Merge tag 'v1.51' into mm

This reverts commit 8e9144e1af.

.github/workflows/main.yml

@@ -0,0 +1,25 @@
+name: test
+on: [push]
+
+jobs:
+  build:
+    name: test
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+    - run: lscpu
+    - run: sudo apt install valgrind
+    - run: make bin/pairing.exe -j4
+    - run: valgrind bin/pairing.exe
+    - run: make clean
+#    - run: wget https://software.intel.com/content/dam/develop/external/us/en/documents/downloads/sde-external-8.63.0-2021-01-18-lin.tar.bz2
+#    - run: bzip2 -dc sde-external-8.63.0-2021-01-18-lin.tar.bz2 | tar xvf -
+#    - run: make bin/bn_test.exe DEBUG=1 -j4
+#    - run: sde-external-8.63.0-2021-01-18-lin/sde64 -hsw -- bin/bn_test.exe
+    - run: make test_ci DEBUG=1 -j4
+    - run: make clean
+    - run: make test_ci DEBUG=1 -j4 CXX=clang++
+    - run: make clean
+#    - run: make test_go
+#    - run: sudo apt install openjdk-8-jdk
+#    - run: make -C ffi/java test JAVA_INC=-I/usr/lib/jvm/java-8-openjdk-amd64/include

.gitignore

@@ -11,3 +11,6 @@ GTAGS
 *.exe
 *.swp
 .cvsignore
+build/
+external/
+Testing/

.travis.yml

@@ -1,17 +0,0 @@
-sudo: true
-dist: trusty
-language: cpp
-compiler:
-  - gcc
-  - clang
-addons:
-  apt:
-    packages:
-      - libgmp-dev
-script:
-  - make test_ci DEBUG=1 -j3
-  - make clean
-  - make test_ci CFLAGS_USER=-DMCL_DONT_USE_XBYAK -j3
-  - make clean
-  - make test_go
- 

CMakeLists.txt

@@ -1,209 +1,395 @@
-cmake_minimum_required (VERSION 3.0)
+cmake_minimum_required (VERSION 3.8)
-project(mcl CXX ASM)
+
-set(SRCS src/fp.cpp)
+project(mcl
+	VERSION 1.22
+	LANGUAGES CXX C ASM)
+
+set(CMAKE_BUILD_TYPE "Release")
 
 option(
 	MCL_MAX_BIT_SIZE
 	"max bit size for Fp"
-	0
+	384
 )
 option(
-	DOWNLOAD_SOURCE
+	MCL_STATIC_LIB
-	"download cybozulib_ext"
+	"build static library"
 	OFF
 )
+if(MSVC)
+	option(
+		MCL_DOWNLOAD_SOURCE
+		"download cybozulib_ext"
+		OFF
+		)
+endif()
 option(
-	USE_OPENSSL
+	MCL_USE_OPENSSL
 	"use openssl"
-	ON
+	OFF
 )
 option(
-	USE_GMP
+	MCL_USE_GMP
 	"use gmp"
 	ON
 )
 option(
-	USE_ASM
+	MCL_USE_ASM
 	"use asm"
 	ON
 )
 option(
-	USE_XBYAK
+	MCL_USE_XBYAK
 	"use xbyak"
 	ON
 )
 option(
-	USE_LLVM
+	MCL_USE_LLVM
 	"use base64.ll with -DCMAKE_CXX_COMPILER=clang++"
 	OFF
 )
 option(
-	ONLY_LIB
+	MCL_BUILD_SAMPLE
-	"only lib"
+	"Build mcl samples"
 	OFF
 )
+option(
+	MCL_BUILD_TESTING
+	"Build mcl tests"
+	OFF
+)
+if(MSVC)
+	option(
+		MCL_MSVC_RUNTIME_DLL
+		"use dynamic runtime /MD in msvc builds"
+		OFF
+	)
+endif()
+
+
+if(MSVC)
+	set(MCL_CYBOZULIB_EXT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../cybozulib_ext"
+		CACHE PATH "external cybozulib_ext directory")
+endif()
+
+set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if(USE_LLVM)
+if(MCL_USE_LLVM AND NOT CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+	message(WARNING "MCL_USE_LLVM will not be used: requiring clang/clang++.")
+endif()
+
+add_library(mcl SHARED src/fp.cpp)
+add_library(mcl::mcl ALIAS mcl)
+target_compile_definitions(mcl PUBLIC MCL_NO_AUTOLINK MCLBN_NO_AUTOLINK)
+target_include_directories(mcl PUBLIC
+	$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+	$<INSTALL_INTERFACE:$CMAKE_INSTALL_DIR/include>)
+set_target_properties(mcl PROPERTIES
+	POSITION_INDEPENDENT_CODE ON)
+set_target_properties(mcl PROPERTIES
+	OUTPUT_NAME mcl
+	VERSION ${mcl_VERSION}
+	SOVERSION ${mcl_VERSION_MAJOR})
+# For semantics of ABI compatibility including when you must bump SOVERSION, see:
+# https://community.kde.org/Policies/Binary_Compatibility_Issues_With_C%2B%2B#The_Do.27s_and_Don.27ts
+
+add_library(mcl_st STATIC src/fp.cpp)
+add_library(mcl::mcl_st ALIAS mcl_st)
+target_compile_definitions(mcl_st PUBLIC MCL_NO_AUTOLINK MCLBN_NO_AUTOLINK)
+target_include_directories(mcl_st PUBLIC
+	$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+	$<INSTALL_INTERFACE:$CMAKE_INSTALL_DIR/include>)
+set_target_properties(mcl_st PROPERTIES
+	OUTPUT_NAME mcl
+	POSITION_INDEPENDENT_CODE ON)
+#set_target_properties(mcl_st PROPERTIES PREFIX "lib")
+
+if(MCL_USE_LLVM AND CMAKE_CXX_COMPILER_ID MATCHES "Clang")
 	add_executable(gen src/gen.cpp)
-	add_custom_target(base64.ll
+	target_include_directories(gen PRIVATE
-		DEPENDS gen
+		${CMAKE_CURRENT_SOURCE_DIR}/include)
-		SOURCES base64.ll
+	if(MCL_USE_GMP)
-	)
+		find_package(GMP REQUIRED)
+		target_link_libraries(gen PRIVATE GMP::GMPXX GMP::GMP)
+	endif()
+
 	add_custom_command(OUTPUT base64.ll
 		COMMAND gen > base64.ll
-	)
+		DEPENDS gen
-	add_custom_target(base64.o
+		WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
-		DEPENDS base64.ll
+	add_custom_target(gen_base64.ll
-		SOURCES base64.o
+		SOURCES ${CMAKE_CURRENT_BINARY_DIR}/base64.ll)
-	)
+
 	add_custom_command(OUTPUT base64.o
 		COMMAND ${CMAKE_CXX_COMPILER} -c -o base64.o base64.ll -O3 -fPIC
-	)
+		DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/base64.ll
+		WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+	add_custom_target(gen_base64.o
+		SOURCES ${CMAKE_CURRENT_BINARY_DIR}/base64.o)
+
+	target_link_libraries(mcl PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/base64.o)
+	add_dependencies(mcl gen_base64.o)
+	target_link_libraries(mcl_st PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/base64.o)
+	add_dependencies(mcl_st gen_base64.o)
 endif()
 
 if(MSVC)
-	set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS} /MT /W4 /Oy /Ox /EHsc /GS- /Zi /DNDEBUG /DNOMINMAX")
+	if(MCL_MSVC_RUNTIME_DLL)
-	set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} /MTd /W4 /DNOMINMAX")
+		set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS} /MD /Oy /Ox /EHsc /GS- /Zi /DNDEBUG")
-	link_directories(${CMAKE_SOURCE_DIR}/../cybozulib_ext/lib)
+		set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} /MDd")
-	link_directories(${CMAKE_SOURCE_DIR}/lib)
+	else()
-else()
+		set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS} /MT /Oy /Ox /EHsc /GS- /Zi /DNDEBUG")
-	if("${CFLAGS_OPT_USER}" STREQUAL "")
+		set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} /MTd")
-		set(CFLAGS_OPT_USER "-O3 -DNDEBUG -march=native")
 	endif()
-	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -Wextra -Wformat=2 -Wcast-qual -Wcast-align -Wwrite-strings -Wfloat-equal -Wpointer-arith ${CFLAGS_OPT_USER}")
+	target_compile_definitions(mcl PUBLIC NOMINMAX)
+	target_compile_definitions(mcl_st PUBLIC NOMINMAX)
+	# set compiler flags for warnings level
+	set(MCL_COMPILE_OPTIONS /W4)
+	target_compile_options(mcl PRIVATE ${MCL_COMPILE_OPTIONS})
+	target_compile_options(mcl_st PRIVATE ${MCL_COMPILE_OPTIONS})
+else()
+	# Set compiler flags for warnings
+	set(MCL_COMPILE_OPTIONS -Wall -Wextra -Wformat=2 -Wcast-qual -Wcast-align
+		-Wwrite-strings -Wfloat-equal -Wpointer-arith -DNDEBUG -O3 -fPIC)
+
+	target_compile_options(mcl PRIVATE ${MCL_COMPILE_OPTIONS})
+	target_compile_options(mcl_st PRIVATE ${MCL_COMPILE_OPTIONS})
+	set_target_properties(mcl PROPERTIES
+		CXX_STANDARD 11
+		CXX_STANDARD_REQUIRED YES
+		CXX_EXTENSIONS NO)
+	set_target_properties(mcl_st PROPERTIES
+		CXX_STANDARD 11
+		CXX_STANDARD_REQUIRED YES
+		CXX_EXTENSIONS NO)
+	target_compile_features(mcl PUBLIC cxx_std_11)
+	target_compile_features(mcl_st PUBLIC cxx_std_11)
 
 	if(${MCL_MAX_BIT_SIZE} GREATER 0)
-		add_definitions(-DMCL_MAX_BIT_SIZE=${MCL_MAX_BIT_SIZE})
+		target_compile_definitions(mcl PUBLIC MCL_MAX_BIT_SIZE=${MCL_MAX_BIT_SIZE})
+		target_compile_definitions(mcl_st PUBLIC MCL_MAX_BIT_SIZE=${MCL_MAX_BIT_SIZE})
 	endif()
 
-	if(USE_LLVM)
+	if(MCL_USE_LLVM AND CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-		add_definitions(-DMCL_USE_LLVM=1 -DMCL_LLVM_BMI2=0)
+		target_compile_definitions(mcl PUBLIC MCL_USE_LLVM=1 MCL_LLVM_BMI2=0)
-	elseif(USE_ASM)
+		target_compile_definitions(mcl_st PUBLIC MCL_USE_LLVM=1 MCL_LLVM_BMI2=0)
+	elseif(MCL_USE_ASM)
 		if(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "aarch64")
-			add_definitions(-DMCL_USE_LLVM=1)
+			target_compile_definitions(mcl PUBLIC MCL_USE_LLVM=1)
-			set(SRCS ${SRCS} src/asm/aarch64.s)
+			target_compile_definitions(mcl_st PUBLIC MCL_USE_LLVM=1)
+			target_sources(mcl PRIVATE src/asm/aarch64.s)
+			target_sources(mcl_st PRIVATE src/asm/aarch64.s)
 			set(CPU arch64)
 		elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "^arm")
-			add_definitions(-DMCL_USE_LLVM=1)
+			target_compile_definitions(mcl PUBLIC MCL_USE_LLVM=1)
-			set(SRCS ${SRCS} src/asm/arm.s)
+			target_compile_definitions(mcl_st PUBLIC MCL_USE_LLVM=1)
+			target_sources(mcl PRIVATE src/asm/arm.s)
+			target_sources(mcl_st PRIVATE src/asm/arm.s)
 			set(CPU arm)
 		elseif(APPLE)
-			add_definitions(-DMCL_USE_LLVM=1)
+			target_compile_definitions(mcl PUBLIC MCL_USE_LLVM=1)
-			set(SRCS ${SRCS} src/asm/x86-64mac.s src/asm/x86-64mac.bmi2.s)
+			target_compile_definitions(mcl_st PUBLIC MCL_USE_LLVM=1)
+			target_sources(mcl PRIVATE src/asm/x86-64mac.s src/asm/x86-64mac.bmi2.s)
+			target_sources(mcl_st PRIVATE src/asm/x86-64mac.s src/asm/x86-64mac.bmi2.s)
 			set(CPU x86-64)
 		elseif(UNIX)
-			add_definitions(-DMCL_USE_LLVM=1)
+			target_compile_definitions(mcl PUBLIC MCL_USE_LLVM=1)
-			set(SRCS ${SRCS} src/asm/x86-64.s src/asm/x86-64.bmi2.s)
+			target_compile_definitions(mcl_st PUBLIC MCL_USE_LLVM=1)
+			target_sources(mcl PRIVATE src/asm/x86-64.s src/asm/x86-64.bmi2.s)
+			target_sources(mcl_st PRIVATE src/asm/x86-64.s src/asm/x86-64.bmi2.s)
 			set(CPU x86-64)
 		endif()
 	endif()
-	if(USE_GMP)
+	if(MCL_USE_GMP)
-		set(EXT_LIBS ${EXT_LIBS} gmp gmpxx)
+		find_package(GMP REQUIRED)
+		target_link_libraries(mcl PUBLIC GMP::GMPXX GMP::GMP)
+		target_link_libraries(mcl_st PUBLIC GMP::GMPXX GMP::GMP)
 	endif()
-	if(USE_OPENSSL)
+	if(MCL_USE_OPENSSL)
-		set(EXT_LIBS ${EXT_LIBS} crypto)
+		find_package(OpenSSL REQUIRED)
+		target_link_libraries(mcl PUBLIC OpenSSL::Crypto)
+		target_link_libraries(mcl_st PUBLIC OpenSSL::Crypto)
 	endif()
 endif()
 
-if(NOT USE_GMP)
+if(NOT MCL_USE_GMP)
-	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMCL_USE_VINT -DMCL_VINT_FIXED_BUFFER")
+	target_compile_definitions(mcl PUBLIC MCL_USE_VINT MCL_VINT_FIXED_BUFFER)
+	target_compile_definitions(mcl_st PUBLIC MCL_USE_VINT MCL_VINT_FIXED_BUFFER)
 endif()
-if(NOT USE_OPENSSL)
+if(NOT MCL_USE_OPENSSL)
-	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMCL_DONT_USE_OPENSSL")
+	target_compile_definitions(mcl PUBLIC MCL_DONT_USE_OPENSSL)
+	target_compile_definitions(mcl_st PUBLIC MCL_DONT_USE_OPENSSL)
 endif()
-if(NOT USE_XBYAK)
+if(NOT MCL_USE_XBYAK)
-	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMCL_DONT_USE_XBYAK")
+	target_compile_definitions(mcl PUBLIC MCL_DONT_USE_XBYAK)
+	target_compile_definitions(mcl_st PUBLIC MCL_DONT_USE_XBYAK)
 endif()
 
-if(DOWNLOAD_SOURCE)
+if(MCL_DOWNLOAD_SOURCE)
 	if(MSVC)
+		set(CYBOZULIB_EXT_DOWNLOAD_DIR ${CMAKE_CURRENT_SOURCE_DIR}/external/cybozulib_ext)
 		set(CYBOZULIB_EXT_TAG release20170521)
 		set(FILES config.h gmp-impl.h gmp-mparam.h gmp.h gmpxx.h longlong.h mpir.h mpirxx.h)
 		foreach(file IN ITEMS ${FILES})
-			file(DOWNLOAD https://raw.githubusercontent.com/herumi/cybozulib_ext/${CYBOZULIB_EXT_TAG}/include/${file} ${mcl_SOURCE_DIR}/include/cybozulib_ext/${file})
+			file(DOWNLOAD https://raw.githubusercontent.com/herumi/cybozulib_ext/${CYBOZULIB_EXT_TAG}/include/${file} ${CYBOZULIB_EXT_DOWNLOAD_DIR}/include/${file})
 			message("download cybozulib_ext/" ${file})
 		endforeach()
 		set(FILES aes.h applink.c asn1.h asn1_mac.h asn1t.h bio.h blowfish.h bn.h buffer.h camellia.h cast.h cmac.h cms.h comp.h conf.h conf_api.h crypto.h des.h des_old.h dh.h dsa.h dso.h dtls1.h e_os2.h ebcdic.h ec.h ecdh.h ecdsa.h engine.h err.h evp.h hmac.h idea.h krb5_asn.h kssl.h lhash.h md4.h md5.h mdc2.h modes.h obj_mac.h objects.h ocsp.h opensslconf.h opensslv.h ossl_typ.h pem.h pem2.h pkcs12.h pkcs7.h pqueue.h rand.h rc2.h rc4.h ripemd.h rsa.h safestack.h seed.h sha.h srp.h srtp.h ssl.h ssl2.h ssl23.h ssl3.h stack.h symhacks.h tls1.h ts.h txt_db.h ui.h ui_compat.h whrlpool.h x509.h x509_vfy.h x509v3.h)
 		foreach(file IN ITEMS ${FILES})
-			file(DOWNLOAD https://raw.githubusercontent.com/herumi/cybozulib_ext/${CYBOZULIB_EXT_TAG}/include/openssl/${file} ${mcl_SOURCE_DIR}/include/cybozulib_ext/openssl/${file})
+			file(DOWNLOAD
+				https://raw.githubusercontent.com/herumi/cybozulib_ext/${CYBOZULIB_EXT_TAG}/include/openssl/${file} ${CYBOZULIB_EXT_DOWNLOAD_DIR}/include/openssl/${file})
 			message("download cybozulib_ext/openssl/" ${file})
 		endforeach()
 		set(FILES mpir.lib mpirxx.lib mpirxx.pdb ssleay32.lib libeay32.lib mpir.pdb)
 		foreach(file IN ITEMS ${FILES})
-			file(DOWNLOAD https://raw.githubusercontent.com/herumi/cybozulib_ext/${CYBOZULIB_EXT_TAG}/lib/mt/14/${file} ${mcl_SOURCE_DIR}/lib/mt/14/${file})
+			file(DOWNLOAD
+			    https://raw.githubusercontent.com/herumi/cybozulib_ext/${CYBOZULIB_EXT_TAG}/lib/mt/14/${file} ${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib/mt/14/${file})
 			message("download lib/mt/14/" ${file})
 		endforeach()
-		if(MSVC)
+
-			include_directories(
+		# mpir
-				${mcl_SOURCE_DIR}/include/cybozulib_ext
+		add_library(cybozulib_ext::mpir STATIC IMPORTED)
-			)
+		set_target_properties(cybozulib_ext::mpir PROPERTIES
-		endif()
+			INTERFACE_INCLUDE_DIRECTORIES ${CYBOZULIB_EXT_DOWNLOAD_DIR}/include
+			IMPORTED_LOCATION ${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib/mt/14/mpir.lib)
+		# mpirxx
+		add_library(cybozulib_ext::mpirxx STATIC IMPORTED)
+		set_target_properties(cybozulib_ext::mpirxx PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${CYBOZULIB_EXT_DOWNLOAD_DIR}/include
+			IMPORTED_LOCATION ${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib/mt/14/mpirxx.lib)
+		# libeay32
+		add_library(cybozulib_ext::libeay32 STATIC IMPORTED)
+		set_target_properties(cybozulib_ext::libeay32 PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${CYBOZULIB_EXT_DOWNLOAD_DIR}/include
+			IMPORTED_LOCATION ${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib/mt/14/libeay32.lib)
+		# ssleay32
+		add_library(cybozulib_ext::ssleay32 STATIC IMPORTED)
+		set_target_properties(cybozulib_ext::ssleay32 PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${CYBOZULIB_EXT_DOWNLOAD_DIR}/include
+			IMPORTED_LOCATION ${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib/mt/14/ssleay32.lib)
+		# abstracted cybozulib_ext libraries
+		add_library(windows_specific INTERFACE)
+		add_library(mcl::windows_specific ALIAS windows_specific)
+		target_link_libraries(windows_specific INTERFACE
+			-LIBPATH:${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib
+			-LIBPATH:${CYBOZULIB_EXT_DOWNLOAD_DIR}/lib/mt/14
+			cybozulib_ext::mpir
+			cybozulib_ext::mpirxx
+			cybozulib_ext::libeay32
+			cybozulib_ext::ssleay32)
+
+		target_link_libraries(mcl PUBLIC mcl::windows_specific)
+		target_link_libraries(mcl_st PUBLIC mcl::windows_specific)
 	endif()
 else()
 	if(MSVC)
-		include_directories(
+		# mpir
-			${mcl_SOURCE_DIR}/../cybozulib_ext/include
+		add_library(cybozulib_ext::mpir STATIC IMPORTED)
-		)
+		set_target_properties(cybozulib_ext::mpir PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${MCL_CYBOZULIB_EXT_DIR}/include
+			IMPORTED_LOCATION ${MCL_CYBOZULIB_EXT_DIR}/lib/mt/14/mpir.lib)
+		# mpirxx
+		add_library(cybozulib_ext::mpirxx STATIC IMPORTED)
+		set_target_properties(cybozulib_ext::mpirxx PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${MCL_CYBOZULIB_EXT_DIR}/include
+			IMPORTED_LOCATION ${MCL_CYBOZULIB_EXT_DIR}/lib/mt/14/mpirxx.lib)
+		# libeay32
+		add_library(cybozulib_ext::libeay32 STATIC IMPORTED)
+		set_target_properties(cybozulib_ext::libeay32 PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${MCL_CYBOZULIB_EXT_DIR}/include
+			IMPORTED_LOCATION ${MCL_CYBOZULIB_EXT_DIR}/lib/mt/14/libeay32.lib)
+		# ssleay32
+		add_library(cybozulib_ext::ssleay32 STATIC IMPORTED)
+		set_target_properties(cybozulib_ext::ssleay32 PROPERTIES
+			INTERFACE_INCLUDE_DIRECTORIES ${MCL_CYBOZULIB_EXT_DIR}/include
+			IMPORTED_LOCATION ${MCL_CYBOZULIB_EXT_DIR}/lib/mt/14/ssleay32.lib)
+		 # abstracted cybozulib_ext libraries
+		add_library(windows_specific INTERFACE)
+		add_library(mcl::windows_specific ALIAS windows_specific)
+		target_link_libraries(windows_specific INTERFACE
+			-LIBPATH:${MCL_CYBOZULIB_EXT_DIR}/lib
+			-LIBPATH:${MCL_CYBOZULIB_EXT_DIR}/lib/mt/14
+			cybozulib_ext::mpir
+			cybozulib_ext::mpirxx
+			cybozulib_ext::libeay32
+			cybozulib_ext::ssleay32)
+
+		target_link_libraries(mcl PUBLIC mcl::windows_specific)
+		target_link_libraries(mcl_st PUBLIC mcl::windows_specific)
 	endif()
 endif()
 
-include_directories(
+# mclbnXXX
-	${mcl_SOURCE_DIR}/include
+foreach(bit IN ITEMS 256 384 384_256)
-)
+	if (MCL_STATIC_LIB)
+		add_library(mclbn${bit} STATIC src/bn_c${bit}.cpp)
+	else()
+		add_library(mclbn${bit} SHARED src/bn_c${bit}.cpp)
+	endif()
+	add_library(mcl::mclbn${bit} ALIAS mclbn${bit})
+	set_target_properties(mclbn${bit} PROPERTIES
+		CXX_STANDARD 11
+		CXX_STANDARD_REQUIRED YES
+		CXX_EXTENSIONS NO)
+	target_compile_options(mclbn${bit} PRIVATE ${MCL_COMPILE_OPTIONS})
+	target_compile_definitions(mclbn${bit}
+		PUBLIC MCL_NO_AUTOLINK MCLBN_NO_AUTOLINK)
+	target_link_libraries(mclbn${bit} PUBLIC mcl::mcl)
+	set_target_properties(mclbn${bit} PROPERTIES
+		VERSION ${mcl_VERSION}
+		SOVERSION ${mcl_VERSION_MAJOR})
+endforeach()
 
