btree.h safe_btree_map.h safe_btree_set.h)
+# set_target_properties(cppbtree PROPERTIES LINKER_LANGUAGE CXX)
+
+if(build_tests)
+  enable_testing()
+  include_directories($ENV{GTEST_ROOT}/include)
+  link_directories($ENV{GTEST_ROOT})
+  include_directories($ENV{GFLAGS_ROOT}/include)
+  link_directories($ENV{GFLAGS_ROOT}/lib)
+  add_executable(btree_test btree_test.cc btree_test_flags.cc)
+  add_executable(safe_btree_test safe_btree_test.cc btree_test_flags.cc)
+  add_executable(btree_bench btree_bench.cc btree_test_flags.cc)
+  target_link_libraries(btree_test gtest_main gtest gflags)
+  target_link_libraries(safe_btree_test gtest_main gtest gflags)
+  target_link_libraries(btree_bench gflags gtest)
+endif()
diff --git a/xdelta3/cpp-btree/COPYING b/xdelta3/cpp-btree/COPYING
new file mode 100644
index 0000000..d645695
--- /dev/null
+++ b/xdelta3/cpp-btree/COPYING
@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/xdelta3/cpp-btree/README b/xdelta3/cpp-btree/README
new file mode 100644
index 0000000..319fe9b
--- /dev/null
+++ b/xdelta3/cpp-btree/README
@@ -0,0 +1,31 @@
+This library is a C++ template library and, as such, there is no
+library to build and install.  Copy the .h files and use them!
+
+See http://code.google.com/p/cpp-btree/wiki/UsageInstructions for
+details.
+
+----
+
+To build and run the provided tests, however, you will need to install
+CMake, the Google C++ Test framework, and the Google flags package.
+
+Download and install CMake from http://www.cmake.org
+
+Download and build the GoogleTest framework from
+http://code.google.com/p/googletest
+
+Download and install gflags from https://code.google.com/p/gflags
+
+Set GTEST_ROOT to the directory where GTEST was built.
+Set GFLAGS_ROOT to the directory prefix where GFLAGS is installed.
+
+export GTEST_ROOT=/path/for/gtest-x.y
+export GFLAGS_ROOT=/opt
+
+cmake . -Dbuild_tests=ON
+
+For example, to build on a Unix system with the clang++ compiler,
+
+export GTEST_ROOT=$(HOME)/src/googletest
+export GFLAGS_ROOT=/opt
+cmake . -G "Unix Makefiles" -Dbuild_tests=ON -DCMAKE_CXX_COMPILER=clang++
diff --git a/xdelta3/cpp-btree/btree.h b/xdelta3/cpp-btree/btree.h
new file mode 100644
index 0000000..cdd2b52
--- /dev/null
+++ b/xdelta3/cpp-btree/btree.h
@@ -0,0 +1,2394 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// A btree implementation of the STL set and map interfaces. A btree is both
+// smaller and faster than STL set/map. The red-black tree implementation of
+// STL set/map has an overhead of 3 pointers (left, right and parent) plus the
+// node color information for each stored value. So a set<int32> consumes 20
+// bytes for each value stored. This btree implementation stores multiple
+// values on fixed size nodes (usually 256 bytes) and doesn't store child
+// pointers for leaf nodes. The result is that a btree_set<int32> may use much
+// less memory per stored value. For the random insertion benchmark in
+// btree_test.cc, a btree_set<int32> with node-size of 256 uses 4.9 bytes per
+// stored value.
+//
+// The packing of multiple values on to each node of a btree has another effect
+// besides better space utilization: better cache locality due to fewer cache
+// lines being accessed. Better cache locality translates into faster
+// operations.
+//
+// CAVEATS
+//
+// Insertions and deletions on a btree can cause splitting, merging or
+// rebalancing of btree nodes. And even without these operations, insertions
+// and deletions on a btree will move values around within a node. In both
+// cases, the result is that insertions and deletions can invalidate iterators
+// pointing to values other than the one being inserted/deleted. This is
+// notably different from STL set/map which takes care to not invalidate
+// iterators on insert/erase except, of course, for iterators pointing to the
+// value being erased.  A partial workaround when erasing is available:
+// erase() returns an iterator pointing to the item just after the one that was
+// erased (or end() if none exists).  See also safe_btree.
+
+// PERFORMANCE
+//
+//   btree_bench --benchmarks=. 2>&1 | ./benchmarks.awk
+//
+// Run on pmattis-warp.nyc (4 X 2200 MHz CPUs); 2010/03/04-15:23:06
+// Benchmark                 STL(ns) B-Tree(ns) @    <size>
+// --------------------------------------------------------
+// BM_set_int32_insert        1516      608  +59.89%  <256>    [40.0,  5.2]
+// BM_set_int32_lookup        1160      414  +64.31%  <256>    [40.0,  5.2]
+// BM_set_int32_fulllookup     960      410  +57.29%  <256>    [40.0,  4.4]
+// BM_set_int32_delete        1741      528  +69.67%  <256>    [40.0,  5.2]
+// BM_set_int32_queueaddrem   3078     1046  +66.02%  <256>    [40.0,  5.5]
+// BM_set_int32_mixedaddrem   3600     1384  +61.56%  <256>    [40.0,  5.3]
+// BM_set_int32_fifo           227      113  +50.22%  <256>    [40.0,  4.4]
+// BM_set_int32_fwditer        158       26  +83.54%  <256>    [40.0,  5.2]
+// BM_map_int32_insert        1551      636  +58.99%  <256>    [48.0, 10.5]
+// BM_map_int32_lookup        1200      508  +57.67%  <256>    [48.0, 10.5]
+// BM_map_int32_fulllookup     989      487  +50.76%  <256>    [48.0,  8.8]
+// BM_map_int32_delete        1794      628  +64.99%  <256>    [48.0, 10.5]
+// BM_map_int32_queueaddrem   3189     1266  +60.30%  <256>    [48.0, 11.6]
+// BM_map_int32_mixedaddrem   3822     1623  +57.54%  <256>    [48.0, 10.9]
+// BM_map_int32_fifo           151      134  +11.26%  <256>    [48.0,  8.8]
+// BM_map_int32_fwditer        161       32  +80.12%  <256>    [48.0, 10.5]
+// BM_set_int64_insert        1546      636  +58.86%  <256>    [40.0, 10.5]
+// BM_set_int64_lookup        1200      512  +57.33%  <256>    [40.0, 10.5]
+// BM_set_int64_fulllookup     971      487  +49.85%  <256>    [40.0,  8.8]
+// BM_set_int64_delete        1745      616  +64.70%  <256>    [40.0, 10.5]
+// BM_set_int64_queueaddrem   3163     1195  +62.22%  <256>    [40.0, 11.6]
+// BM_set_int64_mixedaddrem   3760     1564  +58.40%  <256>    [40.0, 10.9]
+// BM_set_int64_fifo           146      103  +29.45%  <256>    [40.0,  8.8]
+// BM_set_int64_fwditer        162       31  +80.86%  <256>    [40.0, 10.5]
+// BM_map_int64_insert        1551      720  +53.58%  <256>    [48.0, 20.7]
+// BM_map_int64_lookup        1214      612  +49.59%  <256>    [48.0, 20.7]
+// BM_map_int64_fulllookup     994      592  +40.44%  <256>    [48.0, 17.2]
+// BM_map_int64_delete        1778      764  +57.03%  <256>    [48.0, 20.7]
+// BM_map_int64_queueaddrem   3189     1547  +51.49%  <256>    [48.0, 20.9]
+// BM_map_int64_mixedaddrem   3779     1887  +50.07%  <256>    [48.0, 21.6]
+// BM_map_int64_fifo           147      145   +1.36%  <256>    [48.0, 17.2]
+// BM_map_int64_fwditer        162       41  +74.69%  <256>    [48.0, 20.7]
+// BM_set_string_insert       1989     1966   +1.16%  <256>    [64.0, 44.5]
+// BM_set_string_lookup       1709     1600   +6.38%  <256>    [64.0, 44.5]
+// BM_set_string_fulllookup   1573     1529   +2.80%  <256>    [64.0, 35.4]
+// BM_set_string_delete       2520     1920  +23.81%  <256>    [64.0, 44.5]
+// BM_set_string_queueaddrem  4706     4309   +8.44%  <256>    [64.0, 48.3]
+// BM_set_string_mixedaddrem  5080     4654   +8.39%  <256>    [64.0, 46.7]
+// BM_set_string_fifo          318      512  -61.01%  <256>    [64.0, 35.4]
+// BM_set_string_fwditer       182       93  +48.90%  <256>    [64.0, 44.5]
+// BM_map_string_insert       2600     2227  +14.35%  <256>    [72.0, 55.8]
+// BM_map_string_lookup       2068     1730  +16.34%  <256>    [72.0, 55.8]
+// BM_map_string_fulllookup   1859     1618  +12.96%  <256>    [72.0, 44.0]
+// BM_map_string_delete       3168     2080  +34.34%  <256>    [72.0, 55.8]
+// BM_map_string_queueaddrem  5840     4701  +19.50%  <256>    [72.0, 59.4]
+// BM_map_string_mixedaddrem  6400     5200  +18.75%  <256>    [72.0, 57.8]
+// BM_map_string_fifo          398      596  -49.75%  <256>    [72.0, 44.0]
+// BM_map_string_fwditer       243      113  +53.50%  <256>    [72.0, 55.8]
+
+#ifndef UTIL_BTREE_BTREE_H__
+#define UTIL_BTREE_BTREE_H__
+
+#include <assert.h>
+#include <stddef.h>
+#include <string.h>
+#include <sys/types.h>
+#include <algorithm>
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <limits>
+#include <type_traits>
+#include <new>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#ifndef NDEBUG
+#define NDEBUG 1
+#endif
+
+namespace btree {
+
+// Inside a btree method, if we just call swap(), it will choose the
+// btree::swap method, which we don't want. And we can't say ::swap
+// because then MSVC won't pickup any std::swap() implementations. We
+// can't just use std::swap() directly because then we don't get the
+// specialization for types outside the std namespace. So the solution
+// is to have a special swap helper function whose name doesn't
+// collide with other swap functions defined by the btree classes.
+template <typename T>
+inline void btree_swap_helper(T &a, T &b) {
+  using std::swap;
+  swap(a, b);
+}
+
+// A template helper used to select A or B based on a condition.
+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;
+};
+
+// Types small_ and big_ are promise that sizeof(small_) < sizeof(big_)
+typedef char small_;
+
+struct big_ {
+  char dummy[2];
+};
+
+// A compile-time assertion.
+template <bool>
+struct CompileAssert {
+};
+
+#define COMPILE_ASSERT(expr, msg) \
+  typedef CompileAssert<(bool(expr))> msg[bool(expr) ? 1 : -1]
+
+// A helper type used to indicate that a key-compare-to functor has been
+// provided. A user can specify a key-compare-to functor by doing:
+//
+//  struct MyStringComparer
+//      : public util::btree::btree_key_compare_to_tag {
+//    int operator()(const string &a, const string &b) const {
+//      return a.compare(b);
+//    }
+//  };
+//
+// Note that the return type is an int and not a bool. There is a
+// COMPILE_ASSERT which enforces this return type.
+struct btree_key_compare_to_tag {
+};
+
+// A helper class that indicates if the Compare parameter is derived from
+// btree_key_compare_to_tag.
+template <typename Compare>
+struct btree_is_key_compare_to
+    : public std::is_convertible<Compare, btree_key_compare_to_tag> {
+};
+
+// A helper class to convert a boolean comparison into a three-way
+// "compare-to" comparison that returns a negative value to indicate
+// less-than, zero to indicate equality and a positive value to
+// indicate greater-than. This helper class is specialized for
+// less<string> and greater<string>. The btree_key_compare_to_adapter
+// class is provided so that btree users automatically get the more
+// efficient compare-to code when using common google string types
+// with common comparison functors.
+template <typename Compare>
+struct btree_key_compare_to_adapter : Compare {
+  btree_key_compare_to_adapter() { }
+  btree_key_compare_to_adapter(const Compare &c) : Compare(c) { }
+  btree_key_compare_to_adapter(const btree_key_compare_to_adapter<Compare> &c)
+      : Compare(c) {
+  }
+};
+
+template <>
+struct btree_key_compare_to_adapter<std::less<std::string> >
+    : public btree_key_compare_to_tag {
+  btree_key_compare_to_adapter() {}
+  btree_key_compare_to_adapter(const std::less<std::string>&) {}
+  btree_key_compare_to_adapter(
+      const btree_key_compare_to_adapter<std::less<std::string> >&) {}
+  int operator()(const std::string &a, const std::string &b) const {
+    return a.compare(b);
+  }
+};
+
+template <>
+struct btree_key_compare_to_adapter<std::greater<std::string> >
+    : public btree_key_compare_to_tag {
+  btree_key_compare_to_adapter() {}
+  btree_key_compare_to_adapter(const std::greater<std::string>&) {}
+  btree_key_compare_to_adapter(
+      const btree_key_compare_to_adapter<std::greater<std::string> >&) {}
+  int operator()(const std::string &a, const std::string &b) const {
+    return b.compare(a);
+  }
+};
+
+// A helper class that allows a compare-to functor to behave like a plain
+// compare functor. This specialization is used when we do not have a
+// compare-to functor.
+template <typename Key, typename Compare, bool HaveCompareTo>
+struct btree_key_comparer {
+  btree_key_comparer() {}
+  btree_key_comparer(Compare c) : comp(c) {}
+  static bool bool_compare(const Compare &comp, const Key &x, const Key &y) {
+    return comp(x, y);
+  }
+  bool operator()(const Key &x, const Key &y) const {
+    return bool_compare(comp, x, y);
+  }
+  Compare comp;
+};
+
+// A specialization of btree_key_comparer when a compare-to functor is
+// present. We need a plain (boolean) comparison in some parts of the btree
+// code, such as insert-with-hint.
+template <typename Key, typename Compare>
+struct btree_key_comparer<Key, Compare, true> {
+  btree_key_comparer() {}
+  btree_key_comparer(Compare c) : comp(c) {}
+  static bool bool_compare(const Compare &comp, const Key &x, const Key &y) {
+    return comp(x, y) < 0;
+  }
+  bool operator()(const Key &x, const Key &y) const {
+    return bool_compare(comp, x, y);
+  }
+  Compare comp;
+};
+
+// A helper function to compare to keys using the specified compare
+// functor. This dispatches to the appropriate btree_key_comparer comparison,
+// depending on whether we have a compare-to functor or not (which depends on
+// whether Compare is derived from btree_key_compare_to_tag).
+template <typename Key, typename Compare>
+static bool btree_compare_keys(
+    const Compare &comp, const Key &x, const Key &y) {
+  typedef btree_key_comparer<Key, Compare,
+      btree_is_key_compare_to<Compare>::value> key_comparer;
+  return key_comparer::bool_compare(comp, x, y);
+}
+
+template <typename Key, typename Compare,
+          typename Alloc, int TargetNodeSize, int ValueSize>
+struct btree_common_params {
+  // If Compare is derived from btree_key_compare_to_tag then use it as the
+  // key_compare type. Otherwise, use btree_key_compare_to_adapter<> which will
+  // fall-back to Compare if we don't have an appropriate specialization.
+  typedef typename if_<
+    btree_is_key_compare_to<Compare>::value,
+    Compare, btree_key_compare_to_adapter<Compare> >::type key_compare;
+  // A type which indicates if we have a key-compare-to functor or a plain old
+  // key-compare functor.
+  typedef btree_is_key_compare_to<key_compare> is_key_compare_to;
+
+  typedef Alloc allocator_type;
+  typedef Key key_type;
+  typedef ssize_t size_type;
+  typedef ptrdiff_t difference_type;
+
+  enum {
+    kTargetNodeSize = TargetNodeSize,
+
+    // Available space for values.  This is largest for leaf nodes,
+    // which has overhead no fewer than two pointers.
+    kNodeValueSpace = TargetNodeSize - 2 * sizeof(void*),
+  };
+
+  // This is an integral type large enough to hold as many
+  // ValueSize-values as will fit a node of TargetNodeSize bytes.
+  typedef typename if_<
+    (kNodeValueSpace / ValueSize) >= 256,
+    uint16_t,
+    uint8_t>::type node_count_type;
+};
+
+// A parameters structure for holding the type parameters for a btree_map.
+template <typename Key, typename Data, typename Compare,
+          typename Alloc, int TargetNodeSize>
+struct btree_map_params
+    : public btree_common_params<Key, Compare, Alloc, TargetNodeSize,
+                                 sizeof(Key) + sizeof(Data)> {
+  typedef Data data_type;
+  typedef Data mapped_type;
+  typedef std::pair<const Key, data_type> value_type;
+  typedef std::pair<Key, data_type> mutable_value_type;
+  typedef value_type* pointer;
+  typedef const value_type* const_pointer;
+  typedef value_type& reference;
+  typedef const value_type& const_reference;
+
+  enum {
+    kValueSize = sizeof(Key) + sizeof(data_type),
+  };
+
+  static const Key& key(const value_type &x) { return x.first; }
+  static const Key& key(const mutable_value_type &x) { return x.first; }
+  static void swap(mutable_value_type *a, mutable_value_type *b) {
+    btree_swap_helper(a->first, b->first);
+    btree_swap_helper(a->second, b->second);
+  }
+};
+
+// A parameters structure for holding the type parameters for a btree_set.
+template <typename Key, typename Compare, typename Alloc, int TargetNodeSize>
+struct btree_set_params
+    : public btree_common_params<Key, Compare, Alloc, TargetNodeSize,
+                                 sizeof(Key)> {
+  typedef std::false_type data_type;
+  typedef std::false_type mapped_type;
+  typedef Key value_type;
+  typedef value_type mutable_value_type;
+  typedef value_type* pointer;
+  typedef const value_type* const_pointer;
+  typedef value_type& reference;
+  typedef const value_type& const_reference;
+
+  enum {
+    kValueSize = sizeof(Key),
+  };
+
+  static const Key& key(const value_type &x) { return x; }
+  static void swap(mutable_value_type *a, mutable_value_type *b) {
+    btree_swap_helper<mutable_value_type>(*a, *b);
+  }
+};
+
+// An adapter class that converts a lower-bound compare into an upper-bound
+// compare.
+template <typename Key, typename Compare>
+struct btree_upper_bound_adapter : public Compare {
+  btree_upper_bound_adapter(Compare c) : Compare(c) {}
+  bool operator()(const Key &a, const Key &b) const {
+    return !static_cast<const Compare&>(*this)(b, a);
+  }
+};
+
+template <typename Key, typename CompareTo>
+struct btree_upper_bound_compare_to_adapter : public CompareTo {
+  btree_upper_bound_compare_to_adapter(CompareTo c) : CompareTo(c) {}
+  int operator()(const Key &a, const Key &b) const {
+    return static_cast<const CompareTo&>(*this)(b, a);
+  }
+};
+
+// Dispatch helper class for using linear search with plain compare.
+template <typename K, typename N, typename Compare>
+struct btree_linear_search_plain_compare {
+  static int lower_bound(const K &k, const N &n, Compare comp)  {
+    return n.linear_search_plain_compare(k, 0, n.count(), comp);
+  }
+  static int upper_bound(const K &k, const N &n, Compare comp)  {
+    typedef btree_upper_bound_adapter<K, Compare> upper_compare;
+    return n.linear_search_plain_compare(k, 0, n.count(), upper_compare(comp));
+  }
+};
+
+// Dispatch helper class for using linear search with compare-to
+template <typename K, typename N, typename CompareTo>
+struct btree_linear_search_compare_to {
+  static int lower_bound(const K &k, const N &n, CompareTo comp)  {
+    return n.linear_search_compare_to(k, 0, n.count(), comp);
+  }
+  static int upper_bound(const K &k, const N &n, CompareTo comp)  {
+    typedef btree_upper_bound_adapter<K,
+        btree_key_comparer<K, CompareTo, true> > upper_compare;
+    return n.linear_search_plain_compare(k, 0, n.count(), upper_compare(comp));
+  }
+};
+
+// Dispatch helper class for using binary search with plain compare.
+template <typename K, typename N, typename Compare>
+struct btree_binary_search_plain_compare {
+  static int lower_bound(const K &k, const N &n, Compare comp)  {
+    return n.binary_search_plain_compare(k, 0, n.count(), comp);
+  }
+  static int upper_bound(const K &k, const N &n, Compare comp)  {
+    typedef btree_upper_bound_adapter<K, Compare> upper_compare;
+    return n.binary_search_plain_compare(k, 0, n.count(), upper_compare(comp));
+  }
+};
+
+// Dispatch helper class for using binary search with compare-to.
+template <typename K, typename N, typename CompareTo>
+struct btree_binary_search_compare_to {
+  static int lower_bound(const K &k, const N &n, CompareTo comp)  {
+    return n.binary_search_compare_to(k, 0, n.count(), CompareTo());
+  }
+  static int upper_bound(const K &k, const N &n, CompareTo comp)  {
+    typedef btree_upper_bound_adapter<K,
+        btree_key_comparer<K, CompareTo, true> > upper_compare;
+    return n.linear_search_plain_compare(k, 0, n.count(), upper_compare(comp));
+  }
+};
+
+// A node in the btree holding. The same node type is used for both internal
+// and leaf nodes in the btree, though the nodes are allocated in such a way
+// that the children array is only valid in internal nodes.
+template <typename Params>
+class btree_node {
+ public:
+  typedef Params params_type;
+  typedef btree_node<Params> self_type;
+  typedef typename Params::key_type key_type;
+  typedef typename Params::data_type data_type;
+  typedef typename Params::value_type value_type;
+  typedef typename Params::mutable_value_type mutable_value_type;
+  typedef typename Params::pointer pointer;
+  typedef typename Params::const_pointer const_pointer;
+  typedef typename Params::reference reference;
+  typedef typename Params::const_reference const_reference;
+  typedef typename Params::key_compare key_compare;
+  typedef typename Params::size_type size_type;
+  typedef typename Params::difference_type difference_type;
+  // Typedefs for the various types of node searches.
+  typedef btree_linear_search_plain_compare<
+    key_type, self_type, key_compare> linear_search_plain_compare_type;
+  typedef btree_linear_search_compare_to<
+    key_type, self_type, key_compare> linear_search_compare_to_type;
+  typedef btree_binary_search_plain_compare<
+    key_type, self_type, key_compare> binary_search_plain_compare_type;
+  typedef btree_binary_search_compare_to<
+    key_type, self_type, key_compare> binary_search_compare_to_type;
+  // If we have a valid key-compare-to type, use linear_search_compare_to,
+  // otherwise use linear_search_plain_compare.
+  typedef typename if_<
+    Params::is_key_compare_to::value,
+    linear_search_compare_to_type,
+    linear_search_plain_compare_type>::type linear_search_type;
+  // If we have a valid key-compare-to type, use binary_search_compare_to,
+  // otherwise use binary_search_plain_compare.
+  typedef typename if_<
+    Params::is_key_compare_to::value,
+    binary_search_compare_to_type,
+    binary_search_plain_compare_type>::type binary_search_type;
+  // If the key is an integral or floating point type, use linear search which
+  // is faster than binary search for such types. Might be wise to also
+  // configure linear search based on node-size.
+  typedef typename if_<
+    std::is_integral<key_type>::value ||
+    std::is_floating_point<key_type>::value,
+    linear_search_type, binary_search_type>::type search_type;
+
+  struct base_fields {
+    typedef typename Params::node_count_type field_type;
+
+    // A boolean indicating whether the node is a leaf or not.
+    bool leaf;
+    // The position of the node in the node's parent.
+    field_type position;
+    // The maximum number of values the node can hold.
+    field_type max_count;
+    // The count of the number of values in the node.
+    field_type count;
+    // A pointer to the node's parent.
+    btree_node *parent;
+  };
+
+  enum {
+    kValueSize = params_type::kValueSize,
+    kTargetNodeSize = params_type::kTargetNodeSize,
+
+    // Compute how many values we can fit onto a leaf node.
+    kNodeTargetValues = (kTargetNodeSize - sizeof(base_fields)) / kValueSize,
+    // We need a minimum of 3 values per internal node in order to perform
+    // splitting (1 value for the two nodes involved in the split and 1 value
+    // propagated to the parent as the delimiter for the split).
+    kNodeValues = kNodeTargetValues >= 3 ? kNodeTargetValues : 3,
+
+    kExactMatch = 1 << 30,
+    kMatchMask = kExactMatch - 1,
+  };
+
+  struct leaf_fields : public base_fields {
+    // The array of values. Only the first count of these values have been
+    // constructed and are valid.
+    mutable_value_type values[kNodeValues];
+  };
+
+  struct internal_fields : public leaf_fields {
+    // The array of child pointers. The keys in children_[i] are all less than
+    // key(i). The keys in children_[i + 1] are all greater than key(i). There
+    // are always count + 1 children.
+    btree_node *children[kNodeValues + 1];
+  };
+
+  struct root_fields : public internal_fields {
+    btree_node *rightmost;
+    size_type size;
+  };
+
+ public:
+  // Getter/setter for whether this is a leaf node or not. This value doesn't
+  // change after the node is created.
+  bool leaf() const { return fields_.leaf; }
+
+  // Getter for the position of this node in its parent.
+  int position() const { return fields_.position; }
+  void set_position(int v) { fields_.position = v; }
+
+  // Getter/setter for the number of values stored in this node.
+  int count() const { return fields_.count; }
+  void set_count(int v) { fields_.count = v; }
+  int max_count() const { return fields_.max_count; }
+
+  // Getter for the parent of this node.
+  btree_node* parent() const { return fields_.parent; }
+  // Getter for whether the node is the root of the tree. The parent of the
+  // root of the tree is the leftmost node in the tree which is guaranteed to
+  // be a leaf.
+  bool is_root() const { return parent()->leaf(); }
+  void make_root() {
+    assert(parent()->is_root());
+    fields_.parent = fields_.parent->parent();
+  }
+
+  // Getter for the rightmost root node field. Only valid on the root node.
+  btree_node* rightmost() const { return fields_.rightmost; }
+  btree_node** mutable_rightmost() { return &fields_.rightmost; }
+
+  // Getter for the size root node field. Only valid on the root node.
+  size_type size() const { return fields_.size; }
+  size_type* mutable_size() { return &fields_.size; }
+
+  // Getters for the key/value at position i in the node.
+  const key_type& key(int i) const {
+    return params_type::key(fields_.values[i]);
+  }
+  reference value(int i) {
+    return reinterpret_cast<reference>(fields_.values[i]);
+  }
+  const_reference value(int i) const {
+    return reinterpret_cast<const_reference>(fields_.values[i]);
+  }
+  mutable_value_type* mutable_value(int i) {
+    return &fields_.values[i];
+  }
+
+  // Swap value i in this node with value j in node x.
+  void value_swap(int i, btree_node *x, int j) {
+    params_type::swap(mutable_value(i), x->mutable_value(j));
+  }
+
+  // Getters/setter for the child at position i in the node.
+  btree_node* child(int i) const { return fields_.children[i]; }
+  btree_node** mutable_child(int i) { return &fields_.children[i]; }
+  void set_child(int i, btree_node *c) {
+    *mutable_child(i) = c;
+    c->fields_.parent = this;
+    c->fields_.position = i;
+  }
+
+  // Returns the position of the first value whose key is not less than k.
+  template <typename Compare>
+  int lower_bound(const key_type &k, const Compare &comp) const {
+    return search_type::lower_bound(k, *this, comp);
+  }
+  // Returns the position of the first value whose key is greater than k.
+  template <typename Compare>
+  int upper_bound(const key_type &k, const Compare &comp) const {
+    return search_type::upper_bound(k, *this, comp);
+  }
+
+  // Returns the position of the first value whose key is not less than k using
+  // linear search performed using plain compare.
+  template <typename Compare>
+  int linear_search_plain_compare(
+      const key_type &k, int s, int e, const Compare &comp) const {
+    while (s < e) {
+      if (!btree_compare_keys(comp, key(s), k)) {
+        break;
+      }
+      ++s;
+    }
+    return s;
+  }
+
+  // Returns the position of the first value whose key is not less than k using
+  // linear search performed using compare-to.
+  template <typename Compare>
+  int linear_search_compare_to(
+      const key_type &k, int s, int e, const Compare &comp) const {
+    while (s < e) {
+      int c = comp(key(s), k);
+      if (c == 0) {
+        return s | kExactMatch;
+      } else if (c > 0) {
+        break;
+      }
+      ++s;
+    }
+    return s;
+  }
+
+  // Returns the position of the first value whose key is not less than k using
+  // binary search performed using plain compare.
+  template <typename Compare>
+  int binary_search_plain_compare(
+      const key_type &k, int s, int e, const Compare &comp) const {
+    while (s != e) {
+      int mid = (s + e) / 2;
+      if (btree_compare_keys(comp, key(mid), k)) {
+        s = mid + 1;
+      } else {
+        e = mid;
+      }
+    }
+    return s;
+  }
+
+  // Returns the position of the first value whose key is not less than k using
+  // binary search performed using compare-to.
+  template <typename CompareTo>
+  int binary_search_compare_to(
+      const key_type &k, int s, int e, const CompareTo &comp) const {
+    while (s != e) {
+      int mid = (s + e) / 2;
+      int c = comp(key(mid), k);
+      if (c < 0) {
+        s = mid + 1;
+      } else if (c > 0) {
+        e = mid;
+      } else {
+        // Need to return the first value whose key is not less than k, which
+        // requires continuing the binary search. Note that we are guaranteed
+        // that the result is an exact match because if "key(mid-1) < k" the
+        // call to binary_search_compare_to() will return "mid".
+        s = binary_search_compare_to(k, s, mid, comp);
+        return s | kExactMatch;
+      }
+    }
+    return s;
+  }
+
+  // Inserts the value x at position i, shifting all existing values and
+  // children at positions >= i to the right by 1.
+  void insert_value(int i, const value_type &x);
+
+  // Removes the value at position i, shifting all existing values and children
+  // at positions > i to the left by 1.
+  void remove_value(int i);
+
+  // Rebalances a node with its right sibling.
+  void rebalance_right_to_left(btree_node *sibling, int to_move);
+  void rebalance_left_to_right(btree_node *sibling, int to_move);
+
+  // Splits a node, moving a portion of the node's values to its right sibling.
+  void split(btree_node *sibling, int insert_position);
+
+  // Merges a node with its right sibling, moving all of the values and the
+  // delimiting key in the parent node onto itself.
+  void merge(btree_node *sibling);
+
+  // Swap the contents of "this" and "src".
+  void swap(btree_node *src);
+
+  // Node allocation/deletion routines.
+  static btree_node* init_leaf(
+      leaf_fields *f, btree_node *parent, int max_count) {
+    btree_node *n = reinterpret_cast<btree_node*>(f);
+    f->leaf = 1;
+    f->position = 0;
+    f->max_count = max_count;
+    f->count = 0;
+    f->parent = parent;
+    if (!NDEBUG) {
+      memset(&f->values, 0, max_count * sizeof(value_type));
+    }
+    return n;
+  }
+  static btree_node* init_internal(internal_fields *f, btree_node *parent) {
+    btree_node *n = init_leaf(f, parent, kNodeValues);
+    f->leaf = 0;
+    if (!NDEBUG) {
+      memset(f->children, 0, sizeof(f->children));
+    }
+    return n;
+  }
+  static btree_node* init_root(root_fields *f, btree_node *parent) {
+    btree_node *n = init_internal(f, parent);
+    f->rightmost = parent;
+    f->size = parent->count();
+    return n;
+  }
+  void destroy() {
+    for (int i = 0; i < count(); ++i) {
+      value_destroy(i);
+    }
+  }
+
+ private:
+  void value_init(int i) {
+    new (&fields_.values[i]) mutable_value_type;
+  }
+  void value_init(int i, const value_type &x) {
+    new (&fields_.values[i]) mutable_value_type(x);
+  }
+  void value_destroy(int i) {
+    fields_.values[i].~mutable_value_type();
+  }
+
+ private:
+  root_fields fields_;
+
+ private:
+  btree_node(const btree_node&);
+  void operator=(const btree_node&);
+};
+
+template <typename Node, typename Reference, typename Pointer>
+struct btree_iterator {
+  typedef typename Node::key_type key_type;
+  typedef typename Node::size_type size_type;
+  typedef typename Node::difference_type difference_type;
+  typedef typename Node::params_type params_type;
+
+  typedef Node node_type;
+  typedef typename std::remove_const<Node>::type normal_node;
+  typedef const Node const_node;
+  typedef typename params_type::value_type value_type;
+  typedef typename params_type::pointer normal_pointer;
+  typedef typename params_type::reference normal_reference;
+  typedef typename params_type::const_pointer const_pointer;
+  typedef typename params_type::const_reference const_reference;
+
+  typedef Pointer pointer;
+  typedef Reference reference;
+  typedef std::bidirectional_iterator_tag iterator_category;
+
+  typedef btree_iterator<
+    normal_node, normal_reference, normal_pointer> iterator;
+  typedef btree_iterator<
+    const_node, const_reference, const_pointer> const_iterator;
+  typedef btree_iterator<Node, Reference, Pointer> self_type;
+
+  btree_iterator()
+      : node(NULL),
+        position(-1) {
+  }
+  btree_iterator(Node *n, int p)
+      : node(n),
+        position(p) {
+  }
+  btree_iterator(const iterator &x)
+      : node(x.node),
+        position(x.position) {
+  }
+
+  // Increment/decrement the iterator.
+  void increment() {
+    if (node->leaf() && ++position < node->count()) {
+      return;
+    }
+    increment_slow();
+  }
+  void increment_by(int count);
+  void increment_slow();
+
+  void decrement() {
+    if (node->leaf() && --position >= 0) {
+      return;
+    }
+    decrement_slow();
+  }
+  void decrement_slow();
+
+  bool operator==(const const_iterator &x) const {
+    return node == x.node && position == x.position;
+  }
+  bool operator!=(const const_iterator &x) const {
+    return node != x.node || position != x.position;
+  }
+
+  // Accessors for the key/value the iterator is pointing at.
+  const key_type& key() const {
+    return node->key(position);
+  }
+  reference operator*() const {
+    return node->value(position);
+  }
+  pointer operator->() const {
+    return &node->value(position);
+  }
+
+  self_type& operator++() {
+    increment();
+    return *this;
+  }
+  self_type& operator--() {
+    decrement();
+    return *this;
+  }
+  self_type operator++(int) {
+    self_type tmp = *this;
+    ++*this;
+    return tmp;
+  }
+  self_type operator--(int) {
+    self_type tmp = *this;
+    --*this;
+    return tmp;
+  }
+
+  // The node in the tree the iterator is pointing at.
+  Node *node;
+  // The position within the node of the tree the iterator is pointing at.
+  int position;
+};
+
+// Dispatch helper class for using btree::internal_locate with plain compare.
+struct btree_internal_locate_plain_compare {
+  template <typename K, typename T, typename Iter>
+  static std::pair<Iter, int> dispatch(const K &k, const T &t, Iter iter) {
+    return t.internal_locate_plain_compare(k, iter);
+  }
+};
+
+// Dispatch helper class for using btree::internal_locate with compare-to.
+struct btree_internal_locate_compare_to {
+  template <typename K, typename T, typename Iter>
+  static std::pair<Iter, int> dispatch(const K &k, const T &t, Iter iter) {
+    return t.internal_locate_compare_to(k, iter);
+  }
+};
+
+template <typename Params>
+class btree : public Params::key_compare {
+  typedef btree<Params> self_type;
+  typedef btree_node<Params> node_type;
+  typedef typename node_type::base_fields base_fields;
+  typedef typename node_type::leaf_fields leaf_fields;
+  typedef typename node_type::internal_fields internal_fields;
+  typedef typename node_type::root_fields root_fields;
+  typedef typename Params::is_key_compare_to is_key_compare_to;
+
+  friend struct btree_internal_locate_plain_compare;
+  friend struct btree_internal_locate_compare_to;
+  typedef typename if_<
+    is_key_compare_to::value,
+    btree_internal_locate_compare_to,
+    btree_internal_locate_plain_compare>::type internal_locate_type;
+
+  enum {
+    kNodeValues = node_type::kNodeValues,
+    kMinNodeValues = kNodeValues / 2,
+    kValueSize = node_type::kValueSize,
+    kExactMatch = node_type::kExactMatch,
+    kMatchMask = node_type::kMatchMask,
+  };
+
+  // A helper class to get the empty base class optimization for 0-size
+  // allocators. Base is internal_allocator_type.
+  // (e.g. empty_base_handle<internal_allocator_type, node_type*>). If Base is
+  // 0-size, the compiler doesn't have to reserve any space for it and
+  // sizeof(empty_base_handle) will simply be sizeof(Data). Google [empty base
+  // class optimization] for more details.
+  template <typename Base, typename Data>
+  struct empty_base_handle : public Base {
+    empty_base_handle(const Base &b, const Data &d)
+        : Base(b),
+          data(d) {
+    }
+    Data data;
+  };
+
+  struct node_stats {
+    node_stats(ssize_t l, ssize_t i)
+        : leaf_nodes(l),
+          internal_nodes(i) {
+    }
+
+    node_stats& operator+=(const node_stats &x) {
+      leaf_nodes += x.leaf_nodes;
+      internal_nodes += x.internal_nodes;
+      return *this;
+    }
+
+    ssize_t leaf_nodes;
+    ssize_t internal_nodes;
+  };
+
+ public:
+  typedef Params params_type;
+  typedef typename Params::key_type key_type;
+  typedef typename Params::data_type data_type;
+  typedef typename Params::mapped_type mapped_type;
+  typedef typename Params::value_type value_type;
+  typedef typename Params::key_compare key_compare;
+  typedef typename Params::pointer pointer;
+  typedef typename Params::const_pointer const_pointer;
+  typedef typename Params::reference reference;
+  typedef typename Params::const_reference const_reference;
+  typedef typename Params::size_type size_type;
+  typedef typename Params::difference_type difference_type;
+  typedef btree_iterator<node_type, reference, pointer> iterator;
+  typedef typename iterator::const_iterator const_iterator;
+  typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
+  typedef std::reverse_iterator<iterator> reverse_iterator;
+
+  typedef typename Params::allocator_type allocator_type;
+  typedef typename allocator_type::template rebind<char>::other
+    internal_allocator_type;
+
+ public:
+  // Default constructor.
+  btree(const key_compare &comp, const allocator_type &alloc);
+
+  // Copy constructor.
+  btree(const self_type &x);
+
+  // Destructor.
+  ~btree() {
+    clear();
+  }
+
+  // Iterator routines.
+  iterator begin() {
+    return iterator(leftmost(), 0);
+  }
+  const_iterator begin() const {
+    return const_iterator(leftmost(), 0);
+  }
+  iterator end() {
+    return iterator(rightmost(), rightmost() ? rightmost()->count() : 0);
+  }
+  const_iterator end() const {
+    return const_iterator(rightmost(), rightmost() ? rightmost()->count() : 0);
+  }
+  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());
+  }
+
+  // Finds the first element whose key is not less than key.
+  iterator lower_bound(const key_type &key) {
+    return internal_end(
+        internal_lower_bound(key, iterator(root(), 0)));
+  }
+  const_iterator lower_bound(const key_type &key) const {
+    return internal_end(
+        internal_lower_bound(key, const_iterator(root(), 0)));
+  }
+
+  // Finds the first element whose key is greater than key.
+  iterator upper_bound(const key_type &key) {
+    return internal_end(
+        internal_upper_bound(key, iterator(root(), 0)));
+  }
+  const_iterator upper_bound(const key_type &key) const {
+    return internal_end(
+        internal_upper_bound(key, const_iterator(root(), 0)));
+  }
+
+  // Finds the range of values which compare equal to key. The first member of
+  // the returned pair is equal to lower_bound(key). The second member pair of
+  // the pair is equal to upper_bound(key).
+  std::pair<iterator,iterator> equal_range(const key_type &key) {
+    return std::make_pair(lower_bound(key), upper_bound(key));
+  }
+  std::pair<const_iterator,const_iterator> equal_range(const key_type &key) const {
+    return std::make_pair(lower_bound(key), upper_bound(key));
+  }
+
+  // Inserts a value into the btree only if it does not already exist. The
+  // boolean return value indicates whether insertion succeeded or failed. The
+  // ValuePointer type is used to avoid instatiating the value unless the key
+  // is being inserted. Value is not dereferenced if the key already exists in
+  // the btree. See btree_map::operator[].
+  template <typename ValuePointer>
+  std::pair<iterator,bool> insert_unique(const key_type &key, ValuePointer value);
+
+  // Inserts a value into the btree only if it does not already exist. The
+  // boolean return value indicates whether insertion succeeded or failed.
+  std::pair<iterator,bool> insert_unique(const value_type &v) {
+    return insert_unique(params_type::key(v), &v);
+  }
+
+  // Insert with hint. Check to see if the value should be placed immediately
+  // before position in the tree. If it does, then the insertion will take
+  // amortized constant time. If not, the insertion will take amortized
+  // logarithmic time as if a call to insert_unique(v) were made.
+  iterator insert_unique(iterator position, const value_type &v);
+
+  // Insert a range of values into the btree.
+  template <typename InputIterator>
+  void insert_unique(InputIterator b, InputIterator e);
+
+  // Inserts a value into the btree. The ValuePointer type is used to avoid
+  // instatiating the value unless the key is being inserted. Value is not
+  // dereferenced if the key already exists in the btree. See
+  // btree_map::operator[].
+  template <typename ValuePointer>
+  iterator insert_multi(const key_type &key, ValuePointer value);
+
+  // Inserts a value into the btree.
+  iterator insert_multi(const value_type &v) {
+    return insert_multi(params_type::key(v), &v);
+  }
+
+  // Insert with hint. Check to see if the value should be placed immediately
+  // before position in the tree. If it does, then the insertion will take
+  // amortized constant time. If not, the insertion will take amortized
+  // logarithmic time as if a call to insert_multi(v) were made.
+  iterator insert_multi(iterator position, const value_type &v);
+
+  // Insert a range of values into the btree.
+  template <typename InputIterator>
+  void insert_multi(InputIterator b, InputIterator e);
+
+  void assign(const self_type &x);
+
+  // Erase the specified iterator from the btree. The iterator must be valid
+  // (i.e. not equal to end()).  Return an iterator pointing to the node after
+  // the one that was erased (or end() if none exists).
+  iterator erase(iterator iter);
+
+  // Erases range. Returns the number of keys erased.
+  int erase(iterator begin, iterator end);
+
+  // Erases the specified key from the btree. Returns 1 if an element was
+  // erased and 0 otherwise.
+  int erase_unique(const key_type &key);
+
+  // Erases all of the entries matching the specified key from the
+  // btree. Returns the number of elements erased.
+  int erase_multi(const key_type &key);
+
+  // Finds the iterator corresponding to a key or returns end() if the key is
+  // not present.
+  iterator find_unique(const key_type &key) {
+    return internal_end(
+        internal_find_unique(key, iterator(root(), 0)));
+  }
+  const_iterator find_unique(const key_type &key) const {
+    return internal_end(
+        internal_find_unique(key, const_iterator(root(), 0)));
+  }
+  iterator find_multi(const key_type &key) {
+    return internal_end(
+        internal_find_multi(key, iterator(root(), 0)));
+  }
+  const_iterator find_multi(const key_type &key) const {
+    return internal_end(
+        internal_find_multi(key, const_iterator(root(), 0)));
+  }
+
+  // Returns a count of the number of times the key appears in the btree.
+  size_type count_unique(const key_type &key) const {
+    const_iterator b = internal_find_unique(
+        key, const_iterator(root(), 0));
+    if (!b.node) {
+      // The key doesn't exist in the tree.
+      return 0;
+    }
+    return 1;
+  }
+  // Returns a count of the number of times the key appears in the btree.
+  size_type count_multi(const key_type &key) const {
+    return distance(lower_bound(key), upper_bound(key));
+  }
+
+  // Clear the btree, deleting all of the values it contains.
+  void clear();
+
+  // Swap the contents of *this and x.
+  void swap(self_type &x);
+
+  // Assign the contents of x to *this.
+  self_type& operator=(const self_type &x) {
+    if (&x == this) {
+      // Don't copy onto ourselves.
+      return *this;
+    }
+    assign(x);
+    return *this;
+  }
+
+  key_compare* mutable_key_comp() {
+    return this;
+  }
+  const key_compare& key_comp() const {
+    return *this;
+  }
+  bool compare_keys(const key_type &x, const key_type &y) const {
+    return btree_compare_keys(key_comp(), x, y);
+  }
+
+  // Dump the btree to the specified ostream. Requires that operator<< is
+  // defined for Key and Value.
+  void dump(std::ostream &os) const {
+    if (root() != NULL) {
+      internal_dump(os, root(), 0);
+    }
+  }
+
+  // Verifies the structure of the btree.
+  void verify() const;
+
+  // Size routines. Note that empty() is slightly faster than doing size()==0.
+  size_type size() const {
+    if (empty()) return 0;
+    if (root()->leaf()) return root()->count();
+    return root()->size();
+  }
+  size_type max_size() const { return std::numeric_limits<size_type>::max(); }
+  bool empty() const { return root() == NULL; }
+
+  // The height of the btree. An empty tree will have height 0.
+  size_type height() const {
+    size_type h = 0;
+    if (root()) {
+      // Count the length of the chain from the leftmost node up to the
+      // root. We actually count from the root back around to the level below
+      // the root, but the calculation is the same because of the circularity
+      // of that traversal.
+      const node_type *n = root();
+      do {
+        ++h;
+        n = n->parent();
+      } while (n != root());
+    }
+    return h;
+  }
+
+  // The number of internal, leaf and total nodes used by the btree.
+  size_type leaf_nodes() const {
+    return internal_stats(root()).leaf_nodes;
+  }
+  size_type internal_nodes() const {
+    return internal_stats(root()).internal_nodes;
+  }
+  size_type nodes() const {
+    node_stats stats = internal_stats(root());
+    return stats.leaf_nodes + stats.internal_nodes;
+  }
+
+  // The total number of bytes used by the btree.
+  size_type bytes_used() const {
+    node_stats stats = internal_stats(root());
+    if (stats.leaf_nodes == 1 && stats.internal_nodes == 0) {
+      return sizeof(*this) +
+          sizeof(base_fields) + root()->max_count() * sizeof(value_type);
+    } else {
+      return sizeof(*this) +
+          sizeof(root_fields) - sizeof(internal_fields) +
+          stats.leaf_nodes * sizeof(leaf_fields) +
+          stats.internal_nodes * sizeof(internal_fields);
+    }
+  }
+
+  // The average number of bytes used per value stored in the btree.
+  static double average_bytes_per_value() {
+    // Returns the number of bytes per value on a leaf node that is 75%
+    // full. Experimentally, this matches up nicely with the computed number of
+    // bytes per value in trees that had their values inserted in random order.
+    return sizeof(leaf_fields) / (kNodeValues * 0.75);
+  }
+
+  // The fullness of the btree. Computed as the number of elements in the btree
+  // divided by the maximum number of elements a tree with the current number
+  // of nodes could hold. A value of 1 indicates perfect space
+  // utilization. Smaller values indicate space wastage.
+  double fullness() const {
+    return double(size()) / (nodes() * kNodeValues);
+  }
+  // The overhead of the btree structure in bytes per node. Computed as the
+  // total number of bytes used by the btree minus the number of bytes used for
+  // storing elements divided by the number of elements.
+  double overhead() const {
+    if (empty()) {
+      return 0.0;
+    }
+    return (bytes_used() - size() * kValueSize) / double(size());
+  }
+
+ private:
+  // Internal accessor routines.
+  node_type* root() { return root_.data; }
+  const node_type* root() const { return root_.data; }
+  node_type** mutable_root() { return &root_.data; }
+
+  // The rightmost node is stored in the root node.
+  node_type* rightmost() {
+    return (!root() || root()->leaf()) ? root() : root()->rightmost();
+  }
+  const node_type* rightmost() const {
+    return (!root() || root()->leaf()) ? root() : root()->rightmost();
+  }
+  node_type** mutable_rightmost() { return root()->mutable_rightmost(); }
+
+  // The leftmost node is stored as the parent of the root node.
+  node_type* leftmost() { return root() ? root()->parent() : NULL; }
+  const node_type* leftmost() const { return root() ? root()->parent() : NULL; }
+
+  // The size of the tree is stored in the root node.
+  size_type* mutable_size() { return root()->mutable_size(); }
+
+  // Allocator routines.
+  internal_allocator_type* mutable_internal_allocator() {
+    return static_cast<internal_allocator_type*>(&root_);
+  }
+  const internal_allocator_type& internal_allocator() const {
+    return *static_cast<const internal_allocator_type*>(&root_);
+  }
+
+  // Node creation/deletion routines.
+  node_type* new_internal_node(node_type *parent) {
+    internal_fields *p = reinterpret_cast<internal_fields*>(
+        mutable_internal_allocator()->allocate(sizeof(internal_fields)));
+    return node_type::init_internal(p, parent);
+  }
+  node_type* new_internal_root_node() {
+    root_fields *p = reinterpret_cast<root_fields*>(
+        mutable_internal_allocator()->allocate(sizeof(root_fields)));
+    return node_type::init_root(p, root()->parent());
+  }
+  node_type* new_leaf_node(node_type *parent) {
+    leaf_fields *p = reinterpret_cast<leaf_fields*>(
+        mutable_internal_allocator()->allocate(sizeof(leaf_fields)));
+    return node_type::init_leaf(p, parent, kNodeValues);
+  }
+  node_type* new_leaf_root_node(int max_count) {
+    leaf_fields *p = reinterpret_cast<leaf_fields*>(
+        mutable_internal_allocator()->allocate(
+            sizeof(base_fields) + max_count * sizeof(value_type)));
+    return node_type::init_leaf(p, reinterpret_cast<node_type*>(p), max_count);
+  }
+  void delete_internal_node(node_type *node) {
+    node->destroy();
+    assert(node != root());
+    mutable_internal_allocator()->deallocate(
+        reinterpret_cast<char*>(node), sizeof(internal_fields));
+  }
+  void delete_internal_root_node() {
+    root()->destroy();
+    mutable_internal_allocator()->deallocate(
+        reinterpret_cast<char*>(root()), sizeof(root_fields));
+  }
+  void delete_leaf_node(node_type *node) {
+    node->destroy();
+    mutable_internal_allocator()->deallocate(
+        reinterpret_cast<char*>(node),
+        sizeof(base_fields) + node->max_count() * sizeof(value_type));
+  }
+
+  // Rebalances or splits the node iter points to.
+  void rebalance_or_split(iterator *iter);
+
+  // Merges the values of left, right and the delimiting key on their parent
+  // onto left, removing the delimiting key and deleting right.
+  void merge_nodes(node_type *left, node_type *right);
+
+  // Tries to merge node with its left or right sibling, and failing that,
+  // rebalance with its left or right sibling. Returns true if a merge
+  // occurred, at which point it is no longer valid to access node. Returns
+  // false if no merging took place.
+  bool try_merge_or_rebalance(iterator *iter);
+
+  // Tries to shrink the height of the tree by 1.
+  void try_shrink();
+
+  iterator internal_end(iterator iter) {
+    return iter.node ? iter : end();
+  }
+  const_iterator internal_end(const_iterator iter) const {
+    return iter.node ? iter : end();
+  }
+
+  // Inserts a value into the btree immediately before iter. Requires that
+  // key(v) <= iter.key() and (--iter).key() <= key(v).
+  iterator internal_insert(iterator iter, const value_type &v);
+
+  // Returns an iterator pointing to the first value >= the value "iter" is
+  // pointing at. Note that "iter" might be pointing to an invalid location as
+  // iter.position == iter.node->count(). This routine simply moves iter up in
+  // the tree to a valid location.
+  template <typename IterType>
+  static IterType internal_last(IterType iter);
+
+  // Returns an iterator pointing to the leaf position at which key would
+  // reside in the tree. We provide 2 versions of internal_locate. The first
+  // version (internal_locate_plain_compare) always returns 0 for the second
+  // field of the pair. The second version (internal_locate_compare_to) is for
+  // the key-compare-to specialization and returns either kExactMatch (if the
+  // key was found in the tree) or -kExactMatch (if it wasn't) in the second
+  // field of the pair. The compare_to specialization allows the caller to
+  // avoid a subsequent comparison to determine if an exact match was made,
+  // speeding up string keys.
+  template <typename IterType>
+  std::pair<IterType, int> internal_locate(
+      const key_type &key, IterType iter) const;
+  template <typename IterType>
+  std::pair<IterType, int> internal_locate_plain_compare(
+      const key_type &key, IterType iter) const;
+  template <typename IterType>
+  std::pair<IterType, int> internal_locate_compare_to(
+      const key_type &key, IterType iter) const;
+
+  // Internal routine which implements lower_bound().
+  template <typename IterType>
+  IterType internal_lower_bound(
+      const key_type &key, IterType iter) const;
+
+  // Internal routine which implements upper_bound().
+  template <typename IterType>
+  IterType internal_upper_bound(
+      const key_type &key, IterType iter) const;
+
+  // Internal routine which implements find_unique().
+  template <typename IterType>
+  IterType internal_find_unique(
+      const key_type &key, IterType iter) const;
+
+  // Internal routine which implements find_multi().
+  template <typename IterType>
+  IterType internal_find_multi(
+      const key_type &key, IterType iter) const;
+
+  // Deletes a node and all of its children.
+  void internal_clear(node_type *node);
+
+  // Dumps a node and all of its children to the specified ostream.
+  void internal_dump(std::ostream &os, const node_type *node, int level) const;
+
+  // Verifies the tree structure of node.
+  int internal_verify(const node_type *node,
+                      const key_type *lo, const key_type *hi) const;
+
+  node_stats internal_stats(const node_type *node) const {
+    if (!node) {
+      return node_stats(0, 0);
+    }
+    if (node->leaf()) {
+      return node_stats(1, 0);
+    }
+    node_stats res(0, 1);
+    for (int i = 0; i <= node->count(); ++i) {
+      res += internal_stats(node->child(i));
+    }
+    return res;
+  }
+
+ private:
+  empty_base_handle<internal_allocator_type, node_type*> root_;
+
+ private:
+  // A never instantiated helper function that returns big_ if we have a
+  // key-compare-to functor or if R is bool and small_ otherwise.
+  template <typename R>
+  static typename if_<
+   if_<is_key_compare_to::value,
+             std::is_same<R, int>,
+             std::is_same<R, bool> >::type::value,
+   big_, small_>::type key_compare_checker(R);
+
+  // A never instantiated helper function that returns the key comparison
+  // functor.
+  static key_compare key_compare_helper();
+
+  // Verify that key_compare returns a bool. This is similar to the way
+  // is_convertible in base/type_traits.h works. Note that key_compare_checker
+  // is never actually invoked. The compiler will select which
+  // key_compare_checker() to instantiate and then figure out the size of the
+  // return type of key_compare_checker() at compile time which we then check
+  // against the sizeof of big_.
+  COMPILE_ASSERT(
+      sizeof(key_compare_checker(key_compare_helper()(key_type(), key_type()))) ==
+      sizeof(big_),
+      key_comparison_function_must_return_bool);
+
+  // Note: We insist on kTargetValues, which is computed from
+  // Params::kTargetNodeSize, must fit the base_fields::field_type.
+  COMPILE_ASSERT(kNodeValues <
+                 (1 << (8 * sizeof(typename base_fields::field_type))),
+                 target_node_size_too_large);
+
+  // Test the assumption made in setting kNodeValueSpace.
+  COMPILE_ASSERT(sizeof(base_fields) >= 2 * sizeof(void*),
+                 node_space_assumption_incorrect);
+};
+
+////
+// btree_node methods
+template <typename P>
+inline void btree_node<P>::insert_value(int i, const value_type &x) {
+  assert(i <= count());
+  value_init(count(), x);
+  for (int j = count(); j > i; --j) {
+    value_swap(j, this, j - 1);
+  }
+  set_count(count() + 1);
+
+  if (!leaf()) {
+    ++i;
+    for (int j = count(); j > i; --j) {
+      *mutable_child(j) = child(j - 1);
+      child(j)->set_position(j);
+    }
+    *mutable_child(i) = NULL;
+  }
+}
+
+template <typename P>
+inline void btree_node<P>::remove_value(int i) {
+  if (!leaf()) {
+    assert(child(i + 1)->count() == 0);
+    for (int j = i + 1; j < count(); ++j) {
+      *mutable_child(j) = child(j + 1);
+      child(j)->set_position(j);
+    }
+    *mutable_child(count()) = NULL;
+  }
+
+  set_count(count() - 1);
+  for (; i < count(); ++i) {
+    value_swap(i, this, i + 1);
+  }
+  value_destroy(i);
+}
+
+template <typename P>
+void btree_node<P>::rebalance_right_to_left(btree_node *src, int to_move) {
+  assert(parent() == src->parent());
+  assert(position() + 1 == src->position());
+  assert(src->count() >= count());
+  assert(to_move >= 1);
+  assert(to_move <= src->count());
+
+  // Make room in the left node for the new values.
+  for (int i = 0; i < to_move; ++i) {
+    value_init(i + count());
+  }
+
+  // Move the delimiting value to the left node and the new delimiting value
+  // from the right node.
+  value_swap(count(), parent(), position());
+  parent()->value_swap(position(), src, to_move - 1);
+
+  // Move the values from the right to the left node.
+  for (int i = 1; i < to_move; ++i) {
+    value_swap(count() + i, src, i - 1);
+  }
+  // Shift the values in the right node to their correct position.
+  for (int i = to_move; i < src->count(); ++i) {
+    src->value_swap(i - to_move, src, i);
+  }
+  for (int i = 1; i <= to_move; ++i) {
+    src->value_destroy(src->count() - i);
+  }
+
+  if (!leaf()) {
+    // Move the child pointers from the right to the left node.
+    for (int i = 0; i < to_move; ++i) {
+      set_child(1 + count() + i, src->child(i));
+    }
+    for (int i = 0; i <= src->count() - to_move; ++i) {
+      assert(i + to_move <= src->max_count());
+      src->set_child(i, src->child(i + to_move));
+      *src->mutable_child(i + to_move) = NULL;
+    }
+  }
+
+  // Fixup the counts on the src and dest nodes.
+  set_count(count() + to_move);
+  src->set_count(src->count() - to_move);
+}
+
+template <typename P>
+void btree_node<P>::rebalance_left_to_right(btree_node *dest, int to_move) {
+  assert(parent() == dest->parent());
+  assert(position() + 1 == dest->position());
+  assert(count() >= dest->count());
+  assert(to_move >= 1);
+  assert(to_move <= count());
+
+  // Make room in the right node for the new values.
+  for (int i = 0; i < to_move; ++i) {
+    dest->value_init(i + dest->count());
+  }
+  for (int i = dest->count() - 1; i >= 0; --i) {
+    dest->value_swap(i, dest, i + to_move);
+  }
+
+  // Move the delimiting value to the right node and the new delimiting value
+  // from the left node.
+  dest->value_swap(to_move - 1, parent(), position());
+  parent()->value_swap(position(), this, count() - to_move);
+  value_destroy(count() - to_move);
+
+  // Move the values from the left to the right node.
+  for (int i = 1; i < to_move; ++i) {
+    value_swap(count() - to_move + i, dest, i - 1);
+    value_destroy(count() - to_move + i);
+  }
+
+  if (!leaf()) {
+    // Move the child pointers from the left to the right node.
+    for (int i = dest->count(); i >= 0; --i) {
+      dest->set_child(i + to_move, dest->child(i));
+      *dest->mutable_child(i) = NULL;
+    }
+    for (int i = 1; i <= to_move; ++i) {
+      dest->set_child(i - 1, child(count() - to_move + i));
+      *mutable_child(count() - to_move + i) = NULL;
+    }
+  }
+
+  // Fixup the counts on the src and dest nodes.
+  set_count(count() - to_move);
+  dest->set_count(dest->count() + to_move);
+}
+
+template <typename P>
+void btree_node<P>::split(btree_node *dest, int insert_position) {
+  assert(dest->count() == 0);
+
+  // We bias the split based on the position being inserted. If we're
+  // inserting at the beginning of the left node then bias the split to put
+  // more values on the right node. If we're inserting at the end of the
+  // right node then bias the split to put more values on the left node.
+  if (insert_position == 0) {
+    dest->set_count(count() - 1);
+  } else if (insert_position == max_count()) {
+    dest->set_count(0);
+  } else {
+    dest->set_count(count() / 2);
+  }
+  set_count(count() - dest->count());
+  assert(count() >= 1);
+
+  // Move values from the left sibling to the right sibling.
+  for (int i = 0; i < dest->count(); ++i) {
+    dest->value_init(i);
+    value_swap(count() + i, dest, i);
+    value_destroy(count() + i);
+  }
+
+  // The split key is the largest value in the left sibling.
+  set_count(count() - 1);
+  parent()->insert_value(position(), value_type());
+  value_swap(count(), parent(), position());
+  value_destroy(count());
+  parent()->set_child(position() + 1, dest);
+
+  if (!leaf()) {
+    for (int i = 0; i <= dest->count(); ++i) {
+      assert(child(count() + i + 1) != NULL);
+      dest->set_child(i, child(count() + i + 1));
+      *mutable_child(count() + i + 1) = NULL;
+    }
+  }
+}
+
+template <typename P>
+void btree_node<P>::merge(btree_node *src) {
+  assert(parent() == src->parent());
+  assert(position() + 1 == src->position());
+
+  // Move the delimiting value to the left node.
+  value_init(count());
+  value_swap(count(), parent(), position());
+
+  // Move the values from the right to the left node.
+  for (int i = 0; i < src->count(); ++i) {
+    value_init(1 + count() + i);
+    value_swap(1 + count() + i, src, i);
+    src->value_destroy(i);
+  }
+
+  if (!leaf()) {
+    // Move the child pointers from the right to the left node.
+    for (int i = 0; i <= src->count(); ++i) {
+      set_child(1 + count() + i, src->child(i));
+      *src->mutable_child(i) = NULL;
+    }
+  }
+
+  // Fixup the counts on the src and dest nodes.
+  set_count(1 + count() + src->count());
+  src->set_count(0);
+
+  // Remove the value on the parent node.
+  parent()->remove_value(position());
+}
+
+template <typename P>
+void btree_node<P>::swap(btree_node *x) {
+  assert(leaf() == x->leaf());
+
+  // Swap the values.
+  for (int i = count(); i < x->count(); ++i) {
+    value_init(i);
+  }
+  for (int i = x->count(); i < count(); ++i) {
+    x->value_init(i);
+  }
+  int n = std::max(count(), x->count());
+  for (int i = 0; i < n; ++i) {
+    value_swap(i, x, i);
+  }
+  for (int i = count(); i < x->count(); ++i) {
+    x->value_destroy(i);
+  }
+  for (int i = x->count(); i < count(); ++i) {
+    value_destroy(i);
+  }
+
+  if (!leaf()) {
+    // Swap the child pointers.
+    for (int i = 0; i <= n; ++i) {
+      btree_swap_helper(*mutable_child(i), *x->mutable_child(i));
+    }
+    for (int i = 0; i <= count(); ++i) {
+      x->child(i)->fields_.parent = x;
+    }
+    for (int i = 0; i <= x->count(); ++i) {
+      child(i)->fields_.parent = this;
+    }
+  }
+
+  // Swap the counts.
+  btree_swap_helper(fields_.count, x->fields_.count);
+}
+
+////
+// btree_iterator methods
+template <typename N, typename R, typename P>
+void btree_iterator<N, R, P>::increment_slow() {
+  if (node->leaf()) {
+    assert(position >= node->count());
+    self_type save(*this);
+    while (position == node->count() && !node->is_root()) {
+      assert(node->parent()->child(node->position()) == node);
+      position = node->position();
+      node = node->parent();
+    }
+    if (position == node->count()) {
+      *this = save;
+    }
+  } else {
+    assert(position < node->count());
+    node = node->child(position + 1);
+    while (!node->leaf()) {
+      node = node->child(0);
+    }
+    position = 0;
+  }
+}
+
+template <typename N, typename R, typename P>
+void btree_iterator<N, R, P>::increment_by(int count) {
+  while (count > 0) {
+    if (node->leaf()) {
+      int rest = node->count() - position;
+      position += std::min(rest, count);
+      count = count - rest;
+      if (position < node->count()) {
+        return;
+      }
+    } else {
+      --count;
+    }
+    increment_slow();
+  }
+}
+
+template <typename N, typename R, typename P>
+void btree_iterator<N, R, P>::decrement_slow() {
+  if (node->leaf()) {
+    assert(position <= -1);
+    self_type save(*this);
+    while (position < 0 && !node->is_root()) {
+      assert(node->parent()->child(node->position()) == node);
+      position = node->position() - 1;
+      node = node->parent();
+    }
+    if (position < 0) {
+      *this = save;
+    }
+  } else {
+    assert(position >= 0);
+    node = node->child(position);
+    while (!node->leaf()) {
+      node = node->child(node->count());
+    }
+    position = node->count() - 1;
+  }
+}
+
+////
+// btree methods
+template <typename P>
+btree<P>::btree(const key_compare &comp, const allocator_type &alloc)
+    : key_compare(comp),
+      root_(alloc, NULL) {
+}
+
+template <typename P>
+btree<P>::btree(const self_type &x)
+    : key_compare(x.key_comp()),
+      root_(x.internal_allocator(), NULL) {
+  assign(x);
+}
+
+template <typename P> template <typename ValuePointer>
+std::pair<typename btree<P>::iterator, bool>
+btree<P>::insert_unique(const key_type &key, ValuePointer value) {
+  if (empty()) {
+    *mutable_root() = new_leaf_root_node(1);
+  }
+
+  std::pair<iterator, int> res = internal_locate(key, iterator(root(), 0));
+  iterator &iter = res.first;
+  if (res.second == kExactMatch) {
+    // The key already exists in the tree, do nothing.
+    return std::make_pair(internal_last(iter), false);
+  } else if (!res.second) {
+    iterator last = internal_last(iter);
+    if (last.node && !compare_keys(key, last.key())) {
+      // The key already exists in the tree, do nothing.
+      return std::make_pair(last, false);
+    }
+  }
+
+  return std::make_pair(internal_insert(iter, *value), true);
+}
+
+template <typename P>
+inline typename btree<P>::iterator
+btree<P>::insert_unique(iterator position, const value_type &v) {
+  if (!empty()) {
+    const key_type &key = params_type::key(v);
+    if (position == end() || compare_keys(key, position.key())) {
+      iterator prev = position;
+      if (position == begin() || compare_keys((--prev).key(), key)) {
+        // prev.key() < key < position.key()
+        return internal_insert(position, v);
+      }
+    } else if (compare_keys(position.key(), key)) {
+      iterator next = position;
+      ++next;
+      if (next == end() || compare_keys(key, next.key())) {
+        // position.key() < key < next.key()
+        return internal_insert(next, v);
+      }
+    } else {
+      // position.key() == key
+      return position;
+    }
+  }
+  return insert_unique(v).first;
+}
+
+template <typename P> template <typename InputIterator>
+void btree<P>::insert_unique(InputIterator b, InputIterator e) {
+  for (; b != e; ++b) {
+    insert_unique(end(), *b);
+  }
+}
+
+template <typename P> template <typename ValuePointer>
+typename btree<P>::iterator
+btree<P>::insert_multi(const key_type &key, ValuePointer value) {
+  if (empty()) {
+    *mutable_root() = new_leaf_root_node(1);
+  }
+
+  iterator iter = internal_upper_bound(key, iterator(root(), 0));
+  if (!iter.node) {
+    iter = end();
+  }
+  return internal_insert(iter, *value);
+}
+
+template <typename P>
+typename btree<P>::iterator
+btree<P>::insert_multi(iterator position, const value_type &v) {
+  if (!empty()) {
+    const key_type &key = params_type::key(v);
+    if (position == end() || !compare_keys(position.key(), key)) {
+      iterator prev = position;
+      if (position == begin() || !compare_keys(key, (--prev).key())) {
+        // prev.key() <= key <= position.key()
+        return internal_insert(position, v);
+      }
+    } else {
+      iterator next = position;
+      ++next;
+      if (next == end() || !compare_keys(next.key(), key)) {
+        // position.key() < key <= next.key()
+        return internal_insert(next, v);
+      }
+    }
+  }
+  return insert_multi(v);
+}
+
+template <typename P> template <typename InputIterator>
+void btree<P>::insert_multi(InputIterator b, InputIterator e) {
+  for (; b != e; ++b) {
+    insert_multi(end(), *b);
+  }
+}
+
+template <typename P>
+void btree<P>::assign(const self_type &x) {
+  clear();
+
+  *mutable_key_comp() = x.key_comp();
+  *mutable_internal_allocator() = x.internal_allocator();
+
+  // Assignment can avoid key comparisons because we know the order of the
+  // values is the same order we'll store them in.
+  for (const_iterator iter = x.begin(); iter != x.end(); ++iter) {
+    if (empty()) {
+      insert_multi(*iter);
+    } else {
+      // If the btree is not empty, we can just insert the new value at the end
+      // of the tree!
+      internal_insert(end(), *iter);
+    }
+  }
+}
+
+template <typename P>
+typename btree<P>::iterator btree<P>::erase(iterator iter) {
+  bool internal_delete = false;
+  if (!iter.node->leaf()) {
+    // Deletion of a value on an internal node. Swap the key with the largest
+    // value of our left child. This is easy, we just decrement iter.
+    iterator tmp_iter(iter--);
+    assert(iter.node->leaf());
+    assert(!compare_keys(tmp_iter.key(), iter.key()));
+    iter.node->value_swap(iter.position, tmp_iter.node, tmp_iter.position);
+    internal_delete = true;
+    --*mutable_size();
+  } else if (!root()->leaf()) {
+    --*mutable_size();
+  }
+
+  // Delete the key from the leaf.
+  iter.node->remove_value(iter.position);
+
+  // We want to return the next value after the one we just erased. If we
+  // erased from an internal node (internal_delete == true), then the next
+  // value is ++(++iter). If we erased from a leaf node (internal_delete ==
+  // false) then the next value is ++iter. Note that ++iter may point to an
+  // internal node and the value in the internal node may move to a leaf node
+  // (iter.node) when rebalancing is performed at the leaf level.
+
+  // Merge/rebalance as we walk back up the tree.
+  iterator res(iter);
+  for (;;) {
+    if (iter.node == root()) {
+      try_shrink();
+      if (empty()) {
+        return end();
+      }
+      break;
+    }
+    if (iter.node->count() >= kMinNodeValues) {
+      break;
+    }
+    bool merged = try_merge_or_rebalance(&iter);
+    if (iter.node->leaf()) {
+      res = iter;
+    }
+    if (!merged) {
+      break;
+    }
+    iter.node = iter.node->parent();
+  }
+
+  // Adjust our return value. If we're pointing at the end of a node, advance
+  // the iterator.
+  if (res.position == res.node->count()) {
+    res.position = res.node->count() - 1;
+    ++res;
+  }
+  // If we erased from an internal node, advance the iterator.
+  if (internal_delete) {
+    ++res;
+  }
+  return res;
+}
+
+template <typename P>
+int btree<P>::erase(iterator b, iterator e) {
+  int count = distance(b, e);
+  for (int i = 0; i < count; i++) {
+    b = erase(b);
+  }
+  return count;
+}
+
+template <typename P>
+int btree<P>::erase_unique(const key_type &key) {
+  iterator iter = internal_find_unique(key, iterator(root(), 0));
+  if (!iter.node) {
+    // The key doesn't exist in the tree, return nothing done.
+    return 0;
+  }
+  erase(iter);
+  return 1;
+}
+
+template <typename P>
+int btree<P>::erase_multi(const key_type &key) {
+  iterator b = internal_lower_bound(key, iterator(root(), 0));
+  if (!b.node) {
+    // The key doesn't exist in the tree, return nothing done.
+    return 0;
+  }
+  // Delete all of the keys between begin and upper_bound(key).
+  iterator e = internal_end(
+      internal_upper_bound(key, iterator(root(), 0)));
+  return erase(b, e);
+}
+
+template <typename P>
+void btree<P>::clear() {
+  if (root() != NULL) {
+    internal_clear(root());
+  }
+  *mutable_root() = NULL;
+}
+
+template <typename P>
+void btree<P>::swap(self_type &x) {
+  std::swap(static_cast<key_compare&>(*this), static_cast<key_compare&>(x));
+  std::swap(root_, x.root_);
+}
+
+template <typename P>
+void btree<P>::verify() const {
+  if (root() != NULL) {
+    assert(size() == internal_verify(root(), NULL, NULL));
+    assert(leftmost() == (++const_iterator(root(), -1)).node);
+    assert(rightmost() == (--const_iterator(root(), root()->count())).node);
+    assert(leftmost()->leaf());
+    assert(rightmost()->leaf());
+  } else {
+    assert(size() == 0);
+    assert(leftmost() == NULL);
+    assert(rightmost() == NULL);
+  }
+}
+
+template <typename P>
+void btree<P>::rebalance_or_split(iterator *iter) {
+  node_type *&node = iter->node;
+  int &insert_position = iter->position;
+  assert(node->count() == node->max_count());
+
+  // First try to make room on the node by rebalancing.
+  node_type *parent = node->parent();
+  if (node != root()) {
+    if (node->position() > 0) {
+      // Try rebalancing with our left sibling.
+      node_type *left = parent->child(node->position() - 1);
+      if (left->count() < left->max_count()) {
+        // We bias rebalancing based on the position being inserted. If we're
+        // inserting at the end of the right node then we bias rebalancing to
+        // fill up the left node.
+        int to_move = (left->max_count() - left->count()) /
+            (1 + (insert_position < left->max_count()));
+        to_move = std::max(1, to_move);
+
+        if (((insert_position - to_move) >= 0) ||
+            ((left->count() + to_move) < left->max_count())) {
+          left->rebalance_right_to_left(node, to_move);
+
+          assert(node->max_count() - node->count() == to_move);
+          insert_position = insert_position - to_move;
+          if (insert_position < 0) {
+            insert_position = insert_position + left->count() + 1;
+            node = left;
+          }
+
+          assert(node->count() < node->max_count());
+          return;
+        }
+      }
+    }
+
+    if (node->position() < parent->count()) {
+      // Try rebalancing with our right sibling.
+      node_type *right = parent->child(node->position() + 1);
+      if (right->count() < right->max_count()) {
+        // We bias rebalancing based on the position being inserted. If we're
+        // inserting at the beginning of the left node then we bias rebalancing
+        // to fill up the right node.
+        int to_move = (right->max_count() - right->count()) /
+            (1 + (insert_position > 0));
+        to_move = std::max(1, to_move);
+
+        if ((insert_position <= (node->count() - to_move)) ||
+            ((right->count() + to_move) < right->max_count())) {