-if(USE_LLVM)
+if(MSVC)
-	add_library(mcl SHARED ${SRCS} base64.o)
+	install(TARGETS mcl mcl_st mclbn256 mclbn384 mclbn384_256 windows_specific
-	add_library(mcl_st STATIC ${SRCS} base64.o)
+		EXPORT mclTargets
-	add_dependencies(mcl base64.o)
+		LIBRARY DESTINATION lib
-	add_dependencies(mcl_st base64.o)
+		ARCHIVE DESTINATION lib
+		RUNTIME DESTINATION lib)
 else()
-	add_library(mcl SHARED ${SRCS})
+	install(TARGETS mcl mcl_st mclbn256 mclbn384 mclbn384_256
-	add_library(mcl_st STATIC ${SRCS})
+		EXPORT mclTargets
+		LIBRARY DESTINATION lib
+		ARCHIVE DESTINATION lib
+		RUNTIME DESTINATION lib)
 endif()
-target_link_libraries(mcl ${EXT_LIBS})
+install(DIRECTORY include/mcl
-target_link_libraries(mcl_st ${EXT_LIBS})
+	DESTINATION include
-set_target_properties(mcl_st PROPERTIES OUTPUT_NAME mcl)
+	FILES_MATCHING PATTERN "*.hpp"
-#set_target_properties(mcl_st PROPERTIES PREFIX "lib")
+	PATTERN "curve_type.h"
-#set_target_properties(mcl PROPERTIES OUTPUT_NAME mcl VERSION 1.0.0 SOVERSION 1)
+	PATTERN "bn.h"
-# For semantics of ABI compatibility including when you must bump SOVERSION, see:
+	PATTERN "bn_c256.h"
-# https://community.kde.org/Policies/Binary_Compatibility_Issues_With_C%2B%2B#The_Do.27s_and_Don.27ts
+	PATTERN "bn_c384_256.h"
+	PATTERN "bn_c384.h")
+install(DIRECTORY include/cybozu
+	DESTINATION include
+	FILES_MATCHING PATTERN "*.hpp")
 
-set(LIBS mcl ${EXT_LIBS})
+install(EXPORT mclTargets
-foreach(bit IN ITEMS 256 384 384_256 512)
+	FILE mclTargets.cmake
-	add_library(mclbn${bit} SHARED src/bn_c${bit}.cpp)
+	NAMESPACE mcl::
-	target_link_libraries(mclbn${bit} ${LIBS})
+	DESTINATION lib/cmake/mcl)
-	add_executable(bn_c${bit}_test test/bn_c${bit}_test.cpp)
-	target_link_libraries(bn_c${bit}_test mclbn${bit})
-endforeach()
 
-file(GLOB MCL_HEADERS include/mcl/*.hpp include/mcl/bn.h include/mcl/curve_type.h)
+# support local build-tree export to allow import from external projects
-file(GLOB CYBOZULIB_HEADERS include/cybozu/*.hpp)
+export(EXPORT mclTargets
-
+	FILE mclTargets.cmake
-install(TARGETS mcl DESTINATION lib)
+	NAMESPACE mcl::)
-install(TARGETS mcl_st DESTINATION lib)
+set(CMAKE_EXPORT_PACKAGE_REGISTRY ON)
-install(TARGETS mclbn256 DESTINATION lib)
+export(PACKAGE mcl)
-install(TARGETS mclbn384 DESTINATION lib)
+
-install(TARGETS mclbn384_256 DESTINATION lib)
+
-install(TARGETS mclbn512 DESTINATION lib)
+# Tests
-install(FILES ${MCL_HEADERS} DESTINATION include/mcl)
+if((CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME OR MCL_BUILD_TESTING)
-install(FILES include/mcl/impl/bn_c_impl.hpp DESTINATION include/mcl/impl)
+	AND BUILD_TESTING)
-install(FILES ${CYBOZULIB_HEADERS} DESTINATION include/cybozu)
+	enable_testing()
-
+	add_subdirectory(test)
-if(NOT ONLY_LIB)
+endif()
-	set(TEST_BASE fp_test ec_test fp_util_test window_method_test elgamal_test fp_tower_test gmp_test bn_test glv_test)
+
-	#set(TEST_BASE bn_test)
+if(MCL_BUILD_SAMPLE)
-	foreach(base IN ITEMS ${TEST_BASE})
+	# sample code
-		add_executable(
+	add_subdirectory(sample)
-			${base}
-			test/${base}.cpp
-		)
-		target_link_libraries(
-			${base}
-			${LIBS}
-		)
-	endforeach()
 endif()

Makefile

@@ -2,14 +2,19 @@ include common.mk
 LIB_DIR=lib
 OBJ_DIR=obj
 EXE_DIR=bin
-SRC_SRC=fp.cpp bn_c256.cpp bn_c384.cpp bn_c512.cpp she_c256.cpp
+SRC_SRC=fp.cpp bn_c256.cpp bn_c384.cpp bn_c384_256.cpp bn_c512.cpp she_c256.cpp
-TEST_SRC=fp_test.cpp ec_test.cpp fp_util_test.cpp window_method_test.cpp elgamal_test.cpp fp_tower_test.cpp gmp_test.cpp bn_test.cpp bn384_test.cpp glv_test.cpp paillier_test.cpp she_test.cpp vint_test.cpp bn512_test.cpp ecdsa_test.cpp conversion_test.cpp
+TEST_SRC=fp_test.cpp ec_test.cpp fp_util_test.cpp window_method_test.cpp elgamal_test.cpp fp_tower_test.cpp gmp_test.cpp bn_test.cpp bn384_test.cpp glv_test.cpp paillier_test.cpp she_test.cpp vint_test.cpp bn512_test.cpp conversion_test.cpp
 TEST_SRC+=bn_c256_test.cpp bn_c384_test.cpp bn_c384_256_test.cpp bn_c512_test.cpp
 TEST_SRC+=she_c256_test.cpp she_c384_test.cpp she_c384_256_test.cpp
 TEST_SRC+=aggregate_sig_test.cpp array_test.cpp
 TEST_SRC+=bls12_test.cpp
-TEST_SRC+=ecdsa_c_test.cpp
+TEST_SRC+=mapto_wb19_test.cpp
 TEST_SRC+=modp_test.cpp
+LIB_OBJ=$(OBJ_DIR)/fp.o
+ifeq ($(MCL_STATIC_CODE),1)
+  LIB_OBJ+=obj/static_code.o
+  TEST_SRC=bls12_test.cpp
+endif
 ifeq ($(CPU),x86-64)
   MCL_USE_XBYAK?=1
   TEST_SRC+=mont_fp_test.cpp sq_test.cpp
@@ -28,6 +33,13 @@ endif
 ifeq ($(MCL_USE_XBYAK),0)
   CFLAGS+=-DMCL_DONT_USE_XBYAK
 endif
+ifeq ($(MCL_USE_PROF),1)
+  CFLAGS+=-DMCL_USE_PROF
+endif
+ifeq ($(MCL_USE_PROF),2)
+  CFLAGS+=-DMCL_USE_PROF -DXBYAK_USE_VTUNE -I /opt/intel/vtune_amplifier/include/
+  LDFLAGS+=-L /opt/intel/vtune_amplifier/lib64 -ljitprofiling -ldl
+endif
 ##################################################################
 MCL_LIB=$(LIB_DIR)/libmcl.a
 MCL_SNAME=mcl
@@ -53,20 +65,19 @@ SHE384_LIB=$(LIB_DIR)/libmclshe384.a
 SHE384_SLIB=$(LIB_DIR)/lib$(SHE384_SNAME).$(LIB_SUF)
 SHE384_256_LIB=$(LIB_DIR)/libmclshe384_256.a
 SHE384_256_SLIB=$(LIB_DIR)/lib$(SHE384_256_SNAME).$(LIB_SUF)
-ECDSA_LIB=$(LIB_DIR)/libmclecdsa.a
 SHE_LIB_ALL=$(SHE256_LIB) $(SHE256_SLIB) $(SHE384_LIB) $(SHE384_SLIB) $(SHE384_256_LIB) $(SHE384_256_SLIB)
-all: $(MCL_LIB) $(MCL_SLIB) $(BN256_LIB) $(BN256_SLIB) $(BN384_LIB) $(BN384_SLIB) $(BN384_256_LIB) $(BN384_256_SLIB) $(BN512_LIB) $(BN512_SLIB) $(SHE_LIB_ALL) $(ECDSA_LIB)
+all: $(MCL_LIB) $(MCL_SLIB) $(BN256_LIB) $(BN256_SLIB) $(BN384_LIB) $(BN384_SLIB) $(BN384_256_LIB) $(BN384_256_SLIB) $(BN512_LIB) $(BN512_SLIB) $(SHE_LIB_ALL)
 
 #LLVM_VER=-3.8
 LLVM_LLC=llc$(LLVM_VER)
 LLVM_OPT=opt$(LLVM_VER)
-LLVM_OPT_VERSION=$(shell $(LLVM_OPT) --version 2>/dev/null | awk '/version/ {print $$3}')
+LLVM_OPT_VERSION=$(shell $(LLVM_OPT) --version 2>/dev/null | awk '/version/ { split($$3,a,"."); print a[1]}')
 GEN_EXE=src/gen
 GEN_EXE_OPT=-u $(BIT)
 # incompatibility between llvm 3.4 and the later version
 ifneq ($(LLVM_OPT_VERSION),)
-ifeq ($(shell expr $(LLVM_OPT_VERSION) \< 3.5.0),1)
+ifeq ($(shell expr $(LLVM_OPT_VERSION) \>= 9),1)
-  GEN_EXE_OPT+=-old
+  GEN_EXE_OPT+=-ver 0x90
 endif
 endif
 ifeq ($(OS),mac)
@@ -82,7 +93,6 @@ ifeq ($(OS),mac-m1)
   ASM_SRC=src/base64.ll
   ASM_OBJ=$(OBJ_DIR)/base64.o
 endif
-LIB_OBJ=$(OBJ_DIR)/fp.o
 BN256_OBJ=$(OBJ_DIR)/bn_c256.o
 BN384_OBJ=$(OBJ_DIR)/bn_c384.o
 BN384_256_OBJ=$(OBJ_DIR)/bn_c384_256.o
@@ -90,7 +100,6 @@ BN512_OBJ=$(OBJ_DIR)/bn_c512.o
 SHE256_OBJ=$(OBJ_DIR)/she_c256.o
 SHE384_OBJ=$(OBJ_DIR)/she_c384.o
 SHE384_256_OBJ=$(OBJ_DIR)/she_c384_256.o
-ECDSA_OBJ=$(OBJ_DIR)/ecdsa_c.o
 FUNC_LIST=src/func.list
 ifeq ($(findstring $(OS),mingw64/cygwin),)
   MCL_USE_LLVM?=1
@@ -102,7 +111,9 @@ ifeq ($(MCL_USE_LLVM),1)
   LIB_OBJ+=$(ASM_OBJ)
   # special case for intel with bmi2
   ifeq ($(INTEL),1)
-    LIB_OBJ+=$(OBJ_DIR)/$(CPU).bmi2.o
+    ifneq ($(MCL_STATIC_CODE),1)
+      LIB_OBJ+=$(OBJ_DIR)/$(CPU).bmi2.o
+    endif
   endif
 endif
 LLVM_SRC=src/base$(BIT).ll
@@ -139,15 +150,6 @@ ifneq ($(findstring $(OS),mac/mingw64),)
   SHE384_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
   SHE384_256_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
 endif
-ifneq ($(findstring $(OS),mac/mac-m1/mingw64),)
-  BN256_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-  BN384_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-  BN384_256_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-  BN512_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-  SHE256_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-  SHE384_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-  SHE384_256_SLIB_LDFLAGS+=-l$(MCL_SNAME) -L./lib
-endif
 ifeq ($(OS),mingw64)
   MCL_SLIB_LDFLAGS+=-Wl,--out-implib,$(LIB_DIR)/lib$(MCL_SNAME).a
   BN256_SLIB_LDFLAGS+=-Wl,--out-implib,$(LIB_DIR)/lib$(BN256_SNAME).a
@@ -186,9 +188,6 @@ $(SHE384_SLIB): $(SHE384_OBJ) $(MCL_LIB)
 $(SHE384_256_SLIB): $(SHE384_256_OBJ) $(MCL_LIB)
 	$(PRE)$(CXX) -o $@ $(SHE384_256_OBJ) $(MCL_LIB) -shared $(LDFLAGS) $(SHE384_256_SLIB_LDFLAGS)
 
-$(ECDSA_LIB): $(ECDSA_OBJ)
-	$(AR) $@ $(ECDSA_OBJ)
-
 $(BN256_SLIB): $(BN256_OBJ) $(MCL_SLIB)
 	$(PRE)$(CXX) -o $@ $(BN256_OBJ) -shared $(LDFLAGS) $(BN256_SLIB_LDFLAGS)
 
@@ -240,8 +239,20 @@ else
 endif
 
 $(GEN_EXE): src/gen.cpp src/llvm_gen.hpp
-	$(CXX) -o $@ $< $(CFLAGS)
+	$(CXX) -o $@ $< $(CFLAGS) -DMCL_USE_VINT -DMCL_VINT_FIXED_BUFFER
 
+src/dump_code: src/dump_code.cpp src/fp.cpp src/fp_generator.hpp
+	$(CXX) -o $@ src/dump_code.cpp src/fp.cpp -g -I include -DMCL_DUMP_JIT -DMCL_MAX_BIT_SIZE=384 -DMCL_DONT_USE_OPENSSL -DMCL_USE_VINT -DMCL_SIZEOF_UNIT=8 -DMCL_VINT_FIXED_BUFFER
+
+src/static_code.asm: src/dump_code
+	$< > $@
+
+obj/static_code.o: src/static_code.asm
+	nasm $(NASM_ELF_OPT) -o $@ $<
+
+bin/static_code_test.exe: test/static_code_test.cpp src/fp.cpp obj/static_code.o
+	$(CXX) -o $@ -O3 $^ -g -DMCL_DONT_USE_XBYAK -DMCL_STATIC_CODE -DMCL_MAX_BIT_SIZE=384 -DMCL_DONT_USE_OPENSSL -DMCL_USE_VINT -DMCL_SIZEOF_UNIT=8 -DMCL_VINT_FIXED_BUFFER -I include -Wall -Wextra
+ 
 asm: $(LLVM_SRC)
 	$(LLVM_OPT) -O3 -o - $(LLVM_SRC) | $(LLVM_LLC) -O3 $(LLVM_FLAGS) -x86-asm-syntax=intel
 
@@ -300,8 +311,8 @@ $(EXE_DIR)/bn_c384_256_test.exe: $(OBJ_DIR)/bn_c384_256_test.o $(BN384_256_LIB)
 $(EXE_DIR)/bn_c512_test.exe: $(OBJ_DIR)/bn_c512_test.o $(BN512_LIB) $(MCL_LIB)
 	$(PRE)$(CXX) $< -o $@ $(BN512_LIB) $(MCL_LIB) $(LDFLAGS)
 
-$(EXE_DIR)/pairing_c.exe: $(OBJ_DIR)/pairing_c.o $(BN256_LIB) $(MCL_LIB)
+$(EXE_DIR)/pairing_c.exe: $(OBJ_DIR)/pairing_c.o $(BN384_256_LIB) $(MCL_LIB)
-	$(PRE)$(CC) $< -o $@ $(BN256_LIB) $(MCL_LIB) $(LDFLAGS) -lstdc++
+	$(PRE)$(CC) $< -o $@ $(BN384_256_LIB) $(MCL_LIB) $(LDFLAGS) -lstdc++
 
 $(EXE_DIR)/she_c256_test.exe: $(OBJ_DIR)/she_c256_test.o $(SHE256_LIB) $(MCL_LIB)
 	$(PRE)$(CXX) $< -o $@ $(SHE256_LIB) $(MCL_LIB) $(LDFLAGS)
@@ -312,9 +323,6 @@ $(EXE_DIR)/she_c384_test.exe: $(OBJ_DIR)/she_c384_test.o $(SHE384_LIB) $(MCL_LIB
 $(EXE_DIR)/she_c384_256_test.exe: $(OBJ_DIR)/she_c384_256_test.o $(SHE384_256_LIB) $(MCL_LIB)
 	$(PRE)$(CXX) $< -o $@ $(SHE384_256_LIB) $(MCL_LIB) $(LDFLAGS)
 
-$(EXE_DIR)/ecdsa_c_test.exe: $(OBJ_DIR)/ecdsa_c_test.o $(ECDSA_LIB) $(MCL_LIB) src/ecdsa_c.cpp include/mcl/ecdsa.hpp include/mcl/ecdsa.h
-	$(PRE)$(CXX) $< -o $@ $(ECDSA_LIB) $(MCL_LIB) $(LDFLAGS)
-
 $(OBJ_DIR)/modp_test.o: test/modp_test.cpp
 	$(PRE)$(CXX) -c $< -o $@ -MMD -MP -MF $(@:.o=.d) -DMCL_USE_VINT -DMCL_MAX_BIT_SIZE=384 -DMCL_VINT_64BIT_PORTABLE -DMCL_SIZEOF_UNIT=8 -DMCL_VINT_FIXED_BUFFER -I./include -O2 $(CFLAGS_WARN)
 
@@ -326,7 +334,7 @@ sample: $(SAMPLE_EXE) $(MCL_LIB)
 
 TEST_EXE=$(addprefix $(EXE_DIR)/,$(TEST_SRC:.cpp=.exe))
 test_ci: $(TEST_EXE)
-	@sh -ec 'for i in $(TEST_EXE); do echo $$i; env LSAN_OPTIONS=verbosity=1:log_threads=1 $$i; done'
+	@sh -ec 'for i in $(TEST_EXE); do echo $$i; env LSAN_OPTIONS=verbosity=0:log_threads=1 $$i; done'
 test: $(TEST_EXE)
 	@echo test $(TEST_EXE)
 	@sh -ec 'for i in $(TEST_EXE); do $$i|grep "ctest:name"; done' > result.txt
@@ -334,7 +342,7 @@ test: $(TEST_EXE)
 
 EMCC_OPT=-I./include -I./src -Wall -Wextra
 EMCC_OPT+=-O3 -DNDEBUG -DMCLSHE_WIN_SIZE=8
-EMCC_OPT+=-s WASM=1 -s NO_EXIT_RUNTIME=1 -s MODULARIZE=1 #-s ASSERTIONS=1
+EMCC_OPT+=-s WASM=1 -s NO_EXIT_RUNTIME=1 -s NODEJS_CATCH_EXIT=0 -s NODEJS_CATCH_REJECTION=0  -s MODULARIZE=1 #-s ASSERTIONS=1
 EMCC_OPT+=-DCYBOZU_MINIMUM_EXCEPTION
 EMCC_OPT+=-s ABORTING_MALLOC=0
 SHE_C_DEP=src/fp.cpp src/she_c_impl.hpp include/mcl/she.hpp include/mcl/fp.hpp include/mcl/op.hpp include/mcl/she.h Makefile
@@ -343,37 +351,18 @@ ifeq ($(MCL_USE_LLVM),2)
   EMCC_OPT+=src/base64m.ll -DMCL_USE_LLVM
   SHE_C_DEP+=src/base64m.ll
 endif
-../she-wasm/she_c.js: src/she_c256.cpp $(SHE_C_DEP)
-	emcc -o $@ src/fp.cpp src/she_c256.cpp $(EMCC_OPT) -DMCL_MAX_BIT_SIZE=256 -s TOTAL_MEMORY=67108864 -s DISABLE_EXCEPTION_CATCHING=0
-
-../she-wasm/she_c384.js: src/she_c384.cpp $(SHE_C_DEP)
-	emcc -o $@ src/fp.cpp src/she_c384.cpp $(EMCC_OPT) -DMCL_MAX_BIT_SIZE=384 -s TOTAL_MEMORY=67108864 -s DISABLE_EXCEPTION_CATCHING=0
-
-../mcl-wasm/mcl_c.js: src/bn_c256.cpp $(MCL_C_DEP)
-	emcc -o $@ src/fp.cpp src/bn_c256.cpp $(EMCC_OPT) -DMCL_MAX_BIT_SIZE=256 -DMCL_USE_WEB_CRYPTO_API -s DISABLE_EXCEPTION_CATCHING=1 -DCYBOZU_DONT_USE_EXCEPTION -DCYBOZU_DONT_USE_STRING -fno-exceptions -MD -MP -MF obj/mcl_c.d
-
-../mcl-wasm/mcl_c512.js: src/bn_c512.cpp $(MCL_C_DEP)
-	emcc -o $@ src/fp.cpp src/bn_c512.cpp $(EMCC_OPT) -DMCL_MAX_BIT_SIZE=512 -DMCL_USE_WEB_CRYPTO_API -s DISABLE_EXCEPTION_CATCHING=1 -DCYBOZU_DONT_USE_EXCEPTION -DCYBOZU_DONT_USE_STRING -fno-exceptions
-
-../ecdsa-wasm/ecdsa_c.js: src/ecdsa_c.cpp src/fp.cpp include/mcl/ecdsa.hpp include/mcl/ecdsa.h Makefile
-	emcc -o $@ src/fp.cpp src/ecdsa_c.cpp $(EMCC_OPT) -DMCL_MAX_BIT_SIZE=256 -DMCL_USE_WEB_CRYPTO_API -s DISABLE_EXCEPTION_CATCHING=1 -DCYBOZU_DONT_USE_EXCEPTION -DCYBOZU_DONT_USE_STRING -fno-exceptions
-
-mcl-wasm:
-	$(MAKE) ../mcl-wasm/mcl_c.js
-	$(MAKE) ../mcl-wasm/mcl_c512.js
-
-she-wasm:
-	$(MAKE) ../she-wasm/she_c.js
-	$(MAKE) ../she-wasm/she_c384.js
-
-ecdsa-wasm:
-	$(MAKE) ../ecdsa-wasm/ecdsa_c.js
 
 # test
 bin/emu:
 	$(CXX) -g -o $@ src/fp.cpp src/bn_c256.cpp test/bn_c256_test.cpp -DMCL_DONT_USE_XBYAK -DMCL_DONT_USE_OPENSSL -DMCL_USE_VINT -DMCL_SIZEOF_UNIT=8 -DMCL_VINT_64BIT_PORTABLE -DMCL_VINT_FIXED_BUFFER -DMCL_MAX_BIT_SIZE=256 -I./include
 bin/pairing_c_min.exe: sample/pairing_c.c include/mcl/vint.hpp src/fp.cpp include/mcl/bn.hpp
-	$(CXX) -o $@ sample/pairing_c.c src/fp.cpp src/bn_c256.cpp -O3 -g -I./include -fno-threadsafe-statics -DMCL_DONT_USE_XBYAK -DMCL_DONT_USE_OPENSSL -DMCL_USE_VINT -DMCL_SIZEOF_UNIT=8 -DMCL_VINT_FIXED_BUFFER -DMCL_MAX_BIT_SIZE=256 -DMCL_VINT_64BIT_PORTABLE -DCYBOZU_DONT_USE_STRING -DCYBOZU_DONT_USE_EXCEPTION -DNDEBUG # -DMCL_DONT_USE_CSPRNG
+	$(CXX) -std=c++03 -O3 -g -fno-threadsafe-statics -fno-exceptions -fno-rtti -o $@ sample/pairing_c.c src/fp.cpp src/bn_c384_256.cpp -I./include -DXBYAK_NO_EXCEPTION -DMCL_DONT_USE_OPENSSL -DMCL_USE_VINT -DMCL_SIZEOF_UNIT=8 -DMCL_VINT_FIXED_BUFFER -DMCL_MAX_BIT_SIZE=384 -DMCL_VINT_64BIT_PORTABLE -DCYBOZU_DONT_USE_STRING -DCYBOZU_DONT_USE_EXCEPTION -DNDEBUG # -DMCL_DONT_USE_CSPRNG
+
+bin/llvm_test64.exe: test/llvm_test.cpp src/base64.ll
+	clang++$(LLVM_VER) -o $@ -Ofast -DNDEBUG -Wall -Wextra -I ./include test/llvm_test.cpp src/base64.ll
+
+bin/llvm_test32.exe: test/llvm_test.cpp src/base32.ll
+	clang++$(LLVM_VER) -o $@ -Ofast -DNDEBUG -Wall -Wextra -I ./include test/llvm_test.cpp src/base32.ll -m32
 
 make_tbl:
 	$(MAKE) ../bls/src/qcoeff-bn254.hpp
@@ -389,7 +378,7 @@ update_cybozulib:
 	cp -a $(addprefix ../cybozulib/,$(wildcard include/cybozu/*.hpp)) include/cybozu/
 
 clean:
-	$(RM) $(LIB_DIR)/*.a $(LIB_DIR)/*.$(LIB_SUF) $(OBJ_DIR)/*.o $(OBJ_DIR)/*.obj $(OBJ_DIR)/*.d $(EXE_DIR)/*.exe $(GEN_EXE) $(ASM_OBJ) $(LIB_OBJ) $(BN256_OBJ) $(BN384_OBJ) $(BN512_OBJ) $(FUNC_LIST) src/*.ll lib/*.a
+	$(RM) $(LIB_DIR)/*.a $(LIB_DIR)/*.$(LIB_SUF) $(OBJ_DIR)/*.o $(OBJ_DIR)/*.obj $(OBJ_DIR)/*.d $(EXE_DIR)/*.exe $(GEN_EXE) $(ASM_OBJ) $(LIB_OBJ) $(BN256_OBJ) $(BN384_OBJ) $(BN512_OBJ) $(FUNC_LIST) lib/*.a src/static_code.asm src/dump_code
 
 ALL_SRC=$(SRC_SRC) $(TEST_SRC) $(SAMPLE_SRC)
 DEPEND_FILE=$(addprefix $(OBJ_DIR)/, $(addsuffix .d,$(basename $(ALL_SRC))))
@@ -398,12 +387,12 @@ DEPEND_FILE=$(addprefix $(OBJ_DIR)/, $(addsuffix .d,$(basename $(ALL_SRC))))
 PREFIX?=/usr/local
 install: lib/libmcl.a lib/libmcl.$(LIB_SUF)
 	$(MKDIR) $(PREFIX)/include/mcl
-	cp -a include/mcl/ $(PREFIX)/include/
+	cp -a include/mcl $(PREFIX)/include/
-	cp -a include/cybozu/ $(PREFIX)/include/
+	cp -a include/cybozu $(PREFIX)/include/
 	$(MKDIR) $(PREFIX)/lib
 	cp -a lib/libmcl.a lib/libmcl.$(LIB_SUF) $(PREFIX)/lib/
 
-.PHONY: test mcl-wasm she-wasm bin/emu
+.PHONY: test she-wasm bin/emu
 
 # don't remove these files automatically
 .SECONDARY: $(addprefix $(OBJ_DIR)/, $(ALL_SRC:.cpp=.o))

api.md

@@ -0,0 +1,548 @@
+# C API
+
+## News
+
+APIs for old ethreum hash functions are removed.
+`mclBn_setMapToMode` supports only `MCL_MAP_TO_MODE_ETH2`.
+
+## Minimum sample
+
+[sample/pairing_c.c](sample/pairing_c.c) is a sample of how to use BLS12-381 pairing.
+
+```
+cd mcl
+make -j4
+make bin/pairing_c.exe && bin/pairing_c.exe
+```
+
+## Header and libraries
+
+To use BLS12-381, include `mcl/bn_c384_256.h` and link
+- libmclbn384_256.{a,so}
+- libmcl.{a,so} ; core library
+
+`384_256` means the max bit size of `Fp` is 384 and that size of `Fr` is 256.
+
+## Notation
+
+The elliptic equation of a curve E is `E: y^2 = x^3 + b`.
+
+- `Fp` ; a finite field of a prime order `p`, where curves is defined over.
+- `Fr` ; a finite field of a prime order `r`.
+- `Fp2` ; the field extension over Fp with degree 2. Fp[i] / (i^2 + 1).
+- `Fp6` ; the field extension over Fp2 with degree 3. Fp2[v] / (v^3 - Xi) where Xi = i + 1.
+- `Fp12` ; the field extension over Fp6 with degree 2. Fp6[w] / (w^2 - v).
+- `G1` ; the cyclic subgroup of E(Fp).
+- `G2` ; the cyclic subgroup of the inverse image of E'(Fp^2) under a twisting isomorphism from E' to E.
+- `GT` ; the cyclie subgroup of Fp12.
+  - `G1`, `G2` and `GT` have the order `r`.
+
+The pairing e: G1 x G2 -> GT is the optimal ate pairing.
+
+mcl treats `G1` and `G2` as an additive group and `GT` as a multiplicative group.
+
+- `mclSize` ; `unsigned int` if WebAssembly else `size_t`
+
+### Curve Parameter
+r = |G1| = |G2| = |GT|
+
+curveType   | b| r and p |
+------------|--|------------------|
+BN254       | 2|r = 0x2523648240000001ba344d8000000007ff9f800000000010a10000000000000d <br> p = 0x2523648240000001ba344d80000000086121000000000013a700000000000013 |
+BLS12-381   | 4|r = 0x73eda753299d7d483339d80809a1d80553bda402fffe5bfeffffffff00000001 <br> p = 0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffaaab |
+BN381       | 2|r = 0x240026400f3d82b2e42de125b00158405b710818ac000007e0042f008e3e00000000001080046200000000000000000d <br> p = 0x240026400f3d82b2e42de125b00158405b710818ac00000840046200950400000000001380052e000000000000000013 |
+
+## Structures
+
+### `mclBnFp`
+This is a struct of `Fp`. The value is stored as Montgomery representation.
+
+### `mclBnFr`
+This is a struct of `Fr`. The value is stored as Montgomery representation.
+
+### `mclBnFp2`
+This is a struct of `Fp2` which has a member `mclBnFp d[2]`.
+
+An element `x` of `Fp2` is represented as `x = d[0] + d[1] i` where `i^2 = -1`.
+
+### `mclBnG1`
+This is a struct of `G1` which has three members `x`, `y`, `z` of type `mclBnFp`.
+
+An element `P` of `G1` is represented as `P = [x:y:z]` of a Jacobi coordinate.
+
+### `mclBnG2`
+This is a struct of `G2` which has three members `x`, `y`, `z` of type `mclBnFp2`.
+
+An element `Q` of `G2` is represented as `Q = [x:y:z]` of a Jacobi coordinate.
+
+### `mclBnGT`
+
+This is a struct of `GT` which has a member `mclBnFp d[12]`.
+
+### sizeof
+
+library           |MCLBN_FR_UNIT_SIZE|MCLBN_FP_UNIT_SIZE|sizeof Fr|sizeof Fp|
+------------------|------------------|------------------|---------|---------|
+libmclbn256.a     |          4       |         4        |   32    |   32    |
+libmclbn384_256.a |          4       |         6        |   32    |   48    |
+libmclbn384.a     |          6       |         6        |   48    |   48    |
+
+## Thread safety
+All functions except for initialization and changing global setting are thread-safe.
+
+## Initialization
+
+Initialize mcl library. Call this function at first before calling the other functions.
+
+```
+int mclBn_init(int curve, int compiledTimeVar);
+```
+
+- `curve` ; specify the curve type
+  - MCL_BN254 ; BN254 (a little faster if including `mcl/bn_c256.h` and linking `libmclbn256.{a,so}`)
+  - MCL_BN_SNARK1 ; the same parameter used in libsnark
+  - MCL_BLS12_381 ; BLS12-381
+  - MCL_BN381_1 ; BN381 (include `mcl/bn_c384.h` and link `libmclbn384.{a,so}`)
+- `compiledTimeVar` ; set `MCLBN_COMPILED_TIME_VAR`, which macro is used to make sure that
+the values are the same when the library is built and used.
+- return 0 if success.
+- This is not thread safe.
+
+## Global setting
+
+```
+int mclBn_setMapToMode(int mode);
+```
+The map-to-G2 function if `mode = MCL_MAP_TO_MODE_HASH_TO_CURVE`.
+
+### Control to verify that a point of the elliptic curve has the order `r`.
+
+This function affects `setStr()` and `deserialize()` for G1/G2.
+```
+void mclBn_verifyOrderG1(int doVerify);
+void mclBn_verifyOrderG2(int doVerify);
+```
+- verify if `doVerify` is 1 or does not. The default parameter is 0 because the cost of verification is not small.
+- Set `doVerify = 1` if considering subgroup attack is necessary.
+- This is not thread safe.
+
+## Setter / Getter
+
+### Clear
+Set `x` is zero.
+```
+void mclBnFr_clear(mclBnFr *x);
+void mclBnFp_clear(mclBnFp *x);
+void mclBnFp2_clear(mclBnFp2 *x);
+void mclBnG1_clear(mclBnG1 *x);
+void mclBnG2_clear(mclBnG2 *x);
+void mclBnGT_clear(mclBnGT *x);
+```
+
+### Set `x` to `y`.
+```
+void mclBnFp_setInt(mclBnFp *y, mclInt x);
+void mclBnFr_setInt(mclBnFr *y, mclInt x);
+void mclBnGT_setInt(mclBnGT *y, mclInt x);
+```
+
+### Set `buf[0..bufSize-1]` to `x` with masking according to the following way.
+```
+int mclBnFp_setLittleEndian(mclBnFp *x, const void *buf, mclSize bufSize);
+int mclBnFr_setLittleEndian(mclBnFr *x, const void *buf, mclSize bufSize);
+```
+1. set x = buf[0..bufSize-1] as little endian
+2. x &= (1 << bitLen(r)) - 1
+3. if (x >= r) x &= (1 << (bitLen(r) - 1)) - 1
+
+- always return 0
+
+### Set (`buf[0..bufSize-1]` mod `p` or `r`) to `x`.
+```
+int mclBnFp_setLittleEndianMod(mclBnFp *x, const void *buf, mclSize bufSize);
+int mclBnFr_setLittleEndianMod(mclBnFr *x, const void *buf, mclSize bufSize);
+```
+- return 0 if bufSize <= (sizeof(*x) * 8 * 2) else -1
+
+### Get little endian byte sequence `buf` corresponding to `x`
+```
+mclSize mclBnFr_getLittleEndian(void *buf, mclSize maxBufSize, const mclBnFr *x);
+mclSize mclBnFp_getLittleEndian(void *buf, mclSize maxBufSize, const mclBnFp *x);
+```
+- write `x` to `buf` as little endian
+- return the written size if sucess else 0
+- NOTE: `buf[0] = 0` and return 1 if `x` is zero.
+
+### Serialization
+### Serialize
+```
+mclSize mclBnFr_serialize(void *buf, mclSize maxBufSize, const mclBnFr *x);
+mclSize mclBnG1_serialize(void *buf, mclSize maxBufSize, const mclBnG1 *x);
+mclSize mclBnG2_serialize(void *buf, mclSize maxBufSize, const mclBnG2 *x);
+mclSize mclBnGT_serialize(void *buf, mclSize maxBufSize, const mclBnGT *x);
+mclSize mclBnFp_serialize(void *buf, mclSize maxBufSize, const mclBnFp *x);
+mclSize mclBnFp2_serialize(void *buf, mclSize maxBufSize, const mclBnFp2 *x);
+```
+- serialize `x` into `buf[0..maxBufSize-1]`
+- return written byte size if success else 0
+
+### Serialization format
+- `Fp`(resp.  `Fr`) ; a little endian byte sequence with a fixed size
+  - the size is the return value of `mclBn_getFpByteSize()` (resp. `mclBn_getFpByteSize()`).
+- `G1` ; a compressed fixed size
+  - the size is equal to `mclBn_getG1ByteSize()` (=`mclBn_getFpByteSize()`).
+- `G2` ; a compressed fixed size
+  - the size is equal to `mclBn_getG1ByteSize() * 2`.
+
+pseudo-code to serialize of `P` of `G1` (resp. `G2`)
+```
+size = mclBn_getG1ByteSize() # resp. mclBn_getG1ByteSize() * 2
+if P is zero:
+  return [0] * size
+else:
+  P = P.normalize()
+  s = P.x.serialize()
+  # x in Fp2 is odd <=> x.a is odd
+  if P.y is odd: # resp. P.y.d[0] is odd
+    s[byte-length(s) - 1] |= 0x80
+  return s
+```
+
+### Ethereum serialization mode for BLS12-381
+```
+void mclBn_setETHserialization(int ETHserialization);
+```
+- serialize according to [serialization of BLS12-381](https://github.com/zkcrypto/pairing/blob/master/src/bls12_381/README.md#serialization) if BLS12-381 is used and `ETHserialization = 1` (default 0).
+
+### Deserialize
+```
+mclSize mclBnFr_deserialize(mclBnFr *x, const void *buf, mclSize bufSize);
+mclSize mclBnG1_deserialize(mclBnG1 *x, const void *buf, mclSize bufSize);
+mclSize mclBnG2_deserialize(mclBnG2 *x, const void *buf, mclSize bufSize);
+mclSize mclBnGT_deserialize(mclBnGT *x, const void *buf, mclSize bufSize);
+mclSize mclBnFp_deserialize(mclBnFp *x, const void *buf, mclSize bufSize);
+mclSize mclBnFp2_deserialize(mclBnFp2 *x, const void *buf, mclSize bufSize);
+```
+- deserialize `x` from `buf[0..bufSize-1]`
+- return read size if success else 0
+
+## String conversion
+### Get string
+```
+mclSize mclBnFr_getStr(char *buf, mclSize maxBufSize, const mclBnFr *x, int ioMode);
+mclSize mclBnG1_getStr(char *buf, mclSize maxBufSize, const mclBnG1 *x, int ioMode);
+mclSize mclBnG2_getStr(char *buf, mclSize maxBufSize, const mclBnG2 *x, int ioMode);
+mclSize mclBnGT_getStr(char *buf, mclSize maxBufSize, const mclBnGT *x, int ioMode);
+mclSize mclBnFp_getStr(char *buf, mclSize maxBufSize, const mclBnFp *x, int ioMode);
+```
+- write `x` to `buf` according to `ioMode`
+- `ioMode`
+  - 10 ; decimal number
+  - 16 ; hexadecimal number
+  - `MCLBN_IO_EC_PROJ` ; output as Jacobi coordinate
+- return `strlen(buf)` if success else 0.
+
+The meaning of the output of `G1`.
+- `0` ; infinity
+- `1 <x> <y>` ; affine coordinate
+- `4 <x> <y> <z>` ; Jacobi coordinate
+- the element `<x>` of `G2` outputs `d[0] d[1]`.
+
+### Set string
+```
+int mclBnFr_setStr(mclBnFr *x, const char *buf, mclSize bufSize, int ioMode);
+int mclBnG1_setStr(mclBnG1 *x, const char *buf, mclSize bufSize, int ioMode);
+int mclBnG2_setStr(mclBnG2 *x, const char *buf, mclSize bufSize, int ioMode);
+int mclBnGT_setStr(mclBnGT *x, const char *buf, mclSize bufSize, int ioMode);
+int mclBnFp_setStr(mclBnFp *x, const char *buf, mclSize bufSize, int ioMode);
+```
+- set `buf[0..bufSize-1]` to `x` accoring to `ioMode`
+- return 0 if success else -1
+
+If you want to use the same generators of BLS12-381 with [zkcrypto](https://github.com/zkcrypto/pairing/tree/master/src/bls12_381#g2) then,
+
+```
+mclBnG1 P;
+mclBnG1_setStr(&P, "1 3685416753713387016781088315183077757961620795782546409894578378688607592378376318836054947676345821548104185464507 1339506544944476473020471379941921221584933875938349620426543736416511423956333506472724655353366534992391756441569", 10);
+
+mclBnG2 Q;
+mclBnG2_setStr(&Q, "1 352701069587466618187139116011060144890029952792775240219908644239793785735715026873347600343865175952761926303160 3059144344244213709971259814753781636986470325476647558659373206291635324768958432433509563104347017837885763365758 1985150602287291935568054521177171638300868978215655730859378665066344726373823718423869104263333984641494340347905 927553665492332455747201965776037880757740193453592970025027978793976877002675564980949289727957565575433344219582");
+```
+
+
+## Set random value
+Set `x` by cryptographically secure pseudo random number generator.
+```
+int mclBnFr_setByCSPRNG(mclBnFr *x);
+int mclBnFp_setByCSPRNG(mclBnFp *x);
+```
+
+### Change random generator function
+```
+void mclBn_setRandFunc(
+  void *self,
+  unsigned int (*readFunc)(void *self, void *buf, unsigned int bufSize)
+);
+```
+- `self` ; user-defined pointer
+- `readFunc` ; user-defined function, which writes random `bufSize` bytes to `buf` and returns `bufSize` if success else returns 0.
+  - `readFunc` must be thread-safe.
+- Set the default random function if `self == 0` and `readFunc == 0`.
+- This is not thread safe.
+
+## Arithmetic operations
+### neg / inv / sqr / add / sub / mul / div of `Fr`, `Fp`, `Fp2`, `GT`.
+```
+void mclBnFr_neg(mclBnFr *y, const mclBnFr *x);
+void mclBnFr_inv(mclBnFr *y, const mclBnFr *x);
+void mclBnFr_sqr(mclBnFr *y, const mclBnFr *x);
+void mclBnFr_add(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
+void mclBnFr_sub(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
+void mclBnFr_mul(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
+void mclBnFr_div(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
+
+void mclBnFp_neg(mclBnFp *y, const mclBnFp *x);
+void mclBnFp_inv(mclBnFp *y, const mclBnFp *x);
+void mclBnFp_sqr(mclBnFp *y, const mclBnFp *x);
+void mclBnFp_add(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+void mclBnFp_sub(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+void mclBnFp_mul(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+void mclBnFp_div(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+
+void mclBnFp2_neg(mclBnFp2 *y, const mclBnFp2 *x);
+void mclBnFp2_inv(mclBnFp2 *y, const mclBnFp2 *x);
+void mclBnFp2_sqr(mclBnFp2 *y, const mclBnFp2 *x);
+void mclBnFp2_add(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+void mclBnFp2_sub(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+void mclBnFp2_mul(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+void mclBnFp2_div(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+
+void mclBnGT_inv(mclBnGT *y, const mclBnGT *x); // y = a - bw for x = a + bw where Fp12 = Fp6[w]
+void mclBnGT_sqr(mclBnGT *y, const mclBnGT *x);
+void mclBnGT_mul(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
+void mclBnGT_div(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
+```
+- use `mclBnGT_invGeneric` for an element in Fp12 - GT.
+
+- NOTE: The following functions do NOT return a GT element because GT is multiplicative group.
+
+```
+void mclBnGT_neg(mclBnGT *y, const mclBnGT *x);
+void mclBnGT_add(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
+void mclBnGT_sub(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
+```
+
+### Square root of `x`.
+```
+int mclBnFr_squareRoot(mclBnFr *y, const mclBnFr *x);
+int mclBnFp_squareRoot(mclBnFp *y, const mclBnFp *x);
+int mclBnFp2_squareRoot(mclBnFp2 *y, const mclBnFp2 *x);
+```
+- `y` is one of square root of `x` if `y` exists.
+- return 0 if success else -1
+
+### add / sub / dbl / neg for `G1` and `G2`.
+```
+void mclBnG1_neg(mclBnG1 *y, const mclBnG1 *x);
+void mclBnG1_dbl(mclBnG1 *y, const mclBnG1 *x);
+void mclBnG1_add(mclBnG1 *z, const mclBnG1 *x, const mclBnG1 *y);
+void mclBnG1_sub(mclBnG1 *z, const mclBnG1 *x, const mclBnG1 *y);
+
+void mclBnG2_neg(mclBnG2 *y, const mclBnG2 *x);
+void mclBnG2_dbl(mclBnG2 *y, const mclBnG2 *x);
+void mclBnG2_add(mclBnG2 *z, const mclBnG2 *x, const mclBnG2 *y);
+void mclBnG2_sub(mclBnG2 *z, const mclBnG2 *x, const mclBnG2 *y);
+```
+
+### Convert a point from Jacobi coordinate to affine.
+```
+void mclBnG1_normalize(mclBnG1 *y, const mclBnG1 *x);
+void mclBnG2_normalize(mclBnG2 *y, const mclBnG2 *x);
+```
+- convert `[x:y:z]` to `[x:y:1]` if `z != 0` else `[*:*:0]`
+
+### scalar multiplication
+```
+void mclBnG1_mul(mclBnG1 *z, const mclBnG1 *x, const mclBnFr *y);
+void mclBnG2_mul(mclBnG2 *z, const mclBnG2 *x, const mclBnFr *y);
+void mclBnGT_pow(mclBnGT *z, const mclBnGT *x, const mclBnFr *y);
+```
+- z = x * y for G1 / G2
+- z = pow(x, y) for GT
+
+- use `mclBnGT_powGeneric` for an element in Fp12 - GT.
+
+### multi scalar multiplication
+```
+void mclBnG1_mulVec(mclBnG1 *z, const mclBnG1 *x, const mclBnFr *y, mclSize n);
+void mclBnG2_mulVec(mclBnG2 *z, const mclBnG2 *x, const mclBnFr *y, mclSize n);
+void mclBnGT_powVec(mclBnGT *z, const mclBnGT *x, const mclBnFr *y, mclSize n);
+```
+- z = sum_{i=0}^{n-1} mul(x[i], y[i]) for G1 / G2.
+- z = prod_{i=0}^{n-1} pow(x[i], y[i]) for GT.
+
+## hash and mapTo functions
+### Set hash of `buf[0..bufSize-1]` to `x`
+```
+int mclBnFr_setHashOf(mclBnFr *x, const void *buf, mclSize bufSize);
+int mclBnFp_setHashOf(mclBnFp *x, const void *buf, mclSize bufSize);
+```
+- always return 0
+- use SHA-256 if sizeof(*x) <= 256 else SHA-512
+- set accoring to the same way as `setLittleEndian`
+  - support the other wasy if you want in the future
+
+### map `x` to G1 / G2.
+```
+int mclBnFp_mapToG1(mclBnG1 *y, const mclBnFp *x);
+int mclBnFp2_mapToG2(mclBnG2 *y, const mclBnFp2 *x);
+```
+- See `struct MapTo` in `mcl/bn.hpp` for the detail of the algorithm.
+- return 0 if success else -1
+
+### hash and map to G1 / G2.
+```
+int mclBnG1_hashAndMapTo(mclBnG1 *x, const void *buf, mclSize bufSize);
+int mclBnG2_hashAndMapTo(mclBnG2 *x, const void *buf, mclSize bufSize);
+```
+- Combine `setHashOf` and `mapTo` functions
+
+## Pairing operations
+The pairing function `e(P, Q)` is consist of two parts:
+  - `MillerLoop(P, Q)`
+  - `finalExp(x)`
+
+`finalExp` satisfies the following properties:
+  - `e(P, Q) = finalExp(MillerLoop(P, Q))`
+  - `e(P1, Q1) e(P2, Q2) = finalExp(MillerLoop(P1, Q1) MillerLoop(P2, Q2))`
+
+### pairing
+```
+void mclBn_pairing(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y);
+```
+### millerLoop
+```
+void mclBn_millerLoop(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y);
+```
+### finalExp
+```
+void mclBn_finalExp(mclBnGT *y, const mclBnGT *x);
+```
+
+## Variants of MillerLoop
+### multi pairing
+```
+void mclBn_millerLoopVec(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y, mclSize n);
+```
+- This function is for multi-pairing
+  - computes prod_{i=0}^{n-1} MillerLoop(x[i], y[i])
+  - prod_{i=0}^{n-1} e(x[i], y[i]) = finalExp(prod_{i=0}^{n-1} MillerLoop(x[i], y[i]))
+
+### pairing for a fixed point of G2
+```
+int mclBn_getUint64NumToPrecompute(void);
+void mclBn_precomputeG2(uint64_t *Qbuf, const mclBnG2 *Q);
+void mclBn_precomputedMillerLoop(mclBnGT *f, const mclBnG1 *P, const uint64_t *Qbuf);
+```
+These functions is the same computation of `pairing(P, Q);` as the followings:
+```
+uint64_t *Qbuf = (uint64_t*)malloc(mclBn_getUint64NumToPrecompute() * sizeof(uint64_t));
+mclBn_precomputeG2(Qbuf, Q); // precomputing of Q
+mclBn_precomputedMillerLoop(f, P, Qbuf); // pairing of any P of G1 and the fixed Q
+free(p);
+```
+
+```
+void mclBn_precomputedMillerLoop2(
+  mclBnGT *f,
+  const mclBnG1 *P1, const uint64_t *Q1buf,
+  const mclBnG1 *P2, const uint64_t *Q2buf
+);
+```
+- compute `MillerLoop(P1, Q1buf) * MillerLoop(P2, Q2buf)`
+
+
+```
+void mclBn_precomputedMillerLoop2mixed(
+  mclBnGT *f,
+  const mclBnG1 *P1, const mclBnG2 *Q1,
+  const mclBnG1 *P2, const uint64_t *Q2buf
+);
+```
+- compute `MillerLoop(P1, Q2) * MillerLoop(P2, Q2buf)`
+
+## Check value
+### Check validness
+```
+int mclBnFr_isValid(const mclBnFr *x);
+int mclBnFp_isValid(const mclBnFp *x);
+int mclBnG1_isValid(const mclBnG1 *x);
+int mclBnG2_isValid(const mclBnG2 *x);
+```
+- return 1 if true else 0
+
+### Check the order of a point
+```
+int mclBnG1_isValidOrder(const mclBnG1 *x);
+int mclBnG2_isValidOrder(const mclBnG2 *x);
+```
+- Check whether the order of `x` is valid or not
+- return 1 if true else 0
+- This function always cheks according to `mclBn_verifyOrderG1` and `mclBn_verifyOrderG2`.
+
+### Is equal / zero / one / isOdd
+```
+int mclBnFr_isEqual(const mclBnFr *x, const mclBnFr *y);
+int mclBnFr_isZero(const mclBnFr *x);
+int mclBnFr_isOne(const mclBnFr *x);
+int mclBnFr_isOdd(const mclBnFr *x);
+
+int mclBnFp_isEqual(const mclBnFp *x, const mclBnFp *y);
+int mclBnFp_isZero(const mclBnFp *x);
+int mclBnFp_isOne(const mclBnFp *x);
+int mclBnFp_isOdd(const mclBnFp *x);
+
+int mclBnFp2_isEqual(const mclBnFp2 *x, const mclBnFp2 *y);
+int mclBnFp2_isZero(const mclBnFp2 *x);
+int mclBnFp2_isOne(const mclBnFp2 *x);
+
+int mclBnG1_isEqual(const mclBnG1 *x, const mclBnG1 *y);
+int mclBnG1_isZero(const mclBnG1 *x);
+
+int mclBnG2_isEqual(const mclBnG2 *x, const mclBnG2 *y);
+int mclBnG2_isZero(const mclBnG2 *x);
+
+int mclBnGT_isEqual(const mclBnGT *x, const mclBnGT *y);
+int mclBnGT_isZero(const mclBnGT *x);
+int mclBnGT_isOne(const mclBnGT *x);
+```
+- return 1 if true else 0
+
+### isNegative
+```
+int mclBnFr_isNegative(const mclBnFr *x);
+int mclBnFp_isNegative(const mclBnFr *x);
+```
+return 1 if x >= half where half = (r + 1) / 2 (resp. (p + 1) / 2).
+
+## Lagrange interpolation
+
+```
+int mclBn_FrLagrangeInterpolation(mclBnFr *out, const mclBnFr *xVec, const mclBnFr *yVec, mclSize k);
+int mclBn_G1LagrangeInterpolation(mclBnG1 *out, const mclBnFr *xVec, const mclBnG1 *yVec, mclSize k);
+int mclBn_G2LagrangeInterpolation(mclBnG2 *out, const mclBnFr *xVec, const mclBnG2 *yVec, mclSize k);
+```
+- Lagrange interpolation
+- recover out = y(0) from {(xVec[i], yVec[i])} for {i=0..k-1}
+- return 0 if success else -1
+  - satisfy that xVec[i] != 0, xVec[i] != xVec[j] for i != j
+
+```
+int mclBn_FrEvaluatePolynomial(mclBnFr *out, const mclBnFr *cVec, mclSize cSize, const mclBnFr *x);
+int mclBn_G1EvaluatePolynomial(mclBnG1 *out, const mclBnG1 *cVec, mclSize cSize, const mclBnFr *x);
+int mclBn_G2EvaluatePolynomial(mclBnG2 *out, const mclBnG2 *cVec, mclSize cSize, const mclBnFr *x);
+```
+- Evaluate polynomial
+- out = f(x) = c[0] + c[1] * x + ... + c[cSize - 1] * x^{cSize - 1}
+- return 0 if success else -1
+  - satisfy cSize >= 1