+          node->rebalance_left_to_right(right, to_move);
+
+          if (insert_position > node->count()) {
+            insert_position = insert_position - node->count() - 1;
+            node = right;
+          }
+
+          assert(node->count() < node->max_count());
+          return;
+        }
+      }
+    }
+
+    // Rebalancing failed, make sure there is room on the parent node for a new
+    // value.
+    if (parent->count() == parent->max_count()) {
+      iterator parent_iter(node->parent(), node->position());
+      rebalance_or_split(&parent_iter);
+    }
+  } else {
+    // Rebalancing not possible because this is the root node.
+    if (root()->leaf()) {
+      // The root node is currently a leaf node: create a new root node and set
+      // the current root node as the child of the new root.
+      parent = new_internal_root_node();
+      parent->set_child(0, root());
+      *mutable_root() = parent;
+      assert(*mutable_rightmost() == parent->child(0));
+    } else {
+      // The root node is an internal node. We do not want to create a new root
+      // node because the root node is special and holds the size of the tree
+      // and a pointer to the rightmost node. So we create a new internal node
+      // and move all of the items on the current root into the new node.
+      parent = new_internal_node(parent);
+      parent->set_child(0, parent);
+      parent->swap(root());
+      node = parent;
+    }
+  }
+
+  // Split the node.
+  node_type *split_node;
+  if (node->leaf()) {
+    split_node = new_leaf_node(parent);
+    node->split(split_node, insert_position);
+    if (rightmost() == node) {
+      *mutable_rightmost() = split_node;
+    }
+  } else {
+    split_node = new_internal_node(parent);
+    node->split(split_node, insert_position);
+  }
+
+  if (insert_position > node->count()) {
+    insert_position = insert_position - node->count() - 1;
+    node = split_node;
+  }
+}
+
+template <typename P>
+void btree<P>::merge_nodes(node_type *left, node_type *right) {
+  left->merge(right);
+  if (right->leaf()) {
+    if (rightmost() == right) {
+      *mutable_rightmost() = left;
+    }
+    delete_leaf_node(right);
+  } else {
+    delete_internal_node(right);
+  }
+}
+
+template <typename P>
+bool btree<P>::try_merge_or_rebalance(iterator *iter) {
+  node_type *parent = iter->node->parent();
+  if (iter->node->position() > 0) {
+    // Try merging with our left sibling.
+    node_type *left = parent->child(iter->node->position() - 1);
+    if ((1 + left->count() + iter->node->count()) <= left->max_count()) {
+      iter->position += 1 + left->count();
+      merge_nodes(left, iter->node);
+      iter->node = left;
+      return true;
+    }
+  }
+  if (iter->node->position() < parent->count()) {
+    // Try merging with our right sibling.
+    node_type *right = parent->child(iter->node->position() + 1);
+    if ((1 + iter->node->count() + right->count()) <= right->max_count()) {
+      merge_nodes(iter->node, right);
+      return true;
+    }
+    // Try rebalancing with our right sibling. We don't perform rebalancing if
+    // we deleted the first element from iter->node and the node is not
+    // empty. This is a small optimization for the common pattern of deleting
+    // from the front of the tree.
+    if ((right->count() > kMinNodeValues) &&
+        ((iter->node->count() == 0) ||
+         (iter->position > 0))) {
+      int to_move = (right->count() - iter->node->count()) / 2;
+      to_move = std::min(to_move, right->count() - 1);
+      iter->node->rebalance_right_to_left(right, to_move);
+      return false;
+    }
+  }
+  if (iter->node->position() > 0) {
+    // Try rebalancing with our left sibling. We don't perform rebalancing if
+    // we deleted the last element from iter->node and the node is not
+    // empty. This is a small optimization for the common pattern of deleting
+    // from the back of the tree.
+    node_type *left = parent->child(iter->node->position() - 1);
+    if ((left->count() > kMinNodeValues) &&
+        ((iter->node->count() == 0) ||
+         (iter->position < iter->node->count()))) {
+      int to_move = (left->count() - iter->node->count()) / 2;
+      to_move = std::min(to_move, left->count() - 1);
+      left->rebalance_left_to_right(iter->node, to_move);
+      iter->position += to_move;
+      return false;
+    }
+  }
+  return false;
+}
+
+template <typename P>
+void btree<P>::try_shrink() {
+  if (root()->count() > 0) {
+    return;
+  }
+  // Deleted the last item on the root node, shrink the height of the tree.
+  if (root()->leaf()) {
+    assert(size() == 0);
+    delete_leaf_node(root());
+    *mutable_root() = NULL;
+  } else {
+    node_type *child = root()->child(0);
+    if (child->leaf()) {
+      // The child is a leaf node so simply make it the root node in the tree.
+      child->make_root();
+      delete_internal_root_node();
+      *mutable_root() = child;
+    } else {
+      // The child is an internal node. We want to keep the existing root node
+      // so we move all of the values from the child node into the existing
+      // (empty) root node.
+      child->swap(root());
+      delete_internal_node(child);
+    }
+  }
+}
+
+template <typename P> template <typename IterType>
+inline IterType btree<P>::internal_last(IterType iter) {
+  while (iter.node && iter.position == iter.node->count()) {
+    iter.position = iter.node->position();
+    iter.node = iter.node->parent();
+    if (iter.node->leaf()) {
+      iter.node = NULL;
+    }
+  }
+  return iter;
+}
+
+template <typename P>
+inline typename btree<P>::iterator
+btree<P>::internal_insert(iterator iter, const value_type &v) {
+  if (!iter.node->leaf()) {
+    // We can't insert on an internal node. Instead, we'll insert after the
+    // previous value which is guaranteed to be on a leaf node.
+    --iter;
+    ++iter.position;
+  }
+  if (iter.node->count() == iter.node->max_count()) {
+    // Make room in the leaf for the new item.
+    if (iter.node->max_count() < kNodeValues) {
+      // Insertion into the root where the root is smaller that the full node
+      // size. Simply grow the size of the root node.
+      assert(iter.node == root());
+      iter.node = new_leaf_root_node(
+          std::min<int>(kNodeValues, 2 * iter.node->max_count()));
+      iter.node->swap(root());
+      delete_leaf_node(root());
+      *mutable_root() = iter.node;
+    } else {
+      rebalance_or_split(&iter);
+      ++*mutable_size();
+    }
+  } else if (!root()->leaf()) {
+    ++*mutable_size();
+  }
+  iter.node->insert_value(iter.position, v);
+  return iter;
+}
+
+template <typename P> template <typename IterType>
+inline std::pair<IterType, int> btree<P>::internal_locate(
+    const key_type &key, IterType iter) const {
+  return internal_locate_type::dispatch(key, *this, iter);
+}
+
+template <typename P> template <typename IterType>
+inline std::pair<IterType, int> btree<P>::internal_locate_plain_compare(
+    const key_type &key, IterType iter) const {
+  for (;;) {
+    iter.position = iter.node->lower_bound(key, key_comp());
+    if (iter.node->leaf()) {
+      break;
+    }
+    iter.node = iter.node->child(iter.position);
+  }
+  return std::make_pair(iter, 0);
+}
+
+template <typename P> template <typename IterType>
+inline std::pair<IterType, int> btree<P>::internal_locate_compare_to(
+    const key_type &key, IterType iter) const {
+  for (;;) {
+    int res = iter.node->lower_bound(key, key_comp());
+    iter.position = res & kMatchMask;
+    if (res & kExactMatch) {
+      return std::make_pair(iter, static_cast<int>(kExactMatch));
+    }
+    if (iter.node->leaf()) {
+      break;
+    }
+    iter.node = iter.node->child(iter.position);
+  }
+  return std::make_pair(iter, -kExactMatch);
+}
+
+template <typename P> template <typename IterType>
+IterType btree<P>::internal_lower_bound(
+    const key_type &key, IterType iter) const {
+  if (iter.node) {
+    for (;;) {
+      iter.position =
+          iter.node->lower_bound(key, key_comp()) & kMatchMask;
+      if (iter.node->leaf()) {
+        break;
+      }
+      iter.node = iter.node->child(iter.position);
+    }
+    iter = internal_last(iter);
+  }
+  return iter;
+}
+
+template <typename P> template <typename IterType>
+IterType btree<P>::internal_upper_bound(
+    const key_type &key, IterType iter) const {
+  if (iter.node) {
+    for (;;) {
+      iter.position = iter.node->upper_bound(key, key_comp());
+      if (iter.node->leaf()) {
+        break;
+      }
+      iter.node = iter.node->child(iter.position);
+    }
+    iter = internal_last(iter);
+  }
+  return iter;
+}
+
+template <typename P> template <typename IterType>
+IterType btree<P>::internal_find_unique(
+    const key_type &key, IterType iter) const {
+  if (iter.node) {
+    std::pair<IterType, int> res = internal_locate(key, iter);
+    if (res.second == kExactMatch) {
+      return res.first;
+    }
+    if (!res.second) {
+      iter = internal_last(res.first);
+      if (iter.node && !compare_keys(key, iter.key())) {
+        return iter;
+      }
+    }
+  }
+  return IterType(NULL, 0);
+}
+
+template <typename P> template <typename IterType>
+IterType btree<P>::internal_find_multi(
+    const key_type &key, IterType iter) const {
+  if (iter.node) {
+    iter = internal_lower_bound(key, iter);
+    if (iter.node) {
+      iter = internal_last(iter);
+      if (iter.node && !compare_keys(key, iter.key())) {
+        return iter;
+      }
+    }
+  }
+  return IterType(NULL, 0);
+}
+
+template <typename P>
+void btree<P>::internal_clear(node_type *node) {
+  if (!node->leaf()) {
+    for (int i = 0; i <= node->count(); ++i) {
+      internal_clear(node->child(i));
+    }
+    if (node == root()) {
+      delete_internal_root_node();
+    } else {
+      delete_internal_node(node);
+    }
+  } else {
+    delete_leaf_node(node);
+  }
+}
+
+template <typename P>
+void btree<P>::internal_dump(
+    std::ostream &os, const node_type *node, int level) const {
+  for (int i = 0; i < node->count(); ++i) {
+    if (!node->leaf()) {
+      internal_dump(os, node->child(i), level + 1);
+    }
+    for (int j = 0; j < level; ++j) {
+      os << "  ";
+    }
+    os << node->key(i) << " [" << level << "]\n";
+  }
+  if (!node->leaf()) {
+    internal_dump(os, node->child(node->count()), level + 1);
+  }
+}
+
+template <typename P>
+int btree<P>::internal_verify(
+    const node_type *node, const key_type *lo, const key_type *hi) const {
+  assert(node->count() > 0);
+  assert(node->count() <= node->max_count());
+  if (lo) {
+    assert(!compare_keys(node->key(0), *lo));
+  }
+  if (hi) {
+    assert(!compare_keys(*hi, node->key(node->count() - 1)));
+  }
+  for (int i = 1; i < node->count(); ++i) {
+    assert(!compare_keys(node->key(i), node->key(i - 1)));
+  }
+  int count = node->count();
+  if (!node->leaf()) {
+    for (int i = 0; i <= node->count(); ++i) {
+      assert(node->child(i) != NULL);
+      assert(node->child(i)->parent() == node);
+      assert(node->child(i)->position() == i);
+      count += internal_verify(
+          node->child(i),
+          (i == 0) ? lo : &node->key(i - 1),
+          (i == node->count()) ? hi : &node->key(i));
+    }
+  }
+  return count;
+}
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_BTREE_H__
diff --git a/xdelta3/cpp-btree/btree_bench.cc b/xdelta3/cpp-btree/btree_bench.cc
new file mode 100644
index 0000000..6eaed99
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_bench.cc
@@ -0,0 +1,593 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <algorithm>
+#include <functional>
+#include <map>
+#include <set>
+#include <string>
+#include <sys/time.h>
+#include <type_traits>
+#include <vector>
+
+#include "gflags/gflags.h"
+#include "btree_map.h"
+#include "btree_set.h"
+#include "btree_test.h"
+
+DEFINE_int32(test_random_seed, 123456789, "Seed for srand()");
+DEFINE_int32(benchmark_max_iters, 10000000, "Maximum test iterations");
+DEFINE_int32(benchmark_min_iters, 100, "Minimum test iterations");
+DEFINE_int32(benchmark_target_seconds, 1,
+             "Attempt to benchmark for this many seconds");
+
+using std::allocator;
+using std::less;
+using std::map;
+using std::max;
+using std::min;
+using std::multimap;
+using std::multiset;
+using std::set;
+using std::string;
+using std::vector;
+
+namespace btree {
+namespace {
+
+struct RandGen {
+  typedef ptrdiff_t result_type;
+  RandGen(result_type seed) {
+    srand(seed);
+  }
+  result_type operator()(result_type l) {
+    return rand() % l;
+  }
+};
+
+struct BenchmarkRun {
+  BenchmarkRun(const char *name, void (*func)(int));
+  void Run();
+  void Stop();
+  void Start();
+  void Reset();
+
+  BenchmarkRun *next_benchmark;
+  const char *benchmark_name;
+  void (*benchmark_func)(int);
+  int64_t accum_micros;
+  int64_t last_started;
+};
+
+BenchmarkRun *first_benchmark;
+BenchmarkRun *current_benchmark;
+
+int64_t get_micros () {
+  timeval tv;
+  gettimeofday(&tv, NULL);
+  return tv.tv_sec * 1000000 + tv.tv_usec;
+}
+
+BenchmarkRun::BenchmarkRun(const char *name, void (*func)(int))
+  : next_benchmark(first_benchmark),
+    benchmark_name(name),
+    benchmark_func(func),
+    accum_micros(0),
+    last_started(0) {
+  first_benchmark = this;
+}
+
+#define BTREE_BENCHMARK(name) \
+  BTREE_BENCHMARK2(#name, name, __COUNTER__)
+#define BTREE_BENCHMARK2(name, func, counter)   \
+  BTREE_BENCHMARK3(name, func, counter)
+#define BTREE_BENCHMARK3(name, func, counter)   \
+  BenchmarkRun bench ## counter (name, func)
+
+void StopBenchmarkTiming() {
+  current_benchmark->Stop();
+}
+
+void StartBenchmarkTiming() {
+  current_benchmark->Start();
+}
+
+void RunBenchmarks() {
+  for (BenchmarkRun *bench = first_benchmark; bench; 
+       bench = bench->next_benchmark) {
+    bench->Run();
+  }
+}
+
+void BenchmarkRun::Start() {
+  assert(!last_started);
+  last_started = get_micros();
+}
+
+void BenchmarkRun::Stop() {
+  if (last_started == 0) {
+    return;
+  }
+  accum_micros += get_micros() - last_started;
+  last_started = 0;
+}
+
+void BenchmarkRun::Reset() {
+  last_started = 0;
+  accum_micros = 0;
+}
+
+void BenchmarkRun::Run() {
+  assert(current_benchmark == NULL);
+  current_benchmark = this;
+  int iters = FLAGS_benchmark_min_iters;
+  for (;;) {
+    Reset();
+    Start();
+    benchmark_func(iters);
+    Stop();
+    if (accum_micros > FLAGS_benchmark_target_seconds * 1000000 ||
+        iters >= FLAGS_benchmark_max_iters) {
+      break;
+    } else if (accum_micros == 0) {
+      iters *= 100;
+    } else {
+      int64_t target_micros = FLAGS_benchmark_target_seconds * 1000000;
+      iters = target_micros * iters / accum_micros;
+    }
+    iters = min(iters, FLAGS_benchmark_max_iters);
+  }
+  std::cout << benchmark_name << "\t"
+            << accum_micros * 1000 / iters << "\t"
+            << iters;
+  current_benchmark = NULL;
+}
+
+// Used to avoid compiler optimizations for these benchmarks.
+template <typename T>
+void sink(const T& t0) {
+  volatile T t = t0;
+}
+
+// Benchmark insertion of values into a container.
+template <typename T>
+void BM_Insert(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+
+  T container;
+  vector<V> values = GenerateValues<V>(FLAGS_benchmark_values);
+  for (int i = 0; i < values.size(); i++) {
+    container.insert(values[i]);
+  }
+
+  for (int i = 0; i < n; ) {
+    // Remove and re-insert 10% of the keys
+    int m = min(n - i, FLAGS_benchmark_values / 10);
+
+    for (int j = i; j < i + m; j++) {
+      int x = j % FLAGS_benchmark_values;
+      container.erase(key_of_value(values[x]));
+    }
+
+    StartBenchmarkTiming();
+
+    for (int j = i; j < i + m; j++) {
+      int x = j % FLAGS_benchmark_values;
+      container.insert(values[x]);
+    }
+
+    StopBenchmarkTiming();
+
+    i += m;
+  }
+}
+
+// Benchmark lookup of values in a container.
+template <typename T>
+void BM_Lookup(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+
+  T container;
+  vector<V> values = GenerateValues<V>(FLAGS_benchmark_values);
+
+  for (int i = 0; i < values.size(); i++) {
+    container.insert(values[i]);
+  }
+
+  V r = V();
+
+  StartBenchmarkTiming();
+
+  for (int i = 0; i < n; i++) {
+    int m = i % values.size();
+    r = *container.find(key_of_value(values[m]));
+  }
+
+  StopBenchmarkTiming();
+
+  sink(r); // Keep compiler from optimizing away r.
+}
+
+// Benchmark lookup of values in a full container, meaning that values
+// are inserted in-order to take advantage of biased insertion, which
+// yields a full tree.
+template <typename T>
+void BM_FullLookup(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+
+  T container;
+  vector<V> values = GenerateValues<V>(FLAGS_benchmark_values);
+  vector<V> sorted(values);
+  sort(sorted.begin(), sorted.end());
+
+  for (int i = 0; i < sorted.size(); i++) {
+    container.insert(sorted[i]);
+  }
+
+  V r = V();
+
+  StartBenchmarkTiming();
+
+  for (int i = 0; i < n; i++) {
+    int m = i % values.size();
+    r = *container.find(key_of_value(values[m]));
+  }
+
+  StopBenchmarkTiming();
+
+  sink(r); // Keep compiler from optimizing away r.
+}
+
+// Benchmark deletion of values from a container.
+template <typename T>
+void BM_Delete(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+
+  T container;
+  vector<V> values = GenerateValues<V>(FLAGS_benchmark_values);
+  for (int i = 0; i < values.size(); i++) {
+    container.insert(values[i]);
+  }
+
+  for (int i = 0; i < n; ) {
+    // Remove and re-insert 10% of the keys
+    int m = min(n - i, FLAGS_benchmark_values / 10);
+
+    StartBenchmarkTiming();
+
+    for (int j = i; j < i + m; j++) {
+      int x = j % FLAGS_benchmark_values;
+      container.erase(key_of_value(values[x]));
+    }
+
+    StopBenchmarkTiming();
+
+    for (int j = i; j < i + m; j++) {
+      int x = j % FLAGS_benchmark_values;
+      container.insert(values[x]);
+    }
+
+    i += m;
+  }
+}
+
+// Benchmark steady-state insert (into first half of range) and remove
+// (from second second half of range), treating the container
+// approximately like a queue with log-time access for all elements.
+// This benchmark does not test the case where insertion and removal
+// happen in the same region of the tree.  This benchmark counts two
+// value constructors.
+template <typename T>
+void BM_QueueAddRem(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+  assert(FLAGS_benchmark_values % 2 == 0);
+
+  T container;
+
+  const int half = FLAGS_benchmark_values / 2;
+  vector<int> remove_keys(half);
+  vector<int> add_keys(half);
+
+  for (int i = 0; i < half; i++) {
+    remove_keys[i] = i;
+    add_keys[i] = i;
+  }
+
+  RandGen rand(FLAGS_test_random_seed);
+
+  random_shuffle(remove_keys.begin(), remove_keys.end(), rand);
+  random_shuffle(add_keys.begin(), add_keys.end(), rand);
+
+  Generator<V> g(FLAGS_benchmark_values + FLAGS_benchmark_max_iters);
+
+  for (int i = 0; i < half; i++) {
+    container.insert(g(add_keys[i]));
+    container.insert(g(half + remove_keys[i]));
+  }
+
+  // There are three parts each of size "half":
+  // 1 is being deleted from  [offset - half, offset)
+  // 2 is standing            [offset, offset + half)
+  // 3 is being inserted into [offset + half, offset + 2 * half)
+  int offset = 0;
+
+  StartBenchmarkTiming();
+
+  for (int i = 0; i < n; i++) {
+    int idx = i % half;
+
+    if (idx == 0) {
+      StopBenchmarkTiming();
+      random_shuffle(remove_keys.begin(), remove_keys.end(), rand);
+      random_shuffle(add_keys.begin(), add_keys.end(), rand);
+      offset += half;
+      StartBenchmarkTiming();
+    }
+
+    int e = container.erase(key_of_value(g(offset - half + remove_keys[idx])));
+    assert(e == 1);
+    container.insert(g(offset + half + add_keys[idx]));
+  }
+
+  StopBenchmarkTiming();
+}
+
+// Mixed insertion and deletion in the same range using pre-constructed values.
+template <typename T>
+void BM_MixedAddRem(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+  assert(FLAGS_benchmark_values % 2 == 0);
+
+  T container;
+  RandGen rand(FLAGS_test_random_seed);
+
+  vector<V> values = GenerateValues<V>(FLAGS_benchmark_values * 2);
+
+  // Create two random shuffles
+  vector<int> remove_keys(FLAGS_benchmark_values);
+  vector<int> add_keys(FLAGS_benchmark_values);
+
+  // Insert the first half of the values (already in random order)
+  for (int i = 0; i < FLAGS_benchmark_values; i++) {
+    container.insert(values[i]);
+
+    // remove_keys and add_keys will be swapped before each round,
+    // therefore fill add_keys here w/ the keys being inserted, so
+    // they'll be the first to be removed.
+    remove_keys[i] = i + FLAGS_benchmark_values;
+    add_keys[i] = i;
+  }
+
+  StartBenchmarkTiming();
+
+  for (int i = 0; i < n; i++) {
+    int idx = i % FLAGS_benchmark_values;
+
+    if (idx == 0) {
+      StopBenchmarkTiming();
+      remove_keys.swap(add_keys);
+      random_shuffle(remove_keys.begin(), remove_keys.end(), rand);
+      random_shuffle(add_keys.begin(), add_keys.end(), rand);
+      StartBenchmarkTiming();
+    }
+
+    int e = container.erase(key_of_value(values[remove_keys[idx]]));
+    assert(e == 1);
+    container.insert(values[add_keys[idx]]);
+  }
+
+  StopBenchmarkTiming();
+}
+
+// Insertion at end, removal from the beginning.  This benchmark
+// counts two value constructors.
+template <typename T>
+void BM_Fifo(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+
+  T container;
+  Generator<V> g(FLAGS_benchmark_values + FLAGS_benchmark_max_iters);
+
+  for (int i = 0; i < FLAGS_benchmark_values; i++) {
+    container.insert(g(i));
+  }
+
+  StartBenchmarkTiming();
+
+  for (int i = 0; i < n; i++) {
+    container.erase(container.begin());
+    container.insert(container.end(), g(i + FLAGS_benchmark_values));
+  }
+
+  StopBenchmarkTiming();
+}
+
+// Iteration (forward) through the tree
+template <typename T>
+void BM_FwdIter(int n) {
+  typedef typename std::remove_const<typename T::value_type>::type V;
+
+  // Disable timing while we perform some initialization.
+  StopBenchmarkTiming();
+
+  T container;
+  vector<V> values = GenerateValues<V>(FLAGS_benchmark_values);
+
+  for (int i = 0; i < FLAGS_benchmark_values; i++) {
+    container.insert(values[i]);
+  }
+
+  typename T::iterator iter;
+
+  V r = V();
+
+  StartBenchmarkTiming();
+
+  for (int i = 0; i < n; i++) {
+    int idx = i % FLAGS_benchmark_values;
+
+    if (idx == 0) {
+      iter = container.begin();
+    }
+    r = *iter;
+    ++iter;
+  }
+
+  StopBenchmarkTiming();
+
+  sink(r); // Keep compiler from optimizing away r.
+}
+
+typedef set<int32_t> stl_set_int32;
+typedef set<int64_t> stl_set_int64;
+typedef set<string> stl_set_string;
+
+typedef map<int32_t, intptr_t> stl_map_int32;
+typedef map<int64_t, intptr_t> stl_map_int64;
+typedef map<string, intptr_t> stl_map_string;
+
+typedef multiset<int32_t> stl_multiset_int32;
+typedef multiset<int64_t> stl_multiset_int64;
+typedef multiset<string> stl_multiset_string;
+
+typedef multimap<int32_t, intptr_t> stl_multimap_int32;
+typedef multimap<int64_t, intptr_t> stl_multimap_int64;
+typedef multimap<string, intptr_t> stl_multimap_string;
+
+#define MY_BENCHMARK_TYPES2(value, name, size)                                \
+  typedef btree ## _set<value, less<value>, allocator<value>, size>           \
+    btree ## _ ## size ## _set_ ## name;                                      \
+  typedef btree ## _map<value, int, less<value>, allocator<value>, size>      \
+    btree ## _ ## size ## _map_ ## name;                                      \
+  typedef btree ## _multiset<value, less<value>, allocator<value>, size>      \
+    btree ## _ ## size ## _multiset_ ## name;                                 \
+  typedef btree ## _multimap<value, int, less<value>, allocator<value>, size> \
+    btree ## _ ## size ## _multimap_ ## name
+
+#define MY_BENCHMARK_TYPES(value, name)  \
+  MY_BENCHMARK_TYPES2(value, name, 128); \
+  MY_BENCHMARK_TYPES2(value, name, 160); \
+  MY_BENCHMARK_TYPES2(value, name, 192); \
+  MY_BENCHMARK_TYPES2(value, name, 224); \
+  MY_BENCHMARK_TYPES2(value, name, 256); \
+  MY_BENCHMARK_TYPES2(value, name, 288); \
+  MY_BENCHMARK_TYPES2(value, name, 320); \
+  MY_BENCHMARK_TYPES2(value, name, 352); \
+  MY_BENCHMARK_TYPES2(value, name, 384); \
+  MY_BENCHMARK_TYPES2(value, name, 416); \
+  MY_BENCHMARK_TYPES2(value, name, 448); \
+  MY_BENCHMARK_TYPES2(value, name, 480); \
+  MY_BENCHMARK_TYPES2(value, name, 512); \
+  MY_BENCHMARK_TYPES2(value, name, 1024); \
+  MY_BENCHMARK_TYPES2(value, name, 1536); \
+  MY_BENCHMARK_TYPES2(value, name, 2048)
+
+MY_BENCHMARK_TYPES(int32_t, int32);
+MY_BENCHMARK_TYPES(int64_t, int64);
+MY_BENCHMARK_TYPES(string, string);
+
+#define MY_BENCHMARK4(type, name, func)                            \
+  void BM_ ## type ## _ ## name(int n) { BM_ ## func <type>(n); }  \
+  BTREE_BENCHMARK(BM_ ## type ## _ ## name)
+
+// Define NODESIZE_TESTING when running btree_perf.py.
+
+#ifdef NODESIZE_TESTING
+#define MY_BENCHMARK3(tree, type, name, func) \
+  MY_BENCHMARK4(tree ## _128_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _160_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _192_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _224_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _256_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _288_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _320_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _352_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _384_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _416_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _448_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _480_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _512_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _1024_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _1536_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _2048_ ## type, name, func)
+#else
+#define MY_BENCHMARK3(tree, type, name, func) \
+  MY_BENCHMARK4(tree ## _256_ ## type, name, func); \
+  MY_BENCHMARK4(tree ## _2048_ ## type, name, func)
+#endif
+
+#define MY_BENCHMARK2(type, name, func)    \
+  MY_BENCHMARK4(stl_ ## type, name, func); \
+  MY_BENCHMARK3(btree, type, name, func)
+
+#define MY_BENCHMARK(type)                        \
+  MY_BENCHMARK2(type, insert, Insert);            \
+  MY_BENCHMARK2(type, lookup, Lookup);            \
+  MY_BENCHMARK2(type, fulllookup, FullLookup);    \
+  MY_BENCHMARK2(type, delete, Delete);            \
+  MY_BENCHMARK2(type, queueaddrem, QueueAddRem);  \
+  MY_BENCHMARK2(type, mixedaddrem, MixedAddRem);  \
+  MY_BENCHMARK2(type, fifo, Fifo);                \
+  MY_BENCHMARK2(type, fwditer, FwdIter)
+
+MY_BENCHMARK(set_int32);
+MY_BENCHMARK(map_int32);
+MY_BENCHMARK(set_int64);
+MY_BENCHMARK(map_int64);
+MY_BENCHMARK(set_string);
+MY_BENCHMARK(map_string);
+
+MY_BENCHMARK(multiset_int32);
+MY_BENCHMARK(multimap_int32);
+MY_BENCHMARK(multiset_int64);
+MY_BENCHMARK(multimap_int64);
+MY_BENCHMARK(multiset_string);
+MY_BENCHMARK(multimap_string);
+
+} // namespace
+} // namespace btree
+
+int main(int argc, char **argv) {
+  btree::RunBenchmarks();
+  return 0;
+}
diff --git a/xdelta3/cpp-btree/btree_container.h b/xdelta3/cpp-btree/btree_container.h
new file mode 100644
index 0000000..fb617ab
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_container.h
@@ -0,0 +1,349 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef UTIL_BTREE_BTREE_CONTAINER_H__
+#define UTIL_BTREE_BTREE_CONTAINER_H__
+
+#include <iosfwd>
+#include <utility>
+
+#include "btree.h"
+
+namespace btree {
+
+// A common base class for btree_set, btree_map, btree_multiset and
+// btree_multimap.
+template <typename Tree>
+class btree_container {
+  typedef btree_container<Tree> self_type;
+
+ public:
+  typedef typename Tree::params_type params_type;
+  typedef typename Tree::key_type key_type;
+  typedef typename Tree::value_type value_type;
+  typedef typename Tree::key_compare key_compare;
+  typedef typename Tree::allocator_type allocator_type;
+  typedef typename Tree::pointer pointer;
+  typedef typename Tree::const_pointer const_pointer;
+  typedef typename Tree::reference reference;
+  typedef typename Tree::const_reference const_reference;
+  typedef typename Tree::size_type size_type;
+  typedef typename Tree::difference_type difference_type;
+  typedef typename Tree::iterator iterator;
+  typedef typename Tree::const_iterator const_iterator;
+  typedef typename Tree::reverse_iterator reverse_iterator;
+  typedef typename Tree::const_reverse_iterator const_reverse_iterator;
+
+ public:
+  // Default constructor.
+  btree_container(const key_compare &comp, const allocator_type &alloc)
+      : tree_(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_container(const self_type &x)
+      : tree_(x.tree_) {
+  }
+
+  // Iterator routines.
+  iterator begin() { return tree_.begin(); }
+  const_iterator begin() const { return tree_.begin(); }
+  iterator end() { return tree_.end(); }
+  const_iterator end() const { return tree_.end(); }
+  reverse_iterator rbegin() { return tree_.rbegin(); }
+  const_reverse_iterator rbegin() const { return tree_.rbegin(); }
+  reverse_iterator rend() { return tree_.rend(); }
+  const_reverse_iterator rend() const { return tree_.rend(); }
+
+  // Lookup routines.
+  iterator lower_bound(const key_type &key) {
+    return tree_.lower_bound(key);
+  }
+  const_iterator lower_bound(const key_type &key) const {
+    return tree_.lower_bound(key);
+  }
+  iterator upper_bound(const key_type &key) {
+    return tree_.upper_bound(key);
+  }
+  const_iterator upper_bound(const key_type &key) const {
+    return tree_.upper_bound(key);
+  }
+  std::pair<iterator,iterator> equal_range(const key_type &key) {
+    return tree_.equal_range(key);
+  }
+  std::pair<const_iterator,const_iterator> equal_range(const key_type &key) const {
+    return tree_.equal_range(key);
+  }
+
+  // Utility routines.
+  void clear() {
+    tree_.clear();
+  }
+  void swap(self_type &x) {
+    tree_.swap(x.tree_);
+  }
+  void dump(std::ostream &os) const {
+    tree_.dump(os);
+  }
+  void verify() const {
+    tree_.verify();
+  }
+
+  // Size routines.
+  size_type size() const { return tree_.size(); }
+  size_type max_size() const { return tree_.max_size(); }
+  bool empty() const { return tree_.empty(); }
+  size_type height() const { return tree_.height(); }
+  size_type internal_nodes() const { return tree_.internal_nodes(); }
+  size_type leaf_nodes() const { return tree_.leaf_nodes(); }
+  size_type nodes() const { return tree_.nodes(); }
+  size_type bytes_used() const { return tree_.bytes_used(); }
+  static double average_bytes_per_value() {
+    return Tree::average_bytes_per_value();
+  }
+  double fullness() const { return tree_.fullness(); }
+  double overhead() const { return tree_.overhead(); }
+
+  bool operator==(const self_type& x) const {
+    if (size() != x.size()) {
+      return false;
+    }
+    for (const_iterator i = begin(), xi = x.begin(); i != end(); ++i, ++xi) {
+      if (*i != *xi) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  bool operator!=(const self_type& other) const {
+    return !operator==(other);
+  }
+
+
+ protected:
+  Tree tree_;
+};
+
+template <typename T>
+inline std::ostream& operator<<(std::ostream &os, const btree_container<T> &b) {
+  b.dump(os);
+  return os;
+}
+
+// A common base class for btree_set and safe_btree_set.
+template <typename Tree>
+class btree_unique_container : public btree_container<Tree> {
+  typedef btree_unique_container<Tree> self_type;
+  typedef btree_container<Tree> super_type;
+
+ public:
+  typedef typename Tree::key_type key_type;
+  typedef typename Tree::value_type value_type;
+  typedef typename Tree::size_type size_type;
+  typedef typename Tree::key_compare key_compare;
+  typedef typename Tree::allocator_type allocator_type;
+  typedef typename Tree::iterator iterator;
+  typedef typename Tree::const_iterator const_iterator;
+
+ public:
+  // Default constructor.
+  btree_unique_container(const key_compare &comp = key_compare(),
+                         const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_unique_container(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_unique_container(InputIterator b, InputIterator e,
+                         const key_compare &comp = key_compare(),
+                         const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+    insert(b, e);
+  }
+
+  // Lookup routines.
+  iterator find(const key_type &key) {
+    return this->tree_.find_unique(key);
+  }
+  const_iterator find(const key_type &key) const {
+    return this->tree_.find_unique(key);
+  }
+  size_type count(const key_type &key) const {
+    return this->tree_.count_unique(key);
+  }
+
+  // Insertion routines.
+  std::pair<iterator,bool> insert(const value_type &x) {
+    return this->tree_.insert_unique(x);
+  }
+  iterator insert(iterator position, const value_type &x) {
+    return this->tree_.insert_unique(position, x);
+  }
+  template <typename InputIterator>
+  void insert(InputIterator b, InputIterator e) {
+    this->tree_.insert_unique(b, e);
+  }
+
+  // Deletion routines.
+  int erase(const key_type &key) {
+    return this->tree_.erase_unique(key);
+  }
+  // Erase the specified iterator from the btree. The iterator must be valid
+  // (i.e. not equal to end()).  Return an iterator pointing to the node after
+  // the one that was erased (or end() if none exists).
+  iterator erase(const iterator &iter) {
+    return this->tree_.erase(iter);
+  }
+  void erase(const iterator &first, const iterator &last) {
+    this->tree_.erase(first, last);
+  }
+};
+
+// A common base class for btree_map and safe_btree_map.
+template <typename Tree>
+class btree_map_container : public btree_unique_container<Tree> {
+  typedef btree_map_container<Tree> self_type;
+  typedef btree_unique_container<Tree> super_type;
+
+ public:
+  typedef typename Tree::key_type key_type;
+  typedef typename Tree::data_type data_type;
+  typedef typename Tree::value_type value_type;
+  typedef typename Tree::mapped_type mapped_type;
+  typedef typename Tree::key_compare key_compare;
+  typedef typename Tree::allocator_type allocator_type;
+
+ private:
+  // A pointer-like object which only generates its value when
+  // dereferenced. Used by operator[] to avoid constructing an empty data_type
+  // if the key already exists in the map.
+  struct generate_value {
+    generate_value(const key_type &k)
+        : key(k) {
+    }
+    value_type operator*() const {
+      return std::make_pair(key, data_type());
+    }
+    const key_type &key;
+  };
+
+ public:
+  // Default constructor.
+  btree_map_container(const key_compare &comp = key_compare(),
+                      const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_map_container(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_map_container(InputIterator b, InputIterator e,
+                      const key_compare &comp = key_compare(),
+                      const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+
+  // Insertion routines.
+  data_type& operator[](const key_type &key) {
+    return this->tree_.insert_unique(key, generate_value(key)).first->second;
+  }
+};
+
+// A common base class for btree_multiset and btree_multimap.
+template <typename Tree>
+class btree_multi_container : public btree_container<Tree> {
+  typedef btree_multi_container<Tree> self_type;
+  typedef btree_container<Tree> super_type;
+
+ public:
+  typedef typename Tree::key_type key_type;
+  typedef typename Tree::value_type value_type;
+  typedef typename Tree::size_type size_type;
+  typedef typename Tree::key_compare key_compare;
+  typedef typename Tree::allocator_type allocator_type;
+  typedef typename Tree::iterator iterator;
+  typedef typename Tree::const_iterator const_iterator;
+
+ public:
+  // Default constructor.
+  btree_multi_container(const key_compare &comp = key_compare(),
+                        const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_multi_container(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_multi_container(InputIterator b, InputIterator e,
+                        const key_compare &comp = key_compare(),
+                        const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+    insert(b, e);
+  }
+
+  // Lookup routines.
+  iterator find(const key_type &key) {
+    return this->tree_.find_multi(key);
+  }
+  const_iterator find(const key_type &key) const {
+    return this->tree_.find_multi(key);
+  }
+  size_type count(const key_type &key) const {
+    return this->tree_.count_multi(key);
+  }
+
+  // Insertion routines.
+  iterator insert(const value_type &x) {
+    return this->tree_.insert_multi(x);
+  }
+  iterator insert(iterator position, const value_type &x) {
+    return this->tree_.insert_multi(position, x);
+  }
+  template <typename InputIterator>
+  void insert(InputIterator b, InputIterator e) {
+    this->tree_.insert_multi(b, e);
+  }
+
+  // Deletion routines.
+  int erase(const key_type &key) {
+    return this->tree_.erase_multi(key);
+  }
+  // Erase the specified iterator from the btree. The iterator must be valid
+  // (i.e. not equal to end()).  Return an iterator pointing to the node after
+  // the one that was erased (or end() if none exists).
+  iterator erase(const iterator &iter) {
+    return this->tree_.erase(iter);
+  }
+  void erase(const iterator &first, const iterator &last) {
+    this->tree_.erase(first, last);
+  }
+};
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_BTREE_CONTAINER_H__
diff --git a/xdelta3/cpp-btree/btree_map.h b/xdelta3/cpp-btree/btree_map.h
new file mode 100644
index 0000000..b83489f
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_map.h
@@ -0,0 +1,130 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// A btree_map<> implements the STL unique sorted associative container
+// interface and the pair associative container interface (a.k.a map<>) using a
+// btree. A btree_multimap<> implements the STL multiple sorted associative
+// container interface and the pair associtive container interface (a.k.a
+// multimap<>) using a btree. See btree.h for details of the btree
+// implementation and caveats.
+
+#ifndef UTIL_BTREE_BTREE_MAP_H__
+#define UTIL_BTREE_BTREE_MAP_H__
+
+#include <algorithm>
+#include <functional>
+#include <memory>
+#include <string>
+#include <utility>
+
+#include "btree.h"
+#include "btree_container.h"
+
+namespace btree {
+
+// The btree_map class is needed mainly for its constructors.
+template <typename Key, typename Value,
+          typename Compare = std::less<Key>,
+          typename Alloc = std::allocator<std::pair<const Key, Value> >,
+          int TargetNodeSize = 256>
+class btree_map : public btree_map_container<
+  btree<btree_map_params<Key, Value, Compare, Alloc, TargetNodeSize> > > {
+
+  typedef btree_map<Key, Value, Compare, Alloc, TargetNodeSize> self_type;
+  typedef btree_map_params<
+    Key, Value, Compare, Alloc, TargetNodeSize> params_type;
+  typedef btree<params_type> btree_type;
+  typedef btree_map_container<btree_type> super_type;
+
+ public:
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+
+ public:
+  // Default constructor.
+  btree_map(const key_compare &comp = key_compare(),
+            const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_map(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_map(InputIterator b, InputIterator e,
+            const key_compare &comp = key_compare(),
+            const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+};
+
+template <typename K, typename V, typename C, typename A, int N>
+inline void swap(btree_map<K, V, C, A, N> &x,
+                 btree_map<K, V, C, A, N> &y) {
+  x.swap(y);
+}
+
+// The btree_multimap class is needed mainly for its constructors.
+template <typename Key, typename Value,
+          typename Compare = std::less<Key>,
+          typename Alloc = std::allocator<std::pair<const Key, Value> >,
+          int TargetNodeSize = 256>
+class btree_multimap : public btree_multi_container<
+  btree<btree_map_params<Key, Value, Compare, Alloc, TargetNodeSize> > > {
+
+  typedef btree_multimap<Key, Value, Compare, Alloc, TargetNodeSize> self_type;
+  typedef btree_map_params<
+    Key, Value, Compare, Alloc, TargetNodeSize> params_type;
+  typedef btree<params_type> btree_type;
+  typedef btree_multi_container<btree_type> super_type;
+
+ public:
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+  typedef typename btree_type::data_type data_type;
+  typedef typename btree_type::mapped_type mapped_type;
+
+ public:
+  // Default constructor.
+  btree_multimap(const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_multimap(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_multimap(InputIterator b, InputIterator e,
+                 const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+};
+
+template <typename K, typename V, typename C, typename A, int N>
+inline void swap(btree_multimap<K, V, C, A, N> &x,
+                 btree_multimap<K, V, C, A, N> &y) {
+  x.swap(y);
+}
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_BTREE_MAP_H__
diff --git a/xdelta3/cpp-btree/btree_set.h b/xdelta3/cpp-btree/btree_set.h
new file mode 100644
index 0000000..f9b2e75
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_set.h
@@ -0,0 +1,121 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// A btree_set<> implements the STL unique sorted associative container
+// interface (a.k.a set<>) using a btree. A btree_multiset<> implements the STL
+// multiple sorted associative container interface (a.k.a multiset<>) using a
+// btree. See btree.h for details of the btree implementation and caveats.
+
+#ifndef UTIL_BTREE_BTREE_SET_H__
+#define UTIL_BTREE_BTREE_SET_H__
+
+#include <functional>
+#include <memory>
+#include <string>
+
+#include "btree.h"
+#include "btree_container.h"
+
+namespace btree {
+
+// The btree_set class is needed mainly for its constructors.
+template <typename Key,
+          typename Compare = std::less<Key>,
+          typename Alloc = std::allocator<Key>,
+          int TargetNodeSize = 256>
+class btree_set : public btree_unique_container<
+  btree<btree_set_params<Key, Compare, Alloc, TargetNodeSize> > > {
+
+  typedef btree_set<Key, Compare, Alloc, TargetNodeSize> self_type;
+  typedef btree_set_params<Key, Compare, Alloc, TargetNodeSize> params_type;
+  typedef btree<params_type> btree_type;
+  typedef btree_unique_container<btree_type> super_type;
+
+ public:
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+
+ public:
+  // Default constructor.
+  btree_set(const key_compare &comp = key_compare(),
+            const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_set(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_set(InputIterator b, InputIterator e,
+            const key_compare &comp = key_compare(),
+            const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+};
+
+template <typename K, typename C, typename A, int N>
+inline void swap(btree_set<K, C, A, N> &x, btree_set<K, C, A, N> &y) {
+  x.swap(y);
+}
+
+// The btree_multiset class is needed mainly for its constructors.
+template <typename Key,
+          typename Compare = std::less<Key>,
+          typename Alloc = std::allocator<Key>,
+          int TargetNodeSize = 256>
+class btree_multiset : public btree_multi_container<
+  btree<btree_set_params<Key, Compare, Alloc, TargetNodeSize> > > {
+
+  typedef btree_multiset<Key, Compare, Alloc, TargetNodeSize> self_type;
+  typedef btree_set_params<Key, Compare, Alloc, TargetNodeSize> params_type;
+  typedef btree<params_type> btree_type;
+  typedef btree_multi_container<btree_type> super_type;
+
+ public:
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+
+ public:
+  // Default constructor.
+  btree_multiset(const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  btree_multiset(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  btree_multiset(InputIterator b, InputIterator e,
+                 const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+};
+
+template <typename K, typename C, typename A, int N>
+inline void swap(btree_multiset<K, C, A, N> &x,
+                 btree_multiset<K, C, A, N> &y) {
+  x.swap(y);
+}
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_BTREE_SET_H__
diff --git a/xdelta3/cpp-btree/btree_test.cc b/xdelta3/cpp-btree/btree_test.cc
new file mode 100644
index 0000000..6b1837d
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_test.cc
@@ -0,0 +1,270 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "gtest/gtest.h"
+#include "btree_map.h"
+#include "btree_set.h"
+#include "btree_test.h"
+
+namespace btree {
+namespace {
+
+template <typename K, int N>
+void SetTest() {
+  typedef TestAllocator<K> TestAlloc;
+  ASSERT_EQ(sizeof(btree_set<K>), sizeof(void*));
+  BtreeTest<btree_set<K, std::less<K>, std::allocator<K>, N>, std::set<K> >();
+  BtreeAllocatorTest<btree_set<K, std::less<K>, TestAlloc, N> >();
+}
+
+template <typename K, int N>
+void MapTest() {
+  typedef TestAllocator<K> TestAlloc;
+  ASSERT_EQ(sizeof(btree_map<K, K>), sizeof(void*));
+  BtreeTest<btree_map<K, K, std::less<K>, std::allocator<K>, N>, std::map<K, K> >();
+  BtreeAllocatorTest<btree_map<K, K, std::less<K>, TestAlloc, N> >();
+  BtreeMapTest<btree_map<K, K, std::less<K>, std::allocator<K>, N> >();
+}
+
+TEST(Btree, set_int32_32)   { SetTest<int32_t, 32>(); }
+TEST(Btree, set_int32_64)   { SetTest<int32_t, 64>(); }
+TEST(Btree, set_int32_128)  { SetTest<int32_t, 128>(); }
+TEST(Btree, set_int32_256)  { SetTest<int32_t, 256>(); }
+TEST(Btree, set_int64_256)  { SetTest<int64_t, 256>(); }
+TEST(Btree, set_string_256) { SetTest<std::string, 256>(); }
+TEST(Btree, set_pair_256)   { SetTest<std::pair<int, int>, 256>(); }
+TEST(Btree, map_int32_256)  { MapTest<int32_t, 256>(); }
+TEST(Btree, map_int64_256)  { MapTest<int64_t, 256>(); }
+TEST(Btree, map_string_256) { MapTest<std::string, 256>(); }
+TEST(Btree, map_pair_256)   { MapTest<std::pair<int, int>, 256>(); }
+
+// Large-node tests
+TEST(Btree, map_int32_1024)   { MapTest<int32_t, 1024>(); }
+TEST(Btree, map_int32_1032)   { MapTest<int32_t, 1032>(); }
+TEST(Btree, map_int32_1040)   { MapTest<int32_t, 1040>(); }
+TEST(Btree, map_int32_1048)   { MapTest<int32_t, 1048>(); }
+TEST(Btree, map_int32_1056)   { MapTest<int32_t, 1056>(); }
+
+TEST(Btree, map_int32_2048)   { MapTest<int32_t, 2048>(); }
+TEST(Btree, map_int32_4096)   { MapTest<int32_t, 4096>(); }
+TEST(Btree, set_int32_1024)   { SetTest<int32_t, 1024>(); }
+TEST(Btree, set_int32_2048)   { SetTest<int32_t, 2048>(); }
+TEST(Btree, set_int32_4096)   { SetTest<int32_t, 4096>(); }
+TEST(Btree, map_string_1024)   { MapTest<std::string, 1024>(); }
+TEST(Btree, map_string_2048)   { MapTest<std::string, 2048>(); }
+TEST(Btree, map_string_4096)   { MapTest<std::string, 4096>(); }
+TEST(Btree, set_string_1024)   { SetTest<std::string, 1024>(); }
+TEST(Btree, set_string_2048)   { SetTest<std::string, 2048>(); }
+TEST(Btree, set_string_4096)   { SetTest<std::string, 4096>(); }
+
+template <typename K, int N>
+void MultiSetTest() {
+  typedef TestAllocator<K> TestAlloc;
+  ASSERT_EQ(sizeof(btree_multiset<K>), sizeof(void*));
+  BtreeMultiTest<btree_multiset<K, std::less<K>, std::allocator<K>, N>,
+      std::multiset<K> >();
+  BtreeAllocatorTest<btree_multiset<K, std::less<K>, TestAlloc, N> >();
+}
+
+template <typename K, int N>
+void MultiMapTest() {
+  typedef TestAllocator<K> TestAlloc;
+  ASSERT_EQ(sizeof(btree_multimap<K, K>), sizeof(void*));
+  BtreeMultiTest<btree_multimap<K, K, std::less<K>, std::allocator<K>, N>,
+      std::multimap<K, K> >();
+  BtreeMultiMapTest<btree_multimap<K, K, std::less<K>, std::allocator<K>, N> >();
+  BtreeAllocatorTest<btree_multimap<K, K, std::less<K>, TestAlloc, N> >();
+}
+
+TEST(Btree, multiset_int32_256)  { MultiSetTest<int32_t, 256>(); }
+TEST(Btree, multiset_int64_256)  { MultiSetTest<int64_t, 256>(); }
+TEST(Btree, multiset_string_256) { MultiSetTest<std::string, 256>(); }
+TEST(Btree, multiset_pair_256)   { MultiSetTest<std::pair<int, int>, 256>(); }
+TEST(Btree, multimap_int32_256)  { MultiMapTest<int32_t, 256>(); }
+TEST(Btree, multimap_int64_256)  { MultiMapTest<int64_t, 256>(); }
+TEST(Btree, multimap_string_256) { MultiMapTest<std::string, 256>(); }
+TEST(Btree, multimap_pair_256)   { MultiMapTest<std::pair<int, int>, 256>(); }
+
+// Large-node tests
+TEST(Btree, multimap_int32_1024)   { MultiMapTest<int32_t, 1024>(); }
+TEST(Btree, multimap_int32_2048)   { MultiMapTest<int32_t, 2048>(); }
+TEST(Btree, multimap_int32_4096)   { MultiMapTest<int32_t, 4096>(); }
+TEST(Btree, multiset_int32_1024)   { MultiSetTest<int32_t, 1024>(); }
+TEST(Btree, multiset_int32_2048)   { MultiSetTest<int32_t, 2048>(); }
+TEST(Btree, multiset_int32_4096)   { MultiSetTest<int32_t, 4096>(); }
+TEST(Btree, multimap_string_1024)   { MultiMapTest<std::string, 1024>(); }
+TEST(Btree, multimap_string_2048)   { MultiMapTest<std::string, 2048>(); }
+TEST(Btree, multimap_string_4096)   { MultiMapTest<std::string, 4096>(); }
+TEST(Btree, multiset_string_1024)   { MultiSetTest<std::string, 1024>(); }
+TEST(Btree, multiset_string_2048)   { MultiSetTest<std::string, 2048>(); }
+TEST(Btree, multiset_string_4096)   { MultiSetTest<std::string, 4096>(); }
+
+// Verify that swapping btrees swaps the key comparision functors.
+struct SubstringLess {
+  SubstringLess() : n(2) {}
+  SubstringLess(size_t length)
+      : n(length) {
+  }
+  bool operator()(const std::string &a, const std::string &b) const {
+    std::string as(a.data(), std::min(n, a.size()));
+    std::string bs(b.data(), std::min(n, b.size()));
+    return as < bs;
+  }
+  size_t n;
+};
+
+TEST(Btree, SwapKeyCompare) {
+  typedef btree_set<std::string, SubstringLess> SubstringSet;
+  SubstringSet s1(SubstringLess(1), SubstringSet::allocator_type());
+  SubstringSet s2(SubstringLess(2), SubstringSet::allocator_type());
+
+  ASSERT_TRUE(s1.insert("a").second);
+  ASSERT_FALSE(s1.insert("aa").second);
+
+  ASSERT_TRUE(s2.insert("a").second);
+  ASSERT_TRUE(s2.insert("aa").second);
+  ASSERT_FALSE(s2.insert("aaa").second);
+
+  swap(s1, s2);
+
+  ASSERT_TRUE(s1.insert("b").second);
+  ASSERT_TRUE(s1.insert("bb").second);
+  ASSERT_FALSE(s1.insert("bbb").second);
+
+  ASSERT_TRUE(s2.insert("b").second);
+  ASSERT_FALSE(s2.insert("bb").second);
+}
+
+TEST(Btree, UpperBoundRegression) {
+  // Regress a bug where upper_bound would default-construct a new key_compare
+  // instead of copying the existing one.
+  typedef btree_set<std::string, SubstringLess> SubstringSet;
+  SubstringSet my_set(SubstringLess(3));
+  my_set.insert("aab");
+  my_set.insert("abb");
+  // We call upper_bound("aaa").  If this correctly uses the length 3
+  // comparator, aaa < aab < abb, so we should get aab as the result.
+  // If it instead uses the default-constructed length 2 comparator,
+  // aa == aa < ab, so we'll get abb as our result.
+  SubstringSet::iterator it = my_set.upper_bound("aaa");
+  ASSERT_TRUE(it != my_set.end());
+  EXPECT_EQ("aab", *it);
+}
+
+
+TEST(Btree, IteratorIncrementBy) {
+  // Test that increment_by returns the same position as increment.
+  const int kSetSize = 2341;
+  btree_set<int32_t> my_set;
+  for (int i = 0; i < kSetSize; ++i) {
+    my_set.insert(i);
+  }
+
+  {
+    // Simple increment vs. increment by.
+    btree_set<int32_t>::iterator a = my_set.begin();
+    btree_set<int32_t>::iterator b = my_set.begin();
+    a.increment();
+    b.increment_by(1);
+    EXPECT_EQ(*a, *b);
+  }
+
+  btree_set<int32_t>::iterator a = my_set.begin();
+  for (int i = 1; i < kSetSize; ++i) {
+    ++a;
+    // increment_by
+    btree_set<int32_t>::iterator b = my_set.begin();
+    b.increment_by(i);
+    EXPECT_EQ(*a, *b) << ": i=" << i;
+  }
+}
+
+TEST(Btree, Comparison) {
+  const int kSetSize = 1201;
+  btree_set<int64_t> my_set;
+  for (int i = 0; i < kSetSize; ++i) {
+    my_set.insert(i);
+  }
+  btree_set<int64_t> my_set_copy(my_set);
+  EXPECT_TRUE(my_set_copy == my_set);
+  EXPECT_TRUE(my_set == my_set_copy);
+  EXPECT_FALSE(my_set_copy != my_set);
+  EXPECT_FALSE(my_set != my_set_copy);
+
+  my_set.insert(kSetSize);
+  EXPECT_FALSE(my_set_copy == my_set);
+  EXPECT_FALSE(my_set == my_set_copy);
+  EXPECT_TRUE(my_set_copy != my_set);
+  EXPECT_TRUE(my_set != my_set_copy);
+
+  my_set.erase(kSetSize - 1);
+  EXPECT_FALSE(my_set_copy == my_set);
+  EXPECT_FALSE(my_set == my_set_copy);
+  EXPECT_TRUE(my_set_copy != my_set);
+  EXPECT_TRUE(my_set != my_set_copy);
+
+  btree_map<std::string, int64_t> my_map;
+  for (int i = 0; i < kSetSize; ++i) {
+    my_map[std::string(i, 'a')] = i;
+  }
+  btree_map<std::string, int64_t> my_map_copy(my_map);
+  EXPECT_TRUE(my_map_copy == my_map);
+  EXPECT_TRUE(my_map == my_map_copy);
+  EXPECT_FALSE(my_map_copy != my_map);
+  EXPECT_FALSE(my_map != my_map_copy);
+
+  ++my_map_copy[std::string(7, 'a')];
+  EXPECT_FALSE(my_map_copy == my_map);
+  EXPECT_FALSE(my_map == my_map_copy);
+  EXPECT_TRUE(my_map_copy != my_map);
+  EXPECT_TRUE(my_map != my_map_copy);
+
+  my_map_copy = my_map;
+  my_map["hello"] = kSetSize;
+  EXPECT_FALSE(my_map_copy == my_map);
+  EXPECT_FALSE(my_map == my_map_copy);
+  EXPECT_TRUE(my_map_copy != my_map);
+  EXPECT_TRUE(my_map != my_map_copy);
+
+  my_map.erase(std::string(kSetSize - 1, 'a'));
+  EXPECT_FALSE(my_map_copy == my_map);
+  EXPECT_FALSE(my_map == my_map_copy);
+  EXPECT_TRUE(my_map_copy != my_map);
+  EXPECT_TRUE(my_map != my_map_copy);
+}
+
+TEST(Btree, RangeCtorSanity) {
+  typedef btree_set<int, std::less<int>, std::allocator<int>, 256> test_set;
+  typedef btree_map<int, int, std::less<int>, std::allocator<int>, 256> 
+      test_map;
+  typedef btree_multiset<int, std::less<int>, std::allocator<int>, 256> 
+      test_mset;
+  typedef btree_multimap<int, int, std::less<int>, std::allocator<int>, 256> 
+      test_mmap;
+  std::vector<int> ivec;
+  ivec.push_back(1);
+  std::map<int, int> imap;
+  imap.insert(std::make_pair(1, 2));
+  test_mset tmset(ivec.begin(), ivec.end());
+  test_mmap tmmap(imap.begin(), imap.end());
+  test_set tset(ivec.begin(), ivec.end());
+  test_map tmap(imap.begin(), imap.end());
+  EXPECT_EQ(1, tmset.size());
+  EXPECT_EQ(1, tmmap.size());
+  EXPECT_EQ(1, tset.size());
+  EXPECT_EQ(1, tmap.size());
+}
+
+} // namespace
+} // namespace btree
diff --git a/xdelta3/cpp-btree/btree_test.h b/xdelta3/cpp-btree/btree_test.h
new file mode 100644
index 0000000..413dc3c
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_test.h
@@ -0,0 +1,940 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef UTIL_BTREE_BTREE_TEST_H__
+#define UTIL_BTREE_BTREE_TEST_H__
+
+#include <stdio.h>
+#include <algorithm>
+#include <functional>
+#include <type_traits>
+#include <iosfwd>
+#include <map>
+#include <set>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "gflags/gflags.h"
+#include "btree_container.h"
+
+DECLARE_int32(test_values);
+DECLARE_int32(benchmark_values);
+
+namespace std {
+
+// Provide operator<< support for std::pair<T, U>.
+template <typename T, typename U>
+ostream& operator<<(ostream &os, const std::pair<T, U> &p) {
+  os << "(" << p.first << "," << p.second << ")";
+  return os;
+}
+
+// Provide pair equality testing that works as long as x.first is comparable to
+// y.first and x.second is comparable to y.second. Needed in the test for
+// comparing std::pair<T, U> to std::pair<const T, U>.
+template <typename T, typename U, typename V, typename W>
+bool operator==(const std::pair<T, U> &x, const std::pair<V, W> &y) {
+  return x.first == y.first && x.second == y.second;
+}
+
+// Partial specialization of remove_const that propagates the removal through
+// std::pair.
+template <typename T, typename U>
+struct remove_const<pair<T, U> > {
+  typedef pair<typename remove_const<T>::type,
+               typename remove_const<U>::type> type;
+};
+
+} // namespace std
+
+namespace btree {
+
+// Select the first member of a pair.
+template <class _Pair>
+struct select1st : public std::unary_function<_Pair, typename _Pair::first_type> {
+  const typename _Pair::first_type& operator()(const _Pair& __x) const {
+    return __x.first;
+  }
+};
+
+// Utility class to provide an accessor for a key given a value. The default
+// behavior is to treat the value as a pair and return the first element.
+template <typename K, typename V>
+struct KeyOfValue {
+  typedef select1st<V> type;
+};
+
+template <typename T>
+struct identity {
+  inline const T& operator()(const T& t) const { return t; }
+};
+
+// Partial specialization of KeyOfValue class for when the key and value are
+// the same type such as in set<> and btree_set<>.
+template <typename K>
+struct KeyOfValue<K, K> {
+  typedef identity<K> type;
+};
+
+// Counts the number of occurances of "c" in a buffer.
+inline ptrdiff_t strcount(const char* buf_begin, const char* buf_end, char c) {
+  if (buf_begin == NULL)
+    return 0;
+  if (buf_end <= buf_begin)
+    return 0;
+  ptrdiff_t num = 0;
+  for (const char* bp = buf_begin; bp != buf_end; bp++) {
+    if (*bp == c)
+      num++;
+  }
+  return num;
+}
+
+// for when the string is not null-terminated.
+inline ptrdiff_t strcount(const char* buf, size_t len, char c) {
+  return strcount(buf, buf + len, c);
+}
+
+inline ptrdiff_t strcount(const std::string& buf, char c) {
+  return strcount(buf.c_str(), buf.size(), c);
+}
+
+// The base class for a sorted associative container checker. TreeType is the
+// container type to check and CheckerType is the container type to check
+// against. TreeType is expected to be btree_{set,map,multiset,multimap} and
+// CheckerType is expected to be {set,map,multiset,multimap}.
+template <typename TreeType, typename CheckerType>
+class base_checker {
+  typedef base_checker<TreeType, CheckerType> self_type;
+
+ public:
+  typedef typename TreeType::key_type key_type;
+  typedef typename TreeType::value_type value_type;
+  typedef typename TreeType::key_compare key_compare;
+  typedef typename TreeType::pointer pointer;
+  typedef typename TreeType::const_pointer const_pointer;
+  typedef typename TreeType::reference reference;
+  typedef typename TreeType::const_reference const_reference;
+  typedef typename TreeType::size_type size_type;
+  typedef typename TreeType::difference_type difference_type;
+  typedef typename TreeType::iterator iterator;
+  typedef typename TreeType::const_iterator const_iterator;
+  typedef typename TreeType::reverse_iterator reverse_iterator;
+  typedef typename TreeType::const_reverse_iterator const_reverse_iterator;
+
+ public:
+  // Default constructor.
+  base_checker()
+      : const_tree_(tree_) {
+  }
+  // Copy constructor.
+  base_checker(const self_type &x)
+      : tree_(x.tree_),
+        const_tree_(tree_),
+        checker_(x.checker_) {
+  }
+  // Range constructor.
+  template <typename InputIterator>
+  base_checker(InputIterator b, InputIterator e)
+      : tree_(b, e),
+        const_tree_(tree_),
+        checker_(b, e) {
+  }
+
+  // Iterator routines.
+  iterator begin() { return tree_.begin(); }
+  const_iterator begin() const { return tree_.begin(); }
+  iterator end() { return tree_.end(); }
+  const_iterator end() const { return tree_.end(); }
+  reverse_iterator rbegin() { return tree_.rbegin(); }
+  const_reverse_iterator rbegin() const { return tree_.rbegin(); }
+  reverse_iterator rend() { return tree_.rend(); }
+  const_reverse_iterator rend() const { return tree_.rend(); }
+
+  // Helper routines.
+  template <typename IterType, typename CheckerIterType>
+  IterType iter_check(
+      IterType tree_iter, CheckerIterType checker_iter) const {
+    if (tree_iter == tree_.end()) {
+      EXPECT_EQ(checker_iter, checker_.end());
+    } else {
+      EXPECT_EQ(*tree_iter, *checker_iter);
+    }
+    return tree_iter;
+  }
+  template <typename IterType, typename CheckerIterType>
+  IterType riter_check(
+      IterType tree_iter, CheckerIterType checker_iter) const {
+    if (tree_iter == tree_.rend()) {
+      EXPECT_EQ(checker_iter, checker_.rend());
+    } else {
+      EXPECT_EQ(*tree_iter, *checker_iter);
+    }
+    return tree_iter;
+  }
+  void value_check(const value_type &x) {
+    typename KeyOfValue<typename TreeType::key_type,
+        typename TreeType::value_type>::type key_of_value;
+    const key_type &key = key_of_value(x);
+    EXPECT_EQ(*find(key), x);
+    lower_bound(key);
+    upper_bound(key);
+    equal_range(key);
+    count(key);
+  }
+  void erase_check(const key_type &key) {
+    EXPECT_TRUE(tree_.find(key) == const_tree_.end());
+    EXPECT_TRUE(const_tree_.find(key) == tree_.end());
+    EXPECT_TRUE(tree_.equal_range(key).first ==
+                const_tree_.equal_range(key).second);
+  }
+
+  // Lookup routines.
+  iterator lower_bound(const key_type &key) {
+    return iter_check(tree_.lower_bound(key), checker_.lower_bound(key));
+  }
+  const_iterator lower_bound(const key_type &key) const {
+    return iter_check(tree_.lower_bound(key), checker_.lower_bound(key));
+  }
+  iterator upper_bound(const key_type &key) {
+    return iter_check(tree_.upper_bound(key), checker_.upper_bound(key));
+  }
+  const_iterator upper_bound(const key_type &key) const {
+    return iter_check(tree_.upper_bound(key), checker_.upper_bound(key));
+  }
+  std::pair<iterator,iterator> equal_range(const key_type &key) {
+    std::pair<typename CheckerType::iterator,
+        typename CheckerType::iterator> checker_res =
+        checker_.equal_range(key);
+    std::pair<iterator, iterator> tree_res = tree_.equal_range(key);
+    iter_check(tree_res.first, checker_res.first);
+    iter_check(tree_res.second, checker_res.second);
+    return tree_res;
+  }
+  std::pair<const_iterator,const_iterator> equal_range(const key_type &key) const {
+    std::pair<typename CheckerType::const_iterator,
+        typename CheckerType::const_iterator> checker_res =
+        checker_.equal_range(key);
+    std::pair<const_iterator, const_iterator> tree_res = tree_.equal_range(key);
+    iter_check(tree_res.first, checker_res.first);
+    iter_check(tree_res.second, checker_res.second);
+    return tree_res;
+  }
+  iterator find(const key_type &key) {
+    return iter_check(tree_.find(key), checker_.find(key));
+  }
+  const_iterator find(const key_type &key) const {
+    return iter_check(tree_.find(key), checker_.find(key));
+  }
+  size_type count(const key_type &key) const {
+    size_type res = checker_.count(key);
+    EXPECT_EQ(res, tree_.count(key));
+    return res;
+  }
+
+  // Assignment operator.
+  self_type& operator=(const self_type &x) {
+    tree_ = x.tree_;
+    checker_ = x.checker_;
+    return *this;
+  }
+
+  // Deletion routines.
+  int erase(const key_type &key) {
+    int size = tree_.size();
+    int res = checker_.erase(key);
+    EXPECT_EQ(res, tree_.count(key));
+    EXPECT_EQ(res, tree_.erase(key));
+    EXPECT_EQ(tree_.count(key), 0);
+    EXPECT_EQ(tree_.size(), size - res);
+    erase_check(key);
+    return res;
+  }
+  iterator erase(iterator iter) {
+    key_type key = iter.key();
+    int size = tree_.size();
+    int count = tree_.count(key);
+    typename CheckerType::iterator checker_iter = checker_.find(key);
+    for (iterator tmp(tree_.find(key)); tmp != iter; ++tmp) {
+      ++checker_iter;
+    }
+    typename CheckerType::iterator checker_next = checker_iter;
+    ++checker_next;
+    checker_.erase(checker_iter);
+    iter = tree_.erase(iter);
+    EXPECT_EQ(tree_.size(), checker_.size());
+    EXPECT_EQ(tree_.size(), size - 1);
+    EXPECT_EQ(tree_.count(key), count - 1);
+    if (count == 1) {
+      erase_check(key);
+    }
+    return iter_check(iter, checker_next);
+  }
+
+  void erase(iterator begin, iterator end) {
+    int size = tree_.size();
+    int count = distance(begin, end);
+    typename CheckerType::iterator checker_begin = checker_.find(begin.key());
+    for (iterator tmp(tree_.find(begin.key())); tmp != begin; ++tmp) {
+      ++checker_begin;
+    }
+    typename CheckerType::iterator checker_end =
+        end == tree_.end() ? checker_.end() : checker_.find(end.key());
+    if (end != tree_.end()) {
+      for (iterator tmp(tree_.find(end.key())); tmp != end; ++tmp) {
+        ++checker_end;
+      }
+    }
+    checker_.erase(checker_begin, checker_end);
+    tree_.erase(begin, end);
+    EXPECT_EQ(tree_.size(), checker_.size());
+    EXPECT_EQ(tree_.size(), size - count);
+  }
+
+  // Utility routines.
+  void clear() {
+    tree_.clear();
+    checker_.clear();
+  }
+  void swap(self_type &x) {
+    tree_.swap(x.tree_);
+    checker_.swap(x.checker_);
+  }
+
+  void verify() const {
+    tree_.verify();
+    EXPECT_EQ(tree_.size(), checker_.size());
+
+    // Move through the forward iterators using increment.
+    typename CheckerType::const_iterator
+        checker_iter(checker_.begin());
+    const_iterator tree_iter(tree_.begin());
+    for (; tree_iter != tree_.end();
+         ++tree_iter, ++checker_iter) {
+      EXPECT_EQ(*tree_iter, *checker_iter);
+    }
+
+    // Move through the forward iterators using decrement.
+    for (int n = tree_.size() - 1; n >= 0; --n) {
+      iter_check(tree_iter, checker_iter);
+      --tree_iter;
+      --checker_iter;
+    }
+    EXPECT_TRUE(tree_iter == tree_.begin());
+    EXPECT_TRUE(checker_iter == checker_.begin());
+
+    // Move through the reverse iterators using increment.
+    typename CheckerType::const_reverse_iterator
+        checker_riter(checker_.rbegin());
+    const_reverse_iterator tree_riter(tree_.rbegin());
+    for (; tree_riter != tree_.rend();
+         ++tree_riter, ++checker_riter) {
+      EXPECT_EQ(*tree_riter, *checker_riter);
+    }
+
+    // Move through the reverse iterators using decrement.
+    for (int n = tree_.size() - 1; n >= 0; --n) {
+      riter_check(tree_riter, checker_riter);
+      --tree_riter;
+      --checker_riter;
+    }
+    EXPECT_EQ(tree_riter, tree_.rbegin());
+    EXPECT_EQ(checker_riter, checker_.rbegin());
+  }
+
+  // Access to the underlying btree.
+  const TreeType& tree() const { return tree_; }
+
+  // Size routines.
+  size_type size() const {
+    EXPECT_EQ(tree_.size(), checker_.size());
+    return tree_.size();
+  }
+  size_type max_size() const { return tree_.max_size(); }
+  bool empty() const {
+    EXPECT_EQ(tree_.empty(), checker_.empty());
+    return tree_.empty();
+  }
+  size_type height() const { return tree_.height(); }
+  size_type internal_nodes() const { return tree_.internal_nodes(); }
+  size_type leaf_nodes() const { return tree_.leaf_nodes(); }
+  size_type nodes() const { return tree_.nodes(); }
+  size_type bytes_used() const { return tree_.bytes_used(); }
+  double fullness() const { return tree_.fullness(); }
+  double overhead() const { return tree_.overhead(); }
+
+ protected:
+  TreeType tree_;
+  const TreeType &const_tree_;
+  CheckerType checker_;
+};
+
+// A checker for unique sorted associative containers. TreeType is expected to
+// be btree_{set,map} and CheckerType is expected to be {set,map}.
+template <typename TreeType, typename CheckerType>
+class unique_checker : public base_checker<TreeType, CheckerType> {
+  typedef base_checker<TreeType, CheckerType> super_type;
+  typedef unique_checker<TreeType, CheckerType> self_type;
+
+ public:
+  typedef typename super_type::iterator iterator;
+  typedef typename super_type::value_type value_type;
+
+ public:
+  // Default constructor.
+  unique_checker()
+      : super_type() {
+  }
+  // Copy constructor.
+  unique_checker(const self_type &x)
+      : super_type(x) {
+  }
+  // Range constructor.
+  template <class InputIterator>
+  unique_checker(InputIterator b, InputIterator e)
+      : super_type(b, e) {
+  }
+
+  // Insertion routines.
+  std::pair<iterator,bool> insert(const value_type &x) {
+    int size = this->tree_.size();
+    std::pair<typename CheckerType::iterator,bool> checker_res =
+        this->checker_.insert(x);
+    std::pair<iterator,bool> tree_res = this->tree_.insert(x);
+    EXPECT_EQ(*tree_res.first, *checker_res.first);
+    EXPECT_EQ(tree_res.second, checker_res.second);
+    EXPECT_EQ(this->tree_.size(), this->checker_.size());
+    EXPECT_EQ(this->tree_.size(), size + tree_res.second);
+    return tree_res;
+  }
+  iterator insert(iterator position, const value_type &x) {
+    int size = this->tree_.size();
+    std::pair<typename CheckerType::iterator,bool> checker_res =
+        this->checker_.insert(x);
+    iterator tree_res = this->tree_.insert(position, x);
+    EXPECT_EQ(*tree_res, *checker_res.first);
+    EXPECT_EQ(this->tree_.size(), this->checker_.size());
+    EXPECT_EQ(this->tree_.size(), size + checker_res.second);
+    return tree_res;
+  }
+  template <typename InputIterator>
+  void insert(InputIterator b, InputIterator e) {
+    for (; b != e; ++b) {
+      insert(*b);
+    }
+  }
+};
+
+// A checker for multiple sorted associative containers. TreeType is expected
+// to be btree_{multiset,multimap} and CheckerType is expected to be
+// {multiset,multimap}.
+template <typename TreeType, typename CheckerType>
+class multi_checker : public base_checker<TreeType, CheckerType> {
+  typedef base_checker<TreeType, CheckerType> super_type;
+  typedef multi_checker<TreeType, CheckerType> self_type;
+
+ public:
+  typedef typename super_type::iterator iterator;
+  typedef typename super_type::value_type value_type;
+
+ public:
+  // Default constructor.
+  multi_checker()
+      : super_type() {
+  }
+  // Copy constructor.
+  multi_checker(const self_type &x)
+      : super_type(x) {
+  }
+  // Range constructor.
+  template <class InputIterator>
+  multi_checker(InputIterator b, InputIterator e)
+      : super_type(b, e) {
+  }
+
+  // Insertion routines.
+  iterator insert(const value_type &x) {
+    int size = this->tree_.size();
+    typename CheckerType::iterator checker_res = this->checker_.insert(x);
+    iterator tree_res = this->tree_.insert(x);
+    EXPECT_EQ(*tree_res, *checker_res);
+    EXPECT_EQ(this->tree_.size(), this->checker_.size());
+    EXPECT_EQ(this->tree_.size(), size + 1);
+    return tree_res;
+  }
+  iterator insert(iterator position, const value_type &x) {
+    int size = this->tree_.size();
+    typename CheckerType::iterator checker_res = this->checker_.insert(x);
+    iterator tree_res = this->tree_.insert(position, x);
+    EXPECT_EQ(*tree_res, *checker_res);
+    EXPECT_EQ(this->tree_.size(), this->checker_.size());
+    EXPECT_EQ(this->tree_.size(), size + 1);
+    return tree_res;
+  }
+  template <typename InputIterator>
+  void insert(InputIterator b, InputIterator e) {
+    for (; b != e; ++b) {
+      insert(*b);
+    }
+  }
+};
+
+char* GenerateDigits(char buf[16], int val, int maxval) {
+  EXPECT_LE(val, maxval);
+  int p = 15;
+  buf[p--] = 0;
+  while (maxval > 0) {
+    buf[p--] = '0' + (val % 10);
+    val /= 10;
+    maxval /= 10;
+  }
+  return buf + p + 1;
+}
+
+template <typename K>
+struct Generator {
+  int maxval;
+  Generator(int m)
+      : maxval(m) {
+  }
+  K operator()(int i) const {
+    EXPECT_LE(i, maxval);
+    return i;
+  }
+};
+
+template <>
+struct Generator<std::string> {
+  int maxval;
+  Generator(int m)
+      : maxval(m) {
+  }
+  std::string operator()(int i) const {
+    char buf[16];
+    return GenerateDigits(buf, i, maxval);
+  }
+};
+
+template <typename T, typename U>
+struct Generator<std::pair<T, U> > {
+  Generator<typename std::remove_const<T>::type> tgen;
+  Generator<typename std::remove_const<U>::type> ugen;
+
+  Generator(int m)
+      : tgen(m),
+        ugen(m) {
+  }
+  std::pair<T, U> operator()(int i) const {
+    return std::make_pair(tgen(i), ugen(i));
+  }
+};
+
+// Generate values for our tests and benchmarks. Value range is [0, maxval].
+const std::vector<int>& GenerateNumbers(int n, int maxval) {
+  static std::vector<int> values;
+  static std::set<int> unique_values;
+
+  if (values.size() < n) {
+
+    for (int i = values.size(); i < n; i++) {
+      int value;
+      do {
+        value = rand() % (maxval + 1);
+      } while (unique_values.find(value) != unique_values.end());
+
+      values.push_back(value);
+      unique_values.insert(value);
+    }
+  }
+
+  return values;
+}
+
+// Generates values in the range
+// [0, 4 * min(FLAGS_benchmark_values, FLAGS_test_values)]
+template <typename V>
+std::vector<V> GenerateValues(int n) {
+  int two_times_max = 2 * std::max(FLAGS_benchmark_values, FLAGS_test_values);
+  int four_times_max = 2 * two_times_max;
+  EXPECT_LE(n, two_times_max);
+  const std::vector<int> &nums = GenerateNumbers(n, four_times_max);
+  Generator<V> gen(four_times_max);
+  std::vector<V> vec;
+
+  for (int i = 0; i < n; i++) {
+    vec.push_back(gen(nums[i]));
+  }
+
+  return vec;
+}
+
+template <typename T, typename V>
+void DoTest(const char *name, T *b, const std::vector<V> &values) {
+  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
+
+  T &mutable_b = *b;
+  const T &const_b = *b;
+
+  // Test insert.
+  for (int i = 0; i < values.size(); ++i) {
+    mutable_b.insert(values[i]);
+    mutable_b.value_check(values[i]);
+  }
+  assert(mutable_b.size() == values.size());
+
+  const_b.verify();
+  printf("    %s fullness=%0.2f  overhead=%0.2f  bytes-per-value=%0.2f\n",
+         name, const_b.fullness(), const_b.overhead(),
+         double(const_b.bytes_used()) / const_b.size());
+
+  // Test copy constructor.
+  T b_copy(const_b);
+  EXPECT_EQ(b_copy.size(), const_b.size());
+  EXPECT_LE(b_copy.height(), const_b.height());
+  EXPECT_LE(b_copy.internal_nodes(), const_b.internal_nodes());
+  EXPECT_LE(b_copy.leaf_nodes(), const_b.leaf_nodes());
+  for (int i = 0; i < values.size(); ++i) {
+    EXPECT_EQ(*b_copy.find(key_of_value(values[i])), values[i]);
+  }
+
+  // Test range constructor.
+  T b_range(const_b.begin(), const_b.end());
+  EXPECT_EQ(b_range.size(), const_b.size());
+  EXPECT_LE(b_range.height(), const_b.height());
+  EXPECT_LE(b_range.internal_nodes(), const_b.internal_nodes());
+  EXPECT_LE(b_range.leaf_nodes(), const_b.leaf_nodes());
+  for (int i = 0; i < values.size(); ++i) {
+    EXPECT_EQ(*b_range.find(key_of_value(values[i])), values[i]);
+  }
+
+  // Test range insertion for values that already exist.
+  b_range.insert(b_copy.begin(), b_copy.end());
+  b_range.verify();
+
+  // Test range insertion for new values.
+  b_range.clear();
+  b_range.insert(b_copy.begin(), b_copy.end());
+  EXPECT_EQ(b_range.size(), b_copy.size());
+  EXPECT_EQ(b_range.height(), b_copy.height());
+  EXPECT_EQ(b_range.internal_nodes(), b_copy.internal_nodes());
+  EXPECT_EQ(b_range.leaf_nodes(), b_copy.leaf_nodes());
+  for (int i = 0; i < values.size(); ++i) {
+    EXPECT_EQ(*b_range.find(key_of_value(values[i])), values[i]);
+  }
+
+  // Test assignment to self. Nothing should change.
+  b_range.operator=(b_range);
+  EXPECT_EQ(b_range.size(), b_copy.size());
+  EXPECT_EQ(b_range.height(), b_copy.height());
+  EXPECT_EQ(b_range.internal_nodes(), b_copy.internal_nodes());
+  EXPECT_EQ(b_range.leaf_nodes(), b_copy.leaf_nodes());
+
+  // Test assignment of new values.
+  b_range.clear();
+  b_range = b_copy;
+  EXPECT_EQ(b_range.size(), b_copy.size());
+  EXPECT_EQ(b_range.height(), b_copy.height());
+  EXPECT_EQ(b_range.internal_nodes(), b_copy.internal_nodes());
+  EXPECT_EQ(b_range.leaf_nodes(), b_copy.leaf_nodes());
+
+  // Test swap.
+  b_range.clear();
+  b_range.swap(b_copy);
+  EXPECT_EQ(b_copy.size(), 0);
+  EXPECT_EQ(b_range.size(), const_b.size());
+  for (int i = 0; i < values.size(); ++i) {
+    EXPECT_EQ(*b_range.find(key_of_value(values[i])), values[i]);
+  }
+  b_range.swap(b_copy);
+
+  // Test erase via values.
+  for (int i = 0; i < values.size(); ++i) {
+    mutable_b.erase(key_of_value(values[i]));
+    // Erasing a non-existent key should have no effect.
+    EXPECT_EQ(mutable_b.erase(key_of_value(values[i])), 0);
+  }
+
+  const_b.verify();
+  EXPECT_EQ(const_b.internal_nodes(), 0);
+  EXPECT_EQ(const_b.leaf_nodes(), 0);
+  EXPECT_EQ(const_b.size(), 0);
+
+  // Test erase via iterators.
+  mutable_b = b_copy;
+  for (int i = 0; i < values.size(); ++i) {
+    mutable_b.erase(mutable_b.find(key_of_value(values[i])));
+  }
+
+  const_b.verify();
+  EXPECT_EQ(const_b.internal_nodes(), 0);
+  EXPECT_EQ(const_b.leaf_nodes(), 0);
+  EXPECT_EQ(const_b.size(), 0);
+
+  // Test insert with hint.
+  for (int i = 0; i < values.size(); i++) {
+    mutable_b.insert(mutable_b.upper_bound(key_of_value(values[i])), values[i]);
+  }
+
+  const_b.verify();
+
+  // Test dumping of the btree to an ostream. There should be 1 line for each
+  // value.
+  std::stringstream strm;
+  strm << mutable_b.tree();
+  EXPECT_EQ(mutable_b.size(), strcount(strm.str(), '\n'));
+
+  // Test range erase.
+  mutable_b.erase(mutable_b.begin(), mutable_b.end());
+  EXPECT_EQ(mutable_b.size(), 0);
+  const_b.verify();
+
+  // First half.
+  mutable_b = b_copy;
+  typename T::iterator mutable_iter_end = mutable_b.begin();
+  for (int i = 0; i < values.size() / 2; ++i) ++mutable_iter_end;
+  mutable_b.erase(mutable_b.begin(), mutable_iter_end);
+  EXPECT_EQ(mutable_b.size(), values.size() - values.size() / 2);
+  const_b.verify();
+
+  // Second half.
+  mutable_b = b_copy;
+  typename T::iterator mutable_iter_begin = mutable_b.begin();
+  for (int i = 0; i < values.size() / 2; ++i) ++mutable_iter_begin;
+  mutable_b.erase(mutable_iter_begin, mutable_b.end());
+  EXPECT_EQ(mutable_b.size(), values.size() / 2);
+  const_b.verify();
+
+  // Second quarter.
+  mutable_b = b_copy;
+  mutable_iter_begin = mutable_b.begin();
+  for (int i = 0; i < values.size() / 4; ++i) ++mutable_iter_begin;
+  mutable_iter_end = mutable_iter_begin;
+  for (int i = 0; i < values.size() / 4; ++i) ++mutable_iter_end;
+  mutable_b.erase(mutable_iter_begin, mutable_iter_end);
+  EXPECT_EQ(mutable_b.size(), values.size() - values.size() / 4);
+  const_b.verify();
+
+  mutable_b.clear();
+}
+
+template <typename T>
+void ConstTest() {
+  typedef typename T::value_type value_type;
+  typename KeyOfValue<typename T::key_type, value_type>::type key_of_value;
+
+  T mutable_b;
+  const T &const_b = mutable_b;
+
+  // Insert a single value into the container and test looking it up.
+  value_type value = Generator<value_type>(2)(2);
+  mutable_b.insert(value);
+  EXPECT_TRUE(mutable_b.find(key_of_value(value)) != const_b.end());
+  EXPECT_TRUE(const_b.find(key_of_value(value)) != mutable_b.end());
+  EXPECT_EQ(*const_b.lower_bound(key_of_value(value)), value);
+  EXPECT_TRUE(const_b.upper_bound(key_of_value(value)) == const_b.end());
+  EXPECT_EQ(*const_b.equal_range(key_of_value(value)).first, value);
+
+  // We can only create a non-const iterator from a non-const container.
+  typename T::iterator mutable_iter(mutable_b.begin());
+  EXPECT_TRUE(mutable_iter == const_b.begin());
+  EXPECT_TRUE(mutable_iter != const_b.end());
+  EXPECT_TRUE(const_b.begin() == mutable_iter);
+  EXPECT_TRUE(const_b.end() != mutable_iter);
+  typename T::reverse_iterator mutable_riter(mutable_b.rbegin());
+  EXPECT_TRUE(mutable_riter == const_b.rbegin());
+  EXPECT_TRUE(mutable_riter != const_b.rend());
+  EXPECT_TRUE(const_b.rbegin() == mutable_riter);
+  EXPECT_TRUE(const_b.rend() != mutable_riter);
+
+  // We can create a const iterator from a non-const iterator.
+  typename T::const_iterator const_iter(mutable_iter);
+  EXPECT_TRUE(const_iter == mutable_b.begin());
+  EXPECT_TRUE(const_iter != mutable_b.end());
+  EXPECT_TRUE(mutable_b.begin() == const_iter);
+  EXPECT_TRUE(mutable_b.end() != const_iter);
+  typename T::const_reverse_iterator const_riter(mutable_riter);
+  EXPECT_EQ(const_riter, mutable_b.rbegin());
+  EXPECT_TRUE(const_riter != mutable_b.rend());
+  EXPECT_EQ(mutable_b.rbegin(), const_riter);
+  EXPECT_TRUE(mutable_b.rend() != const_riter);
+
+  // Make sure various methods can be invoked on a const container.
+  const_b.verify();
+  EXPECT_FALSE(const_b.empty());
+  EXPECT_EQ(const_b.size(), 1);
+  EXPECT_GT(const_b.max_size(), 0);
+  EXPECT_EQ(const_b.height(), 1);
+  EXPECT_EQ(const_b.count(key_of_value(value)), 1);
+  EXPECT_EQ(const_b.internal_nodes(), 0);
+  EXPECT_EQ(const_b.leaf_nodes(), 1);
+  EXPECT_EQ(const_b.nodes(), 1);
+  EXPECT_GT(const_b.bytes_used(), 0);
+  EXPECT_GT(const_b.fullness(), 0);
+  EXPECT_GT(const_b.overhead(), 0);
+}
+
+template <typename T, typename C>
+void BtreeTest() {
+  ConstTest<T>();
+
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  std::vector<V> random_values = GenerateValues<V>(FLAGS_test_values);
+
+  unique_checker<T, C> container;
+
+  // Test key insertion/deletion in sorted order.
+  std::vector<V> sorted_values(random_values);
+  sort(sorted_values.begin(), sorted_values.end());
+  DoTest("sorted:    ", &container, sorted_values);
+
+  // Test key insertion/deletion in reverse sorted order.
+  reverse(sorted_values.begin(), sorted_values.end());
+  DoTest("rsorted:   ", &container, sorted_values);
+
+  // Test key insertion/deletion in random order.
+  DoTest("random:    ", &container, random_values);
+}
+
+template <typename T, typename C>
+void BtreeMultiTest() {
+  ConstTest<T>();
+
+  typedef typename std::remove_const<typename T::value_type>::type V;
+  const std::vector<V>& random_values = GenerateValues<V>(FLAGS_test_values);
+
+  multi_checker<T, C> container;
+
+  // Test keys in sorted order.
+  std::vector<V> sorted_values(random_values);
+  sort(sorted_values.begin(), sorted_values.end());
+  DoTest("sorted:    ", &container, sorted_values);
+
+  // Test keys in reverse sorted order.
+  reverse(sorted_values.begin(), sorted_values.end());
+  DoTest("rsorted:   ", &container, sorted_values);
+
+  // Test keys in random order.
+  DoTest("random:    ", &container, random_values);
+
+  // Test keys in random order w/ duplicates.
+  std::vector<V> duplicate_values(random_values);
+  duplicate_values.insert(
+      duplicate_values.end(), random_values.begin(), random_values.end());
+  DoTest("duplicates:", &container, duplicate_values);
+
+  // Test all identical keys.
+  std::vector<V> identical_values(100);
+  fill(identical_values.begin(), identical_values.end(), Generator<V>(2)(2));
+  DoTest("identical: ", &container, identical_values);
+}
+
+template <typename T, typename Alloc = std::allocator<T> >
+class TestAllocator : public Alloc {
+ public:
+  typedef typename Alloc::pointer pointer;
+  typedef typename Alloc::size_type size_type;
+
+  TestAllocator() : bytes_used_(NULL) { }
+  TestAllocator(int64_t *bytes_used) : bytes_used_(bytes_used) { }
+
+  // Constructor used for rebinding
+  template <class U>
+  TestAllocator(const TestAllocator<U>& x)
+      : Alloc(x),
+        bytes_used_(x.bytes_used()) {
+  }
+
+  pointer allocate(size_type n, std::allocator<void>::const_pointer hint = 0) {
+    EXPECT_TRUE(bytes_used_ != NULL);
+    *bytes_used_ += n * sizeof(T);
+    return Alloc::allocate(n, hint);
+  }
+
+  void deallocate(pointer p, size_type n) {
+    Alloc::deallocate(p, n);
+    EXPECT_TRUE(bytes_used_ != NULL);
+    *bytes_used_ -= n * sizeof(T);
+  }
+
+  // Rebind allows an allocator<T> to be used for a different type
+  template <class U> struct rebind {
+    typedef TestAllocator<U, typename Alloc::template rebind<U>::other> other;
+  };
+
+  int64_t* bytes_used() const { return bytes_used_; }
+
+ private:
+  int64_t *bytes_used_;
+};
+
+template <typename T>
+void BtreeAllocatorTest() {
+  typedef typename T::value_type value_type;
+
+  int64_t alloc1 = 0;
+  int64_t alloc2 = 0;
+  T b1(typename T::key_compare(), &alloc1);
+  T b2(typename T::key_compare(), &alloc2);
+
+  // This should swap the allocators!
+  swap(b1, b2);
+
+  for (int i = 0; i < 1000; i++) {
+    b1.insert(Generator<value_type>(1000)(i));
+  }
+
+  // We should have allocated out of alloc2!
+  EXPECT_LE(b1.bytes_used(), alloc2 + sizeof(b1));
+  EXPECT_GT(alloc2, alloc1);
+}
+
+template <typename T>
+void BtreeMapTest() {
+  typedef typename T::value_type value_type;
+  typedef typename T::mapped_type mapped_type;
+
+  mapped_type m = Generator<mapped_type>(0)(0);
+  (void) m;
+
+  T b;
+
+  // Verify we can insert using operator[].
+  for (int i = 0; i < 1000; i++) {
+    value_type v = Generator<value_type>(1000)(i);
+    b[v.first] = v.second;
+  }
+  EXPECT_EQ(b.size(), 1000);
+
+  // Test whether we can use the "->" operator on iterators and
+  // reverse_iterators. This stresses the btree_map_params::pair_pointer
+  // mechanism.
+  EXPECT_EQ(b.begin()->first, Generator<value_type>(1000)(0).first);
+  EXPECT_EQ(b.begin()->second, Generator<value_type>(1000)(0).second);
+  EXPECT_EQ(b.rbegin()->first, Generator<value_type>(1000)(999).first);
+  EXPECT_EQ(b.rbegin()->second, Generator<value_type>(1000)(999).second);
+}
+
+template <typename T>
+void BtreeMultiMapTest() {
+  typedef typename T::mapped_type mapped_type;
+  mapped_type m = Generator<mapped_type>(0)(0);
+  (void) m;
+}
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_BTREE_TEST_H__
diff --git a/xdelta3/cpp-btree/btree_test_flags.cc b/xdelta3/cpp-btree/btree_test_flags.cc
new file mode 100644
index 0000000..bf608a9
--- /dev/null
+++ b/xdelta3/cpp-btree/btree_test_flags.cc
@@ -0,0 +1,20 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "gflags/gflags.h"
+
+DEFINE_int32(test_values, 10000,
+             "The number of values to use for tests.");
+DEFINE_int32(benchmark_values, 1000000,
+             "The number of values to use for benchmarks.");
diff --git a/xdelta3/cpp-btree/safe_btree.h b/xdelta3/cpp-btree/safe_btree.h
new file mode 100644
index 0000000..2d85c70
--- /dev/null
+++ b/xdelta3/cpp-btree/safe_btree.h
@@ -0,0 +1,395 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// A safe_btree<> wraps around a btree<> and removes the caveat that insertion
+// and deletion invalidate iterators. A safe_btree<> maintains a generation
+// number that is incremented on every mutation. A safe_btree<>::iterator keeps
+// a pointer to the safe_btree<> it came from, the generation of the tree when
+// it was last validated and the key the underlying btree<>::iterator points
+// to. If an iterator is accessed and its generation differs from the tree
+// generation it is revalidated.
+//
+// References and pointers returned by safe_btree iterators are not safe.
+//
+// See the incorrect usage examples mentioned in safe_btree_set.h and
+// safe_btree_map.h.
+
+#ifndef UTIL_BTREE_SAFE_BTREE_H__
+#define UTIL_BTREE_SAFE_BTREE_H__
+
+#include <stddef.h>
+#include <iosfwd>
+#include <utility>
+
+#include "btree.h"
+
+namespace btree {
+
+template <typename Tree, typename Iterator>
+class safe_btree_iterator {
+ public:
+  typedef typename Iterator::key_type key_type;
+  typedef typename Iterator::value_type value_type;
+  typedef typename Iterator::size_type size_type;
+  typedef typename Iterator::difference_type difference_type;
+  typedef typename Iterator::pointer pointer;
+  typedef typename Iterator::reference reference;
+  typedef typename Iterator::const_pointer const_pointer;
+  typedef typename Iterator::const_reference const_reference;
+  typedef typename Iterator::iterator_category iterator_category;
+  typedef typename Tree::iterator iterator;
+  typedef typename Tree::const_iterator const_iterator;
+  typedef safe_btree_iterator<Tree, Iterator> self_type;
+
+  void update() const {
+    if (iter_ != tree_->internal_btree()->end()) {
+      // A positive generation indicates a valid key.
+      generation_ = tree_->generation();
+      key_ = iter_.key();
+    } else {
+      // Use a negative generation to indicate iter_ points to end().
+      generation_ = -tree_->generation();
+    }
+  }
+
+ public:
+  safe_btree_iterator()
+      : generation_(0),
+        key_(),
+        iter_(),
+        tree_(NULL) {
+  }
+  safe_btree_iterator(const iterator &x)
+      : generation_(x.generation()),
+        key_(x.key()),
+        iter_(x.iter()),
+        tree_(x.tree()) {
+  }
+  safe_btree_iterator(Tree *tree, const Iterator &iter)
+      : generation_(),
+        key_(),
+        iter_(iter),
+        tree_(tree) {
+    update();
+  }
+
+  Tree* tree() const { return tree_; }
+  int64_t generation() const { return generation_; }
+
+  Iterator* mutable_iter() const {
+    if (generation_ != tree_->generation()) {
+      if (generation_ > 0) {
+        // This does the wrong thing for a multi{set,map}. If my iter was
+        // pointing to the 2nd of 2 values with the same key, then this will
+        // reset it to point to the first. This is why we don't provide a
+        // safe_btree_multi{set,map}.
+        iter_ = tree_->internal_btree()->lower_bound(key_);
+        update();
+      } else if (-generation_ != tree_->generation()) {
+        iter_ = tree_->internal_btree()->end();
+        generation_ = -tree_->generation();
+      }
+    }
+    return &iter_;
+  }
+  const Iterator& iter() const {
+    return *mutable_iter();
+  }
+
+  // Equality/inequality operators.
+  bool operator==(const const_iterator &x) const {
+    return iter() == x.iter();
+  }
+  bool operator!=(const const_iterator &x) const {
+    return iter() != x.iter();
+  }
+
+  // Accessors for the key/value the iterator is pointing at.
+  const key_type& key() const {
+    return key_;
+  }
+  // This reference value is potentially invalidated by any non-const
+  // method on the tree; it is NOT safe.
+  reference operator*() const {
+    assert(generation_ > 0);
+    return iter().operator*();
+  }
+  // This pointer value is potentially invalidated by any non-const
+  // method on the tree; it is NOT safe.
+  pointer operator->() const {
+    assert(generation_ > 0);
+    return iter().operator->();
+  }
+
+  // Increment/decrement operators.
+  self_type& operator++() {
+    ++(*mutable_iter());
+    update();
+    return *this;
+  }
+  self_type& operator--() {
+    --(*mutable_iter());
+    update();
+    return *this;
+  }
+  self_type operator++(int) {
+    self_type tmp = *this;
+    ++*this;
+    return tmp;
+  }
+  self_type operator--(int) {
+    self_type tmp = *this;
+    --*this;
+    return tmp;
+  }
+
+ private:
+  // The generation of the tree when "iter" was updated.
+  mutable int64_t generation_;
+  // The key the iterator points to.
+  mutable key_type key_;
+  // The underlying iterator.
+  mutable Iterator iter_;
+  // The tree the iterator is associated with.
+  Tree *tree_;
+};
+
+template <typename Params>
+class safe_btree {
+  typedef safe_btree<Params> self_type;
+
+  typedef btree<Params> btree_type;
+  typedef typename btree_type::iterator tree_iterator;
+  typedef typename btree_type::const_iterator tree_const_iterator;
+
+ public:
+  typedef typename btree_type::params_type params_type;
+  typedef typename btree_type::key_type key_type;
+  typedef typename btree_type::data_type data_type;
+  typedef typename btree_type::mapped_type mapped_type;
+  typedef typename btree_type::value_type value_type;
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+  typedef typename btree_type::pointer pointer;
+  typedef typename btree_type::const_pointer const_pointer;
+  typedef typename btree_type::reference reference;
+  typedef typename btree_type::const_reference const_reference;
+  typedef typename btree_type::size_type size_type;
+  typedef typename btree_type::difference_type difference_type;
+  typedef safe_btree_iterator<self_type, tree_iterator> iterator;
+  typedef safe_btree_iterator<
+    const self_type, tree_const_iterator> const_iterator;
+  typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
+  typedef std::reverse_iterator<iterator> reverse_iterator;
+
+ public:
+  // Default constructor.
+  safe_btree(const key_compare &comp, const allocator_type &alloc)
+      : tree_(comp, alloc),
+        generation_(1) {
+  }
+
+  // Copy constructor.
+  safe_btree(const self_type &x)
+      : tree_(x.tree_),
+        generation_(1) {
+  }
+
+  iterator begin() {
+    return iterator(this, tree_.begin());
+  }
+  const_iterator begin() const {
+    return const_iterator(this, tree_.begin());
+  }
+  iterator end() {
+    return iterator(this, tree_.end());
+  }
+  const_iterator end() const {
+    return const_iterator(this, tree_.end());
+  }
+  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());
+  }
+
+  // Lookup routines.
+  iterator lower_bound(const key_type &key) {
+    return iterator(this, tree_.lower_bound(key));
+  }
+  const_iterator lower_bound(const key_type &key) const {
+    return const_iterator(this, tree_.lower_bound(key));
+  }
+  iterator upper_bound(const key_type &key) {
+    return iterator(this, tree_.upper_bound(key));
+  }
+  const_iterator upper_bound(const key_type &key) const {
+    return const_iterator(this, tree_.upper_bound(key));
+  }
+  std::pair<iterator, iterator> equal_range(const key_type &key) {
+    std::pair<tree_iterator, tree_iterator> p = tree_.equal_range(key);
+    return std::make_pair(iterator(this, p.first),
+                     iterator(this, p.second));
+  }
+  std::pair<const_iterator, const_iterator> equal_range(const key_type &key) const {
+    std::pair<tree_const_iterator, tree_const_iterator> p = tree_.equal_range(key);
+    return std::make_pair(const_iterator(this, p.first),
+                     const_iterator(this, p.second));
+  }
+  iterator find_unique(const key_type &key) {
+    return iterator(this, tree_.find_unique(key));
+  }
+  const_iterator find_unique(const key_type &key) const {
+    return const_iterator(this, tree_.find_unique(key));
+  }
+  iterator find_multi(const key_type &key) {
+    return iterator(this, tree_.find_multi(key));
+  }
+  const_iterator find_multi(const key_type &key) const {
+    return const_iterator(this, tree_.find_multi(key));
+  }
+  size_type count_unique(const key_type &key) const {
+    return tree_.count_unique(key);
+  }
+  size_type count_multi(const key_type &key) const {
+    return tree_.count_multi(key);
+  }
+
+  // Insertion routines.
+  template <typename ValuePointer>
+  std::pair<iterator, bool> insert_unique(const key_type &key, ValuePointer value) {
+    std::pair<tree_iterator, bool> p = tree_.insert_unique(key, value);
+    generation_ += p.second;
+    return std::make_pair(iterator(this, p.first), p.second);
+  }
+  std::pair<iterator, bool> insert_unique(const value_type &v) {
+    std::pair<tree_iterator, bool> p = tree_.insert_unique(v);
+    generation_ += p.second;
+    return std::make_pair(iterator(this, p.first), p.second);
+  }
+  iterator insert_unique(iterator position, const value_type &v) {
+    tree_iterator tree_pos = position.iter();
+    ++generation_;
+    return iterator(this, tree_.insert_unique(tree_pos, v));
+  }
+  template <typename InputIterator>
+  void insert_unique(InputIterator b, InputIterator e) {
+    for (; b != e; ++b) {
+      insert_unique(*b);
+    }
+  }
+  iterator insert_multi(const value_type &v) {
+    ++generation_;
+    return iterator(this, tree_.insert_multi(v));
+  }
+  iterator insert_multi(iterator position, const value_type &v) {
+    tree_iterator tree_pos = position.iter();
+    ++generation_;
+    return iterator(this, tree_.insert_multi(tree_pos, v));
+  }
+  template <typename InputIterator>
+  void insert_multi(InputIterator b, InputIterator e) {
+    for (; b != e; ++b) {
+      insert_multi(*b);
+    }
+  }
+  self_type& operator=(const self_type &x) {
+    if (&x == this) {
+      // Don't copy onto ourselves.
+      return *this;
+    }
+    ++generation_;
+    tree_ = x.tree_;
+    return *this;
+  }
+
+  // Deletion routines.
+  void erase(const iterator &begin, const iterator &end) {
+    tree_.erase(begin.iter(), end.iter());
+    ++generation_;
+  }
+  // Erase the specified iterator from the btree. The iterator must be valid
+  // (i.e. not equal to end()).  Return an iterator pointing to the node after
+  // the one that was erased (or end() if none exists).
+  iterator erase(iterator iter) {
+    tree_iterator res = tree_.erase(iter.iter());
+    ++generation_;
+    return iterator(this, res);
+  }
+  int erase_unique(const key_type &key) {
+    int res = tree_.erase_unique(key);
+    generation_ += res;
+    return res;
+  }
+  int erase_multi(const key_type &key) {
+    int res = tree_.erase_multi(key);
+    generation_ += res;
+    return res;
+  }
+
+  // Access to the underlying btree.
+  btree_type* internal_btree() { return &tree_; }
+  const btree_type* internal_btree() const { return &tree_; }
+
+  // Utility routines.
+  void clear() {
+    ++generation_;
+    tree_.clear();
+  }
+  void swap(self_type &x) {
+    ++generation_;
+    ++x.generation_;
+    tree_.swap(x.tree_);
+  }
+  void dump(std::ostream &os) const {
+    tree_.dump(os);
+  }
+  void verify() const {
+    tree_.verify();
+  }
+  int64_t generation() const {
+    return generation_;
+  }
+  key_compare key_comp() const { return tree_.key_comp(); }
+
+  // Size routines.
+  size_type size() const { return tree_.size(); }
+  size_type max_size() const { return tree_.max_size(); }
+  bool empty() const { return tree_.empty(); }
+  size_type height() const { return tree_.height(); }
+  size_type internal_nodes() const { return tree_.internal_nodes(); }
+  size_type leaf_nodes() const { return tree_.leaf_nodes(); }
+  size_type nodes() const { return tree_.nodes(); }
+  size_type bytes_used() const { return tree_.bytes_used(); }
+  static double average_bytes_per_value() {
+    return btree_type::average_bytes_per_value();
+  }
+  double fullness() const { return tree_.fullness(); }
+  double overhead() const { return tree_.overhead(); }
+
+ private:
+  btree_type tree_;
+  int64_t generation_;
+};
+
+}  // namespace btree
+
+#endif  // UTIL_BTREE_SAFE_BTREE_H__
diff --git a/xdelta3/cpp-btree/safe_btree_map.h b/xdelta3/cpp-btree/safe_btree_map.h
new file mode 100644
index 0000000..a0668f1
--- /dev/null
+++ b/xdelta3/cpp-btree/safe_btree_map.h
@@ -0,0 +1,89 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// The safe_btree_map<> is like btree_map<> except that it removes the caveat
+// about insertion and deletion invalidating existing iterators at a small cost
+// in making iterators larger and slower.
+//
+// Revalidation occurs whenever an iterator is accessed.  References
+// and pointers returned by safe_btree_map<> iterators are not stable,
+// they are potentially invalidated by any non-const method on the map.
+//
+// BEGIN INCORRECT EXAMPLE
+//   for (auto i = safe_map->begin(); i != safe_map->end(); ++i) {
+//     const T *value = &i->second;  // DO NOT DO THIS
+//     [code that modifies safe_map and uses value];
+//   }
+// END INCORRECT EXAMPLE
+#ifndef UTIL_BTREE_SAFE_BTREE_MAP_H__
+#define UTIL_BTREE_SAFE_BTREE_MAP_H__
+
+#include <functional>
+#include <memory>
+#include <utility>
+
+#include "btree_container.h"
+#include "btree_map.h"
+#include "safe_btree.h"
+
+namespace btree {
+
+// The safe_btree_map class is needed mainly for its constructors.
+template <typename Key, typename Value,
+          typename Compare = std::less<Key>,
+          typename Alloc = std::allocator<std::pair<const Key, Value> >,
+          int TargetNodeSize = 256>
+class safe_btree_map : public btree_map_container<
+  safe_btree<btree_map_params<Key, Value, Compare, Alloc, TargetNodeSize> > > {
+
+  typedef safe_btree_map<Key, Value, Compare, Alloc, TargetNodeSize> self_type;
+  typedef btree_map_params<
+    Key, Value, Compare, Alloc, TargetNodeSize> params_type;
+  typedef safe_btree<params_type> btree_type;
+  typedef btree_map_container<btree_type> super_type;
+
+ public:
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+
+ public:
+  // Default constructor.
+  safe_btree_map(const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  safe_btree_map(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  safe_btree_map(InputIterator b, InputIterator e,
+                 const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+};
+
+template <typename K, typename V, typename C, typename A, int N>
+inline void swap(safe_btree_map<K, V, C, A, N> &x,
+                 safe_btree_map<K, V, C, A, N> &y) {
+  x.swap(y);
+}
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_SAFE_BTREE_MAP_H__
diff --git a/xdelta3/cpp-btree/safe_btree_set.h b/xdelta3/cpp-btree/safe_btree_set.h
new file mode 100644
index 0000000..a6cd541
--- /dev/null
+++ b/xdelta3/cpp-btree/safe_btree_set.h
@@ -0,0 +1,88 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// The safe_btree_set<> is like btree_set<> except that it removes the caveat
+// about insertion and deletion invalidating existing iterators at a small cost
+// in making iterators larger and slower.
+//
+// Revalidation occurs whenever an iterator is accessed.  References
+// and pointers returned by safe_btree_map<> iterators are not stable,
+// they are potentially invalidated by any non-const method on the set.
+//
+// BEGIN INCORRECT EXAMPLE
+//   for (auto i = safe_set->begin(); i != safe_set->end(); ++i) {
+//     const T &value = *i;  // DO NOT DO THIS
+//     [code that modifies safe_set and uses value];
+//   }
+// END INCORRECT EXAMPLE
+
+#ifndef UTIL_BTREE_SAFE_BTREE_SET_H__
+#define UTIL_BTREE_SAFE_BTREE_SET_H__
+
+#include <functional>
+#include <memory>
+
+#include "btree_container.h"
+#include "btree_set.h"
+#include "safe_btree.h"
+
+namespace btree {
+
+// The safe_btree_set class is needed mainly for its constructors.
+template <typename Key,
+          typename Compare = std::less<Key>,
+          typename Alloc = std::allocator<Key>,
+          int TargetNodeSize = 256>
+class safe_btree_set : public btree_unique_container<
+  safe_btree<btree_set_params<Key, Compare, Alloc, TargetNodeSize> > > {
+
+  typedef safe_btree_set<Key, Compare, Alloc, TargetNodeSize> self_type;
+  typedef btree_set_params<Key, Compare, Alloc, TargetNodeSize> params_type;
+  typedef safe_btree<params_type> btree_type;
+  typedef btree_unique_container<btree_type> super_type;
+
+ public:
+  typedef typename btree_type::key_compare key_compare;
+  typedef typename btree_type::allocator_type allocator_type;
+
+ public:
+  // Default constructor.
+  safe_btree_set(const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(comp, alloc) {
+  }
+
+  // Copy constructor.
+  safe_btree_set(const self_type &x)
+      : super_type(x) {
+  }
+
+  // Range constructor.
+  template <class InputIterator>
+  safe_btree_set(InputIterator b, InputIterator e,
+                 const key_compare &comp = key_compare(),
+                 const allocator_type &alloc = allocator_type())
+      : super_type(b, e, comp, alloc) {
+  }
+};
+
+template <typename K, typename C, typename A, int N>
+inline void swap(safe_btree_set<K, C, A, N> &x,
+                 safe_btree_set<K, C, A, N> &y) {
+  x.swap(y);
+}
+
+} // namespace btree
+
+#endif  // UTIL_BTREE_SAFE_BTREE_SET_H__
diff --git a/xdelta3/cpp-btree/safe_btree_test.cc b/xdelta3/cpp-btree/safe_btree_test.cc
new file mode 100644
index 0000000..0d77ae0
--- /dev/null
+++ b/xdelta3/cpp-btree/safe_btree_test.cc
@@ -0,0 +1,116 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// TODO(pmattis): Add some tests that iterators are not invalidated by
+// insertion and deletion.
+
+#include <functional>
+#include <map>
+#include <set>
+#include <string>
+#include <utility>
+
+#include "gtest/gtest.h"
+#include "btree_test.h"
+#include "safe_btree_map.h"
+#include "safe_btree_set.h"
+
+class UnsafeArena;
+
+namespace btree {
+namespace {
+
+template <typename K, int N>
+void SetTest() {
+  typedef TestAllocator<K> TestAlloc;
+  BtreeTest<safe_btree_set<K, std::less<K>, std::allocator<K>, N>, std::set<K> >();
+  BtreeAllocatorTest<safe_btree_set<K, std::less<K>, TestAlloc, N> >();
+}
+
+template <typename K, int N>
+void MapTest() {
+  typedef TestAllocator<K> TestAlloc;
+  BtreeTest<safe_btree_map<K, K, std::less<K>, std::allocator<K>, N>, std::map<K, K> >();
+  BtreeAllocatorTest<safe_btree_map<K, K, std::less<K>, TestAlloc, N> >();
+  BtreeMapTest<safe_btree_map<K, K, std::less<K>, std::allocator<K>, N> >();
+}
+
+TEST(SafeBtree, set_int32_32)   { SetTest<int32_t, 32>(); }
+TEST(SafeBtree, set_int32_64)   { SetTest<int32_t, 64>(); }
+TEST(SafeBtree, set_int32_128)  { SetTest<int32_t, 128>(); }
+TEST(SafeBtree, set_int32_256)  { SetTest<int32_t, 256>(); }
+TEST(SafeBtree, set_int64_256)  { SetTest<int64_t, 256>(); }
+TEST(SafeBtree, set_string_256) { SetTest<std::string, 256>(); }
+TEST(SafeBtree, set_pair_256)   { SetTest<std::pair<int, int>, 256>(); }
+TEST(SafeBtree, map_int32_256)  { MapTest<int32_t, 256>(); }
+TEST(SafeBtree, map_int64_256)  { MapTest<int64_t, 256>(); }
+TEST(SafeBtree, map_string_256) { MapTest<std::string, 256>(); }
+TEST(SafeBtree, map_pair_256)   { MapTest<std::pair<int, int>, 256>(); }
+
+TEST(SafeBtree, Comparison) {
+  const int kSetSize = 1201;
+  safe_btree_set<int64_t> my_set;
+  for (int i = 0; i < kSetSize; ++i) {
+    my_set.insert(i);
+  }
+  safe_btree_set<int64_t> my_set_copy(my_set);
+  EXPECT_TRUE(my_set_copy == my_set);
+  EXPECT_TRUE(my_set == my_set_copy);
+  EXPECT_FALSE(my_set_copy != my_set);
+  EXPECT_FALSE(my_set != my_set_copy);
+
+  my_set.insert(kSetSize);
+  EXPECT_FALSE(my_set_copy == my_set);
+  EXPECT_FALSE(my_set == my_set_copy);
+  EXPECT_TRUE(my_set_copy != my_set);
+  EXPECT_TRUE(my_set != my_set_copy);
+
+  my_set.erase(kSetSize - 1);
+  EXPECT_FALSE(my_set_copy == my_set);
+  EXPECT_FALSE(my_set == my_set_copy);
+  EXPECT_TRUE(my_set_copy != my_set);
+  EXPECT_TRUE(my_set != my_set_copy);
+
+  safe_btree_map<std::string, int64_t> my_map;
+  for (int i = 0; i < kSetSize; ++i) {
+    my_map[std::string(i, 'a')] = i;
+  }
+  safe_btree_map<std::string, int64_t> my_map_copy(my_map);
+  EXPECT_TRUE(my_map_copy == my_map);
+  EXPECT_TRUE(my_map == my_map_copy);
+  EXPECT_FALSE(my_map_copy != my_map);
+  EXPECT_FALSE(my_map != my_map_copy);
+
+  ++my_map_copy[std::string(7, 'a')];
+  EXPECT_FALSE(my_map_copy == my_map);
+  EXPECT_FALSE(my_map == my_map_copy);
+  EXPECT_TRUE(my_map_copy != my_map);
+  EXPECT_TRUE(my_map != my_map_copy);
+
+  my_map_copy = my_map;
+  my_map["hello"] = kSetSize;
+  EXPECT_FALSE(my_map_copy == my_map);
+  EXPECT_FALSE(my_map == my_map_copy);
+  EXPECT_TRUE(my_map_copy != my_map);
+  EXPECT_TRUE(my_map != my_map_copy);
+
+  my_map.erase(std::string(kSetSize - 1, 'a'));
+  EXPECT_FALSE(my_map_copy == my_map);
+  EXPECT_FALSE(my_map == my_map_copy);
+  EXPECT_TRUE(my_map_copy != my_map);
+  EXPECT_TRUE(my_map != my_map_copy);
+}
+
+} // namespace
+} // namespace btree
diff --git a/xdelta3/examples/checksum_test.cc b/xdelta3/examples/checksum_test.cc
deleted file mode 100644
index 028922b..0000000
--- a/xdelta3/examples/checksum_test.cc
+++ /dev/null
@@ -1,732 +0,0 @@
-/* Copyright (C) 2007 Josh MacDonald */
-
-extern "C" {
-#include "test.h"
-#include <assert.h>
-}
-
-#include <list>
-#include <vector>
-#include <map>
-#include <algorithm>
-
-using std::list;
-using std::map;
-using std::vector;
-
-// MLCG parameters
-// a, a*
-uint32_t good_32bit_values[] = {
-    1597334677U, // ...
-    741103597U, 887987685U,
-};
-
-// a, a*
-uint64_t good_64bit_values[] = {
-    1181783497276652981ULL, 4292484099903637661ULL,
-    7664345821815920749ULL, // ...
-};
-
-struct true_type { };
-struct false_type { };
-
-template <typename Word>
-int bitsof();
-
-template<>
-int bitsof<uint32_t>() {
-    return 32;
-}
-
-template<>
-int bitsof<uint64_t>() {
-    return 64;
-}
-
-struct plain {
-    int operator()(const uint8_t &c) {
-        return c;
-    }
-};
-
-template <typename Word>
-struct hhash {  // take "h" of the high-bits as a hash value for this
-                // checksum, which are the most "distant" in terms of the
-                // spectral test for the rabin_karp MLCG.  For short windows,
-                // the high bits aren't enough, XOR "mask" worth of these in.
-    Word operator()(const Word& t, const int &h, const int &mask) {
-        return (t >> h) ^ (t & mask);
-    }
-};
-
-template <typename Word>
-Word good_word();
-
-template<>
-uint32_t good_word<uint32_t>() {
-    return good_32bit_values[0];
-}
-
-template<>
-uint64_t good_word<uint64_t>() {
-    return good_64bit_values[0];
-}
-
-// CLASSES
-
-#define SELF Word, CksumSize, CksumSkip, Permute, Hash, Compaction
-#define MEMBER template <typename Word, \
-                         int CksumSize, \
-                         int CksumSkip, \
-                         typename Permute, \
-                         typename Hash, \
-                         int Compaction>
-
-MEMBER
-struct cksum_params {
-    typedef Word word_type;
-    typedef Permute permute_type;
-    typedef Hash hash_type;
-
-    enum { cksum_size = CksumSize,
-           cksum_skip = CksumSkip,
-           compaction = Compaction,
-    };
-};
-
-
-MEMBER
-struct rabin_karp {
-    typedef Word word_type;
-    typedef Permute permute_type;
-    typedef Hash hash_type;
-
-    enum { cksum_size = CksumSize,
-           cksum_skip = CksumSkip, 
-           compaction = Compaction,
-    };
-
-    // (a^cksum_size-1 c_0) + (a^cksum_size-2 c_1) ...
-    rabin_karp() {
-        multiplier = good_word<Word>();
-        powers = new Word[cksum_size];
-        powers[cksum_size - 1] = 1;
-        for (int i = cksum_size - 2; i >= 0; i--) {
-            powers[i] = powers[i + 1] * multiplier;
-        }
-        product = powers[0] * multiplier;
-    }
-
-    ~rabin_karp() {
-        delete [] powers;
-    }
-
-    Word step(const uint8_t *ptr) {
-        Word h = 0;
-        for (int i = 0; i < cksum_size; i++) {
-            h += permute_type()(ptr[i]) * powers[i];
-        }
-        return h;
-    }
-
-    Word state0(const uint8_t *ptr) {
-        incr_state = step(ptr);
-        return incr_state;
-    }
-
-    Word incr(const uint8_t *ptr) {
-        incr_state = multiplier * incr_state -
-            product * permute_type()(ptr[-1]) +
-            permute_type()(ptr[cksum_size - 1]);
-        return incr_state;
-    }
-
-    Word *powers;
-    Word  product;
-    Word  multiplier;
-    Word  incr_state;
-};
-
-MEMBER
-struct adler32_cksum {
-    typedef Word word_type;
-    typedef Permute permute_type;
-    typedef Hash hash_type;
-
-    enum { cksum_size = CksumSize,
-           cksum_skip = CksumSkip, 
-           compaction = Compaction,
-    };
-
-    Word step(const uint8_t *ptr) {
-        return xd3_lcksum (ptr, cksum_size);
-    }
-
-    Word state0(const uint8_t *ptr) {
-        incr_state = step(ptr);
-        return incr_state;
-    }
-
-    Word incr(const uint8_t *ptr) {
-        incr_state = xd3_large_cksum_update (incr_state, ptr - 1, cksum_size);
-        return incr_state;
-    }
-
-    Word  incr_state;
-};
-
-// TESTS
-
-template <typename Word>
-struct file_stats {
-    typedef list<const uint8_t*> ptr_list;
-    typedef Word word_type;
-    typedef map<word_type, ptr_list> table_type;
-    typedef typename table_type::iterator table_iterator;
-    typedef typename ptr_list::iterator ptr_iterator;
-
-    int cksum_size;
-    int cksum_skip;
-    int unique;
-    int unique_values;
-    int count;
-    table_type table;
-
-    file_stats(int size, int skip)
-        : cksum_size(size),
-          cksum_skip(skip),
-          unique(0),
-          unique_values(0),
-          count(0) {
-    }
-
-    void reset() {
-        unique = 0;
-        unique_values = 0;
-        count = 0;
-        table.clear();
-    }
-
-    void update(const word_type &word, const uint8_t *ptr) {
-        table_iterator t_i = table.find(word);
-
-        count++;
-
-        if (t_i == table.end()) {
-            table.insert(make_pair(word, ptr_list()));
-        }
-
-        ptr_list &pl = table[word];
-
-        for (ptr_iterator p_i = pl.begin();
-             p_i != pl.end();
-             ++p_i) {
-            if (memcmp(*p_i, ptr, cksum_size) == 0) {
-                return;
-            }
-        }
-
-        unique++;
-        pl.push_back(ptr);
-    }
-
-    void freeze() {
-        unique_values = table.size();
-        table.clear();
-    }
-};
-
-struct test_result_base;
-
-static vector<test_result_base*> all_tests;
-
-struct test_result_base {
-    virtual ~test_result_base() {
-    }
-    virtual void reset() = 0;
-    virtual void print() = 0;
-    virtual void get(const uint8_t* buf, const int buf_size, int iters) = 0;
-    virtual void stat() = 0;
-    virtual int count() = 0;
-    virtual int dups() = 0;
-    virtual double uniqueness() = 0;
-    virtual double fullness() = 0;
-    virtual double collisions() = 0;
-    virtual double coverage() = 0;
-    virtual double compression() = 0;
-    virtual double time() = 0;
-    virtual double score() = 0;
-    virtual void set_score(double min_dups_frac, double min_time) = 0;
-    virtual double total_time() = 0;
-    virtual int total_count() = 0;
-    virtual int total_dups() = 0;
-};
-
-struct compare_h {
-    bool operator()(test_result_base *a,
-                    test_result_base *b) {
-        return a->score() < b->score();
-    }
-};
-
-MEMBER
-struct test_result : public test_result_base {
-    typedef Word word_type;
-    typedef Permute permute_type;
-    typedef Hash hash_type;
-
-    enum { cksum_size = CksumSize,
-           cksum_skip = CksumSkip, 
-           compaction = Compaction,
-    };
-
-    const char *test_name;
-    file_stats<Word> fstats;
-    int test_size;
-    int n_steps;
-    int n_incrs;
-    int s_bits;
-    int s_mask;
-    int t_entries;
-    int h_bits;
-    int h_buckets_full;
-    double h_score;
-    char *hash_table;
-    long accum_millis;
-    int accum_iters;
-
-    // These are not reset
-    double accum_time;
-    int accum_count;
-    int accum_dups;
-    int accum_colls;
-    int accum_size;
-
-    test_result(const char *name)
-        : test_name(name),
-          fstats(cksum_size, cksum_skip),
-          hash_table(NULL),
-          accum_millis(0),
-          accum_iters(0),
-          accum_time(0.0),
-          accum_count(0),
-          accum_dups(0),
-          accum_colls(0),
-          accum_size(0) {
-        all_tests.push_back(this);
-    }
-
-    ~test_result() {
-        reset();
-    }
-
-    void reset() {
-        // size of file
-        test_size = -1;
-
-        // count
-        n_steps = -1;
-        n_incrs = -1;
-
-        // four values used by new_table()/summarize_table()
-        s_bits = -1;
-        s_mask = -1;
-        t_entries = -1;
-        h_bits = -1;
-        h_buckets_full = -1;
-
-        accum_millis = 0;
-        accum_iters = 0;
-
-        fstats.reset();
-
-        // temporary
-        if (hash_table) {
-            delete(hash_table);
-            hash_table = NULL;
-        }
-    }
-
-    int count() {
-        if (cksum_skip == 1) {
-            return n_incrs;
-        } else {
-            return n_steps;
-        }
-    }
-
-    int dups() {
-        return fstats.count - fstats.unique;
-    }
-
-    int colls() {
-        return fstats.unique - fstats.unique_values;
-    }
-
-    double uniqueness() {
-        return 1.0 - (double) dups() / count();
-    }
-
-    double fullness() {
-        return (double) h_buckets_full / (1 << h_bits);
-    }
-
-    double collisions() {
-        return (double) colls() / fstats.unique;
-    }
-
-    double coverage() {
-        return (double) h_buckets_full / uniqueness() / count();
-    }
-
-    double compression() {
-        return 1.0 - coverage();
-    }
-
-    double time() {
-        return (double) accum_millis / accum_iters;
-    }
-
-    double score() {
-        return h_score;
-    }
-
-    void set_score(double min_compression, double min_time) {
-        h_score = (compression() - 0.99 * min_compression)
-                * (time() - 0.99 * min_time);
-    }
-
-    double total_time() {
-        return accum_time;
-    }
-
-    int total_count() {
-        return accum_count;
-    }
-
-    int total_dups() {
-        return accum_dups;
-    }
-
-    int total_colls() {
-        return accum_dups;
-    }
-
-    void stat() {
-        accum_time += time();
-        accum_count += count();
-        accum_dups += dups();
-        accum_colls += colls();
-        accum_size += test_size;
-    }
-
-    void print() {
-        if (fstats.count != count()) {
-            fprintf(stderr, "internal error: %d != %d\n", fstats.count, count());
-            abort();
-        }
-        printf("%s: (%u#%u) count %u uniq %0.2f%% full %u (%0.4f%% coll %0.4f%%) covers %0.2f%% w/ 2^%d @ %.4f MB/s %u iters\n",
-               test_name,
-               cksum_size,
-               cksum_skip,
-               count(),
-               100.0 * uniqueness(),
-               h_buckets_full,
-               100.0 * fullness(),
-               100.0 * collisions(),
-               100.0 * coverage(),
-               h_bits,
-               0.001 * accum_iters * test_size / accum_millis,
-               accum_iters);
-    }
-
-    int size_log2 (int slots)
-    {
-        int bits = bitsof<word_type>() - 1;
-        int i;
-
-        for (i = 3; i <= bits; i += 1) {
-            if (slots <= (1 << i)) {
-                return i - compaction;
-            }
-        }
-
-        return bits;
-    }
-
-    void new_table(int entries) {
-        t_entries = entries;
-        h_bits = size_log2(entries);
-
-        int n = 1 << h_bits;
-
-        s_bits = bitsof<word_type>() - h_bits;
-        s_mask = n - 1;
-
-        hash_table = new char[n / 8];
-        memset(hash_table, 0, n / 8);
-    }
-
-    int get_table_bit(int i) {
-        return hash_table[i/8] & (1 << i%8);
-    }
-
-    int set_table_bit(int i) {
-        return hash_table[i/8] |= (1 << i%8);
-    }
-
-    void summarize_table() {
-        int n = 1 << h_bits;
-        int f = 0;
-        for (int i = 0; i < n; i++) {
-            if (get_table_bit(i)) {
-                f++;
-            }
-        }
-        h_buckets_full = f;
-    }
-
-    void get(const uint8_t* buf, const int buf_size, int test_iters) {
-        rabin_karp<SELF> test;
-        //adler32_cksum<SELF> test;
-        hash_type hash;
-        const uint8_t *ptr;
-        const uint8_t *end;
-        int last_offset;
-        int periods;
-        int stop;
-
-        test_size = buf_size;
-        last_offset = buf_size - cksum_size;
-
-        if (last_offset < 0) {
-            periods = 0;
-            n_steps = 0;
-            n_incrs = 0;
-            stop = -cksum_size;
-        } else {
-            periods = last_offset / cksum_skip;
-            n_steps = periods + 1;
-            n_incrs = last_offset + 1;
-            stop = last_offset - (periods + 1) * cksum_skip;
-        }
-
-        // Compute file stats once.
-        if (fstats.unique_values == 0) {
-            if (cksum_skip == 1) {
-                for (int i = 0; i <= buf_size - cksum_size; i++) {
-                    fstats.update(hash(test.step(buf + i), s_bits, s_mask), buf + i);
-                }
-            } else {
-                ptr = buf + last_offset;
-                end = buf + stop;
-                
-                for (; ptr != end; ptr -= cksum_skip) {
-                    fstats.update(hash(test.step(ptr), s_bits, s_mask), ptr);
-                }
-            }
-            fstats.freeze();
-        }
-
-        long start_test = get_millisecs_now();
-
-        if (cksum_skip != 1) {
-            new_table(n_steps);
-
-            for (int i = 0; i < test_iters; i++) {
-                ptr = buf + last_offset;
-                end = buf + stop;
-
-                for (; ptr != end; ptr -= cksum_skip) {
-                    set_table_bit(hash(test.step(ptr), s_bits, s_mask));
-                }
-            }
-
-            summarize_table();
-        }
-
-        stop = buf_size - cksum_size + 1;
-        if (stop < 0) {
-            stop = 0;
-        }
-
-        if (cksum_skip == 1) {
-
-            new_table(n_incrs);
-
-            for (int i = 0; i < test_iters; i++) {
-                ptr = buf;
-                end = buf + stop;
-
-                if (ptr != end) {
-                    set_table_bit(hash(test.state0(ptr++), s_bits, s_mask));
-                }
-
-                for (; ptr != end; ptr++) {
-                    Word w = test.incr(ptr);
-                    assert(w == test.step(ptr));
-                    set_table_bit(hash(w, s_bits, s_mask));
-                }
-            }
-
-            summarize_table();
-        }
-
-        accum_iters += test_iters;
-        accum_millis += get_millisecs_now() - start_test;
-    }
-};
-
-template <typename Word>
-void print_array(const char *tname) {
-    printf("static const %s hash_multiplier[64] = {\n", tname);
-    Word p = 1;
-    for (int i = 0; i < 64; i++) {
-        printf("  %uU,\n", p);
-        p *= good_word<Word>();
-    }
-    printf("};\n", tname);
-}
-
-int main(int argc, char** argv) {
-  int i;
-  uint8_t *buf = NULL;
-  size_t buf_len = 0;
-  int ret;
-
-  if (argc <= 1) {
-    fprintf(stderr, "usage: %s file ...\n", argv[0]);
-    return 1;
-  }
-
-  //print_array<uint32_t>("uint32_t");
-
-#define TEST(T,Z,S,P,H,C) test_result<T,Z,S,P,H<T>,C> \
-      _ ## T ## _ ## Z ## _ ## S ## _ ## P ## _ ## H ## _ ## C \
-      (#T "_" #Z "_" #S "_" #P "_" #H "_" #C)
-
-#if 0
-
-  TEST(uint32_t, 4, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 2); /* x */ \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 3); /* x */ \
-
-#endif
-
-#define TESTS(SKIP) \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 2); /* x */ \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 3)
-  
-#define TESTS_ALL(SKIP) \
-  TEST(uint32_t, 3, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 3, SKIP, plain, hhash, 1); \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 2); /* x */ \
-  TEST(uint32_t, 4, SKIP, plain, hhash, 3); /* x */ \
-  TEST(uint32_t, 5, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 5, SKIP, plain, hhash, 1); \
-  TEST(uint32_t, 8, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 8, SKIP, plain, hhash, 1); \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 2); /* x */ \
-  TEST(uint32_t, 9, SKIP, plain, hhash, 3); /* x */ \
-  TEST(uint32_t, 11, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 11, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 13, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 13, SKIP, plain, hhash, 1); \
-  TEST(uint32_t, 15, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 15, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 16, SKIP, plain, hhash, 0); /* x */ \
-  TEST(uint32_t, 16, SKIP, plain, hhash, 1); /* x */ \
-  TEST(uint32_t, 21, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 21, SKIP, plain, hhash, 1); \
-  TEST(uint32_t, 34, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 34, SKIP, plain, hhash, 1); \
-  TEST(uint32_t, 55, SKIP, plain, hhash, 0); \
-  TEST(uint32_t, 55, SKIP, plain, hhash, 1)
-
-  TESTS(1); // *
-//   TESTS(2); // *
-//   TESTS(3); // *
-//   TESTS(5); // *
-//   TESTS(8); // *
-//   TESTS(9);
-//   TESTS(11);
-//   TESTS(13); // *
-  TESTS(15);
-//   TESTS(16);
-//   TESTS(21); // *
-//   TESTS(34); // *
-//   TESTS(55); // *
-//   TESTS(89); // *
-
-  for (i = 1; i < argc; i++) {
-    if ((ret = read_whole_file(argv[i],
-                               & buf,
-                               & buf_len))) {
-      return 1;
-    }
-
-    fprintf(stderr, "file %s is %zu bytes\n",
-            argv[i], buf_len);
-
-    double min_time = -1.0;
-    double min_compression = 0.0;
-
-    for (vector<test_result_base*>::iterator i = all_tests.begin();
-         i != all_tests.end(); ++i) {
-        test_result_base *test = *i;
-        test->reset();
-
-        int iters = 100;
-        long start_test = get_millisecs_now();
-
-        do {
-            test->get(buf, buf_len, iters);
-            iters *= 3;
-            iters /= 2;
-        } while (get_millisecs_now() - start_test < 2000);
-
-        test->stat();
-
-        if (min_time < 0.0) {
-            min_compression = test->compression();
-            min_time = test->time();
-        }
-
-        if (min_time > test->time()) {
-            min_time = test->time();
-        }
-
-        if (min_compression > test->compression()) {
-            min_compression = test->compression();
-        }
-
-        test->print();
-    }
-
-//     for (vector<test_result_base*>::iterator i = all_tests.begin();
-//       i != all_tests.end(); ++i) {
-//      test_result_base *test = *i;
-//      test->set_score(min_compression, min_time);
-//     }        
-
-//     sort(all_tests.begin(), all_tests.end(), compare_h());
-    
-//     for (vector<test_result_base*>::iterator i = all_tests.begin();
-//       i != all_tests.end(); ++i) {
-//      test_result_base *test = *i;
-//      test->print();
-//     }        
-    
-    free(buf);
-    buf = NULL;
-  }
-
-  return 0;      
-}
diff --git a/xdelta3/go/src/regtest.go b/xdelta3/go/src/regtest.go
new file mode 100644
index 0000000..4ffcdaf
--- /dev/null
+++ b/xdelta3/go/src/regtest.go
@@ -0,0 +1,274 @@
+package main
+
+import (
+        "fmt"
+        "io"
+        "path"
+        "os"
+        "sort"
+        "time"
+
+        "xdelta"
+)
+
+const (
+        xdataset = "/volume/home/jmacd/src/testdata"
+        xcompare = "/volume/home/jmacd/src/xdelta-devel/xdelta3/xdelta3"
+        xdelta3  = "/volume/home/jmacd/src/xdelta-64bithash/xdelta3/xdelta3"
+        seed = 1422253499919909358
+)
+
+type Config struct {
+        srcbuf_size int64
+        window_size int64
+        blocksize   int
+}
+
+func NewC() Config {
+        // TODO make these (and above) flags
+        return Config{1<<26, 1<<22, 1<<16}
+}
+
+func (c Config) smokeTest(t *xdelta.TestGroup, p xdelta.Program) {
+        target := "Hello world!"
+        source := "Hello world, nice to meet you!"
+
+        enc, err := t.Exec("encode", p, true, []string{"-e"})
+        if err != nil {
+                t.Panic(err)
+        }
+        dec, err := t.Exec("decode", p, true, []string{"-d"})
+        if err != nil {
+                t.Panic(err)
+        }
+
+        encodeout := t.Drain(enc.Stdout, "encode.stdout")
+        decodeout := t.Drain(dec.Stdout, "decode.stdout")
+
+        t.Empty(enc.Stderr, "encode")
+        t.Empty(dec.Stderr, "decode")
+
+        t.TestWrite("encode.stdin", enc.Stdin, []byte(target))
+        t.TestWrite("encode.srcin", enc.Srcin, []byte(source))
+
+        t.TestWrite("decode.stdin", dec.Stdin, <-encodeout)
+        t.TestWrite("decode.srcin", dec.Srcin, []byte(source))
+
+        if do := string(<-decodeout); do != target {
+                t.Panic(fmt.Errorf("It's not working! %s\n!=\n%s\n", do, target))
+        }
+        t.Wait(enc, dec)
+}
+
+type PairTest struct {
+        // Input
+        Config
+        program xdelta.Program
+        source, target string
+
+        // Output
+        TestOutput
+}
+
+type TestOutput struct {
+        encoded int64
+        encDuration time.Duration
+        decDuration time.Duration
+        encSysDuration time.Duration
+        decSysDuration time.Duration
+}
+
+func (to *TestOutput) Add(a TestOutput) {
+        to.encoded += a.encoded
+        to.encDuration += a.encDuration
+        to.decDuration += a.decDuration
+        to.encSysDuration += a.encSysDuration
+        to.decSysDuration += a.decSysDuration
+}
+
+func (to *TestOutput) String() string {
+        return fmt.Sprintf("SIZE: %v\tT: %v\tTSYS: %v\tDT: %v\tDTSYS: %v",
+                to.encoded, to.encDuration, to.encSysDuration, to.decDuration, to.encSysDuration)
+}
+
+// P is the test program, Q is the reference version.
+func (cfg Config) datasetTest(t *xdelta.TestGroup, p, q xdelta.Program) {
+        dir, err := os.Open(xdataset)
+        if err != nil {
+                t.Panic(err)
+        }
+        dents, err := dir.Readdir(-1)
+        if err != nil {
+                t.Panic(err)
+        }
+        paths := make([]string, len(dents))
+        var total int64
+        for i, d := range dents {
+                if !d.Mode().IsRegular() {
+                        continue
+                }
+                paths[i] = fmt.Sprint(xdataset, "/", d.Name())
+                total += d.Size()
+        }
+        meansize := total / int64(len(dents))
+        largest  := uint(20)
+        for ; largest <= 31 && 1<<largest < meansize; largest++ {}
+
+        sort.Strings(paths)
+
+        testSum := map[uint]*TestOutput{}
+        compSum := map[uint]*TestOutput{}
+
+        for _, in1 := range paths {
+                for _, in2 := range paths {
+                        if in1 == in2 { continue }
+
+                        // 1/4, 1/2, and 1 of the power-of-2 rounded-up mean size
+                        for b := largest - 2; b <= largest; b++ {
+                                if _, has := testSum[b]; !has {
+                                        testSum[b] = &TestOutput{}
+                                        compSum[b] = &TestOutput{}
+                                }
+                                c1 := cfg
+                                c1.srcbuf_size = 1<<b
+                                ptest := &PairTest{c1, p, in1, in2, TestOutput{-1, 0, 0, 0, 0}}
+                                ptest.datasetPairTest(t, 1<<b);
+                                qtest := &PairTest{c1, q, in1, in2, TestOutput{-1, 0, 0, 0, 0}}
+                                qtest.datasetPairTest(t, 1<<b)
+
+                                testSum[b].Add(ptest.TestOutput)
+                                compSum[b].Add(qtest.TestOutput)
+
+                                fmt.Printf("%s, %s: %.2f%% %+d/%d\n\tE:%.2f%%/%s(%.2f%%/%s) D:%.2f%%/%s(%.2f%%/%s) [B=%d]\n",
+                                        path.Base(in1), path.Base(in2),
+                                        float64(ptest.encoded - qtest.encoded) * 100.0 / float64(qtest.encoded),
+                                        ptest.encoded - qtest.encoded,
+                                        qtest.encoded,
+                                        (ptest.encDuration - qtest.encDuration).Seconds() * 100.0 / qtest.encDuration.Seconds(),
+                                        qtest.encDuration,
+                                        (ptest.decDuration - qtest.decDuration).Seconds() * 100.0 / qtest.decDuration.Seconds(),
+                                        qtest.encDuration,
+                                        (ptest.encSysDuration - qtest.encSysDuration).Seconds() * 100.0 / qtest.encSysDuration.Seconds(),
+                                        qtest.encSysDuration,
+                                        (ptest.decSysDuration - qtest.decSysDuration).Seconds() * 100.0 / qtest.decSysDuration.Seconds(),
+                                        qtest.decSysDuration,
+                                        1<<b)
+                        }
+                }
+        }
+        var keys []uint
+        for k, _ := range testSum {
+                keys = append(keys, k)
+        }
+        for _, k := range keys {                
+                fmt.Printf("B=%v\nTEST: %v\nCOMP: %v\n", 1<<k, testSum[k], compSum[k])
+        }
+}
+
+func (pt *PairTest) datasetPairTest(t *xdelta.TestGroup, meanSize int64) {
+        cfg := pt.Config
+        eargs := []string{"-e", fmt.Sprint("-B", cfg.srcbuf_size), // "-q",
+                fmt.Sprint("-W", cfg.window_size), "-s", pt.source,
+                "-I0", "-S", "none", pt.target}
+        enc, err := t.Exec("encode", pt.program, false, eargs)
+        if err != nil {
+                t.Panic(err)
+        }
+
+        dargs := []string{"-dc", fmt.Sprint("-B", cfg.srcbuf_size), //"-q",
+                fmt.Sprint("-W", cfg.window_size), "-s", pt.source,
+                "-S", "none"}
+
+        dec, err := t.Exec("decode", pt.program, false, dargs)
+        if err != nil {
+                t.Panic(err)
+        }
+        tgt_check, err := os.Open(pt.target)
+        if err != nil {
+                t.Panic(err)
+        }
+        tgt_info, err := tgt_check.Stat()
+        if err != nil {
+                t.Panic(err)
+        }
+        t.Empty(enc.Stderr, "encode")
+        t.Empty(dec.Stderr, "decode")
+        t.CopyStreams(enc.Stdout, dec.Stdin, &pt.encoded)
+        t.CompareStreams(dec.Stdout, tgt_check, tgt_info.Size())
+
+        t.Wait(enc, dec)
+
+        pt.decDuration = dec.Cmd.ProcessState.UserTime()
+        pt.encDuration = enc.Cmd.ProcessState.UserTime()
+        pt.decSysDuration = dec.Cmd.ProcessState.SystemTime()
+        pt.encSysDuration = enc.Cmd.ProcessState.SystemTime()
+}
+
+func (cfg Config) offsetTest(t *xdelta.TestGroup, p xdelta.Program, offset, length int64) {
+        eargs := []string{"-e", "-0", fmt.Sprint("-B", cfg.srcbuf_size), "-q",
+                fmt.Sprint("-W", cfg.window_size)}
+        enc, err := t.Exec("encode", p, true, eargs)
+        if err != nil {
+                t.Panic(err)
+        }
+        
+        dargs := []string{"-d", fmt.Sprint("-B", cfg.srcbuf_size), "-q",
+                fmt.Sprint("-W", cfg.window_size)}
+        dec, err := t.Exec("decode", p, true, dargs)
+        if err != nil {
+                t.Panic(err)
+        }
+
+        // The pipe used to read the decoder output and compare
+        // against the target.
+        read, write := io.Pipe()
+
+        t.Empty(enc.Stderr, "encode")
+        t.Empty(dec.Stderr, "decode")
+
+        var encoded_size int64
+        t.CopyStreams(enc.Stdout, dec.Stdin, &encoded_size)
+        t.CompareStreams(dec.Stdout, read, length)
+
+        // The decoder output ("read", above) is compared with the
+        // test-provided output ("write", below).  The following
+        // generates two identical inputs.
+        t.WriteRstreams("encode", seed, offset, length, enc.Srcin, enc.Stdin)
+        t.WriteRstreams("decode", seed, offset, length, dec.Srcin, write)
+        t.Wait(enc, dec)
+
+        expect := cfg.srcbuf_size - offset
+        if float64(encoded_size) < (0.95 * float64(expect)) ||
+                float64(encoded_size) > (1.05 * float64(expect)) {
+                t.Fail("encoded size should be ~=", expect, ", actual ", encoded_size)
+        }
+}
+
+func main() {
+        r, err := xdelta.NewRunner()
+        if err != nil {
+                panic(err)
+        }
+        defer r.Cleanup()
+
+        cfg := NewC()
+
+        prog := xdelta.Program{xdelta3}
+
+        r.RunTest("smoketest", func(t *xdelta.TestGroup) { cfg.smokeTest(t, prog) })
+
+        for i := uint(29); i <= 33; i += 1 {
+                // The arguments to offsetTest are offset, source
+                // window size, and file size. The source window size
+                // is (2 << i) and (in the 3.0x release branch) is
+                // limited to 2^31, so the the greatest value of i is
+                // 30.
+                cfg.srcbuf_size = 2 << i
+                r.RunTest(fmt.Sprint("offset", i), func(t *xdelta.TestGroup) {
+                        cfg.offsetTest(t, prog, 1 << i, 3 << i) })
+        }
+        
+        comp := xdelta.Program{xcompare}
+
+        r.RunTest("dataset", func(t *xdelta.TestGroup) { cfg.datasetTest(t, prog, comp) })
+}
diff --git a/xdelta3/go/src/xdelta/rstream.go b/xdelta3/go/src/xdelta/rstream.go
new file mode 100644
index 0000000..99c3d17
--- /dev/null
+++ b/xdelta3/go/src/xdelta/rstream.go
@@ -0,0 +1,71 @@
+package xdelta
+
+
+import (
+        "io"
+        "math/rand"
+)
+
+const (
+        blocksize = 1<<17
+)
+
+func (t *TestGroup) WriteRstreams(desc string, seed, offset, len int64,
+        src, tgt io.WriteCloser) {
+        t.Go("src-write:"+desc, func (g *Goroutine) {
+                writeOne(g, seed, 0, len, tgt, false)
+        })
+        t.Go("tgt-write:"+desc, func (g *Goroutine) {
+                writeOne(g, seed, offset, len, src, true)
+        })
+}
+
+func writeOne(g *Goroutine, seed, offset, len int64, stream io.WriteCloser, readall bool) {
+        if !readall {
+                // Allow the source-read to fail or block until the process terminates.
+                // This behavior is reserved for the decoder, which is not required to
+                // read the entire source.
+                g.OK()
+        }
+        if offset != 0 {
+                // Fill with other random data until the offset
+                if err := writeRand(g, rand.New(rand.NewSource(^seed)), offset, stream); err != nil {
+                        g.Panic(err)
+                }
+        }
+        if err := writeRand(g, rand.New(rand.NewSource(seed)),
+                len - offset, stream); err != nil {
+                g.Panic(err)
+        }
+        if err := stream.Close(); err != nil {
+                g.Panic(err)
+        }
+        g.OK()
+}
+
+func writeRand(g *Goroutine, r *rand.Rand, len int64, s io.Writer) error {
+        blk := make([]byte, blocksize)
+        for len > 0 {
+                fillRand(r, blk)
+                c := blocksize
+                if len < blocksize {
+                        c = int(len)
+                }
+                if _, err := s.Write(blk[0:c]); err != nil {
+                        return err
+                }
+                len -= int64(c)
+        }
+        return nil
+}
+
+func fillRand(r *rand.Rand, blk []byte) {
+        for p := 0; p < len(blk); {
+                v := r.Int63()
+                for i := 7; i != 0 && p < len(blk); i-- {
+                        blk[p] = byte(v)
+                        p++
+                        v >>= 8
+                }
+        }
+}
diff --git a/xdelta3/go/src/xdelta/run.go b/xdelta3/go/src/xdelta/run.go
new file mode 100644
index 0000000..448fabe
--- /dev/null
+++ b/xdelta3/go/src/xdelta/run.go
@@ -0,0 +1,71 @@
+package xdelta
+
+import (
+        "fmt"
+        "io"
+        "io/ioutil"
+        "os"
+        "os/exec"
+)
+
+type Program struct {
+        Path string
+}
+
+type Run struct {
+        Cmd exec.Cmd
+        Srcfile string
+        Stdin io.WriteCloser
+        Srcin io.WriteCloser
+        Stdout io.ReadCloser
+        Stderr io.ReadCloser
+}
+
+type Runner struct {
+        Testdir string
+}
+
+func (r *Run) Wait() error {
+        return r.Cmd.Wait()
+}
+
+func NewRunner() (*Runner, error) {
+        if dir, err := ioutil.TempDir(tmpDir, "xrt"); err != nil {
+                return nil, err
+        } else {
+                return &Runner{dir}, nil
+        }
+}
+
+func (r *Runner) newTestGroup(name string) (*TestGroup) {
+        tg := &TestGroup{Runner: r}
+        tg.WaitGroup.Add(1)
+        g0 := &Goroutine{tg, name, false}
+        tg.running = append(tg.running, g0)
+        tg.main = g0
+        return tg
+}
+
+func (r *Runner) Cleanup() {
+        os.RemoveAll(r.Testdir)
+}
+
+func (r *Runner) RunTest(name string, f func (t *TestGroup)) {
+        t := r.newTestGroup(name)
+        c := make(chan interface{})
+        go func() {
+                defer func() {
+                        rec := recover()
+                        c <- rec
+                }()
+                fmt.Println("Testing", name, "...")
+                f(t)
+                c <- nil
+        }()
+        rec := <- c
+        if t.errors == nil && rec == nil {
+                fmt.Println("Success:", name)
+        } else {
+                fmt.Println("FAILED:", name, t.errors, rec)
+        }
+}
diff --git a/xdelta3/go/src/xdelta/test.go b/xdelta3/go/src/xdelta/test.go
new file mode 100644
index 0000000..7210698
--- /dev/null
+++ b/xdelta3/go/src/xdelta/test.go
@@ -0,0 +1,164 @@
+package xdelta
+
+import (
+        "bufio"
+        "bytes"
+        "errors"
+        "fmt"
+        "io"
+        "io/ioutil"
+        "os"
+        "path"
+        "sync/atomic"
+
+        "golang.org/x/sys/unix"
+)
+
+var (
+        tmpDir = "/tmp"
+        srcSeq int64
+)
+
+func (t *TestGroup) Drain(f io.ReadCloser, desc string) <-chan []byte {
+        c := make(chan []byte)
+        t.Go(desc, func(g *Goroutine) {
+                if b, err := ioutil.ReadAll(f); err != nil {
+                        g.Panic(err)
+                } else {
+                        c <- b
+                }
+                g.OK()
+        })
+        return c
+}
+
+func (t *TestGroup) Empty(f io.ReadCloser, desc string) *Goroutine {
+        return t.Go("empty:"+desc, func (g *Goroutine) {
+                s := bufio.NewScanner(f)
+                for s.Scan() {
+                        os.Stderr.Write([]byte(fmt.Sprint(desc, ": ", s.Text(), "\n")))
+                }
+                err := s.Err()
+                f.Close()
+                if err != nil {
+                        g.Panic(err)
+                }
+                g.OK()
+        })
+}
+
+func (t *TestGroup) TestWrite(what string, f io.WriteCloser, b []byte) *Goroutine {
+        return t.Go("write", func(g *Goroutine) {
+                if _, err := f.Write(b); err != nil {
+                        g.Panic(err)
+                }
+                if err := f.Close(); err != nil {
+                        g.Panic(err)
+                }
+                g.OK()
+        })
+}
+
+func (t *TestGroup) CopyStreams(r io.ReadCloser, w io.WriteCloser, written *int64) *Goroutine {
+        return t.Go("copy", func(g *Goroutine) {
+                nwrite, err := io.Copy(w, r)
+                if err != nil {
+                        g.Panic(err)
+                }
+                err = r.Close()
+                if err != nil {
+                        g.Panic(err)
+                }
+                err = w.Close()
+                if err != nil {
+                        g.Panic(err)
+                }
+                g.OK()
+                *written = nwrite
+        })
+}
+
+func (t *TestGroup) CompareStreams(r1 io.ReadCloser, r2 io.ReadCloser, length int64) *Goroutine {
+        return t.Go("compare", func(g *Goroutine) {
+                b1 := make([]byte, blocksize)
+                b2 := make([]byte, blocksize)
+                var idx int64
+                for length > 0 {
+                        c := blocksize
+                        if length < blocksize {
+                                c = int(length)
+                        }
+                        if _, err := io.ReadFull(r1, b1[0:c]); err != nil {
+                                g.Panic(err)
+                        }
+                        if _, err := io.ReadFull(r2, b2[0:c]); err != nil {
+                                g.Panic(err)
+                        }
+                        if bytes.Compare(b1[0:c], b2[0:c]) != 0 {
+                                fmt.Println("B1 is", string(b1[0:c]))
+                                fmt.Println("B2 is", string(b2[0:c]))                   
+                                g.Panic(errors.New(fmt.Sprint("Bytes do not compare at ", idx)))
+                        }
+                        length -= int64(c)
+                        idx += int64(c)
+                }
+                g.OK()
+        })
+}
+
+func (t *TestGroup) Exec(desc string, p Program, srcfifo bool, flags []string) (*Run, error) {
+        var err error
+        run := &Run{}
+        args := []string{p.Path}
+        if srcfifo {
+                num := atomic.AddInt64(&srcSeq, 1)
+                run.Srcfile = path.Join(t.Runner.Testdir, fmt.Sprint("source", num))
+                if err = unix.Mkfifo(run.Srcfile, 0600); err != nil {
+                        return nil, err
+                }
+                read, write := io.Pipe()
+                t.writeFifo(run.Srcfile, read)
+                run.Srcin = write
+                args = append(args, "-s")
+                args = append(args, run.Srcfile)
+        }
+        if run.Stdin, err = run.Cmd.StdinPipe(); err != nil {
+                return nil, err
+        }
+        if run.Stdout, err = run.Cmd.StdoutPipe(); err != nil {
+                return nil, err
+        }
+        if run.Stderr, err = run.Cmd.StderrPipe(); err != nil {
+                return nil, err
+        }
+
+        run.Cmd.Path = p.Path
+        run.Cmd.Args = append(args, flags...)
+        run.Cmd.Dir = t.Runner.Testdir
+        if serr := run.Cmd.Start(); serr != nil {
+                return nil, serr
+        }
+        return run, nil
+}
+
+func (t *TestGroup) Fail(v ...interface{}) {
+        panic(fmt.Sprintln(v...))
+}
+
+func (t *TestGroup) writeFifo(srcfile string, read io.Reader) *Goroutine {
+        return t.Go("compare", func(g *Goroutine) {
+                fifo, err := os.OpenFile(srcfile, os.O_WRONLY, 0600)
+                if err != nil {
+                        fifo.Close()
+                        g.Panic(err)
+                }
+                if _, err := io.Copy(fifo, read); err != nil {
+                        fifo.Close()
+                        g.Panic(err)
+                }
+                if err := fifo.Close(); err != nil {
+                        g.Panic(err)
+                }
+                g.OK()
+        })
+}
diff --git a/xdelta3/go/src/xdelta/tgroup.go b/xdelta3/go/src/xdelta/tgroup.go
new file mode 100644
index 0000000..602b1e1
--- /dev/null
+++ b/xdelta3/go/src/xdelta/tgroup.go
@@ -0,0 +1,97 @@
+package xdelta
+
+import (
+        "fmt"
+        "runtime"
+        "sync"
+)
+
+type TestGroup struct {
+        *Runner
+        main *Goroutine
+        sync.Mutex
+        sync.WaitGroup
+        running []*Goroutine
+        errors []error
+        nonerrors []error  // For tolerated / expected conditions
+}
+
+type Goroutine struct {
+        *TestGroup
+        name string
+        done bool
+}
+
+func (g *Goroutine) String() string {
+        return fmt.Sprint("[", g.name, "]")
+}
+
+func (g *Goroutine) finish(err error) {
+        wait := false
+        tg := g.TestGroup
+        sbuf := make([]byte, 4096)
+        sbuf = sbuf[0:runtime.Stack(sbuf, false)]
+        if err != nil {
+                err = fmt.Errorf("%v:%v:%v", g.name, err, string(sbuf))
+        }
+        tg.Lock()
+        if g.done {
+                if err != nil {
+                        tg.nonerrors = append(tg.nonerrors, err)
+                }
+        } else {
+                wait = true
+                g.done = true
+                if err != nil {
+                        tg.errors = append(tg.errors, err)
+                }
+        }
+        tg.Unlock()
+        if wait {
+                tg.WaitGroup.Done()
+        }
+}
+
+func (g *Goroutine) OK() {
+        g.finish(nil)
+}
+
+func (g *Goroutine) Panic(err error) {
+        g.finish(err)
+        if g != g.TestGroup.main {
+                runtime.Goexit()
+        }
+}
+
+func (t *TestGroup) Main() *Goroutine { return t.main }
+
+func (t *TestGroup) Panic(err error) { t.Main().Panic(err) }
+
+func (t *TestGroup) Go(name string, f func(*Goroutine)) *Goroutine {
+        g := &Goroutine{t, name, false}
+        t.Lock()
+        t.WaitGroup.Add(1)
+        t.running = append(t.running, g)
+        t.Unlock()
+        go f(g)
+        return g
+}
+
+func (t *TestGroup) Wait(procs... *Run) {
+        t.Main().OK()
+        t.WaitGroup.Wait()
+        for _, p := range procs {
+                if err := p.Wait(); err != nil {
+                        t.errors = append(t.errors, err)
+                }
+        }
+        for _, err := range t.errors {
+                fmt.Println(":ERROR:", err)
+        }
+        for _, err := range t.nonerrors {
+                fmt.Println("(ERROR)", err)
+        }
+        if len(t.errors) != 0 {
+                t.Fail("Test failed with", len(t.errors), "errors")
+        }
+}
diff --git a/xdelta3/py-compile b/xdelta3/py-compile
deleted file mode 120000
index f90bf99..0000000
--- a/xdelta3/py-compile
+++ /dev/null
@@ -1 +0,0 @@
-/usr/local/share/automake-1.11/py-compile
\ No newline at end of file
diff --git a/xdelta3/run_release.sh b/xdelta3/run_release.sh
index ffd50fb..eb08afc 100755
--- a/xdelta3/run_release.sh
+++ b/xdelta3/run_release.sh
@@ -46,6 +46,7 @@ function setup {
     libtoolize
     automake --add-missing
     aclocal -I m4
+    autoheader
     automake
     autoheader
     autoconf
@@ -101,9 +102,10 @@ function buildlzma {
 function buildit {
     local host=$1
     local march=$2
-    local offsetbits=$3
-    local cargs=$4
-    local afl=$5
+    local usizebits=$3
+    local offsetbits=$4
+    local cargs=$5
+    local afl=$6
     local BM="${host}${march}"
     local USECC="${CC}"
     local USECXX="${CXX}"
@@ -115,8 +117,9 @@ function buildit {
         BM="${BM}-afl"
     fi
 