build.sh

@@ -0,0 +1,67 @@
+#!/bin/bash
+
+SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
+BUILD_DIR=${1:-build}
+
+
+windows_build()
+{
+  if [ -d "${SCRIPT_DIR}/../cybozulib_ext"  ]; then
+      DOWNLOAD_CYBOZULIB_EXT="OFF"
+      CYBOZULIB_EXT_OPTION="-DMCL_CYBOZULIB_EXT_DIR:PATH=${SCRIPT_DIR}/../cybozulib_ext"
+  else
+      DOWNLOAD_CYBOZULIB_EXT="ON"
+      CYBOZULIB_EXT_OPTION=""
+  fi
+
+  cmake -E remove_directory ${BUILD_DIR}
+  cmake -E make_directory ${BUILD_DIR}
+  cmake -H${SCRIPT_DIR} -B${BUILD_DIR} -A x64 \
+    -DBUILD_TESTING=ON \
+    -DMCL_BUILD_SAMPLE=ON \
+    -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}/install \
+    -DMCL_DOWNLOAD_SOURCE=${DOWNLOAD_CYBOZULIB_EXT} ${CYBOZULIB_EXT_OPTION}
+  cmake --build ${BUILD_DIR} --clean-first --config Release --parallel
+}
+
+linux_build()
+{
+  cmake -E remove_directory ${BUILD_DIR}
+  cmake -E make_directory ${BUILD_DIR}
+  cmake -H${SCRIPT_DIR} -B${BUILD_DIR} -DCMAKE_BUILD_TYPE=Release \
+    -DBUILD_TESTING=ON \
+    -DMCL_BUILD_SAMPLE=ON \
+    -DMCL_USE_LLVM=ON \
+    -DMCL_USE_OPENSSL=ON \
+    -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}/install
+  cmake --build ${BUILD_DIR} --clean-first -- -j
+}
+
+osx_build()
+{
+  OPENSSL_ROOT_DIR="/usr/local/opt/openssl"
+
+  cmake -E remove_directory ${BUILD_DIR}
+  cmake -E make_directory ${BUILD_DIR}
+  cmake -H${SCRIPT_DIR} -B${BUILD_DIR} -DCMAKE_BUILD_TYPE=Release \
+    -DBUILD_TESTING=ON \
+    -DMCL_BUILD_SAMPLE=ON \
+    -DMCL_USE_LLVM=ON \
+    -DMCL_USE_OPENSSL=ON \
+    -DOPENSSL_ROOT_DIR="${OPENSSL_ROOT_DIR}" \
+    -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}/install
+  cmake --build ${BUILD_DIR} --clean-first -- -j
+}
+
+os=`uname -s`
+case "${os}" in
+  CYGWIN*|MINGW32*|MSYS*|MINGW*)
+    windows_build
+    ;;
+  Darwin*)
+    osx_build
+    ;;
+  *)
+    linux_build
+    ;;
+esac

cmake/FindGMP.cmake

@@ -0,0 +1,88 @@
+# FindGMP.cmake
+#
+# Finds the GNU Multiple Precision Arithmetic Library (GMP)
+# See http://gmplib.org/
+#
+# This will define the following variables::
+#
+#    GMP_FOUND
+#    GMP_VERSION
+#    GMP_DEFINITIONS
+#    GMP_INCLUDE_DIR
+#    GMP_LIBRARY
+#    GMP_GMPXX_DEFINITIONS
+#    GMP_GMPXX_INCLUDE_DIR
+#    GMP_GMPXX_LIBRARY
+#
+# and the following imported targets::
+#
+#     GMP::GMP
+#     GMP::GMPXX
+
+find_package(PkgConfig QUIET)
+pkg_check_modules(PC_GMP QUIET gmp gmpxx)
+
+set(GMP_VERSION ${PC_GMP_gmp_VERSION})
+
+find_library(GMP_LIBRARY
+	NAMES gmp libgmp
+	HINTS
+		${PC_GMP_gmp_LIBDIR}
+		${PC_GMP_gmp_LIBRARY_DIRS})
+
+find_path(GMP_INCLUDE_DIR
+	NAMES gmp.h
+	HINTS
+		${PC_GMP_gmp_INCLUDEDIR}
+		${PC_GMP_gmp_INCLUDE_DIRS})
+
+find_library(GMP_GMPXX_LIBRARY
+	NAMES gmpxx libgmpxx
+	HINTS
+		${PC_GMP_gmpxx_LIBDIR}
+		${PC_GMP_gmpxx_LIBRARY_DIRS})
+
+find_path(GMP_GMPXX_INCLUDE_DIR
+	NAMES gmpxx.h
+	HINTS
+		${PC_GMP_gmpxx_INCLUDEDIR}
+		${PC_GMP_gmpxx_INCLUDE_DIRS})
+
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(GMP
+	REQUIRED_VARS
+		GMP_INCLUDE_DIR
+		GMP_LIBRARY
+	GMP_GMPXX_INCLUDE_DIR
+	GMP_GMPXX_LIBRARY
+	VERSION_VAR GMP_VERSION)
+
+if(GMP_FOUND)
+	set(GMP_LIBRARIES ${GMP_LIBRARY})
+	set(GMP_INCLUDE_DIRS ${GMP_INCLUDE_DIR})
+	set(GMP_DEFINITIONS ${PC_GMP_gmp_CFLAGS_OTHER})
+	set(GMP_GMPXX_LIBRARIES ${GMP_GMPXX_LIBRARY})
+	set(GMP_GMPXX_INCLUDE_DIRS ${GMP_GMPXX_INCLUDE_DIR})
+	set(GMP_GMPXX_DEFINITIONS ${PC_GMP_gmpxx_CFLAGS_OTHER})
+
+	if(NOT TARGET GMP::GMP)
+		add_library(GMP::GMP UNKNOWN IMPORTED)
+		set_target_properties(GMP::GMP PROPERTIES
+			INTERFACE_COMPILE_OPTIONS "${PC_GMP_gmp_CFLAGS_OTHER}"
+			INTERFACE_INCLUDE_DIRECTORIES "${GMP_INCLUDE_DIR}"
+			IMPORTED_LOCATION "${GMP_LIBRARY}")
+	endif()
+
+	if(NOT TARGET GMP::GMPXX)
+		add_library(GMP::GMPXX UNKNOWN IMPORTED)
+		set_target_properties(GMP::GMPXX PROPERTIES
+			INTERFACE_COMPILE_OPTIONS "${PC_GMP_gmpxx_CFLAGS_OTHER}"
+			INTERFACE_INCLUDE_DIRECTORIES "${GMP_GMPXX_INCLUDE_DIR}"
+			INTERFACE_LINK_LIBRARIES GMP::GMP
+			IMPORTED_LOCATION "${GMP_GMPXX_LIBRARY}")
+	endif()
+endif()
+
+mark_as_advanced(GMP_FOUND GMP_INCLUDE_DIR GMP_LIBRARY)
+mark_as_advanced(GMP_GMPXX_INCLUDE_DIR GMP_GMPXX_LIBRARY)

common.mk

@@ -1,12 +1,14 @@
 GCC_VER=$(shell $(PRE)$(CC) -dumpversion)
 UNAME_S=$(shell uname -s)
 ARCH?=$(shell uname -m)
+NASM_ELF_OPT=-felf64
 ifeq ($(UNAME_S),Linux)
   OS=Linux
 endif
 ifeq ($(findstring MINGW64,$(UNAME_S)),MINGW64)
   OS=mingw64
   CFLAGS+=-D__USE_MINGW_ANSI_STDIO=1
+  NASM_ELF_OPT=-fwin64
 endif
 ifeq ($(findstring CYGWIN,$(UNAME_S)),CYGWIN)
   OS=cygwin
@@ -14,20 +16,27 @@ endif
 ifeq ($(UNAME_S),Darwin)
   ifeq ($(ARCH),x86_64)
     OS=mac
-    GMP_DIR?=/usr/local/opt/gmp
   else
     OS=mac-m1
-    GMP_DIR?=/opt/homebrew/opt/gmp
   endif
   LIB_SUF=dylib
-  OPENSSL_DIR?=/opt/homebrew/opt/openssl@1.1
+  NASM_ELF_OPT=-fmacho64
-  CFLAGS+=-I$(OPENSSL_DIR)/include
-  LDFLAGS+=-L$(OPENSSL_DIR)/lib
-  CFLAGS+=-I$(GMP_DIR)/include
-  LDFLAGS+=-L$(GMP_DIR)/lib
 else
   LIB_SUF=so
 endif
+ifeq ($(UNAME_S),OpenBSD)
+  OS=openbsd
+  CXX?=clang++
+  CFLAGS+=-I/usr/local/include
+  LDFLAGS+=-L/usr/local/lib
+endif
+ifeq ($(UNAME_S),FreeBSD)
+  OS=freebsd
+  CXX?=clang++
+  CFLAGS+=-I/usr/local/include
+  LDFLAGS+=-L/usr/local/lib
+endif
+
 ifneq ($(findstring $(ARCH),x86_64/amd64),)
   CPU=x86-64
   INTEL=1
@@ -46,11 +55,12 @@ ifeq ($(ARCH),x86)
   BIT_OPT=-m32
   #LOW_ASM_SRC=src/asm/low_x86.asm
 endif
-ifeq ($(ARCH),armv7l)
+ifneq ($(findstring $(ARCH),armv7l/armv6l),)
   CPU=arm
   BIT=32
   #LOW_ASM_SRC=src/asm/low_arm.s
 endif
+#ifeq ($(ARCH),aarch64)
 ifneq ($(findstring $(ARCH),aarch64/arm64),)
   CPU=aarch64
   BIT=64
@@ -58,6 +68,10 @@ endif
 ifeq ($(findstring $(OS),mac/mac-m1/mingw64/openbsd),)
   LDFLAGS+=-lrt
 endif
+ifeq ($(ARCH),s390x)
+  CPU=systemz
+  BIT=64
+endif
 
 CP=cp -f
 AR=ar r
@@ -70,12 +84,12 @@ ifeq ($(DEBUG),1)
     LDFLAGS+=-fsanitize=address
   endif
 else
-  CFLAGS_OPT+=-fomit-frame-pointer -DNDEBUG
+  CFLAGS_OPT+=-fomit-frame-pointer -DNDEBUG -fno-stack-protector
   ifeq ($(CXX),clang++)
     CFLAGS_OPT+=-O3
   else
     ifeq ($(shell expr $(GCC_VER) \> 4.6.0),1)
-      CFLAGS_OPT+=-Ofast
+      CFLAGS_OPT+=-O3
     else
       CFLAGS_OPT+=-O3
     endif
@@ -88,7 +102,7 @@ else
     CFLAGS_OPT+=$(MARCH)
   endif
 endif
-CFLAGS_WARN=-Wall -Wextra -Wformat=2 -Wcast-qual -Wcast-align -Wwrite-strings -Wfloat-equal -Wpointer-arith
+CFLAGS_WARN=-Wall -Wextra -Wformat=2 -Wcast-qual -Wcast-align -Wwrite-strings -Wfloat-equal -Wpointer-arith -Wundef
 CFLAGS+=-g3
 INC_OPT=-I include -I test
 CFLAGS+=$(CFLAGS_WARN) $(BIT_OPT) $(INC_OPT)
@@ -99,7 +113,10 @@ CFLAGS+=$(CFLAGS_OPT_USER)
 endif
 CFLAGS+=$(CFLAGS_USER)
 MCL_USE_GMP?=1
-MCL_USE_OPENSSL?=1
+ifneq ($(OS),mac/mac-m1,)
+  MCL_USE_GMP=0
+endif
+MCL_USE_OPENSSL?=0
 ifeq ($(MCL_USE_GMP),0)
   CFLAGS+=-DMCL_USE_VINT
 endif
@@ -111,9 +128,24 @@ ifeq ($(MCL_USE_OPENSSL),0)
 endif
 ifeq ($(MCL_USE_GMP),1)
   GMP_LIB=-lgmp -lgmpxx
+  ifeq ($(UNAME_S),Darwin)
+    GMP_DIR?=/usr/local/opt/gmp
+    CFLAGS+=-I$(GMP_DIR)/include
+    LDFLAGS+=-L$(GMP_DIR)/lib
+  endif
 endif
 ifeq ($(MCL_USE_OPENSSL),1)
   OPENSSL_LIB=-L${OPENSSL_DIR}/lib -lcrypto
+  ifeq ($(UNAME_S),Darwin)
+    OPENSSL_DIR?=/usr/local/opt/openssl
+    CFLAGS+=-I$(OPENSSL_DIR)/include
+    LDFLAGS+=-L$(OPENSSL_DIR)/lib
+  endif
+endif
+ifeq ($(MCL_STATIC_CODE),1)
+  MCL_USE_XBYAK=0
+  MCL_MAX_BIT_SIZE=384
+  CFLAGS+=-DMCL_STATIC_CODE
 endif
 LDFLAGS+=$(GMP_LIB) $(OPENSSL_LIB) $(BIT_OPT) $(LDFLAGS_USER)
 

ffi/cs/bn256.cs

@@ -4,125 +4,126 @@ using System.Runtime.InteropServices;
 
 namespace mcl {
 	public class BN256 {
-		[DllImport("mclBn256.dll")]
+		public const string dllName = "mclBn256";
+		[DllImport(dllName)]
 		public static extern int mclBn_init(int curve, int compiledTimeVar);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_clear(ref Fr x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_setInt(ref Fr y, int x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_setStr(ref Fr x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_isValid(ref Fr x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_isEqual(ref Fr x, ref Fr y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_isZero(ref Fr x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_isOne(ref Fr x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_setByCSPRNG(ref Fr x);
 
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_setHashOf(ref Fr x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnFr_getStr([Out]StringBuilder buf, long maxBufSize, ref Fr x, int ioMode);
 
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_neg(ref Fr y, ref Fr x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_inv(ref Fr y, ref Fr x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_add(ref Fr z, ref Fr x, ref Fr y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_sub(ref Fr z, ref Fr x, ref Fr y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_mul(ref Fr z, ref Fr x, ref Fr y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnFr_div(ref Fr z, ref Fr x, ref Fr y);
 
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG1_clear(ref G1 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG1_setStr(ref G1 x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG1_isValid(ref G1 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG1_isEqual(ref G1 x, ref G1 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG1_isZero(ref G1 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG1_hashAndMapTo(ref G1 x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern long mclBnG1_getStr([Out]StringBuilder buf, long maxBufSize, ref G1 x, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG1_neg(ref G1 y, ref G1 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG1_dbl(ref G1 y, ref G1 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG1_add(ref G1 z, ref G1 x, ref G1 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG1_sub(ref G1 z, ref G1 x, ref G1 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG1_mul(ref G1 z, ref G1 x, ref Fr y);
 
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG2_clear(ref G2 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG2_setStr(ref G2 x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG2_isValid(ref G2 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG2_isEqual(ref G2 x, ref G2 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG2_isZero(ref G2 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnG2_hashAndMapTo(ref G2 x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern long mclBnG2_getStr([Out]StringBuilder buf, long maxBufSize, ref G2 x, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG2_neg(ref G2 y, ref G2 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG2_dbl(ref G2 y, ref G2 x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG2_add(ref G2 z, ref G2 x, ref G2 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG2_sub(ref G2 z, ref G2 x, ref G2 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnG2_mul(ref G2 z, ref G2 x, ref Fr y);
 
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_clear(ref GT x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnGT_setStr(ref GT x, [In][MarshalAs(UnmanagedType.LPStr)] string buf, long bufSize, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnGT_isEqual(ref GT x, ref GT y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnGT_isZero(ref GT x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern int mclBnGT_isOne(ref GT x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern long mclBnGT_getStr([Out]StringBuilder buf, long maxBufSize, ref GT x, int ioMode);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_neg(ref GT y, ref GT x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_inv(ref GT y, ref GT x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_add(ref GT z, ref GT x, ref GT y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_sub(ref GT z, ref GT x, ref GT y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_mul(ref GT z, ref GT x, ref GT y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_div(ref GT z, ref GT x, ref GT y);
 