-    local D="build/${BM}/xoff${offsetbits}"
-    local BMD="${BM}-${offsetbits}"
+    local D="build/${BM}/usize${usizebits}/xoff${offsetbits}"
+    local BMD="${BM}-${usizebits}-${offsetbits}"
+
     local FULLD="${SRCDIR}/${D}"
     local CFLAGS="${march} ${cargs} -I${SRCDIR}/build/lib-${LIBBM}/include"
     local CXXFLAGS="${march} ${cargs} -I${SRCDIR}/build/lib-${LIBBM}/include"
@@ -199,8 +202,9 @@ function buildall {
     echo ""
 
     buildlzma "$1" "$2"
-    buildit "$1" "$2" 32 "-DXD3_USE_LARGEFILE64=0 $3" "$4"
-    buildit "$1" "$2" 64 "-DXD3_USE_LARGEFILE64=1 $3" "$4"
+    buildit "$1" "$2" 32 32 "-DXD3_USE_LARGESIZET=0 -DXD3_USE_LARGEFILE64=0 $3" "$4"
+    buildit "$1" "$2" 32 64 "-DXD3_USE_LARGESIZET=0 -DXD3_USE_LARGEFILE64=1 $3" "$4"
+    buildit "$1" "$2" 64 64 "-DXD3_USE_LARGESIZET=1 -DXD3_USE_LARGEFILE64=1 $3" "$4"
 }
 