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBnGT_pow(ref GT z, ref GT x, ref Fr y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBn_pairing(ref GT z, ref G1 x, ref G2 y);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBn_finalExp(ref GT y, ref GT x);
-		[DllImport("mclBn256.dll")]
+		[DllImport(dllName)]
 		public static extern void mclBn_millerLoop(ref GT z, ref G1 x, ref G2 y);
 
 		public static void init()

ffi/cs/mcl.sln

@@ -0,0 +1,31 @@
+
+Microsoft Visual Studio Solution File, Format Version 12.00
+# Visual Studio Version 16
+VisualStudioVersion = 16.0.30309.148
+MinimumVisualStudioVersion = 15.0.26124.0
+Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "test", "test\test.csproj", "{31ABF32C-3DAC-47EF-8B99-B531F88B6FDC}"
+EndProject
+Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "mcl", "mcl\mcl.csproj", "{73288FB5-7173-4AE9-86C6-F76DF219C37B}"
+EndProject
+Global
+	GlobalSection(SolutionConfigurationPlatforms) = preSolution
+		Debug|x64 = Debug|x64
+		Release|x64 = Release|x64
+	EndGlobalSection
+	GlobalSection(ProjectConfigurationPlatforms) = postSolution
+		{31ABF32C-3DAC-47EF-8B99-B531F88B6FDC}.Debug|x64.ActiveCfg = Debug|Any CPU
+		{31ABF32C-3DAC-47EF-8B99-B531F88B6FDC}.Debug|x64.Build.0 = Debug|Any CPU
+		{31ABF32C-3DAC-47EF-8B99-B531F88B6FDC}.Release|x64.ActiveCfg = Release|Any CPU
+		{31ABF32C-3DAC-47EF-8B99-B531F88B6FDC}.Release|x64.Build.0 = Release|Any CPU
+		{73288FB5-7173-4AE9-86C6-F76DF219C37B}.Debug|x64.ActiveCfg = Debug|Any CPU
+		{73288FB5-7173-4AE9-86C6-F76DF219C37B}.Debug|x64.Build.0 = Debug|Any CPU
+		{73288FB5-7173-4AE9-86C6-F76DF219C37B}.Release|x64.ActiveCfg = Release|Any CPU
+		{73288FB5-7173-4AE9-86C6-F76DF219C37B}.Release|x64.Build.0 = Release|Any CPU
+	EndGlobalSection
+	GlobalSection(SolutionProperties) = preSolution
+		HideSolutionNode = FALSE
+	EndGlobalSection
+	GlobalSection(ExtensibilityGlobals) = postSolution
+		SolutionGuid = {ACBC75BC-5FB9-48DF-A78D-7C6BAD21E647}
+	EndGlobalSection
+EndGlobal

ffi/cs/mcl/mcl.csproj

@@ -0,0 +1,13 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <TargetFramework>netstandard2.1</TargetFramework>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <Compile Remove="CVS\**" />
+    <EmbeddedResource Remove="CVS\**" />
+    <None Remove="CVS\**" />
+  </ItemGroup>
+
+</Project>

ffi/cs/readme.md

@@ -0,0 +1,24 @@
+# C# binding of mcl library
+
+# How to build `bin/mclbn384_256.dll`.
+
+```
+git clone https://github.com/herumi/mcl
+cd mcl
+mklib dll
+```
+
+Open `mcl/ffi/cs/mcl.sln` and Set the directory of `mcl/bin` to `workingDirectory` at `Debug` of test project.
+
+# Remark
+- `bn256.cs` is an old code. It will be removed in the future.
+- `mcl/mcl.cs` is a new version. It support `BN254`, `BN_SNARK` and `BLS12_381` curve, which requires `mclbn384_256.dll`.
+
+# `ETHmode` with `BLS12_381`
+
+If you need the map-to-G1/G2 function defined in [Hashing to Elliptic Curves](https://www.ietf.org/id/draft-irtf-cfrg-hash-to-curve-09.html),
+then initialize this library as the followings:
+```
+MCL.Init(BLS12_381);
+MCL.ETHmode();
+```

ffi/cs/test/test.cs

@@ -0,0 +1,408 @@
+using System;
+
+namespace mcl {
+    using static MCL;
+    class MCLTest {
+        static int err = 0;
+        static void assert(string msg, bool b)
+        {
+            if (b) return;
+            Console.WriteLine("ERR {0}", msg);
+            err++;
+        }
+        static void Main()
+        {
+            err = 0;
+            try {
+                Console.WriteLine("BN254");
+                TestCurve(BN254);
+                Console.WriteLine("BN_SNARK");
+                TestCurve(BN_SNARK);
+                Console.WriteLine("BLS12_381");
+                TestCurve(BLS12_381);
+                Console.WriteLine("BLS12_381 eth");
+                ETHmode();
+                TestETH();
+                if (err == 0) {
+                    Console.WriteLine("all tests succeed");
+                } else {
+                    Console.WriteLine("err={0}", err);
+                }
+            } catch (Exception e) {
+                Console.WriteLine("ERR={0}", e);
+            }
+        }
+
+        static void TestCurve(int curveType)
+        {
+            Init(curveType);
+            TestFr();
+            TestFp();
+            TestG1();
+            TestG2();
+            TestPairing();
+            TestSS();
+        }
+        static void TestFr()
+        {
+            Console.WriteLine("TestFr");
+            Fr x = new Fr();
+            assert("x.isZero", x.IsZero());
+            x.Clear();
+            assert("0", x.GetStr(10) == "0");
+            assert("0.IzZero", x.IsZero());
+            assert("!0.IzOne", !x.IsOne());
+            x.SetInt(1);
+            assert("1", x.GetStr(10) == "1");
+            assert("!1.IzZero", !x.IsZero());
+            assert("1.IzOne", x.IsOne());
+            x.SetInt(3);
+            assert("3", x.GetStr(10) == "3");
+            assert("!3.IzZero", !x.IsZero());
+            assert("!3.IzOne", !x.IsOne());
+            x.SetInt(-5);
+            x = -x;
+            assert("5", x.GetStr(10) == "5");
+            x.SetInt(4);
+            x = x * x;
+            assert("16", x.GetStr(10) == "16");
+            assert("10", x.GetStr(16) == "10");
+            Fr y;
+            y = x;
+            assert("x == y", x.Equals(y));
+            x.SetInt(123);
+            assert("123", x.GetStr(10) == "123");
+            assert("7b", x.GetStr(16) == "7b");
+            assert("y != x", !x.Equals(y));
+            Console.WriteLine("exception test");
+            try {
+                x.SetStr("1234567891234x", 10);
+                Console.WriteLine("x = {0}", x);
+            } catch (Exception e) {
+                Console.WriteLine("OK ; expected exception: {0}", e);
+            }
+            x.SetStr("1234567891234", 10);
+            assert("1234567891234", x.GetStr(10) == "1234567891234");
+            {
+                byte[] buf = x.Serialize();
+                y.Deserialize(buf);
+                assert("x == y", x.Equals(y));
+            }
+        }
+        static void TestFp()
+        {
+            Console.WriteLine("TestFp");
+            Fp x = new Fp();
+            assert("x.isZero", x.IsZero());
+            x.Clear();
+            assert("0", x.GetStr(10) == "0");
+            assert("0.IzZero", x.IsZero());
+            assert("!0.IzOne", !x.IsOne());
+            x.SetInt(1);
+            assert("1", x.GetStr(10) == "1");
+            assert("!1.IzZero", !x.IsZero());
+            assert("1.IzOne", x.IsOne());
+            x.SetInt(3);
+            assert("3", x.GetStr(10) == "3");
+            assert("!3.IzZero", !x.IsZero());
+            assert("!3.IzOne", !x.IsOne());
+            x.SetInt(-5);
+            x = -x;
+            assert("5", x.GetStr(10) == "5");
+            x.SetInt(4);
+            x = x * x;
+            assert("16", x.GetStr(10) == "16");
+            assert("10", x.GetStr(16) == "10");
+            Fp y;
+            y = x;
+            assert("x == y", x.Equals(y));
+            x.SetInt(123);
+            assert("123", x.GetStr(10) == "123");
+            assert("7b", x.GetStr(16) == "7b");
+            assert("y != x", !x.Equals(y));
+            Console.WriteLine("exception test");
+            try {
+                x.SetStr("1234567891234x", 10);
+                Console.WriteLine("x = {0}", x);
+            } catch (Exception e) {
+                Console.WriteLine("OK ; expected exception: {0}", e);
+            }
+            x.SetStr("1234567891234", 10);
+            assert("1234567891234", x.GetStr(10) == "1234567891234");
+            {
+                byte[] buf = x.Serialize();
+                y.Deserialize(buf);
+                assert("x == y", x.Equals(y));
+            }
+        }
+        static void TestG1()
+        {
+            Console.WriteLine("TestG1");
+            G1 P = new G1();
+            assert("P.isZero", P.IsZero());
+            P.Clear();
+            assert("P.IsValid", P.IsValid());
+            assert("P.IsZero", P.IsZero());
+            P.HashAndMapTo("abc");
+            assert("P.IsValid", P.IsValid());
+            assert("!P.IsZero", !P.IsZero());
+            G1 Q = new G1();
+            Q = P;
+            assert("P == Q", Q.Equals(P));
+            Q.Neg(P);
+            Q.Add(Q, P);
+            assert("P = Q", Q.IsZero());
+            Q.Dbl(P);
+            G1 R = new G1();
+            R.Add(P, P);
+            assert("Q == R", Q.Equals(R));
+            Fr x = new Fr();
+            x.SetInt(3);
+            R.Add(R, P);
+            Q.Mul(P, x);
+            assert("Q == R", Q.Equals(R));
+            {
+                byte[] buf = P.Serialize();
+                Q.Clear();
+                Q.Deserialize(buf);
+                assert("P == Q", P.Equals(Q));
+            }
+            {
+                const int n = 5;
+                G1[] xVec = new G1[n];
+                Fr[] yVec = new Fr[n];
+                P.Clear();
+                for (int i = 0; i < n; i++) {
+                    xVec[i].HashAndMapTo(i.ToString());
+                    yVec[i].SetByCSPRNG();
+                    Q.Mul(xVec[i], yVec[i]);
+                    P.Add(P, Q);
+                }
+                MulVec(ref Q, xVec, yVec);
+                assert("mulVecG1", P.Equals(Q));
+            }
+            G1 W = G1.Zero();
+            assert("W.IsZero", W.IsZero());
+        }
+        static void TestG2()
+        {
+            Console.WriteLine("TestG2");
+            G2 P = new G2();
+            assert("P.isZero", P.IsZero());
+            P.Clear();
+            assert("P is valid", P.IsValid());
+            assert("P is zero", P.IsZero());
+            P.HashAndMapTo("abc");
+            assert("P is valid", P.IsValid());
+            assert("P is not zero", !P.IsZero());
+            G2 Q = new G2();
+            Q = P;
+            assert("P == Q", Q.Equals(P));
+            Q.Neg(P);
+            Q.Add(Q, P);
+            assert("Q is zero", Q.IsZero());
+            Q.Dbl(P);
+            G2 R = new G2();
+            R.Add(P, P);
+            assert("Q == R", Q.Equals(R));
+            Fr x = new Fr();
+            x.SetInt(3);
+            R.Add(R, P);
+            Q.Mul(P, x);
+            assert("Q == R", Q.Equals(R));
+            {
+                byte[] buf = P.Serialize();
+                Q.Clear();
+                Q.Deserialize(buf);
+                assert("P == Q", P.Equals(Q));
+            }
+            {
+                const int n = 5;
+                G2[] xVec = new G2[n];
+                Fr[] yVec = new Fr[n];
+                P.Clear();
+                for (int i = 0; i < n; i++) {
+                    xVec[i].HashAndMapTo(i.ToString());
+                    yVec[i].SetByCSPRNG();
+                    Q.Mul(xVec[i], yVec[i]);
+                    P.Add(P, Q);
+                }
+                MulVec(ref Q, xVec, yVec);
+                assert("mulVecG2", P.Equals(Q));
+            }
+            G2 W = G2.Zero();
+            assert("W.IsZero", W.IsZero());
+        }
+        static void TestPairing()
+        {
+            Console.WriteLine("TestG2");
+            G1 P = new G1();
+            P.HashAndMapTo("123");
+            G2 Q = new G2();
+            Q.HashAndMapTo("1");
+            Fr a = new Fr();
+            Fr b = new Fr();
+            a.SetStr("12345678912345673453", 10);
+            b.SetStr("230498230982394243424", 10);
+            G1 aP = new G1();
+            G2 bQ = new G2();
+            aP.Mul(P, a);
+            bQ.Mul(Q, b);
+            GT e1 = new GT();
+            GT e2 = new GT();
+            GT e3 = new GT();
+            e1.Pairing(P, Q);
+            e2.Pairing(aP, Q);
+            e3.Pow(e1, a);
+            assert("e2.Equals(e3)", e2.Equals(e3));
+            e2.Pairing(P, bQ);
+            e3.Pow(e1, b);
+            assert("e2.Equals(e3)", e2.Equals(e3));
+        }
+        static void TestETH_mapToG1()
+        {
+            var tbl = new[] {
+                new {
+                    msg = "asdf",
+                    x = "a72df17570d0eb81260042edbea415ad49bdb94a1bc1ce9d1bf147d0d48268170764bb513a3b994d662e1faba137106",
+                    y = "122b77eca1ed58795b7cd456576362f4f7bd7a572a29334b4817898a42414d31e9c0267f2dc481a4daf8bcf4a460322",
+                },
+           };
+            G1 P = new G1();
+            Fp x = new Fp();
+            Fp y = new Fp();
+            foreach (var v in tbl) {
+                P.HashAndMapTo(v.msg);
+                x.SetStr(v.x, 16);
+                y.SetStr(v.y, 16);
+                Normalize(ref P, P);
+                Console.WriteLine("x={0}", P.x.GetStr(16));
+                Console.WriteLine("y={0}", P.y.GetStr(16));
+                assert("P.x", P.x.Equals(x));
+                assert("P.y", P.y.Equals(y));
+            }
+        }
+        static void TestETH()
+        {
+            TestETH_mapToG1();
+        }
+        static void TestSS_Fr()
+        {
+            const int n = 5;
+            const int k = 3; // can't change because the following loop
+            Fr[] cVec = new Fr[k];
+            // init polynomial coefficient
+            for (int i = 0; i < k; i++) {
+                cVec[i].SetByCSPRNG();
+            }
+
+            Fr[] xVec = new Fr[n];
+            Fr[] yVec = new Fr[n];
+            // share cVec[0] with yVec[0], ..., yVec[n-1]
+            for (int i = 0; i < n; i++) {
+                xVec[i].SetHashOf(i.ToString());
+                MCL.Share(ref yVec[i], cVec, xVec[i]);
+            }
+            // recover cVec[0] from xVecSubset and yVecSubset
+            Fr[] xVecSubset = new Fr[k];
+            Fr[] yVecSubset = new Fr[k];
+            for (int i0 = 0; i0 < n; i0++) {
+                xVecSubset[0] = xVec[i0];
+                yVecSubset[0] = yVec[i0];
+                for (int i1 = i0 + 1; i1 < n; i1++) {
+                    xVecSubset[1] = xVec[i1];
+                    yVecSubset[1] = yVec[i1];
+                    for (int i2 = i1 + 1; i2 < n; i2++) {
+                        xVecSubset[2] = xVec[i2];
+                        yVecSubset[2] = yVec[i2];
+                        Fr s = new Fr();
+                        MCL.Recover(ref s, xVecSubset, yVecSubset);
+                        assert("Recover", s.Equals(cVec[0]));
+                    }
+                }
+            }
+        }
+        static void TestSS_G1()
+        {
+            const int n = 5;
+            const int k = 3; // can't change because the following loop
+            G1[] cVec = new G1[k];
+            // init polynomial coefficient
+            for (int i = 0; i < k; i++) {
+                Fr x = new Fr();
+                x.SetByCSPRNG();
+                cVec[i].SetHashOf(x.GetStr(16));
+            }
+
+            Fr[] xVec = new Fr[n];
+            G1[] yVec = new G1[n];
+            // share cVec[0] with yVec[0], ..., yVec[n-1]
+            for (int i = 0; i < n; i++) {
+                xVec[i].SetHashOf(i.ToString());
+                MCL.Share(ref yVec[i], cVec, xVec[i]);
+            }
+            // recover cVec[0] from xVecSubset and yVecSubset
+            Fr[] xVecSubset = new Fr[k];
+            G1[] yVecSubset = new G1[k];
+            for (int i0 = 0; i0 < n; i0++) {
+                xVecSubset[0] = xVec[i0];
+                yVecSubset[0] = yVec[i0];
+                for (int i1 = i0 + 1; i1 < n; i1++) {
+                    xVecSubset[1] = xVec[i1];
+                    yVecSubset[1] = yVec[i1];
+                    for (int i2 = i1 + 1; i2 < n; i2++) {
+                        xVecSubset[2] = xVec[i2];
+                        yVecSubset[2] = yVec[i2];
+                        G1 s = new G1();
+                        MCL.Recover(ref s, xVecSubset, yVecSubset);
+                        assert("Recover", s.Equals(cVec[0]));
+                    }
+                }
+            }
+        }
+        static void TestSS_G2()
+        {
+            const int n = 5;
+            const int k = 3; // can't change because the following loop
+            G2[] cVec = new G2[k];
+            // init polynomial coefficient
+            for (int i = 0; i < k; i++) {
+                Fr x = new Fr();
+                x.SetByCSPRNG();
+                cVec[i].SetHashOf(x.GetStr(16));
+            }
+
+            Fr[] xVec = new Fr[n];
+            G2[] yVec = new G2[n];
+            // share cVec[0] with yVec[0], ..., yVec[n-1]
+            for (int i = 0; i < n; i++) {
+                xVec[i].SetHashOf(i.ToString());
+                MCL.Share(ref yVec[i], cVec, xVec[i]);
+            }
+            // recover cVec[0] from xVecSubset and yVecSubset
+            Fr[] xVecSubset = new Fr[k];
+            G2[] yVecSubset = new G2[k];
+            for (int i0 = 0; i0 < n; i0++) {
+                xVecSubset[0] = xVec[i0];
+                yVecSubset[0] = yVec[i0];
+                for (int i1 = i0 + 1; i1 < n; i1++) {
+                    xVecSubset[1] = xVec[i1];
+                    yVecSubset[1] = yVec[i1];
+                    for (int i2 = i1 + 1; i2 < n; i2++) {
+                        xVecSubset[2] = xVec[i2];
+                        yVecSubset[2] = yVec[i2];
+                        G2 s = new G2();
+                        MCL.Recover(ref s, xVecSubset, yVecSubset);
+                        assert("Recover", s.Equals(cVec[0]));
+                    }
+                }
+            }
+        }
+        static void TestSS()
+        {
+            TestSS_Fr();
+            TestSS_G1();
+            TestSS_G2();
+        }
+    }
+}

ffi/cs/test/test.csproj

@@ -0,0 +1,28 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <ItemGroup>
+    <Compile Remove="CVS\**" />
+    <EmbeddedResource Remove="CVS\**" />
+    <None Remove="CVS\**" />
+  </ItemGroup>
+
+  <ItemGroup>
+    <ProjectReference Include="..\mcl\mcl.csproj" />
+  </ItemGroup>
+
+  <PropertyGroup>
+    <OutputType>Exe</OutputType>
+    <TargetFramework>netcoreapp3.1</TargetFramework>
+  </PropertyGroup>
+
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|AnyCPU'">
+    <PlatformTarget>x64</PlatformTarget>
+    <OutputPath></OutputPath>
+  </PropertyGroup>
+
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|AnyCPU'">
+    <PlatformTarget>x64</PlatformTarget>
+    <OutputPath></OutputPath>
+  </PropertyGroup>
+
+</Project>

ffi/go/mcl/init.go

@@ -11,6 +11,7 @@ package mcl
 */
 import "C"
 import "fmt"
+
 // Init --
 // call this function before calling all the other operations
 // this function is not thread safe
@@ -21,5 +22,3 @@ func Init(curve int) error {
 	}
 	return nil
 }
-
-

ffi/go/mcl/mcl.go

@@ -19,18 +19,27 @@ const CurveFp382_1 = C.mclBn_CurveFp382_1
 // CurveFp382_2 -- 382 bit curve 2
 const CurveFp382_2 = C.mclBn_CurveFp382_2
 
-// BLS12_381
+// BLS12_381 --
 const BLS12_381 = C.MCL_BLS12_381
 
-// IoSerializeHexStr
+// IoSerializeHexStr --
 const IoSerializeHexStr = C.MCLBN_IO_SERIALIZE_HEX_STR
 
-// GetFrUnitSize() --
+// IO_EC_AFFINE --
+const IO_EC_AFFINE = C.MCLBN_IO_EC_AFFINE
+
+// IO_EC_PROJ --
+const IO_EC_PROJ = C.MCLBN_IO_EC_PROJ
+
+// IRTF -- for SetMapToMode
+const IRTF = 5 /* MCL_MAP_TO_MODE_HASH_TO_CURVE_07 */
+
+// GetFrUnitSize --
 func GetFrUnitSize() int {
 	return int(C.MCLBN_FR_UNIT_SIZE)
 }
 
-// GetFpUnitSize() --
+// GetFpUnitSize --
 // same as GetMaxOpUnitSize()
 func GetFpUnitSize() int {
 	return int(C.MCLBN_FP_UNIT_SIZE)
@@ -47,6 +56,34 @@ func GetOpUnitSize() int {
 	return int(C.mclBn_getOpUnitSize())
 }
 
+// GetFrByteSize -- the serialized size of Fr
+func GetFrByteSize() int {
+	return int(C.mclBn_getFrByteSize())
+}
+
+// GetFpByteSize -- the serialized size of Fp
+func GetFpByteSize() int {
+	return int(C.mclBn_getFpByteSize())
+}
+
+// GetG1ByteSize -- the serialized size of G1
+func GetG1ByteSize() int {
+	return GetFpByteSize()
+}
+
+// GetG2ByteSize -- the serialized size of G2
+func GetG2ByteSize() int {
+	return GetFpByteSize() * 2
+}
+
+// allow zero length byte
+func getPointer(msg []byte) unsafe.Pointer {
+	if len(msg) == 0 {
+		return nil
+	}
+	return unsafe.Pointer(&msg[0])
+}
+
 // GetCurveOrder --
 // return the order of G1
 func GetCurveOrder() string {
@@ -71,6 +108,42 @@ func GetFieldOrder() string {
 	return string(buf[:n])
 }
 
+func bool2Cint(b bool) C.int {
+	if b {
+		return 1
+	} else {
+		return 0
+	}
+}
+
+// VerifyOrderG1 -- verify order if SetString/Deserialize are called
+func VerifyOrderG1(doVerify bool) {
+	// #nosec
+	C.mclBn_verifyOrderG1(bool2Cint(doVerify))
+}
+
+// VerifyOrderG2 -- verify order if SetString/Deserialize are called
+func VerifyOrderG2(doVerify bool) {
+	// #nosec
+	C.mclBn_verifyOrderG2(bool2Cint(doVerify))
+}
+
+// SetETHserialization --
+func SetETHserialization(enable bool) {
+	// #nosec
+	C.mclBn_setETHserialization(bool2Cint(enable))
+}
+
+// SetMapToMode --
+func SetMapToMode(mode int) error {
+	// #nosec
+	err := C.mclBn_setMapToMode((C.int)(mode))
+	if err != 0 {
+		return fmt.Errorf("SetMapToMode mode=%d\n", mode)
+	}
+	return nil
+}
+
 // Fr --
 type Fr struct {
 	v C.mclBnFr
@@ -98,7 +171,7 @@ func (x *Fr) SetInt64(v int64) {
 func (x *Fr) SetString(s string, base int) error {
 	buf := []byte(s)
 	// #nosec
-	err := C.mclBnFr_setStr(x.getPointer(), (*C.char)(unsafe.Pointer(&buf[0])), C.size_t(len(buf)), C.int(base))
+	err := C.mclBnFr_setStr(x.getPointer(), (*C.char)(getPointer(buf)), C.size_t(len(buf)), C.int(base))
 	if err != 0 {
 		return fmt.Errorf("err mclBnFr_setStr %x", err)
 	}
@@ -108,8 +181,8 @@ func (x *Fr) SetString(s string, base int) error {
 // Deserialize --
 func (x *Fr) Deserialize(buf []byte) error {
 	// #nosec
-	err := C.mclBnFr_deserialize(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	n := C.mclBnFr_deserialize(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
-	if err == 0 {
+	if n == 0 || int(n) != len(buf) {
 		return fmt.Errorf("err mclBnFr_deserialize %x", buf)
 	}
 	return nil
@@ -118,13 +191,33 @@ func (x *Fr) Deserialize(buf []byte) error {
 // SetLittleEndian --
 func (x *Fr) SetLittleEndian(buf []byte) error {
 	// #nosec
-	err := C.mclBnFr_setLittleEndian(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	err := C.mclBnFr_setLittleEndian(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
 	if err != 0 {
 		return fmt.Errorf("err mclBnFr_setLittleEndian %x", err)
 	}
 	return nil
 }
 
+// SetLittleEndianMod --
+func (x *Fr) SetLittleEndianMod(buf []byte) error {
+	// #nosec
+	err := C.mclBnFr_setLittleEndianMod(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if err != 0 {
+		return fmt.Errorf("err mclBnFr_setLittleEndianMod %x", err)
+	}
+	return nil
+}
+
+// SetBigEndianMod --
+func (x *Fr) SetBigEndianMod(buf []byte) error {
+	// #nosec
+	err := C.mclBnFr_setBigEndianMod(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if err != 0 {
+		return fmt.Errorf("err mclBnFr_setBigEndianMod %x", err)
+	}
+	return nil
+}
+
 // IsEqual --
 func (x *Fr) IsEqual(rhs *Fr) bool {
 	return C.mclBnFr_isEqual(x.getPointer(), rhs.getPointer()) == 1
@@ -135,11 +228,26 @@ func (x *Fr) IsZero() bool {
 	return C.mclBnFr_isZero(x.getPointer()) == 1
 }
 
+// IsValid --
+func (x *Fr) IsValid() bool {
+	return C.mclBnFr_isValid(x.getPointer()) == 1
+}
+
 // IsOne --
 func (x *Fr) IsOne() bool {
 	return C.mclBnFr_isOne(x.getPointer()) == 1
 }
 
+// IsOdd --
+func (x *Fr) IsOdd() bool {
+	return C.mclBnFr_isOdd(x.getPointer()) == 1
+}
+
+// IsNegative -- true if x >= (r + 1) / 2
+func (x *Fr) IsNegative() bool {
+	return C.mclBnFr_isNegative(x.getPointer()) == 1
+}
+
 // SetByCSPRNG --
 func (x *Fr) SetByCSPRNG() {
 	err := C.mclBnFr_setByCSPRNG(x.getPointer())
@@ -151,7 +259,7 @@ func (x *Fr) SetByCSPRNG() {
 // SetHashOf --
 func (x *Fr) SetHashOf(buf []byte) bool {
 	// #nosec
-	return C.mclBnFr_setHashOf(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf))) == 0
+	return C.mclBnFr_setHashOf(x.getPointer(), getPointer(buf), C.size_t(len(buf))) == 0
 }
 
 // GetString --
@@ -186,6 +294,11 @@ func FrInv(out *Fr, x *Fr) {
 	C.mclBnFr_inv(out.getPointer(), x.getPointer())
 }
 
+// FrSqr --
+func FrSqr(out *Fr, x *Fr) {
+	C.mclBnFr_sqr(out.getPointer(), x.getPointer())
+}
+
 // FrAdd --
 func FrAdd(out *Fr, x *Fr, y *Fr) {
 	C.mclBnFr_add(out.getPointer(), x.getPointer(), y.getPointer())
@@ -206,9 +319,289 @@ func FrDiv(out *Fr, x *Fr, y *Fr) {
 	C.mclBnFr_div(out.getPointer(), x.getPointer(), y.getPointer())
 }
 
+// FrSquareRoot --
+func FrSquareRoot(out *Fr, x *Fr) bool {
+	return C.mclBnFr_squareRoot(out.getPointer(), x.getPointer()) == 0
+}
+
+// Fp --
+type Fp struct {
+	v C.mclBnFp
+}
+
+// getPointer --
+func (x *Fp) getPointer() (p *C.mclBnFp) {
+	// #nosec
+	return (*C.mclBnFp)(unsafe.Pointer(x))
+}
+
+// Clear --
+func (x *Fp) Clear() {
+	// #nosec
+	C.mclBnFp_clear(x.getPointer())
+}
+
+// SetInt64 --
+func (x *Fp) SetInt64(v int64) {
+	// #nosec
+	C.mclBnFp_setInt(x.getPointer(), C.int64_t(v))
+}
+
+// SetString --
+func (x *Fp) SetString(s string, base int) error {
+	buf := []byte(s)
+	// #nosec
+	err := C.mclBnFp_setStr(x.getPointer(), (*C.char)(getPointer(buf)), C.size_t(len(buf)), C.int(base))
+	if err != 0 {
+		return fmt.Errorf("err mclBnFp_setStr %x", err)
+	}
+	return nil
+}
+
+// Deserialize --
+func (x *Fp) Deserialize(buf []byte) error {
+	// #nosec
+	n := C.mclBnFp_deserialize(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if n == 0 || int(n) != len(buf) {
+		return fmt.Errorf("err mclBnFp_deserialize %x", buf)
+	}
+	return nil
+}
+
+// SetLittleEndian --
+func (x *Fp) SetLittleEndian(buf []byte) error {
+	// #nosec
+	err := C.mclBnFp_setLittleEndian(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if err != 0 {
+		return fmt.Errorf("err mclBnFp_setLittleEndian %x", err)
+	}
+	return nil
+}
+
+// SetLittleEndianMod --
+func (x *Fp) SetLittleEndianMod(buf []byte) error {
+	// #nosec
+	err := C.mclBnFp_setLittleEndianMod(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if err != 0 {
+		return fmt.Errorf("err mclBnFp_setLittleEndianMod %x", err)
+	}
+	return nil
+}
+
+// SetBigEndianMod --
+func (x *Fp) SetBigEndianMod(buf []byte) error {
+	// #nosec
+	err := C.mclBnFp_setBigEndianMod(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if err != 0 {
+		return fmt.Errorf("err mclBnFp_setBigEndianMod %x", err)
+	}
+	return nil
+}
+
+// IsEqual --
+func (x *Fp) IsEqual(rhs *Fp) bool {
+	return C.mclBnFp_isEqual(x.getPointer(), rhs.getPointer()) == 1
+}
+
+// IsZero --
+func (x *Fp) IsZero() bool {
+	return C.mclBnFp_isZero(x.getPointer()) == 1
+}
+
+// IsValid --
+func (x *Fp) IsValid() bool {
+	return C.mclBnFp_isValid(x.getPointer()) == 1
+}
+
+// IsOne --
+func (x *Fp) IsOne() bool {
+	return C.mclBnFp_isOne(x.getPointer()) == 1
+}
+
+// IsOdd --
+func (x *Fp) IsOdd() bool {
+	return C.mclBnFp_isOdd(x.getPointer()) == 1
+}
+
+// IsNegative -- true if x >= (p + 1) / 2
+func (x *Fp) IsNegative() bool {
+	return C.mclBnFp_isNegative(x.getPointer()) == 1
+}
+
+// SetByCSPRNG --
+func (x *Fp) SetByCSPRNG() {
+	err := C.mclBnFp_setByCSPRNG(x.getPointer())
+	if err != 0 {
+		panic("err mclBnFp_setByCSPRNG")
+	}
+}
+
+// SetHashOf --
+func (x *Fp) SetHashOf(buf []byte) bool {
+	// #nosec
+	return C.mclBnFp_setHashOf(x.getPointer(), getPointer(buf), C.size_t(len(buf))) == 0
+}
+
+// GetString --
+func (x *Fp) GetString(base int) string {
+	buf := make([]byte, 2048)
+	// #nosec
+	n := C.mclBnFp_getStr((*C.char)(unsafe.Pointer(&buf[0])), C.size_t(len(buf)), x.getPointer(), C.int(base))
+	if n == 0 {
+		panic("err mclBnFp_getStr")
+	}
+	return string(buf[:n])
+}
+
+// Serialize --
+func (x *Fp) Serialize() []byte {
+	buf := make([]byte, 2048)
+	// #nosec
+	n := C.mclBnFp_serialize(unsafe.Pointer(&buf[0]), C.size_t(len(buf)), x.getPointer())
+	if n == 0 {
+		panic("err mclBnFp_serialize")
+	}
+	return buf[:n]
+}
+
+// FpNeg --
+func FpNeg(out *Fp, x *Fp) {
+	C.mclBnFp_neg(out.getPointer(), x.getPointer())
+}
+
+// FpInv --
+func FpInv(out *Fp, x *Fp) {
+	C.mclBnFp_inv(out.getPointer(), x.getPointer())
+}
+
+// FpSqr --
+func FpSqr(out *Fp, x *Fp) {
+	C.mclBnFp_sqr(out.getPointer(), x.getPointer())
+}
+
+// FpAdd --
+func FpAdd(out *Fp, x *Fp, y *Fp) {
+	C.mclBnFp_add(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// FpSub --
+func FpSub(out *Fp, x *Fp, y *Fp) {
+	C.mclBnFp_sub(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// FpMul --
+func FpMul(out *Fp, x *Fp, y *Fp) {
+	C.mclBnFp_mul(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// FpDiv --
+func FpDiv(out *Fp, x *Fp, y *Fp) {
+	C.mclBnFp_div(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// FpSquareRoot --
+func FpSquareRoot(out *Fp, x *Fp) bool {
+	return C.mclBnFp_squareRoot(out.getPointer(), x.getPointer()) == 0
+}
+
+// Fp2 -- x = D[0] + D[1] i where i^2 = -1
+type Fp2 struct {
+	D [2]Fp
+}
+
+// getPointer --
+func (x *Fp2) getPointer() (p *C.mclBnFp2) {
+	// #nosec
+	return (*C.mclBnFp2)(unsafe.Pointer(x))
+}
+
+// Clear --
+func (x *Fp2) Clear() {
+	// #nosec
+	C.mclBnFp2_clear(x.getPointer())
+}
+
+// Deserialize --
+func (x *Fp2) Deserialize(buf []byte) error {
+	// #nosec
+	n := C.mclBnFp2_deserialize(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if n == 0 || int(n) != len(buf) {
+		return fmt.Errorf("err mclBnFp2_deserialize %x", buf)
+	}
+	return nil
+}
+
+// IsEqual --
+func (x *Fp2) IsEqual(rhs *Fp2) bool {
+	return C.mclBnFp2_isEqual(x.getPointer(), rhs.getPointer()) == 1
+}
+
+// IsZero --
+func (x *Fp2) IsZero() bool {
+	return C.mclBnFp2_isZero(x.getPointer()) == 1
+}
+
+// IsOne --
+func (x *Fp2) IsOne() bool {
+	return C.mclBnFp2_isOne(x.getPointer()) == 1
+}
+
+// Serialize --
+func (x *Fp2) Serialize() []byte {
+	buf := make([]byte, 2048)
+	// #nosec
+	n := C.mclBnFp2_serialize(unsafe.Pointer(&buf[0]), C.size_t(len(buf)), x.getPointer())
+	if n == 0 {
+		panic("err mclBnFp2_serialize")
+	}
+	return buf[:n]
+}
+
+// Fp2Neg --
+func Fp2Neg(out *Fp2, x *Fp2) {
+	C.mclBnFp2_neg(out.getPointer(), x.getPointer())
+}
+
+// Fp2Inv --
+func Fp2Inv(out *Fp2, x *Fp2) {
+	C.mclBnFp2_inv(out.getPointer(), x.getPointer())
+}
+
+// Fp2Sqr --
+func Fp2Sqr(out *Fp2, x *Fp2) {
+	C.mclBnFp2_sqr(out.getPointer(), x.getPointer())
+}
+
+// Fp2Add --
+func Fp2Add(out *Fp2, x *Fp2, y *Fp2) {
+	C.mclBnFp2_add(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// Fp2Sub --
+func Fp2Sub(out *Fp2, x *Fp2, y *Fp2) {
+	C.mclBnFp2_sub(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// Fp2Mul --
+func Fp2Mul(out *Fp2, x *Fp2, y *Fp2) {
+	C.mclBnFp2_mul(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// Fp2Div --
+func Fp2Div(out *Fp2, x *Fp2, y *Fp2) {
+	C.mclBnFp2_div(out.getPointer(), x.getPointer(), y.getPointer())
+}
+
+// Fp2SquareRoot --
+func Fp2SquareRoot(out *Fp2, x *Fp2) bool {
+	return C.mclBnFp2_squareRoot(out.getPointer(), x.getPointer()) == 0
+}
+
 // G1 --
 type G1 struct {
-	v C.mclBnG1
+	X Fp
+	Y Fp
+	Z Fp
 }
 
 // getPointer --
@@ -227,7 +620,7 @@ func (x *G1) Clear() {
 func (x *G1) SetString(s string, base int) error {
 	buf := []byte(s)
 	// #nosec
-	err := C.mclBnG1_setStr(x.getPointer(), (*C.char)(unsafe.Pointer(&buf[0])), C.size_t(len(buf)), C.int(base))
+	err := C.mclBnG1_setStr(x.getPointer(), (*C.char)(getPointer(buf)), C.size_t(len(buf)), C.int(base))
 	if err != 0 {
 		return fmt.Errorf("err mclBnG1_setStr %x", err)
 	}
@@ -237,13 +630,54 @@ func (x *G1) SetString(s string, base int) error {
 // Deserialize --
 func (x *G1) Deserialize(buf []byte) error {
 	// #nosec
-	err := C.mclBnG1_deserialize(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	n := C.mclBnG1_deserialize(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
-	if err == 0 {
+	if n == 0 || int(n) != len(buf) {
 		return fmt.Errorf("err mclBnG1_deserialize %x", buf)
 	}
 	return nil
 }
 
+const ZERO_HEADER = 1 << 6
+
+func isZeroFormat(buf []byte, n int) bool {
+	if len(buf) < n {
+		return false
+	}
+	if buf[0] != ZERO_HEADER {
+		return false
+	}
+	for i := 1; i < n; i++ {
+		if buf[i] != 0 {
+			return false
+		}
+	}
+	return true
+}
+
+// DeserializeUncompressed -- x.Deserialize() + y.Deserialize()
+func (x *G1) DeserializeUncompressed(buf []byte) error {
+	if isZeroFormat(buf, GetG1ByteSize()*2) {
+		x.Clear()
+		return nil
+	}
+	// #nosec
+	var n = C.mclBnFp_deserialize(x.X.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if n == 0 {
+		return fmt.Errorf("err UncompressedDeserialize X %x", buf)
+	}
+	buf = buf[n:]
+	// #nosec
+	n = C.mclBnFp_deserialize(x.Y.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if n == 0 {
+		return fmt.Errorf("err UncompressedDeserialize Y %x", buf)
+	}
+	x.Z.SetInt64(1)
+	if !x.IsValid() {
+		return fmt.Errorf("err invalid point")
+	}
+	return nil
+}
+
 // IsEqual --
 func (x *G1) IsEqual(rhs *G1) bool {
 	return C.mclBnG1_isEqual(x.getPointer(), rhs.getPointer()) == 1
@@ -254,10 +688,20 @@ func (x *G1) IsZero() bool {
 	return C.mclBnG1_isZero(x.getPointer()) == 1
 }
 
+// IsValid --
+func (x *G1) IsValid() bool {
+	return C.mclBnG1_isValid(x.getPointer()) == 1
+}
+
+// IsValidOrder --
+func (x *G1) IsValidOrder() bool {
+	return C.mclBnG1_isValidOrder(x.getPointer()) == 1
+}
+
 // HashAndMapTo --
 func (x *G1) HashAndMapTo(buf []byte) error {
 	// #nosec
-	err := C.mclBnG1_hashAndMapTo(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	err := C.mclBnG1_hashAndMapTo(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
 	if err != 0 {
 		return fmt.Errorf("err mclBnG1_hashAndMapTo %x", err)
 	}
@@ -277,13 +721,39 @@ func (x *G1) GetString(base int) string {
 
 // Serialize --
 func (x *G1) Serialize() []byte {
-	buf := make([]byte, 2048)
+	buf := make([]byte, GetG1ByteSize())
 	// #nosec
 	n := C.mclBnG1_serialize(unsafe.Pointer(&buf[0]), C.size_t(len(buf)), x.getPointer())
 	if n == 0 {
 		panic("err mclBnG1_serialize")
 	}
-	return buf[:n]
+	return buf
+}
+
+// SerializeUncompressed -- all zero array if x.IsZero()
+func (x *G1) SerializeUncompressed() []byte {
+	buf := make([]byte, GetG1ByteSize()*2)
+	if x.IsZero() {
+		buf[0] = ZERO_HEADER
+		return buf
+	}
+	var nx G1
+	G1Normalize(&nx, x)
+	// #nosec
+	var n = C.mclBnFp_serialize(unsafe.Pointer(&buf[0]), C.size_t(len(buf)), nx.X.getPointer())
+	if n == 0 {
+		panic("err mclBnFp_serialize X")
+	}
+	n = C.mclBnFp_serialize(unsafe.Pointer(&buf[n]), C.size_t(len(buf))-n, nx.Y.getPointer())
+	if n == 0 {
+		panic("err mclBnFp_serialize Y")
+	}
+	return buf
+}
+
+// G1Normalize --
+func G1Normalize(out *G1, x *G1) {
+	C.mclBnG1_normalize(out.getPointer(), x.getPointer())
 }
 
 // G1Neg --
@@ -311,6 +781,19 @@ func G1Mul(out *G1, x *G1, y *Fr) {
 	C.mclBnG1_mul(out.getPointer(), x.getPointer(), y.getPointer())
 }
 
+// G1MulVec -- multi scalar multiplication out = sum mul(xVec[i], yVec[i])
+func G1MulVec(out *G1, xVec []G1, yVec []Fr) {
+	n := len(xVec)
+	if n != len(yVec) {
+		panic("xVec and yVec have the same size")
+	}
+	if n == 0 {
+		out.Clear()
+		return
+	}
+	C.mclBnG1_mulVec(out.getPointer(), (*C.mclBnG1)(unsafe.Pointer(&xVec[0])), (*C.mclBnFr)(unsafe.Pointer(&yVec[0])), (C.size_t)(n))
+}
+
 // G1MulCT -- constant time (depending on bit lengh of y)
 func G1MulCT(out *G1, x *G1, y *Fr) {
 	C.mclBnG1_mulCT(out.getPointer(), x.getPointer(), y.getPointer())
@@ -318,7 +801,9 @@ func G1MulCT(out *G1, x *G1, y *Fr) {
 
 // G2 --
 type G2 struct {
-	v C.mclBnG2
+	X Fp2
+	Y Fp2
+	Z Fp2
 }
 
 // getPointer --
@@ -337,7 +822,7 @@ func (x *G2) Clear() {
 func (x *G2) SetString(s string, base int) error {
 	buf := []byte(s)
 	// #nosec
-	err := C.mclBnG2_setStr(x.getPointer(), (*C.char)(unsafe.Pointer(&buf[0])), C.size_t(len(buf)), C.int(base))
+	err := C.mclBnG2_setStr(x.getPointer(), (*C.char)(getPointer(buf)), C.size_t(len(buf)), C.int(base))
 	if err != 0 {
 		return fmt.Errorf("err mclBnG2_setStr %x", err)
 	}
@@ -347,13 +832,38 @@ func (x *G2) SetString(s string, base int) error {
 // Deserialize --
 func (x *G2) Deserialize(buf []byte) error {
 	// #nosec
-	err := C.mclBnG2_deserialize(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	n := C.mclBnG2_deserialize(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
-	if err == 0 {
+	if n == 0 || int(n) != len(buf) {
 		return fmt.Errorf("err mclBnG2_deserialize %x", buf)
 	}
 	return nil
 }
 
+// DeserializeUncompressed -- x.Deserialize() + y.Deserialize()
+func (x *G2) DeserializeUncompressed(buf []byte) error {
+	if isZeroFormat(buf, GetG2ByteSize()*2) {
+		x.Clear()
+		return nil
+	}
+	// #nosec
+	var n = C.mclBnFp2_deserialize(x.X.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if n == 0 {
+		return fmt.Errorf("err UncompressedDeserialize X %x", buf)
+	}
+	buf = buf[n:]
+	// #nosec
+	n = C.mclBnFp2_deserialize(x.Y.getPointer(), getPointer(buf), C.size_t(len(buf)))
+	if n == 0 {
+		return fmt.Errorf("err UncompressedDeserialize Y %x", buf)
+	}
+	x.Z.D[0].SetInt64(1)
+	x.Z.D[1].Clear()
+	if !x.IsValid() {
+		return fmt.Errorf("err invalid point")
+	}
+	return nil
+}
+
 // IsEqual --
 func (x *G2) IsEqual(rhs *G2) bool {
 	return C.mclBnG2_isEqual(x.getPointer(), rhs.getPointer()) == 1
@@ -364,10 +874,20 @@ func (x *G2) IsZero() bool {
 	return C.mclBnG2_isZero(x.getPointer()) == 1
 }
 
+// IsValid --
+func (x *G2) IsValid() bool {
+	return C.mclBnG2_isValid(x.getPointer()) == 1
+}
+
+// IsValidOrder --
+func (x *G2) IsValidOrder() bool {
+	return C.mclBnG2_isValidOrder(x.getPointer()) == 1
+}
+
 // HashAndMapTo --
 func (x *G2) HashAndMapTo(buf []byte) error {
 	// #nosec
-	err := C.mclBnG2_hashAndMapTo(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	err := C.mclBnG2_hashAndMapTo(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
 	if err != 0 {
 		return fmt.Errorf("err mclBnG2_hashAndMapTo %x", err)
 	}
@@ -396,6 +916,32 @@ func (x *G2) Serialize() []byte {
 	return buf[:n]
 }
 
+// SerializeUncompressed -- all zero array if x.IsZero()
+func (x *G2) SerializeUncompressed() []byte {
+	buf := make([]byte, GetG2ByteSize()*2)
+	if x.IsZero() {
+		buf[0] = ZERO_HEADER
+		return buf
+	}
+	var nx G2
+	G2Normalize(&nx, x)
+	// #nosec
+	var n = C.mclBnFp2_serialize(unsafe.Pointer(&buf[0]), C.size_t(len(buf)), nx.X.getPointer())
+	if n == 0 {
+		panic("err mclBnFp2_serialize X")
+	}
+	n = C.mclBnFp2_serialize(unsafe.Pointer(&buf[n]), C.size_t(len(buf))-n, nx.Y.getPointer())
+	if n == 0 {
+		panic("err mclBnFp2_serialize Y")
+	}
+	return buf
+}
+
+// G2Normalize --
+func G2Normalize(out *G2, x *G2) {
+	C.mclBnG2_normalize(out.getPointer(), x.getPointer())
+}
+
 // G2Neg --
 func G2Neg(out *G2, x *G2) {
 	C.mclBnG2_neg(out.getPointer(), x.getPointer())
@@ -421,6 +967,19 @@ func G2Mul(out *G2, x *G2, y *Fr) {
 	C.mclBnG2_mul(out.getPointer(), x.getPointer(), y.getPointer())
 }
 
+// G2MulVec -- multi scalar multiplication out = sum mul(xVec[i], yVec[i])
+func G2MulVec(out *G2, xVec []G2, yVec []Fr) {
+	n := len(xVec)
+	if n != len(yVec) {
+		panic("xVec and yVec have the same size")
+	}
+	if n == 0 {
+		out.Clear()
+		return
+	}
+	C.mclBnG2_mulVec(out.getPointer(), (*C.mclBnG2)(unsafe.Pointer(&xVec[0])), (*C.mclBnFr)(unsafe.Pointer(&yVec[0])), (C.size_t)(n))
+}
+
 // GT --
 type GT struct {
 	v C.mclBnGT
@@ -448,7 +1007,7 @@ func (x *GT) SetInt64(v int64) {
 func (x *GT) SetString(s string, base int) error {
 	buf := []byte(s)
 	// #nosec
-	err := C.mclBnGT_setStr(x.getPointer(), (*C.char)(unsafe.Pointer(&buf[0])), C.size_t(len(buf)), C.int(base))
+	err := C.mclBnGT_setStr(x.getPointer(), (*C.char)(getPointer(buf)), C.size_t(len(buf)), C.int(base))
 	if err != 0 {
 		return fmt.Errorf("err mclBnGT_setStr %x", err)
 	}
@@ -458,8 +1017,8 @@ func (x *GT) SetString(s string, base int) error {
 // Deserialize --
 func (x *GT) Deserialize(buf []byte) error {
 	// #nosec
-	err := C.mclBnGT_deserialize(x.getPointer(), unsafe.Pointer(&buf[0]), C.size_t(len(buf)))
+	n := C.mclBnGT_deserialize(x.getPointer(), getPointer(buf), C.size_t(len(buf)))
-	if err == 0 {
+	if n == 0 || int(n) != len(buf) {
 		return fmt.Errorf("err mclBnGT_deserialize %x", buf)
 	}
 	return nil
@@ -537,6 +1096,22 @@ func GTPow(out *GT, x *GT, y *Fr) {
 	C.mclBnGT_pow(out.getPointer(), x.getPointer(), y.getPointer())
 }
 
+// MapToG1 --
+func MapToG1(out *G1, x *Fp) error {
+	if C.mclBnFp_mapToG1(out.getPointer(), x.getPointer()) != 0 {
+		return fmt.Errorf("err mclBnFp_mapToG1")
+	}
+	return nil
+}
+
+// MapToG2 --
+func MapToG2(out *G2, x *Fp2) error {
+	if C.mclBnFp2_mapToG2(out.getPointer(), x.getPointer()) != 0 {
+		return fmt.Errorf("err mclBnFp2_mapToG2")
+	}
+	return nil
+}
+
 // Pairing --
 func Pairing(out *GT, x *G1, y *G2) {
 	C.mclBn_pairing(out.getPointer(), x.getPointer(), y.getPointer())
@@ -552,6 +1127,19 @@ func MillerLoop(out *GT, x *G1, y *G2) {
 	C.mclBn_millerLoop(out.getPointer(), x.getPointer(), y.getPointer())
 }
 
+// MillerLoopVec -- multi pairings ; out = prod_i e(xVec[i], yVec[i])
+func MillerLoopVec(out *GT, xVec []G1, yVec []G2) {
+	n := len(xVec)
+	if n != len(yVec) {
+		panic("xVec and yVec have the same size")
+	}
+	if n == 0 {
+		out.SetInt64(1)
+		return
+	}
+	C.mclBn_millerLoopVec(out.getPointer(), (*C.mclBnG1)(unsafe.Pointer(&xVec[0])), (*C.mclBnG2)(unsafe.Pointer(&yVec[0])), (C.size_t)(n))
+}
+
 // GetUint64NumToPrecompute --
 func GetUint64NumToPrecompute() int {
 	return int(C.mclBn_getUint64NumToPrecompute())
@@ -577,8 +1165,13 @@ func PrecomputedMillerLoop2(out *GT, P1 *G1, Q1buf []uint64, P2 *G1, Q2buf []uin
 