 setup
diff --git a/xdelta3/testing/Makefile b/xdelta3/testing/Makefile
index f859f94..d0b9c9e 100644
--- a/xdelta3/testing/Makefile
+++ b/xdelta3/testing/Makefile
@@ -3,3 +3,6 @@ all:
 
 xdelta3regtest:
         (cd .. && make xdelta3regtest)
+
+xdelta3checksum:
+        (cd .. && make xdelta3checksum)
diff --git a/xdelta3/testing/checksum_test.cc b/xdelta3/testing/checksum_test.cc
new file mode 100644
index 0000000..9418d24
--- /dev/null
+++ b/xdelta3/testing/checksum_test.cc
@@ -0,0 +1,756 @@
+/* Copyright (C) 2007 Josh MacDonald */
+
+#include "test.h"
+#include <assert.h>
+#include <list>
+#include <vector>
+#include <algorithm>
+
+#include "../cpp-btree/btree_map.h"
+
+extern "C" {
+uint32_t xd3_large32_cksum_old (xd3_hash_cfg *cfg, const uint8_t *base, const usize_t look);
+uint32_t xd3_large32_cksum_update_old (xd3_hash_cfg *cfg, uint32_t cksum, 
+                                       const uint8_t *base, const usize_t look);
+
+uint64_t xd3_large64_cksum_old (xd3_hash_cfg *cfg, const uint8_t *base, const usize_t look);
+uint64_t xd3_large64_cksum_update_old (xd3_hash_cfg *cfg, uint64_t cksum, 
+                                       const uint8_t *base, const usize_t look);
+}
+
+using btree::btree_map;
+using std::list;
+using std::vector;
+
+// MLCG parameters
+// a, a*
+uint32_t good_32bit_values[] = {
+  1597334677U, // ...
+  741103597U, 887987685U,
+};
+
+// a, a*
+uint64_t good_64bit_values[] = {
+  1181783497276652981ULL, 4292484099903637661ULL,
+  7664345821815920749ULL, // ...
+};
+
+void print_header() {
+  static int hdr_cnt = 0;
+  if (hdr_cnt++ % 20 == 0) {
+    printf("%-32sConf\t\tCount\tUniq\tFull\tCover\tColls"
+           "\tMB/s\tIters\t#Colls\n", "Name");
+  }
+}
+
+struct true_type { };
+struct false_type { };
+
+template <typename Word>
+usize_t bitsof();
+
+template<>
+usize_t bitsof<unsigned int>() {
+  return sizeof(unsigned int) * 8;
+}
+
+template<>
+usize_t bitsof<unsigned long>() {
+  return sizeof(unsigned long) * 8;
+}
+
+template<>
+usize_t bitsof<unsigned long long>() {
+  return sizeof(unsigned long long) * 8;
+}
+
+template <typename Word>
+struct hhash {  // shift "s" bits leaving the high bits as a hash value for
+                // this checksum, which are the most "distant" in terms of the
+                // spectral test for the rabin_karp MLCG.  For short windows,
+                // the high bits aren't enough, XOR "mask" worth of these in.
+  Word operator()(const Word t, const Word s, const Word mask) {
+    return (t >> s) ^ (t & mask);
+  }
+};
+
+template <typename Word>
+Word good_word();
+
+template<>
+uint32_t good_word<uint32_t>() {
+  return good_32bit_values[0];
+}
+
+template<>
+uint64_t good_word<uint64_t>() {
+  return good_64bit_values[0];
+}
+
+// CLASSES
+
+#define SELF Word, CksumSize, CksumSkip, Hash, Compaction
+#define MEMBER template <typename Word,         \
+                         int CksumSize,         \
+                         int CksumSkip,         \
+                         typename Hash,         \
+                         int Compaction>
+
+MEMBER
+struct cksum_params {
+  typedef Word word_type;
+  typedef Hash hash_type;
+
+  static const int cksum_size = CksumSize;
+  static const int cksum_skip = CksumSkip;
+  static const int compaction = Compaction;
+};
+
+MEMBER
+struct rabin_karp : public cksum_params<SELF> {
+  // (a^cksum_size-1 c_0) + (a^cksum_size-2 c_1) ...
+  rabin_karp()
+    : powers(make_powers()),
+      product(powers[0] * good_word<Word>()),
+      incr_state(0) { }
+
+  static Word* make_powers() {
+    Word *p = new Word[CksumSize];
+    p[CksumSize - 1] = 1;
+    for (int i = CksumSize - 2; i >= 0; i--) {
+      p[i] = p[i + 1] * good_word<Word>();
+    }
+    return p;
+  }
+
+  ~rabin_karp() {
+    delete [] powers;
+  }
+
+  Word step(const uint8_t *ptr) {
+    Word h = 0;
+    for (int i = 0; i < CksumSize; i++) {
+      h += (ptr[i]) * powers[i];
+    }
+    return h;
+  }
+
+  Word state0(const uint8_t *ptr) {
+    incr_state = step(ptr);
+    return incr_state;
+  }
+
+  Word incr(const uint8_t *ptr) {
+    incr_state = good_word<Word>() * incr_state -
+      product * (ptr[-1]) + (ptr[CksumSize - 1]);
+    return incr_state;
+  }
+
+  const Word *const powers;
+  const Word  product;
+  Word        incr_state;
+};
+
+MEMBER
+struct with_stream : public cksum_params<SELF> {
+  xd3_stream stream;
+
+  with_stream()
+  {
+    xd3_config cfg;
+    memset (&stream, 0, sizeof (stream));
+    xd3_init_config (&cfg, 0);
+    cfg.smatch_cfg = XD3_SMATCH_SOFT;
+    cfg.smatcher_soft.large_look = CksumSize;
+    cfg.smatcher_soft.large_step = CksumSkip;
+    cfg.smatcher_soft.small_look = 4;
+    cfg.smatcher_soft.small_chain = 4;
+    cfg.smatcher_soft.small_lchain = 4;
+    cfg.smatcher_soft.max_lazy = 4;
+    cfg.smatcher_soft.long_enough = 4;
+    CHECK_EQ(0, xd3_config_stream (&stream, &cfg));
+
+    CHECK_EQ(0, xd3_size_hashtable (&stream,
+                                    1<<10 /* ignored */,
+                                    stream.smatcher.large_look,
+                                    & stream.large_hash));
+  }
+  ~with_stream() 
+  {
+    xd3_free_stream (&stream);
+  }
+};
+
+MEMBER
+struct large_cksum : public with_stream<SELF> {
+  Word step(const uint8_t *ptr) {
+    return xd3_large_cksum (&this->stream.large_hash, ptr, CksumSize);
+  }
+
+  Word state0(const uint8_t *ptr) {
+    incr_state = step(ptr);
+    return incr_state;
+  }
+
+  Word incr(const uint8_t *ptr) {
+    incr_state = xd3_large_cksum_update (&this->stream.large_hash, 
+                                         incr_state, ptr - 1, CksumSize);
+    return incr_state;
+  }
+
+  Word incr_state;
+};
+
+#if SIZEOF_USIZE_T == 4
+#define xd3_large_cksum_old         xd3_large32_cksum_old
+#define xd3_large_cksum_update_old  xd3_large32_cksum_update_old
+#elif SIZEOF_USIZE_T == 8
+#define xd3_large_cksum_old         xd3_large64_cksum_old
+#define xd3_large_cksum_update_old  xd3_large64_cksum_update_old
+#endif
+
+MEMBER
+struct large_cksum_old : public with_stream<SELF> {
+  Word step(const uint8_t *ptr) {
+    return xd3_large_cksum_old (&this->stream.large_hash, ptr, CksumSize);
+  }
+
+  Word state0(const uint8_t *ptr) {
+    incr_state = step(ptr);
+    return incr_state;
+  }
+
+  Word incr(const uint8_t *ptr) {
+    incr_state = xd3_large_cksum_update_old (&this->stream.large_hash, 
+                                             incr_state, ptr - 1, CksumSize);
+    return incr_state;
+  }
+
+  Word incr_state;
+};
+
+// TESTS
+
+template <typename Word>
+struct file_stats {
+  typedef const uint8_t* ptr_type;
+  typedef Word word_type;
+  typedef btree::btree_multimap<word_type, ptr_type> table_type;
+  typedef typename table_type::iterator table_iterator;
+
+  usize_t cksum_size;
+  usize_t cksum_skip;
+  usize_t unique;
+  usize_t unique_values;
+  usize_t count;
+  table_type table;
+
+  file_stats(usize_t size, usize_t skip)
+    : cksum_size(size),
+      cksum_skip(skip),
+      unique(0),
+      unique_values(0),
+      count(0) {
+  }
+
+  void reset() {
+    unique = 0;
+    unique_values = 0;
+    count = 0;
+    table.clear();
+  }
+
+  void update(word_type word, ptr_type ptr) {
+    table_iterator t_i = table.find(word);
+
+    count++;
+    if (t_i != table.end()) {
+      int collisions = 0;
+      for (table_iterator p_i = t_i;
+           p_i != table.end() && p_i->first == word;
+           ++p_i) {
+        if (memcmp(p_i->second, ptr, cksum_size) == 0) {
+          return;
+        }
+        collisions++;
+      }
+      if (collisions >= 1000) {
+        fprintf(stderr, "Something is not right, lots of collisions=%d\n", 
+                collisions);
+        abort();
+      }
+    } else {
+      unique_values++;
+    }
+    unique++;
+    table.insert(std::make_pair(word, ptr));
+    return;
+  }
+
+  void freeze() {
+    table.clear();
+  }
+};
+
+struct test_result_base;
+
+static vector<test_result_base*> all_tests;
+
+struct test_result_base {
+  virtual ~test_result_base() {
+  }
+  virtual void reset() = 0;
+  virtual void print() = 0;
+  virtual void get(const uint8_t* buf, const size_t buf_size, 
+                   usize_t iters) = 0;
+  virtual void stat() = 0;
+  virtual usize_t count() = 0;
+  virtual usize_t dups() = 0;
+  virtual double uniqueness() = 0;
+  virtual double fullness() = 0;
+  virtual double collisions() = 0;
+  virtual double coverage() = 0;
+  virtual double compression() = 0;
+  virtual double time() = 0;
+  virtual double total_time() = 0;
+  virtual usize_t total_count() = 0;
+  virtual usize_t total_dups() = 0;
+};
+
+template <typename Checksum>
+struct test_result : public test_result_base {
+  Checksum cksum;
+  const char *test_name;
+  file_stats<typename Checksum::word_type> fstats;
+  usize_t test_size;
+  usize_t n_steps;
+  usize_t n_incrs;
+  typename Checksum::word_type s_bits;
+  typename Checksum::word_type s_mask;
+  usize_t t_entries;
+  usize_t h_bits;
+  usize_t h_buckets_full;
+  char *hash_table;
+  long accum_millis;
+  usize_t accum_iters;
+
+  // These are not reset
+  double accum_time;
+  usize_t accum_count;
+  usize_t accum_dups;
+  usize_t accum_colls;
+  size_t accum_size;
+
+  test_result(const char *name)
+    : test_name(name),
+      fstats(Checksum::cksum_size, Checksum::cksum_skip),
+      hash_table(NULL),
+      accum_millis(0),
+      accum_iters(0),
+      accum_time(0.0),
+      accum_count(0),
+      accum_dups(0),
+      accum_colls(0),
+      accum_size(0) {
+    all_tests.push_back(this);
+  }
+
+  ~test_result() {
+    reset();
+  }
+
+  void reset() {
+    // size of file
+    test_size = 0;
+
+    // count
+    n_steps = 0;
+    n_incrs = 0;
+
+    // four values used by new_table()/summarize_table()
+    s_bits = 0;
+    s_mask = 0;
+    t_entries = 0;
+    h_bits = 0;
+    h_buckets_full = 0;
+
+    accum_millis = 0;
+    accum_iters = 0;
+
+    fstats.reset();
+
+    // temporary
+    if (hash_table) {
+      delete(hash_table);
+      hash_table = NULL;
+    }
+  }
+
+  usize_t count() {
+    if (Checksum::cksum_skip == 1) {
+      return n_incrs;
+    } else {
+      return n_steps;
+    }
+  }
+
+  usize_t dups() {
+    return fstats.count - fstats.unique;
+  }
+
+  /* Fraction of distinct strings of length cksum_size which are not
+   * represented in the hash table. */
+  double collisions() {
+    return (fstats.unique - fstats.unique_values) / (double) fstats.unique;
+  }
+  usize_t colls() {
+    return (fstats.unique - fstats.unique_values);
+  }
+
+  double uniqueness() {
+    return 1.0 - (double) dups() / count();
+  }
+
+  double fullness() {
+    return (double) h_buckets_full / (1 << h_bits);
+  }
+
+  double coverage() {
+    return (double) h_buckets_full / uniqueness() / count();
+  }
+
+  double compression() {
+    return 1.0 - coverage();
+  }
+
+  double time() {
+    return (double) accum_millis / accum_iters;
+  }
+
+  double total_time() {
+    return accum_time;
+  }
+
+  usize_t total_count() {
+    return accum_count;
+  }
+
+  usize_t total_dups() {
+    return accum_dups;
+  }
+
+  usize_t total_colls() {
+    return accum_dups;
+  }
+
+  void stat() {
+    accum_time += time();
+    accum_count += count();
+    accum_dups += dups();
+    accum_colls += colls();
+    accum_size += test_size;
+  }
+
+  void print() {
+    if (fstats.count != count()) {
+      fprintf(stderr, "internal error: %" W "d != %" W "d\n", fstats.count, count());
+      abort();
+    }
+    print_header();
+    printf("%-32s%d/%d 2^%" W "u\t%" W "u\t%0.4f\t%.4f\t%.4f\t%.1e\t%.2f\t"
+           "%" W "u\t%" W "u\n",
+           test_name,
+           Checksum::cksum_size,
+           Checksum::cksum_skip,
+           h_bits,
+           count(),
+           uniqueness(),
+           fullness(),
+           coverage(),
+           collisions(),
+           0.001 * accum_iters * test_size / accum_millis,
+           accum_iters,
+           colls());
+  }
+
+  usize_t size_log2 (usize_t slots) {
+    usize_t bits = bitsof<typename Checksum::word_type>() - 1;
+    usize_t i;
+
+    for (i = 3; i <= bits; i += 1) {
+      if (slots <= (1U << i)) {
+        return i - Checksum::compaction;
+      }
+    }
+
+    return bits;
+  }
+
+  void new_table(usize_t entries) {
+    t_entries = entries;
+    h_bits = size_log2(entries);
+
+    usize_t n = 1 << h_bits;
+
+    s_bits = bitsof<typename Checksum::word_type>() - h_bits;
+    s_mask = n - 1U;
+
+    hash_table = new char[n / 8];
+    memset(hash_table, 0, n / 8);
+  }
+
+  int get_table_bit(usize_t i) {
+    return hash_table[i/8] & (1 << i%8);
+  }
+
+  int set_table_bit(usize_t i) {
+    return hash_table[i/8] |= (1 << i%8);
+  }
+
+  void summarize_table() {
+    usize_t n = 1 << h_bits;
+    usize_t f = 0;
+    for (usize_t i = 0; i < n; i++) {
+      if (get_table_bit(i)) {
+        f++;
+      }
+    }
+    h_buckets_full = f;
+  }
+
+  void get(const uint8_t* buf, const size_t buf_size, usize_t test_iters) {
+    typename Checksum::hash_type hash;
+    const uint8_t *ptr;
+    const uint8_t *end;
+    usize_t periods;
+    int64_t last_offset;
+    int64_t stop;
+
+    test_size = buf_size;
+    last_offset = buf_size - Checksum::cksum_size;
+
+    if (last_offset < 0) {
+      periods = 0;
+      n_steps = 0;
+      n_incrs = 0;
+      stop = -Checksum::cksum_size;
+    } else {
+      periods = last_offset / Checksum::cksum_skip;
+      n_steps = periods + 1;
+      n_incrs = last_offset + 1;
+      stop = last_offset - (periods + 1) * Checksum::cksum_skip;
+    }
+
+    // Compute file stats once.
+    if (fstats.unique_values == 0) {
+      if (Checksum::cksum_skip == 1) {
+        for (size_t i = 0; i <= buf_size - Checksum::cksum_size; i++) {
+          fstats.update(hash(cksum.step(buf + i), s_bits, s_mask), buf + i);
+        }
+      } else {
+        ptr = buf + last_offset;
+        end = buf + stop;
+
+        for (; ptr != end; ptr -= Checksum::cksum_skip) {
+          fstats.update(hash(cksum.step(ptr), s_bits, s_mask), ptr);
+        }
+      }
+      fstats.freeze();
+    }
+
+    long start_test = get_millisecs_now();
+
+    if (Checksum::cksum_skip != 1) {
+      new_table(n_steps);
+
+      for (usize_t i = 0; i < test_iters; i++) {
+        ptr = buf + last_offset;
+        end = buf + stop;
+
+        for (; ptr != end; ptr -= Checksum::cksum_skip) {
+          set_table_bit(hash(cksum.step(ptr), s_bits, s_mask));
+        }
+      }
+
+      summarize_table();
+    }
+
+    stop = buf_size - Checksum::cksum_size + 1;
+    if (stop < 0) {
+      stop = 0;
+    }
+
+    if (Checksum::cksum_skip == 1) {
+      new_table(n_incrs);
+
+      for (usize_t i = 0; i < test_iters; i++) {
+        ptr = buf;
+        end = buf + stop;
+
+        if (ptr != end) {
+          set_table_bit(hash(cksum.state0(ptr++), s_bits, s_mask));
+        }
+
+        for (; ptr != end; ptr++) {
+          typename Checksum::word_type w = cksum.incr(ptr);
+          CHECK_EQ(w, cksum.step(ptr));
+          set_table_bit(hash(w, s_bits, s_mask));
+        }
+      }
+
+      summarize_table();
+    }
+
+    accum_iters += test_iters;
+    accum_millis += get_millisecs_now() - start_test;
+  }
+};
+
+static int read_whole_file(const char *name,
+                           uint8_t **buf_ptr,
+                           size_t *buf_len) {
+  main_file file;
+  int ret;
+  xoff_t len;
+  size_t nread;
+  main_file_init(&file);
+  file.filename = name;
+  ret = main_file_open(&file, name, XO_READ);
+  if (ret != 0) {
+    fprintf(stderr, "open failed\n");
+    goto exit;
+  }
+  ret = main_file_stat(&file, &len);
+  if (ret != 0) {
+    fprintf(stderr, "stat failed\n");
+    goto exit;
+  }
+  
+  (*buf_len) = (size_t)len;
+  (*buf_ptr) = (uint8_t*) main_malloc(*buf_len);
+  ret = main_file_read(&file, *buf_ptr, *buf_len, &nread,
+                       "read failed");
+  if (ret == 0 && *buf_len == nread) {
+    ret = 0;
+  } else {
+    fprintf(stderr, "invalid read\n");
+    ret = XD3_INTERNAL;
+  }
+ exit:
+  main_file_cleanup(&file);
+  return ret;
+}
+
+int main(int argc, char** argv) {
+  int i;
+  uint8_t *buf = NULL;
+  size_t buf_len = 0;
+  int ret;
+
+  if (argc <= 1) {
+    fprintf(stderr, "usage: %s file ...\n", argv[0]);
+    return 1;
+  }
+
+// TODO: The xdelta3-hash.h code is identical now; add sameness test.
+// using rabin_karp<> template.
+#define TEST(T,Z,S,C)                                   \
+  test_result<large_cksum<T,Z,S,hhash<T>,C>>            \
+    _xck_ ## T ## _ ## Z ## _ ## S ## _ ## C            \
+    ("xck_" #T "_" #Z "_" #S "_" #C);                   \
+  test_result<large_cksum_old<T,Z,S,hhash<T>,C>>        \
+    _old_ ## T ## _ ## Z ## _ ## S ## _ ## C            \
+    ("old_" #T "_" #Z "_" #S "_" #C)
+
+#define TESTS(SIZE, SKIP)        \
+  TEST(usize_t, SIZE, SKIP, 1);  \
+  TEST(usize_t, SIZE, SKIP, 2)
+   
+  TESTS(5, 1);
+  TESTS(6, 1);
+  TESTS(7, 1);
+  TESTS(8, 1);
+  TESTS(9, 1);
+  TESTS(10, 1);
+  TESTS(11, 1);
+  TESTS(12, 1);
+  TESTS(13, 1);
+  TESTS(14, 1);
+  TESTS(15, 1);
+  TESTS(16, 1);
+  TESTS(17, 1);
+  TESTS(18, 1);
+  TESTS(19, 1);
+  TESTS(20, 1);
+  TESTS(21, 1);
+  TESTS(22, 1);
+  TESTS(23, 1);
+  TESTS(24, 1);
+  TESTS(25, 1);
+  TESTS(26, 1);
+  TESTS(27, 1);
+  TESTS(28, 1);
+  TESTS(29, 1);
+  TESTS(30, 1);
+  TESTS(31, 1);
+  TESTS(32, 1);
+  TESTS(33, 1);
+  TESTS(34, 1);
+  TESTS(35, 1);
+  TESTS(36, 1);
+  TESTS(37, 1);
+  TESTS(38, 1);
+  TESTS(39, 1);
+
+
+  for (i = 1; i < argc; i++) {
+    if ((ret = read_whole_file(argv[i],
+                               & buf,
+                               & buf_len))) {
+      return 1;
+    }
+
+    fprintf(stderr, "file %s is %zu bytes\n",
+            argv[i], buf_len);
+
+    double min_time = -1.0;
+    double min_compression = 0.0;
+
+    for (vector<test_result_base*>::iterator iter = all_tests.begin();
+         iter != all_tests.end(); ++iter) {
+      test_result_base *test = *iter;
+      test->reset();
+
+      usize_t iters = 1;
+      long start_test = get_millisecs_now();
+
+      do {
+        test->get(buf, buf_len, iters);
+        iters *= 3;
+        iters /= 2;
+      } while (get_millisecs_now() - start_test < 2000);
+
+      test->stat();
+
+      if (min_time < 0.0) {
+        min_compression = test->compression();
+        min_time = test->time();
+      }
+
+      if (min_time > test->time()) {
+        min_time = test->time();
+      }
+
+      if (min_compression > test->compression()) {
+        min_compression = test->compression();
+      }
+
+      test->print();
+    }
+
+    main_free(buf);
+    buf = NULL;
+  }
+
+  return 0;      
+}
diff --git a/xdelta3/testing/checksum_test_c.c b/xdelta3/testing/checksum_test_c.c
new file mode 100644
index 0000000..8f0507a
--- /dev/null
+++ b/xdelta3/testing/checksum_test_c.c
@@ -0,0 +1,174 @@
+#include "../xdelta3.c"
+
+// OLD CHECKSUM CODE
+
+#define PERMUTE32(x) (__single_hash32[x])
+#define PERMUTE64(x) (__single_hash64[x])
+
+const uint16_t __single_hash32[256] =
+{
+  /* This hashes the input alphabet (Scheme SLIB pseudo-random). */
+  0xbcd1, 0xbb65, 0x42c2, 0xdffe, 0x9666, 0x431b, 0x8504, 0xeb46,
+  0x6379, 0xd460, 0xcf14, 0x53cf, 0xdb51, 0xdb08, 0x12c8, 0xf602,
+  0xe766, 0x2394, 0x250d, 0xdcbb, 0xa678, 0x02af, 0xa5c6, 0x7ea6,
+  0xb645, 0xcb4d, 0xc44b, 0xe5dc, 0x9fe6, 0x5b5c, 0x35f5, 0x701a,
+  0x220f, 0x6c38, 0x1a56, 0x4ca3, 0xffc6, 0xb152, 0x8d61, 0x7a58,
+  0x9025, 0x8b3d, 0xbf0f, 0x95a3, 0xe5f4, 0xc127, 0x3bed, 0x320b,
+  0xb7f3, 0x6054, 0x333c, 0xd383, 0x8154, 0x5242, 0x4e0d, 0x0a94,
+  0x7028, 0x8689, 0x3a22, 0x0980, 0x1847, 0xb0f1, 0x9b5c, 0x4176,
+  0xb858, 0xd542, 0x1f6c, 0x2497, 0x6a5a, 0x9fa9, 0x8c5a, 0x7743,
+  0xa8a9, 0x9a02, 0x4918, 0x438c, 0xc388, 0x9e2b, 0x4cad, 0x01b6,
+  0xab19, 0xf777, 0x365f, 0x1eb2, 0x091e, 0x7bf8, 0x7a8e, 0x5227,
+  0xeab1, 0x2074, 0x4523, 0xe781, 0x01a3, 0x163d, 0x3b2e, 0x287d,
+  0x5e7f, 0xa063, 0xb134, 0x8fae, 0x5e8e, 0xb7b7, 0x4548, 0x1f5a,
+  0xfa56, 0x7a24, 0x900f, 0x42dc, 0xcc69, 0x02a0, 0x0b22, 0xdb31,
+  0x71fe, 0x0c7d, 0x1732, 0x1159, 0xcb09, 0xe1d2, 0x1351, 0x52e9,
+  0xf536, 0x5a4f, 0xc316, 0x6bf9, 0x8994, 0xb774, 0x5f3e, 0xf6d6,
+  0x3a61, 0xf82c, 0xcc22, 0x9d06, 0x299c, 0x09e5, 0x1eec, 0x514f,
+  0x8d53, 0xa650, 0x5c6e, 0xc577, 0x7958, 0x71ac, 0x8916, 0x9b4f,
+  0x2c09, 0x5211, 0xf6d8, 0xcaaa, 0xf7ef, 0x287f, 0x7a94, 0xab49,
+  0xfa2c, 0x7222, 0xe457, 0xd71a, 0x00c3, 0x1a76, 0xe98c, 0xc037,
+  0x8208, 0x5c2d, 0xdfda, 0xe5f5, 0x0b45, 0x15ce, 0x8a7e, 0xfcad,
+  0xaa2d, 0x4b5c, 0xd42e, 0xb251, 0x907e, 0x9a47, 0xc9a6, 0xd93f,
+  0x085e, 0x35ce, 0xa153, 0x7e7b, 0x9f0b, 0x25aa, 0x5d9f, 0xc04d,
+  0x8a0e, 0x2875, 0x4a1c, 0x295f, 0x1393, 0xf760, 0x9178, 0x0f5b,
+  0xfa7d, 0x83b4, 0x2082, 0x721d, 0x6462, 0x0368, 0x67e2, 0x8624,
+  0x194d, 0x22f6, 0x78fb, 0x6791, 0xb238, 0xb332, 0x7276, 0xf272,
+  0x47ec, 0x4504, 0xa961, 0x9fc8, 0x3fdc, 0xb413, 0x007a, 0x0806,
+  0x7458, 0x95c6, 0xccaa, 0x18d6, 0xe2ae, 0x1b06, 0xf3f6, 0x5050,
+  0xc8e8, 0xf4ac, 0xc04c, 0xf41c, 0x992f, 0xae44, 0x5f1b, 0x1113,
+  0x1738, 0xd9a8, 0x19ea, 0x2d33, 0x9698, 0x2fe9, 0x323f, 0xcde2,
+  0x6d71, 0xe37d, 0xb697, 0x2c4f, 0x4373, 0x9102, 0x075d, 0x8e25,
+  0x1672, 0xec28, 0x6acb, 0x86cc, 0x186e, 0x9414, 0xd674, 0xd1a5
+};
+
+const uint32_t __single_hash64[256] =
+{
+  /* http://random.org 2014.10.24 */
+  0xd25e9f0a, 0xb1af9d5e, 0xb753dfa2, 0x157050f7,  /* 0 */
+  0xc84b072c, 0xdd14fe7c, 0xf92208c3, 0xdf08a0c0,
+  0x63a5c118, 0x76f5d90f, 0xa2f8b93e, 0xb6c12d22,
+  0xaf074957, 0x966fb7d9, 0x62f7b785, 0xb40e8a09,
+  0x0a811d5d, 0x323a6daa, 0xb62f7c5b, 0xfdcb9a53,
+  0xf25a9067, 0x4506bc7a, 0xff58a74b, 0x5ae62817,
+  0x74097675, 0x722c0fd9, 0x116a2a66, 0x65f76728,
+  0x72c79651, 0xe043cf9d, 0x64b867c7, 0x6604834f,
+  0xcdca58a6, 0x0f164e2d, 0x24515f05, 0x632cdbf8,
+  0x18091d4a, 0x3eff4128, 0x673d1c33, 0xd8e10c71,
+  0x1a3edf11, 0xba52892f, 0xa56949e0, 0xf3e1dd77,  /* 10 */
+  0x86fcbe3e, 0x138d66d0, 0x4fc98359, 0xc22e5dd6,
+  0xc59f2267, 0x6c6dd739, 0xe03da190, 0x07e8469c,
+  0xadcfb02c, 0x00d3b0d9, 0xa1f44918, 0x8bd84d87,
+  0x08ec9ec1, 0xbbcd156f, 0xb57718e3, 0x3177e752,
+  0xf52a4d70, 0xde7aaad9, 0x075f1da0, 0x21ba00c6,
+  0xb9469a5c, 0xcf08d5ba, 0x91ac9edc, 0xc6167b63,
+  0xc1974919, 0xc8c8d195, 0x4b1996dd, 0xeff8991c,
+  0xf7f66c6b, 0x25b012e2, 0x59d12a98, 0xea40d3cc,
+  0x41f9970b, 0xec48101a, 0xa3bdcf90, 0x99f16905,
+  0x27af6c97, 0xc849af37, 0x49cad89b, 0xf48c2278,  /* 20 */
+  0x5529c3d8, 0x9e7d6dce, 0x16feb52d, 0xf1b0aca1,
+  0xaf28fccb, 0x48e4ce3c, 0xc4436617, 0x64524e3e,
+  0x61806681, 0x6384f2d7, 0x1172880f, 0x34a5ef5f,
+  0xcc8cc0a8, 0x66e8f100, 0x2866085f, 0xba9b1b2d,
+  0x51285949, 0x2be4b574, 0x889b1ef5, 0x3dbe920d,
+  0x9277a62f, 0x0584a9f6, 0x085d8fc4, 0x4b5d403d,
+  0x4e46ca78, 0x3294c2f9, 0x29313e70, 0xe4f09b24,
+  0xe73b331c, 0x072f5552, 0x2e390b78, 0xea0021ca,
+  0xd8f40320, 0xed0e16fd, 0x7de9cf7a, 0xf17e3d6c,
+  0x8df1bd85, 0x052cae67, 0x3486e512, 0x3a1c09b8,  /* 30 */
+  0x6c2a7b4e, 0x83455753, 0xbc0353ac, 0x0ffe20b6,
+  0x5fdcef85, 0x010f506c, 0x595ce972, 0xe28680d0,
+  0xa7e216b2, 0xa392ee0f, 0x25b73faa, 0x2b1f4983,
+  0xeeaefe98, 0x1d3d9cbc, 0x6aebe97b, 0x8b7b3584,
+  0x9e6a9a07, 0xd37f1e99, 0x4ac2a441, 0x8ae9a213,
+  0x7d0e27d7, 0x5de54b9a, 0x8621de1f, 0xf0f2f866,
+  0xcb08d275, 0x49c3f87e, 0xd5ee68c1, 0x9802fc77,
+  0x68be6c5e, 0x65aa8c27, 0xf423d5f7, 0x10ec5502,
+  0x9909bce1, 0x509cdf1b, 0x338fea72, 0x2733e9bf,
+  0xf92f4fd7, 0x87738ea2, 0x931a8bbc, 0x0a5c9155,  /* 40 */
+  0xbe5edd9b, 0xadbf5838, 0x0338f8d2, 0x290da210,
+  0x390c37d8, 0xe7cffae8, 0x20617ebe, 0x464322dd,
+  0x7b3c4e78, 0xac142dcb, 0x2d5cef76, 0xd8fe49fc,
+  0x60f4e9a9, 0x7473816f, 0x0dc35f39, 0x5eed80c1,
+  0x0cb55ab6, 0x1d3ac541, 0x13c7f529, 0x7bffdf4a,
+  0xe334785b, 0x85263ec1, 0xd132ae56, 0x7c868b9e,
+  0x47f60638, 0x1012b979, 0x81c31dd3, 0x1af868c8,
+  0x0c5d0742, 0xd1b3e1a2, 0x5873200a, 0xf848465c,
+  0x0fc4d596, 0x609c18af, 0xc9f5a480, 0xd1a94a84,
+  0xa1431a3f, 0x7de8bb1a, 0x25f1256b, 0x1dcc732c,  /* 50 */
+  0x6aa1549a, 0xa2367281, 0x32f2a77e, 0x82e62a0f,
+  0x045cbb56, 0x74b2027c, 0xd71a32d9, 0x022e7cb5,
+  0xe99be177, 0x60222fdf, 0xd69681ca, 0x9008ee2c,
+  0x32923db4, 0xcf82bf97, 0x38960a5b, 0xb3503d5b,
+  0x9bd4c7f2, 0x33c029c8, 0x1ef504a3, 0xdb249d3b,
+  0x91e89676, 0x4ca43b36, 0x9191433c, 0x465d5dc4,
+  0xf4dcb118, 0x9d11dd00, 0xb592f058, 0xdbe5ce30,
+  0x74790d92, 0x779850a8, 0x7180d25b, 0xfa951d99,
+  0x5990935a, 0x921cb022, 0x3b7c39bc, 0x6a38a7c7,
+  0xdc22703b, 0x142bab3b, 0x4e3d9479, 0x44bb8482,  /* 60 */
+  0x8043abce, 0xfebe832a, 0x8e6a2f98, 0x4d43c4fe,
+  0xd192a70a, 0x802f3c3a, 0x5d11bbab, 0x2665d241,
+  0xb3f3a680, 0x3a8d223f, 0xcf82cdb4, 0x4ed28743,
+};
+
+uint64_t
+xd3_large64_cksum_old (xd3_hash_cfg *ignore, const uint8_t *base, const usize_t look)
+{
+  static const uint64_t kBits = 32;
+  static const uint64_t kMask = 0xffffffff;
+  usize_t i = 0;
+  uint64_t low  = 0;
+  uint64_t high = 0;
+
+  for (; i < look; i += 1)
+    {
+      low  += PERMUTE64(*base++);
+      high += low;
+    }
+
+  return ((high & kMask) << kBits) | (low & kMask);
+}
+
+uint64_t
+xd3_large64_cksum_update_old (xd3_hash_cfg *ignore, const uint64_t cksum,
+                              const uint8_t *base, const usize_t look)
+{
+  static const uint64_t kBits = 32;
+  static const uint64_t kMask = 0xffffffff;
+  uint64_t old_c = PERMUTE64(base[0]);
+  uint64_t new_c = PERMUTE64(base[look]);
+  uint64_t low   = ((cksum & kMask) - old_c + new_c) & kMask;
+  uint64_t high  = ((cksum >> kBits) - (old_c * look) + low) & kMask;
+  return (high << kBits) | low;
+}
+
+uint32_t
+xd3_large32_cksum_old (xd3_hash_cfg *ignore, const uint8_t *base, const usize_t look)
+{
+  static const uint32_t kBits = 16;
+  static const uint32_t kMask = 0xffff;
+  usize_t i = 0;
+  uint32_t low  = 0;
+  uint32_t high = 0;
+
+  for (; i < look; i += 1)
+    {
+      low  += PERMUTE32(*base++);
+      high += low;
+    }
+
+  return ((high & kMask) << kBits) | (low & kMask);
+}
+
+uint32_t
+xd3_large32_cksum_update_old (xd3_hash_cfg *ignore, const uint32_t cksum,
+                              const uint8_t *base, const usize_t look)
+{
+  static const uint32_t kBits = 16;
+  static const uint32_t kMask = 0xffff;
+  uint32_t old_c = PERMUTE32(base[0]);
+  uint32_t new_c = PERMUTE32(base[look]);
+  uint32_t low   = ((cksum & kMask) - old_c + new_c) & kMask;
+  uint32_t high  = ((cksum >> kBits) - (old_c * look) + low) & kMask;
+  return (high << kBits) | low;
+}
diff --git a/xdelta3/testing/delta.h b/xdelta3/testing/delta.h
index b0cca7c..f874bf2 100644
--- a/xdelta3/testing/delta.h
+++ b/xdelta3/testing/delta.h
@@ -53,13 +53,13 @@ public:
       xd3_winst &winst = stream_.whole_target.inst[i];
       switch (winst.type) {
       case XD3_RUN: 
-        DP(RINT "%"Q"u run %u\n", winst.position, winst.size);
+        DP(RINT "%"Q" run %u\n", winst.position, winst.size);
         break;
       case XD3_ADD: 
-        DP(RINT "%"Q"u add %u\n", winst.position, winst.size);
+        DP(RINT "%"Q" add %u\n", winst.position, winst.size);
         break;
       default:
-        DP(RINT "%"Q"u copy %u @ %"Q"u (mode %u)\n", 
+        DP(RINT "%"Q" copy %u @ %"Q" (mode %u)\n", 
            winst.position, winst.size, winst.addr, winst.mode);
         break;
       }
diff --git a/xdelta3/testing/regtest.cc b/xdelta3/testing/regtest.cc
index b2cdaa5..e83af4b 100644
--- a/xdelta3/testing/regtest.cc
+++ b/xdelta3/testing/regtest.cc
@@ -102,8 +102,8 @@ public:
 