 // FrEvaluatePolynomial -- y = c[0] + c[1] * x + c[2] * x^2 + ...
 func FrEvaluatePolynomial(y *Fr, c []Fr, x *Fr) error {
+	n := len(c)
+	if n == 0 {
+		y.Clear()
+		return nil
+	}
 	// #nosec
-	err := C.mclBn_FrEvaluatePolynomial(y.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&c[0])), (C.size_t)(len(c)), x.getPointer())
+	err := C.mclBn_FrEvaluatePolynomial(y.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&c[0])), (C.size_t)(n), x.getPointer())
 	if err != 0 {
 		return fmt.Errorf("err mclBn_FrEvaluatePolynomial")
 	}
@@ -587,8 +1180,13 @@ func FrEvaluatePolynomial(y *Fr, c []Fr, x *Fr) error {
 
 // G1EvaluatePolynomial -- y = c[0] + c[1] * x + c[2] * x^2 + ...
 func G1EvaluatePolynomial(y *G1, c []G1, x *Fr) error {
+	n := len(c)
+	if n == 0 {
+		y.Clear()
+		return nil
+	}
 	// #nosec
-	err := C.mclBn_G1EvaluatePolynomial(y.getPointer(), (*C.mclBnG1)(unsafe.Pointer(&c[0])), (C.size_t)(len(c)), x.getPointer())
+	err := C.mclBn_G1EvaluatePolynomial(y.getPointer(), (*C.mclBnG1)(unsafe.Pointer(&c[0])), (C.size_t)(n), x.getPointer())
 	if err != 0 {
 		return fmt.Errorf("err mclBn_G1EvaluatePolynomial")
 	}
@@ -597,8 +1195,13 @@ func G1EvaluatePolynomial(y *G1, c []G1, x *Fr) error {
 
 // G2EvaluatePolynomial -- y = c[0] + c[1] * x + c[2] * x^2 + ...
 func G2EvaluatePolynomial(y *G2, c []G2, x *Fr) error {
+	n := len(c)
+	if n == 0 {
+		y.Clear()
+		return nil
+	}
 	// #nosec
-	err := C.mclBn_G2EvaluatePolynomial(y.getPointer(), (*C.mclBnG2)(unsafe.Pointer(&c[0])), (C.size_t)(len(c)), x.getPointer())
+	err := C.mclBn_G2EvaluatePolynomial(y.getPointer(), (*C.mclBnG2)(unsafe.Pointer(&c[0])), (C.size_t)(n), x.getPointer())
 	if err != 0 {
 		return fmt.Errorf("err mclBn_G2EvaluatePolynomial")
 	}
@@ -607,11 +1210,15 @@ func G2EvaluatePolynomial(y *G2, c []G2, x *Fr) error {
 
 // FrLagrangeInterpolation --
 func FrLagrangeInterpolation(out *Fr, xVec []Fr, yVec []Fr) error {
-	if len(xVec) != len(yVec) {
+	n := len(xVec)
+	if n == 0 {
+		return fmt.Errorf("err FrLagrangeInterpolation:n=0")
+	}
+	if n != len(yVec) {
 		return fmt.Errorf("err FrLagrangeInterpolation:bad size")
 	}
 	// #nosec
-	err := C.mclBn_FrLagrangeInterpolation(out.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&xVec[0])), (*C.mclBnFr)(unsafe.Pointer(&yVec[0])), (C.size_t)(len(xVec)))
+	err := C.mclBn_FrLagrangeInterpolation(out.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&xVec[0])), (*C.mclBnFr)(unsafe.Pointer(&yVec[0])), (C.size_t)(n))
 	if err != 0 {
 		return fmt.Errorf("err FrLagrangeInterpolation")
 	}
@@ -620,11 +1227,15 @@ func FrLagrangeInterpolation(out *Fr, xVec []Fr, yVec []Fr) error {
 
 // G1LagrangeInterpolation --
 func G1LagrangeInterpolation(out *G1, xVec []Fr, yVec []G1) error {
-	if len(xVec) != len(yVec) {
+	n := len(xVec)
+	if n == 0 {
+		return fmt.Errorf("err G1LagrangeInterpolation:n=0")
+	}
+	if n != len(yVec) {
 		return fmt.Errorf("err G1LagrangeInterpolation:bad size")
 	}
 	// #nosec
-	err := C.mclBn_G1LagrangeInterpolation(out.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&xVec[0])), (*C.mclBnG1)(unsafe.Pointer(&yVec[0])), (C.size_t)(len(xVec)))
+	err := C.mclBn_G1LagrangeInterpolation(out.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&xVec[0])), (*C.mclBnG1)(unsafe.Pointer(&yVec[0])), (C.size_t)(n))
 	if err != 0 {
 		return fmt.Errorf("err G1LagrangeInterpolation")
 	}
@@ -633,11 +1244,15 @@ func G1LagrangeInterpolation(out *G1, xVec []Fr, yVec []G1) error {
 
 // G2LagrangeInterpolation --
 func G2LagrangeInterpolation(out *G2, xVec []Fr, yVec []G2) error {
-	if len(xVec) != len(yVec) {
+	n := len(xVec)
+	if n == 0 {
+		return fmt.Errorf("err G2LagrangeInterpolation:n=0")
+	}
+	if n != len(yVec) {
 		return fmt.Errorf("err G2LagrangeInterpolation:bad size")
 	}
 	// #nosec
-	err := C.mclBn_G2LagrangeInterpolation(out.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&xVec[0])), (*C.mclBnG2)(unsafe.Pointer(&yVec[0])), (C.size_t)(len(xVec)))
+	err := C.mclBn_G2LagrangeInterpolation(out.getPointer(), (*C.mclBnFr)(unsafe.Pointer(&xVec[0])), (*C.mclBnG2)(unsafe.Pointer(&yVec[0])), (C.size_t)(n))
 	if err != 0 {
 		return fmt.Errorf("err G2LagrangeInterpolation")
 	}

ffi/go/mcl/mcl_test.go

@@ -2,6 +2,7 @@ package mcl
 
 import "testing"
 import "fmt"
+import "encoding/hex"
 
 func testBadPointOfG2(t *testing.T) {
 	var Q G2
@@ -71,6 +72,71 @@ func testNegAdd(t *testing.T) {
 	}
 }
 
+func testVecG1(t *testing.T) {
+	N := 50
+	xVec := make([]G1, N)
+	yVec := make([]Fr, N)
+	xVec[0].HashAndMapTo([]byte("aa"))
+	var R1, R2 G1
+	for i := 0; i < N; i++ {
+		if i > 0 {
+			G1Dbl(&xVec[i], &xVec[i-1])
+		}
+		yVec[i].SetByCSPRNG()
+		G1Mul(&R1, &xVec[i], &yVec[i])
+		G1Add(&R2, &R2, &R1)
+	}
+	G1MulVec(&R1, xVec, yVec)
+	if !R1.IsEqual(&R2) {
+		t.Errorf("wrong G1MulVec")
+	}
+}
+
+func testVecG2(t *testing.T) {
+	N := 50
+	xVec := make([]G2, N)
+	yVec := make([]Fr, N)
+	xVec[0].HashAndMapTo([]byte("aa"))
+	var R1, R2 G2
+	for i := 0; i < N; i++ {
+		if i > 0 {
+			G2Dbl(&xVec[i], &xVec[i-1])
+		}
+		yVec[i].SetByCSPRNG()
+		G2Mul(&R1, &xVec[i], &yVec[i])
+		G2Add(&R2, &R2, &R1)
+	}
+	G2MulVec(&R1, xVec, yVec)
+	if !R1.IsEqual(&R2) {
+		t.Errorf("wrong G2MulVec")
+	}
+}
+
+func testVecPairing(t *testing.T) {
+	N := 50
+	xVec := make([]G1, N)
+	yVec := make([]G2, N)
+	var e1, e2 GT
+	e1.SetInt64(1)
+	for i := 0; i < N; i++ {
+		xVec[0].HashAndMapTo([]byte("aa"))
+		yVec[0].HashAndMapTo([]byte("aa"))
+		Pairing(&e2, &xVec[i], &yVec[i])
+		GTMul(&e1, &e1, &e2)
+	}
+	MillerLoopVec(&e2, xVec, yVec)
+	FinalExp(&e2, &e2)
+	if !e1.IsEqual(&e2) {
+		t.Errorf("wrong MillerLoopVec")
+	}
+}
+
+func testVec(t *testing.T) {
+	testVecG1(t)
+	testVecG2(t)
+	testVecPairing(t)
+}
+
 func testPairing(t *testing.T) {
 	var a, b, ab Fr
 	err := a.SetString("123", 10)
@@ -130,6 +196,44 @@ func testPairing(t *testing.T) {
 	}
 }
 
+func testSerialize(t *testing.T) {
+	var x, xx Fr
+	var y, yy Fp
+	var P, PP G1
+	var Q, QQ G2
+	var e, ee GT
+	x.SetByCSPRNG()
+	y.SetByCSPRNG()
+	P.HashAndMapTo([]byte("abc"))
+	G1Dbl(&P, &P)
+	Q.HashAndMapTo([]byte("abc"))
+	G2Dbl(&Q, &Q)
+	Pairing(&e, &P, &Q)
+	if xx.Deserialize(x.Serialize()) != nil || !x.IsEqual(&xx) {
+		t.Error("Serialize Fr")
+	}
+	if yy.Deserialize(y.Serialize()) != nil || !y.IsEqual(&yy) {
+		t.Error("Serialize Fp")
+	}
+	if PP.Deserialize(P.Serialize()) != nil || !P.IsEqual(&PP) {
+		t.Error("Serialize G1")
+	}
+	if QQ.Deserialize(Q.Serialize()) != nil || !Q.IsEqual(&QQ) {
+		t.Error("Serialize G2")
+	}
+	if ee.Deserialize(e.Serialize()) != nil || !e.IsEqual(&ee) {
+		t.Error("Serialize GT")
+	}
+	G1Dbl(&PP, &PP)
+	if PP.DeserializeUncompressed(P.SerializeUncompressed()) != nil || !P.IsEqual(&PP) {
+		t.Error("SerializeUncompressed G1")
+	}
+	G2Dbl(&QQ, &QQ)
+	if QQ.DeserializeUncompressed(Q.SerializeUncompressed()) != nil || !Q.IsEqual(&QQ) {
+		t.Error("SerializeUncompressed G2")
+	}
+}
+
 func testMcl(t *testing.T, c int) {
 	err := Init(c)
 	if err != nil {
@@ -138,8 +242,47 @@ func testMcl(t *testing.T, c int) {
 	testHash(t)
 	testNegAdd(t)
 	testPairing(t)
+	testVec(t)
 	testGT(t)
 	testBadPointOfG2(t)
+	testSerialize(t)
+}
+
+func testETHserialize(t *testing.T) {
+	b := make([]byte, 32)
+	b[0] = 0x12
+	b[1] = 0x34
+	var x Fr
+	SetETHserialization(false)
+	x.Deserialize(b)
+	fmt.Printf("AAA x=%s\n", x.GetString(16))
+
+	SetETHserialization(true)
+	x.Deserialize(b)
+	fmt.Printf("AAA x=%s\n", x.GetString(16))
+}
+
+func testMapToG1(t *testing.T) {
+	SetMapToMode(IRTF)
+	fpHex := "0000000000000000000000000000000014406e5bfb9209256a3820879a29ac2f62d6aca82324bf3ae2aa7d3c54792043bd8c791fccdb080c1a52dc68b8b69350"
+	g1Hex := "ad7721bcdb7ce1047557776eb2659a444166dc6dd55c7ca6e240e21ae9aa18f529f04ac31d861b54faf3307692545db7"
+	fpBin, _ := hex.DecodeString(fpHex)
+	var x Fp
+	x.SetBigEndianMod(fpBin)
+	var P1 G1
+	err := MapToG1(&P1, &x)
+	if err != nil {
+		t.Fatal("MapToG1")
+	}
+	g1Str, _ := hex.DecodeString(g1Hex)
+	var P2 G1
+	err = P2.Deserialize(g1Str)
+	if err != nil {
+		t.Fatal("G1.Deserialize")
+	}
+	if !P1.IsEqual(&P2) {
+		t.Fatal("bad MapToG1")
+	}
 }
 
 func TestMclMain(t *testing.T) {
@@ -153,5 +296,7 @@ func TestMclMain(t *testing.T) {
 		}
 		t.Log("BLS12_381")
 		testMcl(t, BLS12_381)
+		testETHserialize(t)
+		testMapToG1(t)
 	}
 }

ffi/java/Bn256Test.java

@@ -1,104 +0,0 @@
-import java.io.*;
-import com.herumi.mcl.*;
-
-/*
-	Bn256Test
-*/
-public class Bn256Test {
-	static {
-		String lib = "mcl_bn256";
-		String libName = System.mapLibraryName(lib);
-		System.out.println("libName : " + libName);
-		System.loadLibrary(lib);
-	}
-	public static void assertEquals(String msg, String x, String y) {
-		if (x.equals(y)) {
-			System.out.println("OK : " + msg);
-		} else {
-			System.out.println("NG : " + msg + ", x = " + x + ", y = " + y);
-		}
-	}
-	public static void assertBool(String msg, boolean b) {
-		if (b) {
-			System.out.println("OK : " + msg);
-		} else {
-			System.out.println("NG : " + msg);
-		}
-	}
-	public static void main(String argv[]) {
-		try {
-			Bn256.SystemInit();
-			Fr x = new Fr(5);
-			Fr y = new Fr(-2);
-			Fr z = new Fr(5);
-			assertBool("x != y", !x.equals(y));
-			assertBool("x == z", x.equals(z));
-			assertEquals("x == 5", x.toString(), "5");
-			Bn256.add(x, x, y);
-			assertEquals("x == 3", x.toString(), "3");
-			Bn256.mul(x, x, x);
-			assertEquals("x == 9", x.toString(), "9");
-			G1 P = new G1();
-			System.out.println("P=" + P);
-			P.set("-1", "1");
-			System.out.println("P=" + P);
-			Bn256.neg(P, P);
-			System.out.println("P=" + P);
-
-			String xa = "12723517038133731887338407189719511622662176727675373276651903807414909099441";
-			String xb = "4168783608814932154536427934509895782246573715297911553964171371032945126671";
-			String ya = "13891744915211034074451795021214165905772212241412891944830863846330766296736";
-			String yb = "7937318970632701341203597196594272556916396164729705624521405069090520231616";
-
-			G2 Q = new G2(xa, xb, ya, yb);
-
-			P.hashAndMapToG1("This is a pen");
-			{
-				String s = P.toString();
-				G1 P1 = new G1();
-				P1.setStr(s);
-				assertBool("P == P1", P1.equals(P));
-			}
-
-			GT e = new GT();
-			Bn256.pairing(e, P, Q);
-			GT e1 = new GT();
-			GT e2 = new GT();
-			Fr c = new Fr("1234567890123234928348230428394234");
-			G2 cQ = new G2(Q);
-			Bn256.mul(cQ, Q, c); // cQ = Q * c
-			Bn256.pairing(e1, P, cQ);
-			Bn256.pow(e2, e, c); // e2 = e^c
-			assertBool("e1 == e2", e1.equals(e2));
-
-			G1 cP = new G1(P);
-			Bn256.mul(cP, P, c); // cP = P * c
-			Bn256.pairing(e1, cP, Q);
-			assertBool("e1 == e2", e1.equals(e2));
-
-			BLSsignature(Q);
-		} catch (RuntimeException e) {
-			System.out.println("unknown exception :" + e);
-		}
-	}
-	public static void BLSsignature(G2 Q)
-	{
-		Fr s = new Fr();
-		s.setRand(); // secret key
-		System.out.println("secret key " + s);
-		G2 pub = new G2();
-		Bn256.mul(pub, Q, s); // public key = sQ
-
-		String m = "signature test";
-		G1 H = new G1();
-		H.hashAndMapToG1(m); // H = Hash(m)
-		G1 sign = new G1();
-		Bn256.mul(sign, H, s); // signature of m = s H
-
-		GT e1 = new GT();
-		GT e2 = new GT();
-		Bn256.pairing(e1, H, pub); // e1 = e(H, s Q)
-		Bn256.pairing(e2, sign, Q); // e2 = e(s H, Q);
-		assertBool("verify signature", e1.equals(e2));
-	}
-}

ffi/java/ElgamalTest.java

@@ -8,7 +8,7 @@ import com.herumi.mcl.*;
 */
 public class ElgamalTest {
 	static {
-		String lib = "mcl_elgamal";
+		String lib = "mclelgamaljava";
 		String libName = System.mapLibraryName(lib);
 		System.out.println("libName : " + libName);
 		System.loadLibrary(lib);

ffi/java/Makefile

@@ -1,46 +1,50 @@
 TOP_DIR=../..
 include $(TOP_DIR)/common.mk
 ifeq ($(UNAME_S),Darwin)
-  JAVA_INC=-I/System/Library/Frameworks/JavaVM.framework/Versions/Current/Headers/
+  JAVA_INC_DIR?=/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk/System/Library/Frameworks/JavaVM.framework/Headers/
+  JAVA_INC?=-I$(JAVA_INC_DIR)
+  CFLAGS+=$(JAVA_INC)/darwin
 else
-  JAVA_INC=-I/usr/lib/jvm/default-java/include
+  JAVA_INC_DIR?=/usr/lib/jvm/default-java/include
+  JAVA_INC?=-I$(JAVA_INC_DIR)
 #JAVA_INC=-I/usr/lib/jvm/java-7-openjdk-amd64/include
+  CFLAGS+=$(JAVA_INC)/linux
   CFLAGS+=-z noexecstack
   LDFLAGS+=-lrt
 endif
-CFLAGS+=$(JAVA_INC) $(JAVA_INC)/linux -I $(TOP_DIR)/include -I $(TOP_DIR)/../xbyak -I $(TOP_DIR)/../cybozulib/include -Wno-strict-aliasing
+CFLAGS+=$(JAVA_INC) -I $(TOP_DIR)/include -I -Wno-strict-aliasing
 MCL_LIB=$(TOP_DIR)/lib/libmcl.a
 
 PACKAGE_NAME=com.herumi.mcl
 PACKAGE_DIR=$(subst .,/,$(PACKAGE_NAME))
 
-ELGAMAL_LIB=$(TOP_DIR)/bin/libmcl_elgamal.$(LIB_SUF)
+ELGAMAL_LIB=$(TOP_DIR)/lib/libmclelgamaljava.$(LIB_SUF)
-BN256_LIB=$(TOP_DIR)/bin/libmcl_bn256.$(LIB_SUF)
+MCLJAVA_LIB=$(TOP_DIR)/lib/libmcljava.$(LIB_SUF)
-JAVA_EXE=cd $(TOP_DIR)/bin && LD_LIBRARY_PATH=./:$(LD_LIBRARY_PATH) java -classpath ../ffi/java
+JAVA_EXE=cd $(TOP_DIR)/lib && LD_LIBRARY_PATH=../lib:$(LD_LIBRARY_PATH) java -classpath ../ffi/java
 all: $(ELGAMAL_LIB)
 
 elgamal_wrap.cxx: elgamal.i elgamal_impl.hpp
 	$(MKDIR) $(PACKAGE_DIR)
 	swig -java -package $(PACKAGE_NAME) -outdir $(PACKAGE_DIR) -c++ -Wall elgamal.i
 
-bn256_wrap.cxx: bn256.i bn256_impl.hpp
+mcl_wrap.cxx: mcl.i mcl_impl.hpp
 	$(MKDIR) $(PACKAGE_DIR)
-	swig -java -package $(PACKAGE_NAME) -outdir $(PACKAGE_DIR) -c++ -Wall bn256.i
+	swig -java -package $(PACKAGE_NAME) -outdir $(PACKAGE_DIR) -c++ -Wall mcl.i
 
 $(MCL_LIB):
-	make -C $(TOP_DIR)
+	make -C $(TOP_DIR) lib/libmcl.a
 
 $(ELGAMAL_LIB): elgamal_wrap.cxx $(MCL_LIB)
-	$(PRE)$(CXX) $< -o $@ $(CFLAGS) $(LDFLAGS) $(MCL_LIB) -shared
+	$(PRE)$(CXX) $< -o $@ $(CFLAGS) $(MCL_LIB) $(LDFLAGS) -shared
 
-$(BN256_LIB): bn256_wrap.cxx $(MCL_LIB)
+$(MCLJAVA_LIB): mcl_wrap.cxx $(MCL_LIB)
-	$(PRE)$(CXX) $< -o $@ $(CFLAGS) $(LDFLAGS) $(MCL_LIB) -shared
+	$(PRE)$(CXX) $< -o $@ $(CFLAGS) $(MCL_LIB) $(LDFLAGS) -shared
 
 %.class: %.java
 	javac $<
 
 ElgamalTest.class: ElgamalTest.java $(ELGAMAL_LIB)
-Bn256Test.class: Bn256Test.java $(BN256_LIB)
+MclTest.class: MclTest.java $(MCLJAVA_LIB)
 
 jar:
 	jar cvf mcl.jar com
@@ -49,16 +53,17 @@ test_elgamal: ElgamalTest.class $(ELGAMAL_LIB)
 	$(JAVA_EXE) ElgamalTest
 	$(JAVA_EXE) ElgamalTest -e NIST_P192
 	$(JAVA_EXE) ElgamalTest -e NIST_P256 -h sha256
+	$(JAVA_EXE) ElgamalTest -e secp256k1 -h sha256
 	$(JAVA_EXE) ElgamalTest -e NIST_P384 -h sha384
-	$(JAVA_EXE) ElgamalTest -e NIST_P521 -h sha512
+#	$(JAVA_EXE) ElgamalTest -e NIST_P521 -h sha512
 