       xoff_t blks = target_iterator.Blocks();
 
-      IF_DEBUG2(XPR(NTR "target in %s: %"Q"u..%"Q"u %"Q"u(%"Q"u) "
-                    "verified %"Q"u\n",
+      IF_DEBUG2(XPR(NTR "target in %s: %"Q"..%"Q" %"Q"(%"Q") "
+                    "verified %"Q"\n",
                    encoding ? "encoding" : "decoding",
                    target_iterator.Offset(),
                    target_iterator.Offset() + target_block.Size(),
diff --git a/xdelta3/testing/run_release.sh b/xdelta3/testing/run_release.sh
new file mode 100755
index 0000000..85ed1f7
--- /dev/null
+++ b/xdelta3/testing/run_release.sh
@@ -0,0 +1,2 @@
+#!/bin/sh
+(cd .. && ./run_release.sh)
diff --git a/xdelta3/testing/test.h b/xdelta3/testing/test.h
index 1eabcbd..7de24fb 100644
--- a/xdelta3/testing/test.h
+++ b/xdelta3/testing/test.h
@@ -18,8 +18,8 @@ extern "C" {
 
 #define CHECK_OP(x,y,OP) \
   do { \
-    typeof(x) _x(x); \
-    typeof(x) _y(y); \
+    __typeof__(x) _x(x); \
+    __typeof__(x) _y(y); \
     if (!(_x OP _y)) { \
       cerr << __FILE__ << ":" << __LINE__ << " Check failed: " << #x " " #OP " " #y << endl; \
       cerr << __FILE__ << ":" << __LINE__ << " {0} " << _x << endl; \
@@ -68,5 +68,3 @@ pair<T, U> make_pair(const T& t, const U& u) {
 
 using std::min;
 using std::max;
-
-
diff --git a/xdelta3/xdelta3-blkcache.h b/xdelta3/xdelta3-blkcache.h
index f8fa8e2..20a2a4a 100644
--- a/xdelta3/xdelta3-blkcache.h
+++ b/xdelta3/xdelta3-blkcache.h
@@ -26,6 +26,9 @@
 typedef struct _main_blklru      main_blklru;
 typedef struct _main_blklru_list main_blklru_list;
 
+
+#define XD3_INVALID_OFFSET XOFF_T_MAX
+
 struct _main_blklru_list
 {
   main_blklru_list  *next;
@@ -125,7 +128,7 @@ main_set_source (xd3_stream *stream, xd3_cmd cmd,
   /* Note: The API requires a power-of-two blocksize and srcwinsz
    * (-B).  The logic here will use a single block if the entire file
    * is known to fit into srcwinsz. */
-  option_srcwinsz = xd3_pow2_roundup (option_srcwinsz);
+  option_srcwinsz = xd3_xoff_roundup (option_srcwinsz);
 
   /* Though called "lru", it is not LRU-specific.  We always allocate
    * a maximum number of source block buffers.  If the entire file
@@ -153,7 +156,7 @@ main_set_source (xd3_stream *stream, xd3_cmd cmd,
    * is the point at which external decompression may begin.  Set the
    * system for a single block. */
   lru_size = 1;
-  lru[0].blkno = (xoff_t) -1;
+  lru[0].blkno = XD3_INVALID_OFFSET;
   blksize = option_srcwinsz;
   main_blklru_list_push_back (& lru_list, & lru[0]);
   XD3_ASSERT (blksize != 0);
@@ -162,7 +165,7 @@ main_set_source (xd3_stream *stream, xd3_cmd cmd,
   source->blksize  = blksize;
   source->name     = sfile->filename;
   source->ioh      = sfile;
-  source->curblkno = (xoff_t) -1;
+  source->curblkno = XD3_INVALID_OFFSET;
   source->curblk   = NULL;
   source->max_winsize = option_srcwinsz;
 
@@ -258,7 +261,7 @@ main_set_source (xd3_stream *stream, xd3_cmd cmd,
 
       if (option_verbose > 1)
         {
-          short_sprintf (nbufs, " #bufs %u", lru_size);
+          short_sprintf (nbufs, " #bufs %"W"u", lru_size);
         }
 
       XPR(NT "source %s %s blksize %s window %s%s%s\n",
@@ -293,7 +296,7 @@ main_getblk_lru (xd3_source *source, xoff_t blkno,
           return 0;
         }
       /* No going backwards in a sequential scan. */
-      if (blru->blkno != (xoff_t) -1 && blru->blkno > blkno)
+      if (blru->blkno != XD3_INVALID_OFFSET && blru->blkno > blkno)
         {
           return XD3_TOOFARBACK;
         }
@@ -333,7 +336,7 @@ main_getblk_lru (xd3_source *source, xoff_t blkno,
   lru_filled += 1;
   (*is_new) = 1;
   (*blrup) = blru;
-  blru->blkno = -1;
+  blru->blkno = XD3_INVALID_OFFSET;
   return 0;
 }
 
@@ -443,7 +446,7 @@ main_read_seek_source (xd3_stream *stream,
           sfile->source_position += nread;
           blru->size = nread;
 
-          IF_DEBUG1 (DP(RINT "[getblk] skip blkno %"Q"u size %u\n",
+          IF_DEBUG1 (DP(RINT "[getblk] skip blkno %"Q" size %u\n",
                         skip_blkno, blru->size));
 
           XD3_ASSERT (sfile->source_position <= pos);
@@ -524,7 +527,7 @@ main_getblk_func (xd3_stream *stream,
 
   if (option_verbose > 3)
     {
-      if (blru->blkno != (xoff_t)-1)
+      if (blru->blkno != XD3_INVALID_OFFSET)
         {
           if (blru->blkno != blkno)
             {
diff --git a/xdelta3/xdelta3-decode.h b/xdelta3/xdelta3-decode.h
index 65877a4..e71b624 100644
--- a/xdelta3/xdelta3-decode.h
+++ b/xdelta3/xdelta3-decode.h
@@ -288,13 +288,14 @@ xd3_decode_parse_halfinst (xd3_stream *stream, xd3_hinst *inst)
     {
       IF_DEBUG2 ({
         static int cnt = 0;
-        XPR(NT "DECODE:%u: COPY at %"Q"u (winoffset %u) size %u winaddr %u\n",
-                 cnt++,
-                 stream->total_out + (stream->dec_position -
-                                      stream->dec_cpylen),
-                 (stream->dec_position - stream->dec_cpylen),
-                 inst->size,
-                 inst->addr);
+        XPR(NT "DECODE:%u: COPY at %"Q" (winoffset %"Z") "
+            "size %"Z" winaddr %"Z"\n",
+            cnt++,
+            stream->total_out + (stream->dec_position -
+                                 stream->dec_cpylen),
+            (stream->dec_position - stream->dec_cpylen),
+            inst->size,
+            inst->addr);
       });
 
       if ((ret = xd3_decode_address (stream,
@@ -329,7 +330,7 @@ xd3_decode_parse_halfinst (xd3_stream *stream, xd3_hinst *inst)
         if (inst->type == XD3_ADD)
           {
             static int cnt;
-            XPR(NT "DECODE:%d: ADD at %"Q"u (winoffset %u) size %u\n",
+            XPR(NT "DECODE:%d: ADD at %"Q" (winoffset %"Z") size %"Z"\n",
                cnt++,
                (stream->total_out + stream->dec_position - stream->dec_cpylen),
                stream->dec_position - stream->dec_cpylen,
@@ -339,7 +340,7 @@ xd3_decode_parse_halfinst (xd3_stream *stream, xd3_hinst *inst)
           {
             static int cnt;
             XD3_ASSERT (inst->type == XD3_RUN);
-            XPR(NT "DECODE:%d: RUN at %"Q"u (winoffset %u) size %u\n",
+            XPR(NT "DECODE:%d: RUN at %"Q" (winoffset %"Z") size %"Z"\n",
                cnt++,
                stream->total_out + stream->dec_position - stream->dec_cpylen,
                stream->dec_position - stream->dec_cpylen,
@@ -525,8 +526,8 @@ xd3_decode_output_halfinst (xd3_stream *stream, xd3_hinst *inst)
                 if ((source->onblk != blksize) &&
                     (blkoff + take > source->onblk))
                   {
-                    IF_DEBUG1 (XPR(NT "[srcfile] short at blkno %"Q"u onblk "
-                                   "%u blksize %u blkoff %u take %u\n",
+                    IF_DEBUG1 (XPR(NT "[srcfile] short at blkno %"Q" onblk "
+                                   "%"Z" blksize %"Z" blkoff %"Z" take %"Z"\n",
                                    block,
                                    source->onblk,
                                    blksize,
@@ -1008,7 +1009,7 @@ xd3_decode_input (xd3_stream *stream)
 
         stream->dec_state = DEC_CPYLEN;
 
-        IF_DEBUG2 (DP(RINT "--------- TARGET WINDOW %"Q"u -----------\n",
+        IF_DEBUG2 (DP(RINT "--------- TARGET WINDOW %"Q" -----------\n",
                       stream->current_window));
       }
 
@@ -1025,7 +1026,6 @@ xd3_decode_input (xd3_stream *stream)
 
     case DEC_CPYOFF:
       /* Copy window offset: only if VCD_SOURCE or VCD_TARGET is set */
-      
       OFFSET_CASE(SRCORTGT (stream->dec_win_ind), stream->dec_cpyoff,
                   DEC_ENCLEN);
 
@@ -1039,7 +1039,7 @@ xd3_decode_input (xd3_stream *stream)
       /* Check copy window bounds: VCD_TARGET window may not exceed
          current position. */
       if ((stream->dec_win_ind & VCD_TARGET) &&
-          (stream->dec_cpyoff + (xoff_t) stream->dec_cpylen >
+          (stream->dec_cpyoff + stream->dec_cpylen >
            stream->dec_winstart))
         {
           stream->msg = "VCD_TARGET window out of bounds";
@@ -1179,7 +1179,7 @@ xd3_decode_input (xd3_stream *stream)
       /* xd3_decode_emit returns XD3_OUTPUT on every success. */
       if ((ret = xd3_decode_emit (stream)) == XD3_OUTPUT)
         {
-          stream->total_out += (xoff_t) stream->avail_out;
+          stream->total_out += stream->avail_out;
         }
 
       return ret;
diff --git a/xdelta3/xdelta3-djw.h b/xdelta3/xdelta3-djw.h
index f69cb1d..414702a 100644
--- a/xdelta3/xdelta3-djw.h
+++ b/xdelta3/xdelta3-djw.h
@@ -343,7 +343,7 @@ static inline void
 djw_update_1_2 (int *mtf_run, usize_t *mtf_i,
                 uint8_t *mtfsym, djw_weight *freq)
 {
-  int code;
+  uint8_t code;
   
   do
     {
@@ -395,10 +395,10 @@ djw_build_prefix (const djw_weight *freq, uint8_t *clen, usize_t asize, usize_t
   usize_t heap_last; /* Index of the last _valid_ heap entry. */
   usize_t ents_size; /* Number of entries, including 0th fake entry */
   usize_t  overflow;  /* Number of code lengths that overflow */
-  uint32_t total_bits;
+  usize_t total_bits;
   usize_t i;
 
-  IF_DEBUG (uint32_t first_bits = 0);
+  IF_DEBUG (usize_t first_bits = 0);
 
   /* Insert real symbol frequences. */
   for (i = 0; i < asize; i += 1)
@@ -444,7 +444,7 @@ djw_build_prefix (const djw_weight *freq, uint8_t *clen, usize_t asize, usize_t
   if (heap_last == 1)
     {
       /* Pick either the first or last symbol. */
-      int s = freq[0] ? asize-1 : 0;
+      usize_t s = freq[0] ? asize-1 : 0;
       ents[s+1].freq = 1;
       goto again;
     }
@@ -493,8 +493,8 @@ djw_build_prefix (const djw_weight *freq, uint8_t *clen, usize_t asize, usize_t
     {
       IF_DEBUG2 (if (first_bits != total_bits)
       {
-        DP(RINT "code length overflow changed %u bits\n",
-           (usize_t)(total_bits - first_bits));
+        DP(RINT "code length overflow changed %"Z" bits\n",
+           total_bits - first_bits);
       });
       return total_bits;
     }
diff --git a/xdelta3/xdelta3-fgk.h b/xdelta3/xdelta3-fgk.h
index 7011500..81d5e7e 100644
--- a/xdelta3/xdelta3-fgk.h
+++ b/xdelta3/xdelta3-fgk.h
@@ -109,7 +109,7 @@ static fgk_stream*     fgk_alloc           (xd3_stream *stream /*, usize_t alpha
 static int             fgk_init            (xd3_stream *stream,
                                             fgk_stream *h, 
                                             int is_encode);
-static int             fgk_encode_data     (fgk_stream *h,
+static usize_t         fgk_encode_data     (fgk_stream *h,
                                             usize_t    n);
 static inline fgk_bit  fgk_get_encoded_bit (fgk_stream *h);
 
@@ -125,7 +125,7 @@ static int             xd3_encode_fgk      (xd3_stream  *stream,
 
 static inline int      fgk_decode_bit      (fgk_stream *h,
                                             fgk_bit     b);
-static int             fgk_decode_data     (fgk_stream *h);
+static usize_t         fgk_decode_data     (fgk_stream *h);
 static void            fgk_destroy         (xd3_stream *stream,
                                             fgk_stream *h);
 
@@ -234,7 +234,7 @@ static void fgk_swap_ptrs(fgk_node **one, fgk_node **two)
 
 /* Takes huffman transmitter h and n, the nth elt in the alphabet, and
  * returns the number of required to encode n. */
-static int fgk_encode_data (fgk_stream* h, usize_t n)
+static usize_t fgk_encode_data (fgk_stream* h, usize_t n)
 {
   fgk_node *target_ptr = h->alphabet + n;
 
@@ -249,8 +249,8 @@ static int fgk_encode_data (fgk_stream* h, usize_t n)
    * is not neccesary to encode these bits. */
   if (IS_ADAPTIVE && target_ptr->weight == 0)
     {
-      unsigned int where, shift;
-      int bits;
+      usize_t where, shift;
+      usize_t bits;
 
       where = fgk_find_nth_zero(h, n);
       shift = 1;
@@ -734,7 +734,7 @@ static usize_t fgk_nth_zero (fgk_stream* h, usize_t n)
  * alphabet otherwise this returns 0, indicating more bits are
  * required.
  */
-static int fgk_decode_data (fgk_stream* h)
+static usize_t fgk_decode_data (fgk_stream* h)
 {
   usize_t elt = (usize_t)(h->decode_ptr - h->alphabet);
 
diff --git a/xdelta3/xdelta3-hash.h b/xdelta3/xdelta3-hash.h
index e359436..c112b5a 100644
--- a/xdelta3/xdelta3-hash.h
+++ b/xdelta3/xdelta3-hash.h
@@ -1,5 +1,6 @@
 /* xdelta 3 - delta compression tools and library
- * Copyright (C) 2001, 2003, 2004, 2005, 2006, 2007.  Joshua P. MacDonald
+ * Copyright (C) 2001, 2003, 2004, 2005, 2006, 2007, 2011, 2012, 2014.
+ * Joshua P. MacDonald
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -19,6 +20,8 @@
 #ifndef _XDELTA3_HASH_H_
 #define _XDELTA3_HASH_H_
 
+#include "xdelta3-internal.h"
+
 #if XD3_DEBUG
 #define SMALL_HASH_DEBUG1(s,inp)                                  \
   uint32_t debug_state;                                           \
@@ -32,45 +35,18 @@
 #define SMALL_HASH_DEBUG2(s,inp)
 #endif /* XD3_DEBUG */
 
-/* This is a good hash multiplier for 32-bit LCGs: see "linear
- * congruential generators of different sizes and good lattice
- * structure" */
-static const uint32_t hash_multiplier = 1597334677U;
-
-/***********************************************************************
- Permute stuff
- ***********************************************************************/
-
-#if HASH_PERMUTE == 0
-#define PERMUTE(x) (x)
-#else
-#define PERMUTE(x) (__single_hash[(uint32_t)x])
-
-extern const uint16_t __single_hash[256];
-#endif
-
-/* Update the checksum state. */
-#if ADLER_LARGE_CKSUM
-inline uint32_t
-xd3_large_cksum_update (uint32_t cksum,
-                        const uint8_t *base,
-                        usize_t look) {
-  uint32_t old_c = PERMUTE(base[0]);
-  uint32_t new_c = PERMUTE(base[look]);
-  uint32_t low   = ((cksum & 0xffff) - old_c + new_c) & 0xffff;
-  uint32_t high  = ((cksum >> 16) - (old_c * look) + low) & 0xffff;
-  return (high << 16) | low;
-}
-#else
-/* TODO: revisit this topic */
-#endif
-
 #if UNALIGNED_OK
 #define UNALIGNED_READ32(dest,src) (*(dest)) = (*(uint32_t*)(src))
 #else
 #define UNALIGNED_READ32(dest,src) memcpy((dest), (src), 4);
 #endif
 
+/* These are good hash multipliers for 32-bit and 64-bit LCGs: see
+ * "linear congruential generators of different sizes and good lattice
+ * structure" */
+#define xd3_hash_multiplier32 1597334677U
+#define xd3_hash_multiplier64 1181783497276652981ULL
+
 /* TODO: small cksum is hard-coded for 4 bytes (i.e., "look" is unused) */
 static inline uint32_t
 xd3_scksum (uint32_t *state,
@@ -78,7 +54,7 @@ xd3_scksum (uint32_t *state,
             const usize_t look)
 {
   UNALIGNED_READ32(state, base);
-  return (*state) * hash_multiplier;
+  return (*state) * xd3_hash_multiplier32;
 }
 static inline uint32_t
 xd3_small_cksum_update (uint32_t *state,
@@ -86,66 +62,67 @@ xd3_small_cksum_update (uint32_t *state,
                         usize_t look)
 {
   UNALIGNED_READ32(state, base+1);
-  return (*state) * hash_multiplier;
+  return (*state) * xd3_hash_multiplier32;
 }
 
-/***********************************************************************
- Ctable stuff
- ***********************************************************************/
-
-static inline usize_t
+#if XD3_ENCODER
+inline usize_t
 xd3_checksum_hash (const xd3_hash_cfg *cfg, const usize_t cksum)
 {
   return (cksum >> cfg->shift) ^ (cksum & cfg->mask);
 }
 
-/***********************************************************************
- Cksum function
- ***********************************************************************/
-
-#if ADLER_LARGE_CKSUM
-static inline uint32_t
-xd3_lcksum (const uint8_t *seg, const usize_t ln)
+#if SIZEOF_USIZE_T == 4
+inline uint32_t
+xd3_large32_cksum (xd3_hash_cfg *cfg, const uint8_t *base, const usize_t look)
 {
-  usize_t i = 0;
-  uint32_t low  = 0;
-  uint32_t high = 0;
-
-  for (; i < ln; i += 1)
-    {
-      low  += PERMUTE(*seg++);
-      high += low;
-    }
+  uint32_t h = 0;
+  for (usize_t i = 0; i < look; i++) {
+    h += base[i] * cfg->powers[i];
+  }
+  return h;
+}
 
-  return ((high & 0xffff) << 16) | (low & 0xffff);
+inline uint32_t
+xd3_large32_cksum_update (xd3_hash_cfg *cfg, const uint32_t cksum,
+                          const uint8_t *base, const usize_t look)
+{
+  return xd3_hash_multiplier32 * cksum - cfg->multiplier * base[0] + base[look];
 }
-#else
-static inline uint32_t
-xd3_lcksum (const uint8_t *seg, const usize_t ln)
+#endif
+
+#if SIZEOF_USIZE_T == 8
+inline uint64_t
+xd3_large64_cksum (xd3_hash_cfg *cfg, const uint8_t *base, const usize_t look)
 {
-  usize_t i, j;
-  uint32_t h = 0;
-  for (i = 0, j = ln - 1; i < ln; ++i, --j) {
-    h += PERMUTE(seg[i]) * hash_multiplier_powers[j];
+  uint64_t h = 0;
+  for (usize_t i = 0; i < look; i++) {
+    h += base[i] * cfg->powers[i];
   }
   return h;
 }
+
+inline uint64_t
+xd3_large64_cksum_update (xd3_hash_cfg *cfg, const uint64_t cksum,
+                          const uint8_t *base, const usize_t look)
+{
+  return xd3_hash_multiplier64 * cksum - cfg->multiplier * base[0] + base[look];
+}
 #endif
 
-#if XD3_ENCODER
 static usize_t
-xd3_size_log2 (usize_t slots)
+xd3_size_hashtable_bits (usize_t slots)
 {
-  int bits = 28; /* This should not be an unreasonable limit. */
-  int i;
+  usize_t bits = (SIZEOF_USIZE_T * 8) - 1;
+  usize_t i;
 
   for (i = 3; i <= bits; i += 1)
     {
       if (slots < (1U << i))
         {
-          /* TODO: this is compaction=1 in checksum_test.cc and maybe should
-           * not be fixed at -1. */
-          bits = i - 1; 
+          /* Note: this is the compaction=1 setting measured in
+           * checksum_test */
+          bits = i - 1;
           break;
         }
     }
@@ -153,18 +130,34 @@ xd3_size_log2 (usize_t slots)
   return bits;
 }
 
-static void
-xd3_size_hashtable (xd3_stream    *stream,
-                    usize_t        slots,
-                    xd3_hash_cfg  *cfg)
+int
+xd3_size_hashtable (xd3_stream   *stream,
+                    usize_t       slots,
+                    usize_t       look,
+                    xd3_hash_cfg *cfg)
 {
-  int bits = xd3_size_log2 (slots);
+  usize_t bits = xd3_size_hashtable_bits (slots);
 
-  /* TODO: there's a 32-bit assumption here */
-  cfg->size  = (1 << bits);
+  cfg->size  = (1U << bits);
   cfg->mask  = (cfg->size - 1);
-  cfg->shift = 32 - bits;
+  cfg->shift = (SIZEOF_USIZE_T * 8) - bits;
+  cfg->look  = look;
+
+  if ((cfg->powers = 
+       (usize_t*) xd3_alloc0 (stream, look, sizeof (usize_t))) == NULL)
+    {
+      return ENOMEM;
+    }
+
+  cfg->powers[look-1] = 1;
+  for (int i = look-2; i >= 0; i--)
+    {
+      cfg->powers[i] = cfg->powers[i+1] * xd3_hash_multiplier;
+    }
+  cfg->multiplier = cfg->powers[0] * xd3_hash_multiplier;
+
+  return 0;
 }
-#endif
 
-#endif
+#endif /* XD3_ENCODER */
+#endif /* _XDELTA3_HASH_H_ */
diff --git a/xdelta3/xdelta3-internal.h b/xdelta3/xdelta3-internal.h
index eb360be..e9fc5f3 100644
--- a/xdelta3/xdelta3-internal.h
+++ b/xdelta3/xdelta3-internal.h
@@ -31,6 +31,7 @@ void main_file_cleanup (main_file *xfile);
 int main_file_isopen (main_file *xfile);
 int main_file_open (main_file *xfile, const char* name, int mode);
 int main_file_exists (main_file *xfile);
+int main_file_stat (main_file *xfile, xoff_t *size);
 int xd3_whole_append_window (xd3_stream *stream);
 int xd3_main_cmdline (int argc, char **argv);
 int main_file_read (main_file  *ifile,
@@ -40,9 +41,13 @@ int main_file_read (main_file  *ifile,
                     const char *msg);
 int main_file_write (main_file *ofile, uint8_t *buf, 
                      usize_t size, const char *msg);
+void* main_malloc (size_t size);
+void main_free (void *ptr);
+
 int test_compare_files (const char* f0, const char* f1);
 usize_t xd3_bytes_on_srcblk (xd3_source *src, xoff_t blkno);
 xoff_t xd3_source_eof(const xd3_source *src);
+
 uint32_t xd3_large_cksum_update (uint32_t cksum,
                                  const uint8_t *base,
                                  usize_t look);
@@ -58,7 +63,8 @@ xd3_output* xd3_alloc_output (xd3_stream *stream,
                               xd3_output *old_output);
 
 int xd3_encode_init_full (xd3_stream *stream);
-size_t xd3_pow2_roundup (size_t x);
+usize_t xd3_pow2_roundup (usize_t x);
+long get_millisecs_now ();
 int xd3_process_stream (int            is_encode,
                         xd3_stream    *stream,
                         int          (*func) (xd3_stream *),
@@ -73,6 +79,10 @@ int xd3_process_stream (int            is_encode,
 int xd3_main_cmdline (int argc, char **argv);
 #endif
 
+#if REGRESSION_TEST
+int xd3_selftest (void);
+#endif
+
 /* main_file->mode values */
 typedef enum
 {
@@ -149,14 +159,13 @@ typedef struct {
  * therefore always define xsnprintf_func. */
 #undef PRINTF_ATTRIBUTE
 #ifdef __GNUC__
-/* Let's just assume no one uses gcc 2.x! */
 #define PRINTF_ATTRIBUTE(x,y) __attribute__ ((__format__ (__printf__, x, y)))
 #else
 #define PRINTF_ATTRIBUTE(x,y)
 #endif
 
 /* Underlying xprintf() */
-int xsnprintf_func (char *str, int n, const char *fmt, ...)
+int xsnprintf_func (char *str, size_t n, const char *fmt, ...)
   PRINTF_ATTRIBUTE(3,4);
 
 /* XPR(NT "", ...) (used by main) prefixes an "xdelta3: " to the output. */
@@ -288,28 +297,11 @@ xd3_read_uint32_t (xd3_stream *stream, const uint8_t **inpp,
 static inline int
 xd3_emit_uint32_t (xd3_stream *stream, xd3_output **output, uint32_t num)
 { EMIT_INTEGER_TYPE (); }
-#endif
-#endif
+#endif  /* XD3_ENCODER */
+#endif  /* USE_UINT32 */
 
 #if USE_UINT64
-static inline int
-xd3_decode_uint64_t (xd3_stream *stream, uint64_t *val)
-{ DECODE_INTEGER_TYPE (stream->dec_64part, UINT64_OFLOW_MASK); }
-
-#if XD3_ENCODER
-static inline int
-xd3_emit_uint64_t (xd3_stream *stream, xd3_output **output, uint64_t num)
-{ EMIT_INTEGER_TYPE (); }
-#endif
-
-/* These are tested but not used */
-#if REGRESSION_TEST
-static int
-xd3_read_uint64_t (xd3_stream *stream, const uint8_t **inpp,
-                   const uint8_t *maxp, uint64_t *valp)
-{ READ_INTEGER_TYPE (uint64_t, UINT64_OFLOW_MASK); }
-
-static uint32_t
+static inline uint32_t
 xd3_sizeof_uint64_t (uint64_t num)
 {
   IF_SIZEOF64(1);
@@ -324,49 +316,110 @@ xd3_sizeof_uint64_t (uint64_t num)
 
   return 10;
 }
-#endif
 
-#endif
+static inline int
+xd3_decode_uint64_t (xd3_stream *stream, uint64_t *val)
+{ DECODE_INTEGER_TYPE (stream->dec_64part, UINT64_OFLOW_MASK); }
+
+static inline int
+xd3_read_uint64_t (xd3_stream *stream, const uint8_t **inpp,
+                   const uint8_t *maxp, uint64_t *valp)
+{ READ_INTEGER_TYPE (uint64_t, UINT64_OFLOW_MASK); }
+
+#if XD3_ENCODER
+static inline int
+xd3_emit_uint64_t (xd3_stream *stream, xd3_output **output, uint64_t num)
+{ EMIT_INTEGER_TYPE (); }
+#endif  /* XD3_ENCODER */
+#endif  /* USE_UINT64 */
 
 #if SIZEOF_USIZE_T == 4
-#define USIZE_T_MAX        UINT32_MAX
-#define USIZE_T_MAXBLKSZ   0x80000000U
-#define xd3_decode_size   xd3_decode_uint32_t
-#define xd3_emit_size     xd3_emit_uint32_t
-#define xd3_sizeof_size   xd3_sizeof_uint32_t
-#define xd3_read_size     xd3_read_uint32_t
+#define USIZE_T_MAX             UINT32_MAX
+#define USIZE_T_MAXBLKSZ        0x80000000U
+#define XD3_MAXSRCWINSZ         (1ULL << 31)
+#define xd3_large_cksum         xd3_large32_cksum
+#define xd3_large_cksum_update  xd3_large32_cksum_update
+#define xd3_hash_multiplier     xd3_hash_multiplier32
+
+static inline uint32_t xd3_sizeof_size (usize_t num)
+{ return xd3_sizeof_uint32_t (num); }
+static inline int xd3_decode_size (xd3_stream *stream, usize_t *valp)
+{ return xd3_decode_uint32_t (stream, (uint32_t*) valp); }
+static inline int xd3_read_size (xd3_stream *stream, const uint8_t **inpp,
+                   const uint8_t *maxp, usize_t *valp)
+{ return xd3_read_uint32_t (stream, inpp, maxp, (uint32_t*) valp); }
+#if XD3_ENCODER
+static inline int xd3_emit_size (xd3_stream *stream, xd3_output **output, usize_t num)
+{ return xd3_emit_uint32_t (stream, output, num); }
+#endif
+
 #elif SIZEOF_USIZE_T == 8
-#define USIZE_T_MAX        UINT64_MAX
-#define USIZE_T_MAXBLKSZ   0x8000000000000000ULL
-#define xd3_decode_size   xd3_decode_uint64_t
-#define xd3_emit_size     xd3_emit_uint64_t
-#define xd3_sizeof_size   xd3_sizeof_uint64_t
-#define xd3_read_size     xd3_read_uint64_t
+#define USIZE_T_MAX             UINT64_MAX
+#define USIZE_T_MAXBLKSZ        0x8000000000000000ULL
+#define XD3_MAXSRCWINSZ         (1ULL << 61)
+#define xd3_large_cksum         xd3_large64_cksum
+#define xd3_large_cksum_update  xd3_large64_cksum_update
+#define xd3_hash_multiplier     xd3_hash_multiplier64
+
+static inline uint32_t xd3_sizeof_size (usize_t num)
+{ return xd3_sizeof_uint64_t (num); }
+static inline int xd3_decode_size (xd3_stream *stream, usize_t *valp)
+{ return xd3_decode_uint64_t (stream, (uint64_t*) valp); }
+static inline int xd3_read_size (xd3_stream *stream, const uint8_t **inpp,
+                   const uint8_t *maxp, usize_t *valp)
+{ return xd3_read_uint64_t (stream, inpp, maxp, (uint64_t*) valp); }
+#if XD3_ENCODER
+static inline int xd3_emit_size (xd3_stream *stream, xd3_output **output, usize_t num)
+{ return xd3_emit_uint64_t (stream, output, num); }
 #endif
 
+#endif /* SIZEOF_USIZE_T */
+
 #if SIZEOF_XOFF_T == 4
 #define XOFF_T_MAX        UINT32_MAX
-#define xd3_emit_offset   xd3_emit_uint32_t
-static inline int
-xd3_decode_offset (xd3_stream *stream, xoff_t *val)
-{
-  return xd3_decode_uint32_t (stream, (uint32_t*) val);
-}
+
+static inline int xd3_decode_offset (xd3_stream *stream, xoff_t *valp)
+{ return xd3_decode_uint32_t (stream, (uint32_t*) valp); }
+#if XD3_ENCODER
+static inline int xd3_emit_offset (xd3_stream *stream, xd3_output **output, xoff_t num)
+{ return xd3_emit_uint32_t (stream, output, num); }
+#endif
+
 #elif SIZEOF_XOFF_T == 8
 #define XOFF_T_MAX        UINT64_MAX
-#define xd3_emit_offset   xd3_emit_uint64_t
-static inline int
-xd3_decode_offset (xd3_stream *stream, xoff_t *val)
-{
-  return xd3_decode_uint64_t (stream, (uint64_t*) val);
-}
+
+static inline int xd3_decode_offset (xd3_stream *stream, xoff_t *valp)
+{ return xd3_decode_uint64_t (stream, (uint64_t*) valp); }
+#if XD3_ENCODER
+static inline int xd3_emit_offset (xd3_stream *stream, xd3_output **output, xoff_t num)
+{ return xd3_emit_uint64_t (stream, output, num); }
+#endif
+
 #endif
 
 #define USIZE_T_OVERFLOW(a,b) ((USIZE_T_MAX - (usize_t) (a)) < (usize_t) (b))
 #define XOFF_T_OVERFLOW(a,b) ((XOFF_T_MAX - (xoff_t) (a)) < (xoff_t) (b))
 
+int xd3_size_hashtable (xd3_stream   *stream,
+                        usize_t       slots,
+                        usize_t       look,
+                        xd3_hash_cfg *cfg);
+
+usize_t xd3_checksum_hash (const xd3_hash_cfg *cfg, const usize_t cksum);
+
+#if USE_UINT32
+uint32_t xd3_large32_cksum (xd3_hash_cfg *cfg, const uint8_t *base, const usize_t look);
+uint32_t xd3_large32_cksum_update (xd3_hash_cfg *cfg, uint32_t cksum, 
+                                   const uint8_t *base, const usize_t look);
+#endif /* USE_UINT32 */
+
+#if USE_UINT64
+uint64_t xd3_large64_cksum (xd3_hash_cfg *cfg, const uint8_t *base, const usize_t look);
+uint64_t xd3_large64_cksum_update (xd3_hash_cfg *cfg, uint64_t cksum, 
+                                   const uint8_t *base, const usize_t look);
+#endif /* USE_UINT64 */
+
 #define MAX_LRU_SIZE 32U
 #define XD3_MINSRCWINSZ (XD3_ALLOCSIZE * MAX_LRU_SIZE)
-#define XD3_MAXSRCWINSZ (1ULL << 31)
 
 #endif // XDELTA3_INTERNAL_H__
diff --git a/xdelta3/xdelta3-list.h b/xdelta3/xdelta3-list.h
index 6e3125f..2c8f3ef 100644
--- a/xdelta3/xdelta3-list.h
+++ b/xdelta3/xdelta3-list.h
@@ -115,7 +115,7 @@ static inline usize_t                                                   \
 LTYPE ## _length (LTYPE *l)                                             \
 {                                                                       \
   LTYPE *p;                                                             \
-  int c = 0;                                                            \
+  usize_t c = 0;                                                        \
                                                                         \
   for (p = l->next; p != l; p = p->next)                                \
     {                                                                   \
diff --git a/xdelta3/xdelta3-lzma.h b/xdelta3/xdelta3-lzma.h
index bcf2ddf..de317ee 100644
--- a/xdelta3/xdelta3-lzma.h
+++ b/xdelta3/xdelta3-lzma.h
@@ -55,7 +55,8 @@ xd3_lzma_init (xd3_stream *stream, xd3_lzma_stream *sec, int is_encode)
 
   if (is_encode)
     {
-      int preset = (stream->flags & XD3_COMPLEVEL_MASK) >> XD3_COMPLEVEL_SHIFT;
+      uint32_t preset = 
+        (stream->flags & XD3_COMPLEVEL_MASK) >> XD3_COMPLEVEL_SHIFT;
 
       if (lzma_lzma_preset(&sec->options, preset)) 
         {
diff --git a/xdelta3/xdelta3-main.h b/xdelta3/xdelta3-main.h
index fac5d8d..1a3d95e 100644
--- a/xdelta3/xdelta3-main.h
+++ b/xdelta3/xdelta3-main.h
@@ -49,13 +49,28 @@
 
 /*********************************************************************/
 
+#include <limits.h>
+
+#ifndef XD3_POSIX
+#define XD3_POSIX 0
+#endif
+#ifndef XD3_STDIO
+#define XD3_STDIO 0
+#endif
+#ifndef XD3_WIN32
+#define XD3_WIN32 0
+#endif
+#ifndef NOT_MAIN
+#define NOT_MAIN 0
+#endif
+
 /* Combines xd3_strerror() and strerror() */
 const char* xd3_mainerror(int err_num);
 
 #include "xdelta3-internal.h"
 
 int
-xsnprintf_func (char *str, int n, const char *fmt, ...)
+xsnprintf_func (char *str, size_t n, const char *fmt, ...)
 {
   va_list a;
   int ret;
@@ -230,9 +245,8 @@ static const char *option_source_filename    = NULL;
 static int         option_level              = XD3_DEFAULT_LEVEL;
 static usize_t     option_iopt_size          = XD3_DEFAULT_IOPT_SIZE;
 static usize_t     option_winsize            = XD3_DEFAULT_WINSIZE;
-/* Note: option_srcwinsz is restricted from [16Kb, 4Gb], because
- * addresses in the large hash checksum are 32 bits.  The flag is read
- * as xoff_t, so that 4Gb != 0. */
+
+/* option_srcwinsz is restricted to [16kB, 2GB] when usize_t is 32 bits. */
 static xoff_t      option_srcwinsz           = XD3_DEFAULT_SRCWINSZ;
 static usize_t     option_sprevsz            = XD3_DEFAULT_SPREVSZ;
 
@@ -288,10 +302,6 @@ static void main_get_appheader (xd3_stream *stream, main_file *ifile,
 static int main_getblk_func (xd3_stream *stream,
                              xd3_source *source,
                              xoff_t      blkno);
-static void main_free (void *ptr);
-static void* main_malloc (size_t size);
-
-static int main_file_stat (main_file *xfile, xoff_t *size);
 static int main_file_seek (main_file *xfile, xoff_t pos);
 static int main_read_primary_input (main_file   *file,
                                     uint8_t     *buf,
@@ -364,6 +374,7 @@ main_config (void)
   XPR(NTR "XD3_STDIO=%d\n", XD3_STDIO);
   XPR(NTR "XD3_WIN32=%d\n", XD3_WIN32);
   XPR(NTR "XD3_USE_LARGEFILE64=%d\n", XD3_USE_LARGEFILE64);
+  XPR(NTR "XD3_USE_LARGESIZET=%d\n", XD3_USE_LARGESIZET);
   XPR(NTR "XD3_DEFAULT_LEVEL=%d\n", XD3_DEFAULT_LEVEL);
   XPR(NTR "XD3_DEFAULT_IOPT_SIZE=%d\n", XD3_DEFAULT_IOPT_SIZE);
   XPR(NTR "XD3_DEFAULT_SPREVSZ=%d\n", XD3_DEFAULT_SPREVSZ);
@@ -440,7 +451,7 @@ void* main_bufalloc (size_t size) {
 #endif
 }
 
-static void*
+void*
 main_malloc (size_t size)
 {
   void *r = main_malloc1 (size);
@@ -462,7 +473,7 @@ main_free1 (void *opaque, void *ptr)
   free (ptr);
 }
 
-static void
+void
 main_free (void *ptr)
 {
   if (ptr)
@@ -533,7 +544,7 @@ xd3_mainerror(int err_num) {
 #endif
 }
 
-static long
+long
 get_millisecs_now (void)
 {
 #ifndef _WIN32
@@ -656,14 +667,17 @@ main_strtoxoff (const char* s, xoff_t *xo, char which)
   XD3_ASSERT(s && *s != 0);
 
   {
-    /* Should check LONG_MIN, LONG_MAX, LLONG_MIN, LLONG_MAX? */
-#if SIZEOF_XOFF_T == 4
-    long xx = strtol (s, &e, 0);
+#if SIZEOF_XOFF_T == SIZEOF_UNSIGNED_LONG_LONG
+    unsigned long long xx = strtoull (s, &e, 0);
+    unsigned long long bad = ULLONG_MAX;
+#elif SIZEOF_XOFF_T <= SIZEOF_UNSIGNED_LONG
+    unsigned long xx = strtoul (s, &e, 0);
+    unsigned long long bad = ULONG_MAX;
 #else
-    long long xx = strtoll (s, &e, 0);
+    /* Something wrong with SIZEOF_XOFF_T, SIZEOF_UNSIGNED_LONG, etc. */
 #endif
 
-    if (xx < 0)
+    if (xx == bad)
       {
         XPR(NT "-%c: negative integer: %s\n", which, s);
         return EXIT_FAILURE;
@@ -1093,7 +1107,7 @@ main_file_write (main_file *ofile, uint8_t *buf, usize_t size, const char *msg)
     }
   else
     {
-      if (option_verbose > 5) { XPR(NT "write %s: %u bytes\n",
+      if (option_verbose > 5) { XPR(NT "write %s: %"W"u bytes\n",
                                     ofile->filename, size); }
       ofile->nwrite += size;
     }
@@ -1231,6 +1245,15 @@ main_set_secondary_flags (xd3_config *config)
         }
     }
 
+  if (option_verbose)
+    {
+      XPR(NT "secondary compression: %s\n",
+          (config->flags | XD3_SEC_LZMA) ? "lzma" :
+          ((config->flags | XD3_SEC_FGK) ? "fgk" :
+           ((config->flags | XD3_SEC_DJW) ? "djw" :
+            "none")));
+    }
+
   return 0;
 }
 
@@ -1291,7 +1314,8 @@ main_print_window (xd3_stream* stream, main_file *xfile)
       addr_bytes = (usize_t)(stream->addr_sect.buf - addr_before);
       inst_bytes = (usize_t)(stream->inst_sect.buf - inst_before);
 
-      VC(UT "  %06"Q"u %03u  %s %6u", stream->dec_winstart + size,
+      VC(UT "  %06"Q"u %03"W"u  %s %6"W"u", 
+         stream->dec_winstart + size,
          option_print_cpymode ? code : 0,
          xd3_rtype_to_string ((xd3_rtype) stream->dec_current1.type,
                               option_print_cpymode),
@@ -1303,7 +1327,7 @@ main_print_window (xd3_stream* stream, main_file *xfile)
             {
               if (stream->dec_current1.addr >= stream->dec_cpylen)
                 {
-                  VC(UT " T@%-6u",
+                  VC(UT " T@%-6"W"u",
                      stream->dec_current1.addr - stream->dec_cpylen)VE;
                 }
               else
@@ -1322,7 +1346,7 @@ main_print_window (xd3_stream* stream, main_file *xfile)
 
       if (stream->dec_current2.type != XD3_NOOP)
         {
-          VC(UT "  %s %6u",
+          VC(UT "  %s %6"W"u",
              xd3_rtype_to_string ((xd3_rtype) stream->dec_current2.type,
                                   option_print_cpymode),
              stream->dec_current2.size)VE;
@@ -1331,7 +1355,7 @@ main_print_window (xd3_stream* stream, main_file *xfile)
             {
               if (stream->dec_current2.addr >= stream->dec_cpylen)
                 {
-                  VC(UT " T@%-6u",
+                  VC(UT " T@%-6"W"u",
                      stream->dec_current2.addr - stream->dec_cpylen)VE;
                 }
               else
@@ -1351,7 +1375,7 @@ main_print_window (xd3_stream* stream, main_file *xfile)
           (stream->dec_current1.type >= XD3_CPY ||
            stream->dec_current2.type >= XD3_CPY))
         {
-          VC(UT "  %06"Q"u (inefficiency) %u encoded as %u bytes\n",
+          VC(UT "  %06"Q"u (inefficiency) %"W"u encoded as %"W"u bytes\n",
              stream->dec_winstart + size_before,
              size - size_before,
              addr_bytes + inst_bytes)VE;
@@ -1419,7 +1443,7 @@ main_print_func (xd3_stream* stream, main_file *xfile)
   if (stream->dec_winstart == 0)
     {
       VC(UT "VCDIFF version:               0\n")VE;
-      VC(UT "VCDIFF header size:           %d\n",
+      VC(UT "VCDIFF header size:           %"W"u\n",
          stream->dec_hdrsize)VE;
       VC(UT "VCDIFF header indicator:      ")VE;
       if ((stream->dec_hdr_ind & VCD_SECONDARY) != 0)
@@ -1485,7 +1509,7 @@ main_print_func (xd3_stream* stream, main_file *xfile)
   if ((stream->dec_win_ind & VCD_ADLER32) != 0)
     {
       VC(UT "VCDIFF adler32 checksum:      %08X\n",
-         (usize_t)stream->dec_adler32)VE;
+         stream->dec_adler32)VE;
     }
 
   if (stream->dec_del_ind != 0)
@@ -1505,22 +1529,22 @@ main_print_func (xd3_stream* stream, main_file *xfile)
 
   if (SRCORTGT (stream->dec_win_ind))
     {
-      VC(UT "VCDIFF copy window length:    %u\n",
-         (usize_t)stream->dec_cpylen)VE;
+      VC(UT "VCDIFF copy window length:    %"W"u\n",
+         stream->dec_cpylen)VE;
       VC(UT "VCDIFF copy window offset:    %"Q"u\n",
          stream->dec_cpyoff)VE;
     }
 
-  VC(UT "VCDIFF delta encoding length: %u\n",
+  VC(UT "VCDIFF delta encoding length: %"W"u\n",
      (usize_t)stream->dec_enclen)VE;
-  VC(UT "VCDIFF target window length:  %u\n",
+  VC(UT "VCDIFF target window length:  %"W"u\n",
      (usize_t)stream->dec_tgtlen)VE;
 
-  VC(UT "VCDIFF data section length:   %u\n",
+  VC(UT "VCDIFF data section length:   %"W"u\n",
      (usize_t)stream->data_sect.size)VE;
-  VC(UT "VCDIFF inst section length:   %u\n",
+  VC(UT "VCDIFF inst section length:   %"W"u\n",
      (usize_t)stream->inst_sect.size)VE;
-  VC(UT "VCDIFF addr section length:   %u\n",
+  VC(UT "VCDIFF addr section length:   %"W"u\n",
      (usize_t)stream->addr_sect.size)VE;
 
   ret = 0;
@@ -1960,8 +1984,11 @@ main_merge_output (xd3_stream *stream, main_file *ofile)
                   XD3_ASSERT (inst->addr >= window_start);
                   addr = inst->addr - window_start;
                 }
-              IF_DEBUG2 (XPR(NTR "[merge copy] winpos %u take %u addr %"Q"u mode %u\n",
-                            window_pos, take, addr, inst->mode));
+              IF_DEBUG2 ({
+                  XPR(NTR "[merge copy] winpos %"W" take %"W" "
+                      "addr %"Q" mode %u\n",
+                      window_pos, take, addr, inst->mode);
+                });
               if ((ret = xd3_found_match (recode_stream, window_pos, take,
                                           addr, inst->mode != 0)))
                 {
@@ -3034,7 +3061,7 @@ main_input (xd3_cmd     cmd,
         {
           const char *s = option_smatch_config;
           char *e;
-          int values[XD3_SOFTCFG_VARCNT];
+          long values[XD3_SOFTCFG_VARCNT];
           int got;
 
           config.smatch_cfg = XD3_SMATCH_SOFT;
@@ -3291,7 +3318,7 @@ main_input (xd3_cmd     cmd,
                         stream.i_slots_used > stream.iopt_size)
                       {
                         XPR(NT "warning: input position %"Q"u overflowed "
-                            "instruction buffer, needed %u (vs. %u), "
+                            "instruction buffer, needed %"W"u (vs. %"W"u), "
                             "consider changing -I\n",
                             stream.current_window * winsize,
                             stream.i_slots_used, stream.iopt_size);
@@ -3416,10 +3443,10 @@ done:
   if (option_verbose > 1 && cmd == CMD_ENCODE)
     {
       XPR(NT "scanner configuration: %s\n", stream.smatcher.name);
-      XPR(NT "target hash table size: %u\n", stream.small_hash.size);
+      XPR(NT "target hash table size: %"W"u\n", stream.small_hash.size);
       if (sfile != NULL && sfile->filename != NULL)
         {
-          XPR(NT "source hash table size: %u\n", stream.large_hash.size);
+          XPR(NT "source hash table size: %"W"u\n", stream.large_hash.size);
         }
     }
 
diff --git a/xdelta3/xdelta3-second.h b/xdelta3/xdelta3-second.h
index e8a2c8f..899902e 100644
--- a/xdelta3/xdelta3-second.h
+++ b/xdelta3/xdelta3-second.h
@@ -122,7 +122,7 @@ xd3_decode_secondary (xd3_stream      *stream,
                       xd3_desect      *sect,
                       xd3_sec_stream **sec_streamp)
 {
-  uint32_t dec_size;
+  usize_t dec_size;
   uint8_t *out_used;
   int ret;
 
diff --git a/xdelta3/xdelta3-test.h b/xdelta3/xdelta3-test.h
index 6d7d85b..335de28 100644
--- a/xdelta3/xdelta3-test.h
+++ b/xdelta3/xdelta3-test.h
@@ -139,7 +139,10 @@ static char   TEST_RECON2_FILE[TESTFILESIZE];
 static char   TEST_COPY_FILE[TESTFILESIZE];
 static char   TEST_NOPERM_FILE[TESTFILESIZE];
 
-#define CHECK(cond) if (!(cond)) { XPR(NT "check failure: " #cond); abort(); }
+#define CHECK(cond)                                             \
+  if (!(cond)) {                                                \
+    XPR(NT __FILE__":%d: check failure: " #cond, __LINE__);     \
+    abort(); }
 
 #if SHELL_TESTS
 /* Use a fixed soft config so that test values are fixed.  See also
@@ -199,7 +202,7 @@ test_random_numbers (xd3_stream *stream, int ignore)
 
   for (i = 0; i < n_rounds; i += 1)
     {
-      sum += mt_exp_rand (mean, USIZE_T_MAX);
+      sum += mt_exp_rand (mean, UINT32_MAX);
     }
 
   average = (double) sum / (double) n_rounds;
@@ -704,13 +707,13 @@ test_forward_match (xd3_stream *stream, int unused)
 
   for (i = 0; i < 256; i++)
     {
-      CHECK(xd3_forward_match(buf1, buf2, i) == (int)i);
+      CHECK(xd3_forward_match(buf1, buf2, i) == i);
     }
 
   for (i = 0; i < 255; i++)
     {
       buf2[i] = 1;
-      CHECK(xd3_forward_match(buf1, buf2, 256) == (int)i);
+      CHECK(xd3_forward_match(buf1, buf2, 256) == i);
       buf2[i] = 0;
     }
 
@@ -759,7 +762,7 @@ test_address_cache (xd3_stream *stream, int unused)
       usize_t prev_i;
       usize_t nearby;
 
-      p         = (mt_random (&static_mtrand) / (double)USIZE_T_MAX);
+      p         = (mt_random (&static_mtrand) / (double)UINT32_MAX);
       prev_i    = mt_random (&static_mtrand) % offset;
       nearby    = (mt_random (&static_mtrand) % 256) % offset;
       nearby    = xd3_max (1U, nearby);
@@ -790,9 +793,13 @@ test_address_cache (xd3_stream *stream, int unused)
 
   for (offset = 1; offset < ADDR_CACHE_ROUNDS; offset += 1)
     {
-      uint32_t addr;
+      usize_t addr;
 
-      if ((ret = xd3_decode_address (stream, offset, modes[offset], & buf, buf_max, & addr))) { return ret; }
+      if ((ret = xd3_decode_address (stream, offset, modes[offset], 
+                                     & buf, buf_max, & addr))) 
+        { 
+          return ret; 
+        }
 
       if (addr != addrs[offset])
         {
@@ -1418,7 +1425,7 @@ test_secondary (xd3_stream *stream, const xd3_sec_type *sec, usize_t groups)
           if ((ret = sec->encode (stream, enc_stream,
                                   in_head, out_head, & cfg)))
             {
-              XPR(NT "test %u: encode: %s", test_i, stream->msg);
+              XPR(NT "test %"W"u: encode: %s", test_i, stream->msg);
               goto fail;
             }
 
@@ -1457,7 +1464,7 @@ test_secondary (xd3_stream *stream, const xd3_sec_type *sec, usize_t groups)
                                             compress_size, dec_input,
                                             dec_correct, dec_output)))
             {
-              XPR(NT "test %u: decode: %s", test_i, stream->msg);
+              XPR(NT "test %"W"u: decode: %s", test_i, stream->msg);
               goto fail;
             }
 
@@ -1577,6 +1584,45 @@ test_choose_instruction (xd3_stream *stream, int ignore)
   return 0;
 }
 
+static int
+test_checksum_step (xd3_stream *stream, int ignore)
+{
+  const int bufsize = 128;
+  uint8_t buf[bufsize];
+  for (int i = 0; i < bufsize; i++)
+    {
+      buf[i] = mt_random (&static_mtrand) & 0xff;
+    }
+
+  for (usize_t cksize = 4; cksize <= 32; cksize += 3)
+    {
+      xd3_hash_cfg h1;
+      usize_t x;
+      int ret;
+
+      if ((ret = xd3_size_hashtable (stream, XD3_ALLOCSIZE, cksize, &h1)) != 0)
+        {
+          return ret;
+        }
+
+      x = xd3_large_cksum (&h1, buf, cksize);
+      for (usize_t pos = 0; pos <= (bufsize - cksize); pos++)
+        {
+          usize_t y = xd3_large_cksum (&h1, buf + pos, cksize);
+          if (x != y)
+            {
+              stream->msg = "checksum != incremental checksum";
+              return XD3_INTERNAL;
+            }
+          x = xd3_large_cksum_update (&h1, x, buf + pos, cksize);
+        }
+
+      xd3_free (stream, h1.powers);
+    }
+
+  return 0;
+}
+
 /***********************************************************************
  64BIT STREAMING
  ***********************************************************************/
@@ -2714,14 +2760,14 @@ test_string_matching (xd3_stream *stream, int ignore)
             default: CHECK(0);
             }
 
-          snprintf_func (rptr, rbuf+TESTBUFSIZE-rptr, "%d/%d",
+          snprintf_func (rptr, rbuf+TESTBUFSIZE-rptr, "%"W"u/%"W"u",
                          inst->pos, inst->size);
           rptr += strlen (rptr);
 
           if (inst->type == XD3_CPY)
             {
               *rptr++ = '@';
-              snprintf_func (rptr, rbuf+TESTBUFSIZE-rptr, "%"Q"d", inst->addr);
+              snprintf_func (rptr, rbuf+TESTBUFSIZE-rptr, "%"Q"u", inst->addr);
               rptr += strlen (rptr);
             }
 
@@ -2737,7 +2783,7 @@ test_string_matching (xd3_stream *stream, int ignore)
 
       if (strcmp (rbuf, test->result) != 0)
         {
-          XPR(NT "test %u: expected %s: got %s", i, test->result, rbuf);
+          XPR(NT "test %"W"u: expected %s: got %s", i, test->result, rbuf);
           stream->msg = "wrong result";
           return XD3_INTERNAL;
         }
@@ -2812,9 +2858,10 @@ test_iopt_flush_instructions (xd3_stream *stream, int ignore)
 /*
  * This tests the 32/64bit ambiguity for source-window matching.
  */
+#if !XD3_USE_LARGESIZET
 static int
 test_source_cksum_offset (xd3_stream *stream, int ignore)
-{
+ {
   xd3_source source;
 
   // Inputs are:
@@ -2848,7 +2895,7 @@ test_source_cksum_offset (xd3_stream *stream, int ignore)
   stream->src = &source;
 
   for (test_ptr = cksum_test; test_ptr->cpos; test_ptr++) {
-        xoff_t r;
+    xoff_t r;
     stream->srcwin_cksum_pos = test_ptr->cpos;
     stream->total_in = test_ptr->ipos;
 
@@ -2857,6 +2904,7 @@ test_source_cksum_offset (xd3_stream *stream, int ignore)
   }
   return 0;
 }
+#endif /* !XD3_USE_LARGESIZET */
 
 static int
 test_in_memory (xd3_stream *stream, int ignore)
@@ -2899,8 +2947,7 @@ test_in_memory (xd3_stream *stream, int ignore)
  TEST MAIN
  ***********************************************************************/
 
-static int
-xd3_selftest (void)
+int xd3_selftest (void)
 {
 #define DO_TEST(fn,flags,arg)                                         \
   do {                                                                \
@@ -2928,8 +2975,8 @@ xd3_selftest (void)
   DO_TEST (encode_decode_uint32_t, 0, 0);
   DO_TEST (encode_decode_uint64_t, 0, 0);
   DO_TEST (usize_t_overflow, 0, 0);
+  DO_TEST (checksum_step, 0, 0);
   DO_TEST (forward_match, 0, 0);
-
   DO_TEST (address_cache, 0, 0);
 
   DO_TEST (string_matching, 0, 0);
@@ -2938,7 +2985,9 @@ xd3_selftest (void)
   DO_TEST (in_memory, 0, 0);
 
   DO_TEST (iopt_flush_instructions, 0, 0);
+#if !XD3_USE_LARGESIZET
   DO_TEST (source_cksum_offset, 0, 0);
+#endif
 
   DO_TEST (decompress_single_bit_error, 0, 3);
   DO_TEST (decompress_single_bit_error, XD3_ADLER32, 3);
diff --git a/xdelta3/xdelta3.c b/xdelta3/xdelta3.c
index 976b219..9253d68 100644
--- a/xdelta3/xdelta3.c
+++ b/xdelta3/xdelta3.c
@@ -350,7 +350,9 @@ typedef unsigned int xd3_rtype;
 
 #include "xdelta3-list.h"
 
+#if XD3_ENCODER
 XD3_MAKELIST(xd3_rlist, xd3_rinst, link);
+#endif
 
 /***********************************************************************/
 
@@ -376,9 +378,6 @@ XD3_MAKELIST(xd3_rlist, xd3_rinst, link);
 #define ALPHABET_SIZE      256  /* Used in test code--size of the secondary
                                  * compressor alphabet. */
 
-#define HASH_PERMUTE       1    /* The input is permuted by random nums */
-#define ADLER_LARGE_CKSUM  1    /* Adler checksum vs. RK checksum */
-
 #define HASH_CKOFFSET      1U   /* Table entries distinguish "no-entry" from
                                  * offset 0 using this offset. */
 
@@ -476,9 +475,9 @@ static void        xd3_free_output (xd3_stream *stream,
                                     xd3_output *output);
 
 static int         xd3_emit_double (xd3_stream *stream, xd3_rinst *first,
-                                    xd3_rinst *second, usize_t code);
+                                    xd3_rinst *second, uint8_t code);
 static int         xd3_emit_single (xd3_stream *stream, xd3_rinst *single,
-                                    usize_t code);
+                                    uint8_t code);
 
 static usize_t      xd3_sizeof_output (xd3_output *output);
 static void        xd3_encode_reset (xd3_stream *stream);
@@ -503,8 +502,6 @@ static int xd3_srcwin_move_point (xd3_stream *stream,
 
 static int xd3_emit_run (xd3_stream *stream, usize_t pos,
                          usize_t size, uint8_t *run_c);
-static usize_t xd3_checksum_hash (const xd3_hash_cfg *cfg,
-                                  const usize_t cksum);
 static xoff_t xd3_source_cksum_offset(xd3_stream *stream, usize_t low);
 static void xd3_scksum_insert (xd3_stream *stream,
                                usize_t inx,
@@ -519,7 +516,7 @@ static void xd3_verify_run_state (xd3_stream    *stream,
                                   uint8_t       *x_run_c);
 static void xd3_verify_large_state (xd3_stream *stream,
                                     const uint8_t *inp,
-                                    uint32_t x_cksum);
+                                    usize_t x_cksum);
 static void xd3_verify_small_state (xd3_stream    *stream,
                                     const uint8_t *inp,
                                     uint32_t       x_cksum);
@@ -533,15 +530,6 @@ static int         xd3_decode_allocate (xd3_stream *stream, usize_t size,
 static void*       xd3_alloc (xd3_stream *stream, usize_t elts, usize_t size);
 static void        xd3_free  (xd3_stream *stream, void *ptr);
 
-static int         xd3_read_uint32_t (xd3_stream *stream, const uint8_t **inpp,
-                                      const uint8_t *max, uint32_t *valp);
-
-#if REGRESSION_TEST
-static int         xd3_selftest      (void);
-#endif
-
-/***********************************************************************/
-
 const char* xd3_strerror (int ret)
 {
   switch (ret)
@@ -570,7 +558,7 @@ const char* xd3_strerror (int ret)
 
 struct _xd3_sec_type
 {
-  int         id;
+  uint8_t       id;
   const char *name;
   xd3_secondary_flags flags;
 
@@ -602,7 +590,7 @@ struct _xd3_sec_type
 typedef struct _bit_state bit_state;
 struct _bit_state
 {
-  usize_t cur_byte;
+  uint8_t cur_byte;
   usize_t cur_mask;
 };
 
@@ -747,44 +735,6 @@ const xd3_sec_type lzma_sec_type =
 #endif /* __XDELTA3_C_HEADER_PASS__ */
 #ifdef __XDELTA3_C_INLINE_PASS__
 
-const uint16_t __single_hash[256] =
-{
-  /* Random numbers generated using SLIB's pseudo-random number generator.
-   * This hashes the input alphabet. */
-  0xbcd1, 0xbb65, 0x42c2, 0xdffe, 0x9666, 0x431b, 0x8504, 0xeb46,
-  0x6379, 0xd460, 0xcf14, 0x53cf, 0xdb51, 0xdb08, 0x12c8, 0xf602,
-  0xe766, 0x2394, 0x250d, 0xdcbb, 0xa678, 0x02af, 0xa5c6, 0x7ea6,
-  0xb645, 0xcb4d, 0xc44b, 0xe5dc, 0x9fe6, 0x5b5c, 0x35f5, 0x701a,
-  0x220f, 0x6c38, 0x1a56, 0x4ca3, 0xffc6, 0xb152, 0x8d61, 0x7a58,
-  0x9025, 0x8b3d, 0xbf0f, 0x95a3, 0xe5f4, 0xc127, 0x3bed, 0x320b,
-  0xb7f3, 0x6054, 0x333c, 0xd383, 0x8154, 0x5242, 0x4e0d, 0x0a94,
-  0x7028, 0x8689, 0x3a22, 0x0980, 0x1847, 0xb0f1, 0x9b5c, 0x4176,
-  0xb858, 0xd542, 0x1f6c, 0x2497, 0x6a5a, 0x9fa9, 0x8c5a, 0x7743,
-  0xa8a9, 0x9a02, 0x4918, 0x438c, 0xc388, 0x9e2b, 0x4cad, 0x01b6,
-  0xab19, 0xf777, 0x365f, 0x1eb2, 0x091e, 0x7bf8, 0x7a8e, 0x5227,
-  0xeab1, 0x2074, 0x4523, 0xe781, 0x01a3, 0x163d, 0x3b2e, 0x287d,
-  0x5e7f, 0xa063, 0xb134, 0x8fae, 0x5e8e, 0xb7b7, 0x4548, 0x1f5a,
-  0xfa56, 0x7a24, 0x900f, 0x42dc, 0xcc69, 0x02a0, 0x0b22, 0xdb31,
-  0x71fe, 0x0c7d, 0x1732, 0x1159, 0xcb09, 0xe1d2, 0x1351, 0x52e9,
-  0xf536, 0x5a4f, 0xc316, 0x6bf9, 0x8994, 0xb774, 0x5f3e, 0xf6d6,
-  0x3a61, 0xf82c, 0xcc22, 0x9d06, 0x299c, 0x09e5, 0x1eec, 0x514f,
-  0x8d53, 0xa650, 0x5c6e, 0xc577, 0x7958, 0x71ac, 0x8916, 0x9b4f,
-  0x2c09, 0x5211, 0xf6d8, 0xcaaa, 0xf7ef, 0x287f, 0x7a94, 0xab49,
-  0xfa2c, 0x7222, 0xe457, 0xd71a, 0x00c3, 0x1a76, 0xe98c, 0xc037,
-  0x8208, 0x5c2d, 0xdfda, 0xe5f5, 0x0b45, 0x15ce, 0x8a7e, 0xfcad,
-  0xaa2d, 0x4b5c, 0xd42e, 0xb251, 0x907e, 0x9a47, 0xc9a6, 0xd93f,
-  0x085e, 0x35ce, 0xa153, 0x7e7b, 0x9f0b, 0x25aa, 0x5d9f, 0xc04d,
-  0x8a0e, 0x2875, 0x4a1c, 0x295f, 0x1393, 0xf760, 0x9178, 0x0f5b,
-  0xfa7d, 0x83b4, 0x2082, 0x721d, 0x6462, 0x0368, 0x67e2, 0x8624,
-  0x194d, 0x22f6, 0x78fb, 0x6791, 0xb238, 0xb332, 0x7276, 0xf272,
-  0x47ec, 0x4504, 0xa961, 0x9fc8, 0x3fdc, 0xb413, 0x007a, 0x0806,
-  0x7458, 0x95c6, 0xccaa, 0x18d6, 0xe2ae, 0x1b06, 0xf3f6, 0x5050,
-  0xc8e8, 0xf4ac, 0xc04c, 0xf41c, 0x992f, 0xae44, 0x5f1b, 0x1113,
-  0x1738, 0xd9a8, 0x19ea, 0x2d33, 0x9698, 0x2fe9, 0x323f, 0xcde2,
-  0x6d71, 0xe37d, 0xb697, 0x2c4f, 0x4373, 0x9102, 0x075d, 0x8e25,
-  0x1672, 0xec28, 0x6acb, 0x86cc, 0x186e, 0x9414, 0xd674, 0xd1a5
-};
-
 /****************************************************************
  Instruction tables
  *****************************************************************/
@@ -841,24 +791,32 @@ struct _xd3_code_table_desc
   /* Assumes a single RUN instruction */
   /* Assumes that MIN_MATCH is 4 */
 
-  uint8_t add_sizes;            /* Number of immediate-size single adds (default 17) */
+  uint8_t add_sizes;            /* Number of immediate-size single
+                                   adds (default 17) */
   uint8_t near_modes;           /* Number of near copy modes (default 4) */
   uint8_t same_modes;           /* Number of same copy modes (default 3) */
-  uint8_t cpy_sizes;            /* Number of immediate-size single copies (default 15) */
-
-  uint8_t addcopy_add_max;      /* Maximum add size for an add-copy double instruction,
-                                   all modes (default 4) */
-  uint8_t addcopy_near_cpy_max; /* Maximum cpy size for an add-copy double instruction,
-                                   up through VCD_NEAR modes (default 6) */
-  uint8_t addcopy_same_cpy_max; /* Maximum cpy size for an add-copy double instruction,
-                                   VCD_SAME modes (default 4) */
-
-  uint8_t copyadd_add_max;      /* Maximum add size for a copy-add double instruction,
-                                   all modes (default 1) */
-  uint8_t copyadd_near_cpy_max; /* Maximum cpy size for a copy-add double instruction,
-                                   up through VCD_NEAR modes (default 4) */
-  uint8_t copyadd_same_cpy_max; /* Maximum cpy size for a copy-add double instruction,
-                                   VCD_SAME modes (default 4) */
+  uint8_t cpy_sizes;            /* Number of immediate-size single
+                                   copies (default 15) */
+
+  uint8_t addcopy_add_max;      /* Maximum add size for an add-copy
+                                   double instruction, all modes
+                                   (default 4) */
+  uint8_t addcopy_near_cpy_max; /* Maximum cpy size for an add-copy
+                                   double instruction, up through
+                                   VCD_NEAR modes (default 6) */
+  uint8_t addcopy_same_cpy_max; /* Maximum cpy size for an add-copy
+                                   double instruction, VCD_SAME modes
+                                   (default 4) */
+
+  uint8_t copyadd_add_max;      /* Maximum add size for a copy-add
+                                   double instruction, all modes
+                                   (default 1) */
+  uint8_t copyadd_near_cpy_max; /* Maximum cpy size for a copy-add
+                                   double instruction, up through
+                                   VCD_NEAR modes (default 4) */
+  uint8_t copyadd_same_cpy_max; /* Maximum cpy size for a copy-add
+                                   double instruction, VCD_SAME modes
+                                   (default 4) */
 
   xd3_code_table_sizes addcopy_max_sizes[MAX_MODES];
   xd3_code_table_sizes copyadd_max_sizes[MAX_MODES];
@@ -880,17 +838,20 @@ static const xd3_code_table_desc __rfc3284_code_table_desc = {
   4,  /* copy-add max cpy, same */
 
   /* addcopy */
-  { {6,163,3},{6,175,3},{6,187,3},{6,199,3},{6,211,3},{6,223,3},{4,235,1},{4,239,1},{4,243,1} },
+  { {6,163,3},{6,175,3},{6,187,3},{6,199,3},{6,211,3},{6,223,3},
+    {4,235,1},{4,239,1},{4,243,1} },
   /* copyadd */
-  { {4,247,1},{4,248,1},{4,249,1},{4,250,1},{4,251,1},{4,252,1},{4,253,1},{4,254,1},{4,255,1} },
+  { {4,247,1},{4,248,1},{4,249,1},{4,250,1},{4,251,1},{4,252,1},
+    {4,253,1},{4,254,1},{4,255,1} },
 };
 
 /* Computes code table entries of TBL using the specified description. */
 static void
 xd3_build_code_table (const xd3_code_table_desc *desc, xd3_dinst *tbl)
 {
-  usize_t size1, size2, mode;
-  usize_t cpy_modes = 2 + desc->near_modes + desc->same_modes;
+  uint8_t size1, size2;
+  uint8_t mode;
+  usize_t cpy_modes = 2U + desc->near_modes + desc->same_modes;
   xd3_dinst *d = tbl;
 
   (d++)->type1 = XD3_RUN;
@@ -906,7 +867,8 @@ xd3_build_code_table (const xd3_code_table_desc *desc, xd3_dinst *tbl)
     {
       (d++)->type1 = XD3_CPY + mode;
 
-      for (size1 = MIN_MATCH; size1 < MIN_MATCH + desc->cpy_sizes; size1 += 1, d += 1)
+      for (size1 = MIN_MATCH; size1 < MIN_MATCH + desc->cpy_sizes;
+           size1 += 1, d += 1)
         {
           d->type1 = XD3_CPY + mode;
           d->size1 = size1;
@@ -998,7 +960,7 @@ xd3_choose_instruction (xd3_rinst *prev, xd3_rinst *inst)
 
     default:
       {
-        int mode = inst->type - XD3_CPY;
+        uint8_t mode = inst->type - XD3_CPY;
 
         XD3_ASSERT (inst->type >= XD3_CPY && inst->type < 12);
 
@@ -1015,8 +977,9 @@ xd3_choose_instruction (xd3_rinst *prev, xd3_rinst *inst)
                 if ( (inst->size <= 6) &&
                      (mode       <= 5) )
                   {
-                    prev->code2 = 163 + (mode * 12) + (3 * (prev->size - 1)) + (inst->size - 4);
-
+                    prev->code2 = (uint8_t)(163 + (mode * 12) +
+                                            (3 * (prev->size - 1)) +
+                                            (inst->size - 4));
                     XD3_ASSERT (prev->code2 <= 234);
                   }
                 else if ( (inst->size == 4) &&
@@ -1077,10 +1040,10 @@ xd3_check_pow2 (xoff_t value, usize_t *logof)
   return XD3_INTERNAL;
 }
 
-size_t
-xd3_pow2_roundup (size_t x)
+usize_t
+xd3_pow2_roundup (usize_t x)
 {
-  size_t i = 1;
+  usize_t i = 1;
   while (x > i) {
     i <<= 1U;
   }
@@ -1122,7 +1085,8 @@ xd3_round_blksize (usize_t sz, usize_t blksz)
  ***********************************************************************/
 
 #define A32_BASE 65521L /* Largest prime smaller than 2^16 */
-#define A32_NMAX 5552   /* NMAX is the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1 */
+#define A32_NMAX 5552   /* NMAX is the largest n such that 255n(n+1)/2
+                           + (n+1)(BASE-1) <= 2^32-1 */
 
 #define A32_DO1(buf,i)  {s1 += buf[i]; s2 += s1;}
 #define A32_DO2(buf,i)  A32_DO1(buf,i); A32_DO1(buf,i+1);
@@ -1130,11 +1094,10 @@ xd3_round_blksize (usize_t sz, usize_t blksz)
 #define A32_DO8(buf,i)  A32_DO4(buf,i); A32_DO4(buf,i+4);
 #define A32_DO16(buf)   A32_DO8(buf,0); A32_DO8(buf,8);
 
-static unsigned long adler32 (unsigned long adler, const uint8_t *buf, 
-                              usize_t len)
+static uint32_t adler32 (uint32_t adler, const uint8_t *buf, usize_t len)
 {
-    unsigned long s1 = adler & 0xffff;
-    unsigned long s2 = (adler >> 16) & 0xffff;
+    uint32_t s1 = adler & 0xffffU;
+    uint32_t s2 = (adler >> 16) & 0xffffU;
     int k;
 
     while (len > 0)
@@ -1329,7 +1292,8 @@ xd3_encode_address (xd3_stream *stream,
                     uint8_t* mode)
 {
   usize_t d, bestd;
-  usize_t i, bestm, ret;
+  usize_t i, bestm;
+  int ret;
   xd3_addr_cache* acache = & stream->acache;
 
 #define SMALLEST_INT(x) do { if (((x) & ~127U) == 0) { goto good; } } while (0)
@@ -1403,7 +1367,7 @@ xd3_encode_address (xd3_stream *stream,
 static int
 xd3_decode_address (xd3_stream *stream, usize_t here,
                     usize_t mode, const uint8_t **inpp,
-                    const uint8_t *max, uint32_t *valp)
+                    const uint8_t *max, usize_t *valp)
 {
   int ret;
   usize_t same_start = 2 + stream->acache.s_near;
@@ -1620,11 +1584,13 @@ xd3_free_stream (xd3_stream *stream)
       xd3_free (stream, tmp);
     }
 
+#if XD3_ENCODER
   xd3_free (stream, stream->large_table);
   xd3_free (stream, stream->small_table);
+  xd3_free (stream, stream->large_hash.powers);
+  xd3_free (stream, stream->small_hash.powers);
   xd3_free (stream, stream->small_prev);
 
-#if XD3_ENCODER
   {
     int i;
     for (i = 0; i < ENC_SECTS; i += 1)
@@ -1993,7 +1959,7 @@ xd3_set_source (xd3_stream *stream,
     }
 
   src->shiftby = shiftby;
-  src->maskby = (1 << shiftby) - 1;
+  src->maskby = (1ULL << shiftby) - 1ULL;
 
   if (xd3_check_pow2 (src->max_winsize, NULL) != 0)
     {
@@ -2012,7 +1978,8 @@ xd3_set_source_and_size (xd3_stream *stream,
   if (ret == 0)
     {
       stream->src->eof_known = 1;
-
+      IF_DEBUG2 (DP(RINT "[set source] size known %"Q"\n",
+                    source_size));
       xd3_blksize_div(source_size,
                       stream->src,
                       &stream->src->max_blkno,
@@ -2192,22 +2159,22 @@ xd3_iopt_finish_encoding (xd3_stream *stream, xd3_rinst *inst)
                 XD3_ASSERT (inst->addr >= src->srcbase);
                 XD3_ASSERT (inst->addr + inst->size <=
                             src->srcbase + src->srclen);
-                addr = (usize_t)(inst->addr - src->srcbase);
+                addr = inst->addr - src->srcbase;
                 stream->n_scpy += 1;
-                stream->l_scpy += (xoff_t) inst->size;
+                stream->l_scpy += inst->size;
               }
             else
               {
                 /* with source window: target copy address is offset
                  * by taroff. */
-                addr = stream->taroff + (usize_t) inst->addr;
+                addr = stream->taroff + inst->addr;
                 stream->n_tcpy += 1;
-                stream->l_tcpy += (xoff_t) inst->size;
+                stream->l_tcpy += inst->size;
               }
           }
         else
           {
-            addr = (usize_t) inst->addr;
+            addr = inst->addr;
             stream->n_tcpy += 1;
             stream->l_tcpy += inst->size;
           }
@@ -2223,7 +2190,7 @@ xd3_iopt_finish_encoding (xd3_stream *stream, xd3_rinst *inst)
 
         IF_DEBUG2 ({
           static int cnt;
-          DP(RINT "[iopt copy:%d] pos %"Q"u-%"Q"u addr %"Q"u-%"Q"u size %u\n",
+          DP(RINT "[iopt copy:%d] pos %"Q"-%"Q" addr %"Q"-%"Q" size %u\n",
                    cnt++,
                    stream->total_in + inst->pos,
                    stream->total_in + inst->pos + inst->size,
@@ -2240,7 +2207,7 @@ xd3_iopt_finish_encoding (xd3_stream *stream, xd3_rinst *inst)
 
         IF_DEBUG2 ({
           static int cnt;
-          DP(RINT "[iopt run:%d] pos %"Q"u size %u\n", cnt++, stream->total_in + inst->pos, inst->size);
+          DP(RINT "[iopt run:%d] pos %"Q" size %u\n", cnt++, stream->total_in + inst->pos, inst->size);
         });
         break;
       }
@@ -2254,7 +2221,7 @@ xd3_iopt_finish_encoding (xd3_stream *stream, xd3_rinst *inst)
 
         IF_DEBUG2 ({
           static int cnt;
-          DP(RINT "[iopt add:%d] pos %"Q"u size %u\n", cnt++, stream->total_in + inst->pos, inst->size);
+          DP(RINT "[iopt add:%d] pos %"Q" size %u\n", cnt++, stream->total_in + inst->pos, inst->size);
         });
 
         break;
@@ -2273,7 +2240,8 @@ xd3_iopt_finish_encoding (xd3_stream *stream, xd3_rinst *inst)
     {
       if (stream->iout->code2 != 0)
         {
-          if ((ret = xd3_emit_double (stream, stream->iout, inst, stream->iout->code2))) { return ret; }
+          if ((ret = xd3_emit_double (stream, stream->iout, inst, 
+                                      stream->iout->code2))) { return ret; }
 
           xd3_iopt_free_nonadd (stream, stream->iout);
           xd3_iopt_free_nonadd (stream, inst);
@@ -2630,7 +2598,7 @@ xd3_iopt_last_matched (xd3_stream *stream)
  ***********************************************************/
 
 static int
-xd3_emit_single (xd3_stream *stream, xd3_rinst *single, usize_t code)
+xd3_emit_single (xd3_stream *stream, xd3_rinst *single, uint8_t code)
 {
   int has_size = stream->code_table[code].size1 == 0;
   int ret;
@@ -2659,7 +2627,7 @@ xd3_emit_single (xd3_stream *stream, xd3_rinst *single, usize_t code)
 
 static int
 xd3_emit_double (xd3_stream *stream, xd3_rinst *first,
-                 xd3_rinst *second, usize_t code)
+                 xd3_rinst *second, uint8_t code)
 {
   int ret;
 
@@ -2729,8 +2697,8 @@ xd3_emit_hdr (xd3_stream *stream)
   int  use_secondary = stream->sec_type != NULL;
   int  use_adler32   = stream->flags & (XD3_ADLER32 | XD3_ADLER32_RECODE);
   int  vcd_source    = xd3_encoder_used_source (stream);
-  usize_t win_ind = 0;
-  usize_t del_ind = 0;
+  uint8_t win_ind = 0;
+  uint8_t del_ind = 0;
   usize_t enc_len;
   usize_t tgt_len;
   usize_t data_len;
@@ -2739,7 +2707,7 @@ xd3_emit_hdr (xd3_stream *stream)
 
   if (stream->current_window == 0)
     {
-      usize_t hdr_ind = 0;
+      uint8_t hdr_ind = 0;
       int use_appheader  = stream->enc_appheader != NULL;
 
       if (use_secondary)  { hdr_ind |= VCD_SECONDARY; }
@@ -2837,7 +2805,7 @@ xd3_emit_hdr (xd3_stream *stream)
   inst_len = xd3_sizeof_output (INST_HEAD (stream));
   addr_len = xd3_sizeof_output (ADDR_HEAD (stream));
 
-  /* The enc_len field is a redundency for future extensions.*/
+  /* The enc_len field is a redundency for future extensions. */
   enc_len = (1 + (xd3_sizeof_size (tgt_len) +
                   xd3_sizeof_size (data_len) +
                   xd3_sizeof_size (inst_len) +
@@ -2978,6 +2946,7 @@ xd3_alloc_iopt (xd3_stream *stream, usize_t elts)
 static int
 xd3_encode_init (xd3_stream *stream, int full_init)
 {
+  int ret;
   int i;
 
   if (full_init)
@@ -2990,12 +2959,17 @@ xd3_encode_init (xd3_stream *stream, int full_init)
        * identical or short inputs require no table allocation. */
       if (large_comp)
         {
+          /* TODO Need to check for overflow here. */
           usize_t hash_values = stream->src->max_winsize /
                                 stream->smatcher.large_step;
 
-          xd3_size_hashtable (stream,
-                              hash_values,
-                              & stream->large_hash);
+          if ((ret = xd3_size_hashtable (stream,
+                                         hash_values,
+                                         stream->smatcher.large_look,
+                                         & stream->large_hash)))
+            {
+              return ret;
+            }
         }
 
       if (small_comp)
@@ -3005,9 +2979,13 @@ xd3_encode_init (xd3_stream *stream, int full_init)
            * also sort of makes sense. @@@ */
           usize_t hash_values = stream->winsize;
 
-          xd3_size_hashtable (stream,
-                              hash_values,
-                              & stream->small_hash);
+          if ((ret = xd3_size_hashtable (stream,
+                                         hash_values,
+                                         stream->smatcher.small_look,
+                                         & stream->small_hash)))
+            {
+              return ret;
+            }
         }
     }
 
@@ -3156,7 +3134,7 @@ xd3_encode_input (xd3_stream *stream)
 
       stream->enc_state = ENC_SEARCH;
 
-      IF_DEBUG2 (DP(RINT "[WINSTART:%"Q"u] input bytes %u offset %"Q"u\n",
+      IF_DEBUG2 (DP(RINT "[WINSTART:%"Q"] input bytes %u offset %"Q"\n",
                     stream->current_window, stream->avail_in,
                     stream->total_in));
       return XD3_WINSTART;
@@ -3270,7 +3248,7 @@ xd3_encode_input (xd3_stream *stream)
       stream->enc_state  = ENC_POSTOUT;
       stream->next_out   = stream->enc_current->base;
       stream->avail_out  = stream->enc_current->next;
-      stream->total_out += (xoff_t) stream->avail_out;
+      stream->total_out += stream->avail_out;
 
       /* If there is any output in this buffer, return it, otherwise
        * fall through to handle the next buffer or finish the window
@@ -3295,10 +3273,10 @@ xd3_encode_input (xd3_stream *stream)
           goto enc_output;
         }
 
-      stream->total_in += (xoff_t) stream->avail_in;
+      stream->total_in += stream->avail_in;
       stream->enc_state = ENC_POSTWIN;
 
-      IF_DEBUG2 (DP(RINT "[WINFINISH:%"Q"u] in=%"Q"u\n",
+      IF_DEBUG2 (DP(RINT "[WINFINISH:%"Q"] in=%"Q"\n",
                     stream->current_window,
                     stream->total_in));
       return XD3_WINFINISH;
@@ -3607,7 +3585,7 @@ xd3_string_match_init (xd3_stream *stream)
   return 0;
 }
 
-#if XD3_USE_LARGEFILE64
+#if XD3_USE_LARGEFILE64 && !XD3_USE_LARGESIZET
 /* This function handles the 32/64bit ambiguity -- file positions are 64bit
  * but the hash table for source-offsets is 32bit. */
 static xoff_t
@@ -3634,7 +3612,7 @@ xd3_source_cksum_offset(xd3_stream *stream, usize_t low)
 static xoff_t
 xd3_source_cksum_offset(xd3_stream *stream, usize_t low)
 {
-  return (xoff_t) low;
+  return low;
 }
 #endif
 
@@ -3668,7 +3646,7 @@ xd3_srcwin_setup (xd3_stream *stream)
    * use smaller windows. */
   length = stream->match_maxaddr - stream->match_minaddr;
 
-  x = (xoff_t) USIZE_T_MAX;
+  x = USIZE_T_MAX;
   if (length > x)
     {
       stream->msg = "source window length overflow (not 64bit)";
@@ -3687,21 +3665,20 @@ xd3_srcwin_setup (xd3_stream *stream)
 
   /* Otherwise, we have to make a guess.  More copies may still be
    * issued, but we have to decide the source window base and length
-   * now.  */
-  /* TODO: This may not working well in practice, more testing needed. */
+   * now.  
+   * TODO: This may not working well in practice, more testing needed. */
   src->srcbase = stream->match_minaddr;
   src->srclen  = xd3_max ((usize_t) length,
                           stream->avail_in + (stream->avail_in >> 2));
 
   if (src->eof_known)
     {
-      /* Note: if the source size is known, we must reduce srclen or 
+      /* Note: if the source size is known, we must reduce srclen or
        * code that expects to pass a single block w/ getblk == NULL
        * will not function, as the code will return GETSRCBLK asking
        * for the second block. */
       src->srclen = xd3_min (src->srclen, xd3_source_eof(src) - src->srcbase);
     }
-  
   IF_DEBUG1 (DP(RINT "[srcwin_setup_constrained] base %"Q"u len %u\n",
                 src->srcbase, src->srclen));
 
@@ -3768,11 +3745,12 @@ xd3_source_match_setup (xd3_stream *stream, xoff_t srcpos)
 
   /* Going backwards, the 1.5-pass algorithm allows some
    * already-matched input may be covered by a longer source match.
-   * The greedy algorithm does not allow this. */
+   * The greedy algorithm does not allow this.
+   * TODO: Measure this. */
   if (stream->flags & XD3_BEGREEDY)
     {
       /* The greedy algorithm allows backward matching to the last
-         matched position. */
+       * matched position. */
       greedy_or_not = xd3_iopt_last_matched (stream);
     }
   else
@@ -3801,7 +3779,8 @@ xd3_source_match_setup (xd3_stream *stream, xoff_t srcpos)
        * 0--src->size.  We compare the usize_t
        * match_maxfwd/match_maxback against the xoff_t
        * src->size/srcpos values and take the min. */
-      if (srcpos < (xoff_t) stream->match_maxback)
+      /* TODO #if XD3_USE_LARGESIZET ? */
+      if (srcpos < stream->match_maxback)
         {
           stream->match_maxback = (usize_t) srcpos;
         }
@@ -3810,14 +3789,14 @@ xd3_source_match_setup (xd3_stream *stream, xoff_t srcpos)
         {
           xoff_t srcavail = xd3_source_eof (src) - srcpos;
 
-          if (srcavail < (xoff_t) stream->match_maxfwd)
+          if (srcavail < stream->match_maxfwd)
             {
               stream->match_maxfwd = (usize_t) srcavail;
             }
         }
 
       IF_DEBUG2(DP(RINT
-                   "[match_setup] srcpos %"Q"u (tgtpos %"Q"u) "
+                   "[match_setup] srcpos %"Q" (tgtpos %"Q") "
                    "unrestricted maxback %u maxfwd %u\n",
                    srcpos,
                    stream->total_in + stream->input_position,
@@ -3831,7 +3810,7 @@ xd3_source_match_setup (xd3_stream *stream, xoff_t srcpos)
 
   /* Restricted case: fail if the srcpos lies outside the source window */
   if ((srcpos < src->srcbase) ||
-      (srcpos > (src->srcbase + (xoff_t) src->srclen)))
+      (srcpos > (src->srcbase + src->srclen)))
     {
       IF_DEBUG1(DP(RINT "[match_setup] restricted source window failure\n"));
       goto bad;
@@ -3846,7 +3825,7 @@ xd3_source_match_setup (xd3_stream *stream, xoff_t srcpos)
           stream->match_maxback = srcavail;
         }
 
-      srcavail = (usize_t) (src->srcbase + (xoff_t) src->srclen - srcpos);
+      srcavail = src->srcbase + src->srclen - srcpos;
       if (srcavail < stream->match_maxfwd)
         {
           stream->match_maxfwd = srcavail;
@@ -3874,19 +3853,19 @@ xd3_source_match_setup (xd3_stream *stream, xoff_t srcpos)
   return 1;
 }
 
-static inline int
-xd3_forward_match(const uint8_t *s1c, const uint8_t *s2c, int n)
+static inline usize_t
+xd3_forward_match(const uint8_t *s1c, const uint8_t *s2c, usize_t n)
 {
-  int i = 0;
+  usize_t i = 0;
 #if UNALIGNED_OK
-  int nint = n / sizeof(int);
+  usize_t nint = n / sizeof(int);
 
   if (nint >> 3)
     {
-      int j = 0;
+      usize_t j = 0;
       const int *s1 = (const int*)s1c;
       const int *s2 = (const int*)s2c;
-      int nint_8 = nint - 8;
+      usize_t nint_8 = nint - 8;
 
       while (i <= nint_8 &&
              s1[i++] == s2[j++] &&
@@ -3933,6 +3912,9 @@ xd3_source_extend_match (xd3_stream *stream)
   usize_t tryrem;    /* tryrem is the number of matchable bytes */
   usize_t matched;
 
+  IF_DEBUG2(DP(RINT "[extend match] srcpos %"Q"\n",
+               stream->match_srcpos));
+
   XD3_ASSERT (src != NULL);
 
   /* Does it make sense to compute backward match AFTER forward match? */
@@ -3980,7 +3962,7 @@ xd3_source_extend_match (xd3_stream *stream)
 
           tryrem = xd3_min (tryoff, stream->match_maxback - stream->match_back);
 
-          IF_DEBUG2(DP(RINT "[maxback] maxback %u trysrc %"Q"u/%u tgt %u tryrem %u\n",
+          IF_DEBUG2(DP(RINT "[maxback] maxback %u trysrc %"Q"/%u tgt %u tryrem %u\n",
                        stream->match_maxback, tryblk, tryoff, streamoff, tryrem));
 
           /* TODO: This code can be optimized similar to xd3_match_forward() */
@@ -4321,7 +4303,7 @@ xd3_smatch (xd3_stream *stream,
 static void
 xd3_verify_small_state (xd3_stream    *stream,
                         const uint8_t *inp,
-                        uint32_t          x_cksum)
+                        uint32_t       x_cksum)
 {
   uint32_t state;
   uint32_t cksum = xd3_scksum (&state, inp, stream->smatcher.small_look);
@@ -4332,9 +4314,9 @@ xd3_verify_small_state (xd3_stream    *stream,
 static void
 xd3_verify_large_state (xd3_stream    *stream,
                         const uint8_t *inp,
-                        uint32_t          x_cksum)
+                        usize_t        x_cksum)
 {
-  uint32_t cksum = xd3_lcksum (inp, stream->smatcher.large_look);
+  usize_t cksum = xd3_large_cksum (&stream->large_hash, inp, stream->smatcher.large_look);
   XD3_ASSERT (cksum == x_cksum);
 }
 static void
@@ -4478,8 +4460,12 @@ xd3_srcwin_move_point (xd3_stream *stream, usize_t *next_move_point)
 
       do
         {
-          uint32_t cksum = xd3_lcksum (stream->src->curblk + blkpos,
-                                       stream->smatcher.large_look);
+          /* TODO: This would be significantly faster if the compiler
+           * knew stream->smatcher.large_look (which the template for
+           * xd3_string_match_* allows). */
+          usize_t cksum = xd3_large_cksum (&stream->large_hash, 
+                                           stream->src->curblk + blkpos,
+                                           stream->smatcher.large_look);
           usize_t hval = xd3_checksum_hash (& stream->large_hash, cksum);
 
           stream->large_table[hval] =
@@ -4495,6 +4481,15 @@ xd3_srcwin_move_point (xd3_stream *stream, usize_t *next_move_point)
       stream->srcwin_cksum_pos = (blkno + 1) * stream->src->blksize;
     }
 
+  IF_DEBUG1 (DP(RINT
+                "[srcwin_move_point] exited loop T=%"Q"{%"Q"} "
+                "S=%"Q" EOF=%"Q" %s\n",
+                stream->total_in + stream->input_position,
+                logical_input_cksum_pos,
+                stream->srcwin_cksum_pos,
+                xd3_source_eof (stream->src),
+                stream->src->eof_known ? "known" : "unknown"));
+
   if (stream->src->eof_known)
     {
       xoff_t source_size = xd3_source_eof (stream->src);
@@ -4574,7 +4569,7 @@ XD3_TEMPLATE(xd3_string_match_) (xd3_stream *stream)
   const uint8_t *inp;
   uint32_t       scksum = 0;
   uint32_t       scksum_state = 0;
-  uint32_t       lcksum = 0;
+  usize_t        lcksum = 0;
   usize_t        sinx;
   usize_t        linx;
   uint8_t        run_c;
@@ -4650,7 +4645,7 @@ XD3_TEMPLATE(xd3_string_match_) (xd3_stream *stream)
           return ret;
         }
 
-      lcksum = xd3_lcksum (inp, LLOOK);
+      lcksum = xd3_large_cksum (&stream->large_hash, inp, LLOOK);
     }
 
   /* TRYLAZYLEN: True if a certain length match should be followed by
@@ -4826,7 +4821,7 @@ XD3_TEMPLATE(xd3_string_match_) (xd3_stream *stream)
 
       if (DO_LARGE && (stream->input_position + LLOOK < stream->avail_in))
         {
-          lcksum = xd3_large_cksum_update (lcksum, inp, LLOOK);
+          lcksum = xd3_large_cksum_update (&stream->large_hash, lcksum, inp, LLOOK);
         }
     }
 
diff --git a/xdelta3/xdelta3.h b/xdelta3/xdelta3.h
index 6cc0f9f..1b134eb 100644
--- a/xdelta3/xdelta3.h
+++ b/xdelta3/xdelta3.h
@@ -17,7 +17,7 @@
  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
-/* To know more about Xdelta, start by reading xdelta3.c.  If you are
+/* To learn more about Xdelta, start by reading xdelta3.c.  If you are
  * ready to use the API, continue reading here.  There are two
  * interfaces -- xd3_encode_input and xd3_decode_input -- plus a dozen
  * or so related calls.  This interface is styled after Zlib. */
@@ -25,7 +25,7 @@
 #ifndef _XDELTA3_H_
 #define _XDELTA3_H_
 
-#define _POSIX_SOURCE
+#define _POSIX_SOURCE 200112L
 #define _ISOC99_SOURCE
 #define _C99_SOURCE
 
@@ -71,7 +71,7 @@
  *
  * 8-16MB is reasonable, probably don't need to go larger. */
 #ifndef XD3_HARDMAXWINSIZE
-#define XD3_HARDMAXWINSIZE (1U<<24)
+#define XD3_HARDMAXWINSIZE (1U<<26)
 #endif
 /* The IOPT_SIZE value sets the size of a buffer used to batch
  * overlapping copy instructions before they are optimized by picking
@@ -87,8 +87,7 @@
 #define XD3_DEFAULT_SPREVSZ (1U<<18)
 #endif
 
-/* The default compression level
- */
+/* The default compression level */
 #ifndef XD3_DEFAULT_LEVEL
 #define XD3_DEFAULT_LEVEL 3
 #endif
@@ -101,6 +100,12 @@
 #define XD3_USE_LARGEFILE64 1
 #endif
 
+/* The source window size is limited to 2GB unless
+ * XD3_USE_LARGESIZET is defined to 1. */
+#ifndef XD3_USE_LARGESIZET
+#define XD3_USE_LARGESIZET 1
+#endif
+
 /* Sizes and addresses within VCDIFF windows are represented as usize_t
  *
  * For source-file offsets and total file sizes, total input and
@@ -109,19 +114,19 @@
  * the 32bit boundary [xdelta3-test.h]).
  */
 #ifndef _WIN32
+#define __STDC_FORMAT_MACROS
+#include <inttypes.h>
 #include <stdint.h>
 #else /* WIN32 case */
 #define WIN32_LEAN_AND_MEAN
 
 #ifndef WINVER
 #if XD3_USE_LARGEFILE64
-/* 64 bit file offsets: uses GetFileSizeEx and SetFilePointerEx.
- * requires Win2000 or newer version of WinNT */
+/* 64 bit file offsets: uses GetFileSizeEx and SetFilePointerEx. */
 #define WINVER          0x0500
 #define _WIN32_WINNT    0x0500
 #else /* xoff_t is 32bit */
-/* 32 bit (DWORD) file offsets: uses GetFileSize and
- * SetFilePointer. compatible with win9x-me and WinNT4 */
+/* 32 bit file offsets: uses GetFileSize and SetFilePointer. */
 #define WINVER          0x0400
 #define _WIN32_WINNT    0x0400
 #endif /* if XD3_USE_LARGEFILE64 */
@@ -129,9 +134,8 @@
 
 #include <windows.h>
 
-/* _MSV_VER is defined by Microsoft tools, not by mingw32 */
+/* _MSV_VER is defined by Microsoft tools, not by Mingw32 */
 #ifdef _MSC_VER
-/*#define inline*/
 typedef signed int     ssize_t;
 #if _MSC_VER < 1600
 typedef unsigned char  uint8_t;
@@ -143,16 +147,16 @@ typedef ULONGLONG      uint64_t;
 #include <stdint.h>
 #endif /* _MSC_VER < 1600 */
 #else /* _MSC_VER not defined  */
-/* mingw32, lcc and watcom provide a proper header */
+/* Mingw32 */
 #include <stdint.h>
 #endif /* _MSC_VER defined */
-#endif /* _WIN32 defined */
 
-typedef uint32_t usize_t;
+#endif /* _WIN32 defined */
 
+/* Settings based on the size of xoff_t (32 vs 64 file offsets) */
 #if XD3_USE_LARGEFILE64
 /* xoff_t is a 64-bit type */
-#define __USE_FILE_OFFSET64 1 /* GLIBC: for 64bit fileops, ... ? */
+#define __USE_FILE_OFFSET64 1 /* GLIBC: for 64bit fileops. */
 
 #ifndef _LARGEFILE_SOURCE
 #define _LARGEFILE_SOURCE
@@ -162,6 +166,7 @@ typedef uint32_t usize_t;
 #define _FILE_OFFSET_BITS 64
 #endif
 
+/* Set a xoff_t typedef and the "Q" printf insert. */
 #if defined(_WIN32)
 typedef uint64_t xoff_t;
 /* Note: The following generates benign warnings in a mingw
@@ -176,7 +181,7 @@ typedef size_t xoff_t;
 #elif SIZEOF_UNSIGNED_LONG_LONG == 8
 typedef unsigned long long xoff_t;
 #define Q "ll"
-#endif /* #define Q */
+#endif /* typedef and #define Q */
 
 #define SIZEOF_XOFF_T 8
 
@@ -194,9 +199,45 @@ typedef uint32_t xoff_t;
 #define Q
 #endif /* 64 vs 32 bit xoff_t */
 
-/* Note: This gets modified in the 64bithash branch. */
+/* Settings based on the size of usize_t (32 and 64 bit window size) */
+#if XD3_USE_LARGESIZET
+
+/* Set a usize_ttypedef and the "W" printf insert. */
+#if defined(_WIN32)
+typedef uint64_t usize_t;
+/* Note: The following generates benign warnings in a mingw
+ * cross-compiler */
+#define W "I64"
+#elif SIZEOF_UNSIGNED_LONG == 8
+typedef unsigned long usize_t;
+#define W "l"
+#elif SIZEOF_SIZE_T == 8
+typedef size_t usize_t;
+#define W "z"
+#elif SIZEOF_UNSIGNED_LONG_LONG == 8
+typedef unsigned long long usize_t;
+#define W "ll"
+#endif /* typedef and #define W */
+
+#define SIZEOF_USIZE_T 8
+
+#else /* XD3_USE_LARGESIZET == 0 */
+
+#if SIZEOF_UNSIGNED_INT == 4
+typedef unsigned int usize_t;
+#elif SIZEOF_UNSIGNED_LONG == 4
+typedef unsigned long usize_t;
+#else
+typedef uint32_t usize_t;
+#endif /* usize_t is 32 bits */
+
 #define SIZEOF_USIZE_T 4
+#define W
 
+#endif /* 64 vs 32 bit usize_t */
+
+/* Settings based on the size of size_t (the system-provided,
+ * usually-but-maybe-not an unsigned type) */
 #if SIZEOF_SIZE_T == 4
 #define Z "z"
 #elif SIZEOF_SIZE_T == 8
@@ -339,9 +380,12 @@ typedef int              (xd3_comp_table_func) (xd3_stream *stream,
 
 
 #if XD3_DEBUG
-#define XD3_ASSERT(x) \
-    do { if (! (x)) { DP(RINT "%s:%d: XD3 assertion failed: %s\n", __FILE__, __LINE__, #x); \
-    abort (); } } while (0)
+#define XD3_ASSERT(x)                                \
+  do {                                               \
+    if (! (x)) {                                     \
+      DP(RINT "%s:%d: XD3 assertion failed: %s\n",   \
+         __FILE__, __LINE__, #x);                    \
+      abort (); } } while (0)
 #else
 #define XD3_ASSERT(x) (void)0
 #endif  /* XD3_DEBUG */
@@ -589,9 +633,9 @@ struct _xd3_dinst
 /* the decoded form of a single (half) instruction. */
 struct _xd3_hinst
 {
-  uint8_t     type;
-  uint32_t    size;  /* TODO: why decode breaks if this is usize_t? */
-  uint32_t    addr;  /* TODO: why decode breaks if this is usize_t? */
+  uint8_t    type;
+  usize_t    size;
+  usize_t    addr;
 };
 
 /* the form of a whole-file instruction */
@@ -618,7 +662,7 @@ struct _xd3_desect
 {
   const uint8_t *buf;
   const uint8_t *buf_max;
-  uint32_t       size;  /* TODO: why decode breaks if this is usize_t? */
+  usize_t        size;
   usize_t        pos;
 
   /* used in xdelta3-decode.h */
@@ -666,9 +710,13 @@ struct _xd3_smatcher
 /* hash table size & power-of-two hash function. */
 struct _xd3_hash_cfg
 {
-  usize_t           size;
-  usize_t           shift;
-  usize_t           mask;
+  usize_t  size;       // Number of buckets
+  usize_t  shift;
+  usize_t  mask;
+  usize_t  look;       // How wide is this checksum
+  usize_t  multiplier; // K * powers[0]
+  usize_t *powers;     // Array of [0,look) where powers[look-1] == 1
+                       // and powers[N] = powers[N+1]*K (Rabin-Karp)
 };
 
 /* the sprev list */
@@ -728,7 +776,7 @@ struct _xd3_config
   xd3_alloc_func    *alloc;
   xd3_free_func     *freef;
   void              *opaque;        /* Not used. */
-  int                flags;         /* stream->flags are initialized
+  uint32_t           flags;         /* stream->flags are initialized
                                      * from xd3_config & never
                                      * modified by the library.  Use
                                      * xd3_set_flags to modify flags
@@ -771,8 +819,8 @@ struct _xd3_source
   usize_t             srclen;        /* length of this source window */
   xoff_t              srcbase;       /* offset of this source window
                                         in the source itself */
-  int                 shiftby;       /* for power-of-two blocksizes */
-  int                 maskby;        /* for power-of-two blocksizes */
+  usize_t             shiftby;       /* for power-of-two blocksizes */
+  usize_t             maskby;        /* for power-of-two blocksizes */
   xoff_t              cpyoff_blocks; /* offset of dec_cpyoff in blocks */
   usize_t             cpyoff_blkoff; /* offset of copy window in
                                         blocks, remainder */
@@ -829,7 +877,7 @@ struct _xd3_stream
   xd3_free_func    *free;             /* free function */
   void*             opaque;           /* private data object passed to
                                          alloc, free, and getblk */
-  int               flags;            /* various options */
+  uint32_t          flags;            /* various options */
 
   /* secondary compressor configuration */
   xd3_sec_cfg       sec_data;         /* Secondary compressor config: data */
@@ -927,13 +975,11 @@ struct _xd3_stream
 
   usize_t           dec_secondid;     /* Optional secondary compressor ID. */
 
-  /* TODO: why decode breaks if this is usize_t? */
-  uint32_t          dec_codetblsz;    /* Optional code table: length. */
+  usize_t           dec_codetblsz;    /* Optional code table: length. */
   uint8_t          *dec_codetbl;      /* Optional code table: storage. */
   usize_t           dec_codetblbytes; /* Optional code table: position. */
 
-  /* TODO: why decode breaks if this is usize_t? */
-  uint32_t          dec_appheadsz;    /* Optional application header:
+  usize_t           dec_appheadsz;    /* Optional application header:
                                          size. */
   uint8_t          *dec_appheader;    /* Optional application header:
                                          storage */
@@ -944,15 +990,12 @@ struct _xd3_stream
   uint8_t           dec_cksum[4];     /* Optional checksum: storage. */
   uint32_t          dec_adler32;      /* Optional checksum: value. */
 
-  /* TODO: why decode breaks if this is usize_t? */
-  uint32_t           dec_cpylen;       /* length of copy window
+  usize_t            dec_cpylen;       /* length of copy window
                                           (VCD_SOURCE or VCD_TARGET) */
   xoff_t             dec_cpyoff;       /* offset of copy window
                                           (VCD_SOURCE or VCD_TARGET) */
-  /* TODO: why decode breaks if this is usize_t? */
-  uint32_t           dec_enclen;       /* length of delta encoding */
-  /* TODO: why decode breaks if this is usize_t? */
-  uint32_t           dec_tgtlen;       /* length of target window */
+  usize_t            dec_enclen;       /* length of delta encoding */
+  usize_t            dec_tgtlen;       /* length of target window */
 
 #if USE_UINT64
   uint64_t          dec_64part;       /* part of a decoded uint64_t */
@@ -1258,7 +1301,7 @@ const char* xd3_strerror (int ret);
    specified flags. */
 static inline
 void    xd3_init_config (xd3_config *config,
-                         int         flags)
+                         uint32_t    flags)
 {
   memset (config, 0, sizeof (*config));
   config->flags = flags;
@@ -1321,7 +1364,7 @@ usize_t xd3_encoder_srclen (xd3_stream *stream) {
 
 /* Checks for legal flag changes. */
 static inline
-void xd3_set_flags (xd3_stream *stream, int flags)
+void xd3_set_flags (xd3_stream *stream, uint32_t flags)
 {
   /* The bitwise difference should contain only XD3_FLUSH or
      XD3_SKIP_WINDOW */
@@ -1346,8 +1389,8 @@ void xd3_blksize_div (const xoff_t offset,
                       const xd3_source *source,
                       xoff_t *blkno,
                       usize_t *blkoff) {
-  *blkno = (xoff_t) (offset >> source->shiftby);
-  *blkoff = (usize_t) (offset & source->maskby);
+  *blkno = offset >> source->shiftby;
+  *blkoff = offset & source->maskby;
   XD3_ASSERT (*blkoff < source->blksize);
 }