-test_bn256: Bn256Test.class $(BN256_LIB)
+test_mcl: MclTest.class $(MCLJAVA_LIB)
-	$(JAVA_EXE) Bn256Test
+	$(JAVA_EXE) MclTest
 
 test:
 	$(MAKE) test_elgamal
-	$(MAKE) test_bn256
+	$(MAKE) test_mcl
 
 clean:
-	rm -rf *.class $(ELGAMAL_LIB) $(PACKAGE_DIR)/*.class *_wrap.cxx
+	rm -rf *.class $(ELGAMAL_LIB) $(MCLJAVA_LIB) $(PACKAGE_DIR)/*.class *_wrap.cxx
 

ffi/java/MclTest.java

@@ -0,0 +1,160 @@
+import java.io.*;
+import com.herumi.mcl.*;
+
+/*
+	MclTest
+*/
+public class MclTest {
+	static {
+		String lib = "mcljava";
+		String libName = System.mapLibraryName(lib);
+		System.out.println("libName : " + libName);
+		System.loadLibrary(lib);
+	}
+	public static int errN = 0;
+	public static void assertEquals(String msg, String x, String y) {
+		if (x.equals(y)) {
+			System.out.println("OK : " + msg);
+		} else {
+			System.out.println("NG : " + msg + ", x = " + x + ", y = " + y);
+			errN++;
+		}
+	}
+	public static void assertBool(String msg, boolean b) {
+		if (b) {
+			System.out.println("OK : " + msg);
+		} else {
+			System.out.println("NG : " + msg);
+			errN++;
+		}
+	}
+	public static void testCurve(int curveType, String name) {
+		try {
+			System.out.println("curve=" + name);
+			Mcl.SystemInit(curveType);
+			Fr x = new Fr(5);
+			Fr y = new Fr(-2);
+			Fr z = new Fr(5);
+			assertBool("x != y", !x.equals(y));
+			assertBool("x == z", x.equals(z));
+			assertEquals("x == 5", x.toString(), "5");
+			Mcl.add(x, x, y);
+			assertEquals("x == 3", x.toString(), "3");
+			Mcl.mul(x, x, x);
+			assertEquals("x == 9", x.toString(), "9");
+			assertEquals("x == 12", (new Fr("12")).toString(), "12");
+			assertEquals("x == 18", (new Fr("12", 16)).toString(), "18");
+			assertEquals("x == ff", (new Fr("255")).toString(16), "ff");
+			Mcl.inv(y, x);
+			Mcl.mul(x, y, x);
+			assertBool("x == 1", x.isOne());
+
+			{
+				byte[] b = x.serialize();
+				Fr t = new Fr();
+				t.deserialize(b);
+				assertBool("serialize", x.equals(t));
+				t.setLittleEndianMod(b);
+				assertBool("setLittleEndianMod", x.equals(t));
+				t.setHashOf(b);
+				assertBool("setHashOf", !x.equals(t));
+				Fr u = new Fr();
+				u.setHashOf(new byte[]{1,2,3});
+				assertBool("setHashOf - different", !u.equals(t));
+			}
+			G1 P = new G1();
+			System.out.println("P=" + P);
+			Mcl.hashAndMapToG1(P, "test".getBytes());
+			System.out.println("P=" + P);
+			byte[] buf = { 1, 2, 3, 4 };
+			Mcl.hashAndMapToG1(P, buf);
+			System.out.println("P=" + P);
+			Mcl.neg(P, P);
+			System.out.println("P=" + P);
+			{
+				byte[] b = P.serialize();
+				G1 t = new G1();
+				t.deserialize(b);
+				assertBool("serialize", P.equals(t));
+			}
+
+			G2 Q = new G2();
+			Mcl.hashAndMapToG2(Q, "abc".getBytes());
+			System.out.println("Q=" + Q);
+
+			Mcl.hashAndMapToG1(P, "This is a pen".getBytes());
+			{
+				String s = P.toString();
+				G1 P1 = new G1();
+				P1.setStr(s);
+				assertBool("P == P1", P1.equals(P));
+			}
+			{
+				byte[] b = Q.serialize();
+				G2 t = new G2();
+				t.deserialize(b);
+				assertBool("serialize", Q.equals(t));
+			}
+
+			GT e = new GT();
+			Mcl.pairing(e, P, Q);
+			GT e1 = new GT();
+			GT e2 = new GT();
+			Fr c = new Fr("1234567890123234928348230428394234");
+			System.out.println("c=" + c);
+			G2 cQ = new G2(Q);
+			Mcl.mul(cQ, Q, c); // cQ = Q * c
+			Mcl.pairing(e1, P, cQ);
+			Mcl.pow(e2, e, c); // e2 = e^c
+			assertBool("e1 == e2", e1.equals(e2));
+			{
+				byte[] b = e1.serialize();
+				GT t = new GT();
+				t.deserialize(b);
+				assertBool("serialize", e1.equals(t));
+			}
+			G1 cP = new G1(P);
+			Mcl.mul(cP, P, c); // cP = P * c
+			Mcl.pairing(e1, cP, Q);
+			assertBool("e1 == e2", e1.equals(e2));
+			Mcl.inv(e1, e1);
+			Mcl.mul(e1, e1, e2);
+			e2.setStr("1 0 0 0 0 0 0 0 0 0 0 0");
+			assertBool("e1 == 1", e1.equals(e2));
+			assertBool("e1 == 1", e1.isOne());
+
+			BLSsignature(Q);
+			if (errN == 0) {
+				System.out.println("all test passed");
+			} else {
+				System.out.println("ERR=" + errN);
+			}
+		} catch (RuntimeException e) {
+			System.out.println("unknown exception :" + e);
+		}
+	}
+	public static void BLSsignature(G2 Q)
+	{
+		Fr s = new Fr();
+		s.setByCSPRNG(); // secret key
+		System.out.println("secret key " + s);
+		G2 pub = new G2();
+		Mcl.mul(pub, Q, s); // public key = sQ
+
+		byte[] m = "signature test".getBytes();
+		G1 H = new G1();
+		Mcl.hashAndMapToG1(H, m); // H = Hash(m)
+		G1 sign = new G1();
+		Mcl.mul(sign, H, s); // signature of m = s H
+
+		GT e1 = new GT();
+		GT e2 = new GT();
+		Mcl.pairing(e1, H, pub); // e1 = e(H, s Q)
+		Mcl.pairing(e2, sign, Q); // e2 = e(s H, Q);
+		assertBool("verify signature", e1.equals(e2));
+	}
+	public static void main(String argv[]) {
+		testCurve(Mcl.BN254, "BN254");
+		testCurve(Mcl.BLS12_381, "BLS12_381");
+	}
+}

ffi/java/android/jni/Android.mk

@@ -0,0 +1,39 @@
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+
+LOCAL_CPP_EXTENSION := .cpp .ll .cxx
+LOCAL_MODULE := mcljava
+
+LOCAL_MCL_DIR := $(LOCAL_PATH)/../../../../
+
+ifeq ($(TARGET_ARCH_ABI),x86_64)
+  MY_BIT := 64
+endif
+ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
+  MY_BIT := 64
+endif
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  MY_BIT := 32
+endif
+ifeq ($(TARGET_ARCH_ABI),x86)
+  MY_BIT := 32
+endif
+ifeq ($(MY_BIT),64)
+  MY_BASE_LL := $(LOCAL_MCL_DIR)/src/base64.ll
+  LOCAL_CPPFLAGS += -DMCL_SIZEOF_UNIT=8
+endif
+ifeq ($(MY_BIT),32)
+  MY_BASE_LL := $(LOCAL_MCL_DIR)/src/base32.ll
+  LOCAL_CPPFLAGS += -DMCL_SIZEOF_UNIT=4
+endif
+LOCAL_SRC_FILES := $(LOCAL_MCL_DIR)/ffi/java/mcl_wrap.cxx $(LOCAL_MCL_DIR)/src/bn_c384_256.cpp $(LOCAL_MCL_DIR)/src/fp.cpp $(MY_BASE_LL)
+LOCAL_C_INCLUDES := $(LOCAL_MCL_DIR)/include $(LOCAL_MCL_DIR)/src $(LOCAL_MCL_DIR)/ffi/java
+LOCAL_CPPFLAGS += -DMCL_DONT_USE_XBYAK
+LOCAL_CPPFLAGS += -O3 -DNDEBUG -fPIC -DMCL_DONT_USE_OPENSSL -DMCL_LLVM_BMI2=0 -DMCL_USE_LLVM=1 -DMCL_USE_VINT -DMCL_VINT_FIXED_BUFFER -DMCL_MAX_BIT_SIZE=384
+LOCAL_CPPFLAGS += -fno-threadsafe-statics
+#LOCAL_CPPFLAGS+=-fno-exceptions -fno-rtti -DCYBOZU_DONT_USE_EXCEPTION -DCYBOZU_DONT_USE_STRING -std=c++03
+
+#LOCAL_CPPFLAGS += -DBLS_ETH
+#LOCAL_LDLIBS := -llog #-Wl,--no-warn-shared-textrel
+#include $(BUILD_STATIC_LIBRARY)
+include $(BUILD_SHARED_LIBRARY)

ffi/java/android/jni/Application.mk

@@ -0,0 +1,4 @@
+APP_ABI := arm64-v8a armeabi-v7a x86_64
+APP_PLATFORM := android-19
+APP_STL := c++_static
+APP_CPPFLAGS := -fexceptions

ffi/java/bn256.i

@@ -1,31 +0,0 @@
-%module Bn256
-
-%include "std_string.i"
-%include "std_except.i"
-
-
-%{
-#include <cybozu/random_generator.hpp>
-#include <cybozu/crypto.hpp>
-#include <mcl/bn256.hpp>
-struct Param {
-    cybozu::RandomGenerator rg;
-    static inline Param& getParam()
-	{
-		static Param p;
-	    return p;
-	}
-};
-
-static void HashAndMapToG1(mcl::bn256::G1& P, const std::string& m)
-{
-	std::string digest = cybozu::crypto::Hash::digest(cybozu::crypto::Hash::N_SHA256, m);
-	mcl::bn256::Fp t;
-	t.setArrayMask(digest.c_str(), digest.size());
-	mcl::bn256::BN::param.mapTo.calcG1(P, t);
-}
-
-#include "bn256_impl.hpp"
-%}
-
-%include "bn256_impl.hpp"

ffi/java/bn256_impl.hpp

@@ -1,249 +0,0 @@
-#include <mcl/bn256.hpp>
-#include <stdint.h>
-#include <sstream>
-
-void SystemInit() throw(std::exception)
-{
-	mcl::bn256::initPairing();
-}
-
-class G1;
-class G2;
-class GT;
-/*
-	Fr = Z / rZ
-*/
-class Fr {
-	mcl::bn256::Fr self_;
-	friend class G1;
-	friend class G2;
-	friend class GT;
-	friend void neg(Fr& y, const Fr& x);
-	friend void add(Fr& z, const Fr& x, const Fr& y);
-	friend void sub(Fr& z, const Fr& x, const Fr& y);
-	friend void mul(Fr& z, const Fr& x, const Fr& y);
-	friend void mul(G1& z, const G1& x, const Fr& y);
-	friend void mul(G2& z, const G2& x, const Fr& y);
-	friend void div(Fr& z, const Fr& x, const Fr& y);
-	friend void pow(GT& z, const GT& x, const Fr& y);
-public:
-	Fr() {}
-	Fr(const Fr& rhs) : self_(rhs.self_) {}
-	Fr(int x) : self_(x) {}
-	Fr(const std::string& str) throw(std::exception)
-		: self_(str) {}
-	bool equals(const Fr& rhs) const { return self_ == rhs.self_; }
-	void setStr(const  std::string& str) throw(std::exception)
-	{
-		self_.setStr(str);
-	}
-	void setInt(int x)
-	{
-		self_ = x;
-	}
-	void clear()
-	{
-		self_.clear();
-	}
-	void setRand()
-	{
-		self_.setRand(Param::getParam().rg);
-	}
-	std::string toString() const throw(std::exception)
-	{
-		return self_.getStr();
-	}
-};
-
-void neg(Fr& y, const Fr& x)
-{
-	mcl::bn256::Fr::neg(y.self_, x.self_);
-}
-
-void add(Fr& z, const Fr& x, const Fr& y)
-{
-	mcl::bn256::Fr::add(z.self_, x.self_, y.self_);
-}
-
-void sub(Fr& z, const Fr& x, const Fr& y)
-{
-	mcl::bn256::Fr::sub(z.self_, x.self_, y.self_);
-}
-
-void mul(Fr& z, const Fr& x, const Fr& y)
-{
-	mcl::bn256::Fr::mul(z.self_, x.self_, y.self_);
-}
-
-void div(Fr& z, const Fr& x, const Fr& y)
-{
-	mcl::bn256::Fr::div(z.self_, x.self_, y.self_);
-}
-
-/*
-	#G1 = r
-*/
-class G1 {
-	mcl::bn256::G1 self_;
-	friend void neg(G1& y, const G1& x);
-	friend void dbl(G1& y, const G1& x);
-	friend void add(G1& z, const G1& x, const G1& y);
-	friend void sub(G1& z, const G1& x, const G1& y);
-	friend void mul(G1& z, const G1& x, const Fr& y);
-	friend void pairing(GT& e, const G1& P, const G2& Q);
-public:
-	G1() {}
-	G1(const G1& rhs) : self_(rhs.self_) {}
-	G1(const std::string& x, const std::string& y) throw(std::exception)
-		: self_(mcl::bn256::Fp(x), mcl::bn256::Fp(y))
-	{
-	}
-	bool equals(const G1& rhs) const { return self_ == rhs.self_; }
-	void set(const std::string& x, const std::string& y)
-	{
-		self_.set(mcl::bn256::Fp(x), mcl::bn256::Fp(y));
-	}
-	void hashAndMapToG1(const std::string& m) throw(std::exception)
-	{
-		HashAndMapToG1(self_, m);
-	}
-	void clear()
-	{
-		self_.clear();
-	}
-	/*
-		compressed format
-	*/
-	void setStr(const std::string& str) throw(std::exception)
-	{
-		self_.setStr(str);
-	}
-	std::string toString() const throw(std::exception)
-	{
-		return self_.getStr();
-	}
-};
-
-void neg(G1& y, const G1& x)
-{
-	mcl::bn256::G1::neg(y.self_, x.self_);
-}
-void dbl(G1& y, const G1& x)
-{
-	mcl::bn256::G1::dbl(y.self_, x.self_);
-}
-void add(G1& z, const G1& x, const G1& y)
-{
-	mcl::bn256::G1::add(z.self_, x.self_, y.self_);
-}
-void sub(G1& z, const G1& x, const G1& y)
-{
-	mcl::bn256::G1::sub(z.self_, x.self_, y.self_);
-}
-void mul(G1& z, const G1& x, const Fr& y)
-{
-	mcl::bn256::G1::mul(z.self_, x.self_, y.self_);
-}
-
-/*
-	#G2 = r
-*/
-class G2 {
-	mcl::bn256::G2 self_;
-	friend void neg(G2& y, const G2& x);
-	friend void dbl(G2& y, const G2& x);
-	friend void add(G2& z, const G2& x, const G2& y);
-	friend void sub(G2& z, const G2& x, const G2& y);
-	friend void mul(G2& z, const G2& x, const Fr& y);
-	friend void pairing(GT& e, const G1& P, const G2& Q);
-public:
-	G2() {}
-	G2(const G2& rhs) : self_(rhs.self_) {}
-	G2(const std::string& xa, const std::string& xb, const std::string& ya, const std::string& yb) throw(std::exception)
-		: self_(mcl::bn256::Fp2(xa, xb), mcl::bn256::Fp2(ya, yb))
-	{
-	}
-	bool equals(const G2& rhs) const { return self_ == rhs.self_; }
-	void set(const std::string& xa, const std::string& xb, const std::string& ya, const std::string& yb)
-	{
-		self_.set(mcl::bn256::Fp2(xa, xb), mcl::bn256::Fp2(ya, yb));
-	}
-	void clear()
-	{
-		self_.clear();
-	}
-	/*
-		compressed format
-	*/
-	void setStr(const std::string& str) throw(std::exception)
-	{
-		self_.setStr(str);
-	}
-	std::string toString() const throw(std::exception)
-	{
-		return self_.getStr();
-	}
-};
-
-void neg(G2& y, const G2& x)
-{
-	mcl::bn256::G2::neg(y.self_, x.self_);
-}
-void dbl(G2& y, const G2& x)
-{
-	mcl::bn256::G2::dbl(y.self_, x.self_);
-}
-void add(G2& z, const G2& x, const G2& y)
-{
-	mcl::bn256::G2::add(z.self_, x.self_, y.self_);
-}
-void sub(G2& z, const G2& x, const G2& y)
-{
-	mcl::bn256::G2::sub(z.self_, x.self_, y.self_);
-}
-void mul(G2& z, const G2& x, const Fr& y)
-{
-	mcl::bn256::G2::mul(z.self_, x.self_, y.self_);
-}
-
-/*
-	#GT = r
-*/
-class GT {
-	mcl::bn256::Fp12 self_;
-	friend void mul(GT& z, const GT& x, const GT& y);
-	friend void pow(GT& z, const GT& x, const Fr& y);
-	friend void pairing(GT& e, const G1& P, const G2& Q);
-public:
-	GT() {}
-	GT(const GT& rhs) : self_(rhs.self_) {}
-	bool equals(const GT& rhs) const { return self_ == rhs.self_; }
-	void clear()
-	{
-		self_.clear();
-	}
-	void setStr(const std::string& str) throw(std::exception)
-	{
-		std::istringstream iss(str);
-		iss >> self_;
-	}
-	std::string toString() const throw(std::exception)
-	{
-		std::ostringstream oss;
-		oss << self_;
-		return oss.str();
-	}
-};
-
-void mul(GT& z, const GT& x, const GT& y)
-{
-	mcl::bn256::Fp12::mul(z.self_, x.self_, y.self_);
-}
-void pow(GT& z, const GT& x, const Fr& y)
-{
-	mcl::bn256::Fp12::pow(z.self_, x.self_, y.self_);
-}
-void pairing(GT& e, const G1& P, const G2& Q)
-{
-	mcl::bn256::pairing(e.self_, P.self_, Q.self_);
-}

ffi/java/com/herumi/mcl/CipherText.java

@@ -0,0 +1,67 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class CipherText {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected CipherText(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(CipherText obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        ElgamalJNI.delete_CipherText(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public String toStr() {
+    return ElgamalJNI.CipherText_toStr(swigCPtr, this);
+  }
+
+  public String toString() {
+    return ElgamalJNI.CipherText_toString(swigCPtr, this);
+  }
+
+  public void fromStr(String str) {
+    ElgamalJNI.CipherText_fromStr(swigCPtr, this, str);
+  }
+
+  public void add(CipherText c) {
+    ElgamalJNI.CipherText_add(swigCPtr, this, CipherText.getCPtr(c), c);
+  }
+
+  public void mul(int m) {
+    ElgamalJNI.CipherText_mul__SWIG_0(swigCPtr, this, m);
+  }
+
+  public void mul(String str) {
+    ElgamalJNI.CipherText_mul__SWIG_1(swigCPtr, this, str);
+  }
+
+  public CipherText() {
+    this(ElgamalJNI.new_CipherText(), true);
+  }
+
+}

ffi/java/com/herumi/mcl/Elgamal.java

@@ -0,0 +1,38 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class Elgamal {
+  public static SWIGTYPE_p_bool new_p_bool() {
+    long cPtr = ElgamalJNI.new_p_bool();
+    return (cPtr == 0) ? null : new SWIGTYPE_p_bool(cPtr, false);
+  }
+
+  public static SWIGTYPE_p_bool copy_p_bool(boolean value) {
+    long cPtr = ElgamalJNI.copy_p_bool(value);
+    return (cPtr == 0) ? null : new SWIGTYPE_p_bool(cPtr, false);
+  }
+
+  public static void delete_p_bool(SWIGTYPE_p_bool obj) {
+    ElgamalJNI.delete_p_bool(SWIGTYPE_p_bool.getCPtr(obj));
+  }
+
+  public static void p_bool_assign(SWIGTYPE_p_bool obj, boolean value) {
+    ElgamalJNI.p_bool_assign(SWIGTYPE_p_bool.getCPtr(obj), value);
+  }
+
+  public static boolean p_bool_value(SWIGTYPE_p_bool obj) {
+    return ElgamalJNI.p_bool_value(SWIGTYPE_p_bool.getCPtr(obj));
+  }
+
+  public static void SystemInit(String param) {
+    ElgamalJNI.SystemInit(param);
+  }
+
+}

ffi/java/com/herumi/mcl/ElgamalJNI.java

@@ -0,0 +1,51 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class ElgamalJNI {
+  public final static native long new_p_bool();
+  public final static native long copy_p_bool(boolean jarg1);
+  public final static native void delete_p_bool(long jarg1);
+  public final static native void p_bool_assign(long jarg1, boolean jarg2);
+  public final static native boolean p_bool_value(long jarg1);
+  public final static native void SystemInit(String jarg1);
+  public final static native String CipherText_toStr(long jarg1, CipherText jarg1_);
+  public final static native String CipherText_toString(long jarg1, CipherText jarg1_);
+  public final static native void CipherText_fromStr(long jarg1, CipherText jarg1_, String jarg2);
+  public final static native void CipherText_add(long jarg1, CipherText jarg1_, long jarg2, CipherText jarg2_);
+  public final static native void CipherText_mul__SWIG_0(long jarg1, CipherText jarg1_, int jarg2);
+  public final static native void CipherText_mul__SWIG_1(long jarg1, CipherText jarg1_, String jarg2);
+  public final static native long new_CipherText();
+  public final static native void delete_CipherText(long jarg1);
+  public final static native String PublicKey_toStr(long jarg1, PublicKey jarg1_);
+  public final static native String PublicKey_toString(long jarg1, PublicKey jarg1_);
+  public final static native void PublicKey_fromStr(long jarg1, PublicKey jarg1_, String jarg2);
+  public final static native void PublicKey_save(long jarg1, PublicKey jarg1_, String jarg2);
+  public final static native void PublicKey_load(long jarg1, PublicKey jarg1_, String jarg2);
+  public final static native void PublicKey_enc__SWIG_0(long jarg1, PublicKey jarg1_, long jarg2, CipherText jarg2_, int jarg3);
+  public final static native void PublicKey_enc__SWIG_1(long jarg1, PublicKey jarg1_, long jarg2, CipherText jarg2_, String jarg3);
+  public final static native void PublicKey_rerandomize(long jarg1, PublicKey jarg1_, long jarg2, CipherText jarg2_);
+  public final static native void PublicKey_add__SWIG_0(long jarg1, PublicKey jarg1_, long jarg2, CipherText jarg2_, int jarg3);
+  public final static native void PublicKey_add__SWIG_1(long jarg1, PublicKey jarg1_, long jarg2, CipherText jarg2_, String jarg3);
+  public final static native long new_PublicKey();
+  public final static native void delete_PublicKey(long jarg1);
+  public final static native String PrivateKey_toStr(long jarg1, PrivateKey jarg1_);
+  public final static native String PrivateKey_toString(long jarg1, PrivateKey jarg1_);
+  public final static native void PrivateKey_fromStr(long jarg1, PrivateKey jarg1_, String jarg2);
+  public final static native void PrivateKey_save(long jarg1, PrivateKey jarg1_, String jarg2);
+  public final static native void PrivateKey_load(long jarg1, PrivateKey jarg1_, String jarg2);
+  public final static native void PrivateKey_init(long jarg1, PrivateKey jarg1_);
+  public final static native long PrivateKey_getPublicKey(long jarg1, PrivateKey jarg1_);
+  public final static native int PrivateKey_dec__SWIG_0(long jarg1, PrivateKey jarg1_, long jarg2, CipherText jarg2_, long jarg3);
+  public final static native int PrivateKey_dec__SWIG_1(long jarg1, PrivateKey jarg1_, long jarg2, CipherText jarg2_);
+  public final static native void PrivateKey_setCache(long jarg1, PrivateKey jarg1_, int jarg2, int jarg3);
+  public final static native void PrivateKey_clearCache(long jarg1, PrivateKey jarg1_);
+  public final static native long new_PrivateKey();
+  public final static native void delete_PrivateKey(long jarg1);
+}

ffi/java/com/herumi/mcl/Fp.java

@@ -0,0 +1,105 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class Fp {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected Fp(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(Fp obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        MclJNI.delete_Fp(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public Fp() {
+    this(MclJNI.new_Fp__SWIG_0(), true);
+  }
+
+  public Fp(Fp rhs) {
+    this(MclJNI.new_Fp__SWIG_1(Fp.getCPtr(rhs), rhs), true);
+  }
+
+  public Fp(int x) {
+    this(MclJNI.new_Fp__SWIG_2(x), true);
+  }
+
+  public Fp(String str, int base) {
+    this(MclJNI.new_Fp__SWIG_3(str, base), true);
+  }
+
+  public Fp(String str) {
+    this(MclJNI.new_Fp__SWIG_4(str), true);
+  }
+
+  public boolean equals(Fp rhs) {
+    return MclJNI.Fp_equals(swigCPtr, this, Fp.getCPtr(rhs), rhs);
+  }
+
+  public boolean isZero() {
+    return MclJNI.Fp_isZero(swigCPtr, this);
+  }
+
+  public boolean isOne() {
+    return MclJNI.Fp_isOne(swigCPtr, this);
+  }
+
+  public void setStr(String str, int base) {
+    MclJNI.Fp_setStr__SWIG_0(swigCPtr, this, str, base);
+  }
+
+  public void setStr(String str) {
+    MclJNI.Fp_setStr__SWIG_1(swigCPtr, this, str);
+  }
+
+  public void setInt(int x) {
+    MclJNI.Fp_setInt(swigCPtr, this, x);
+  }
+
+  public void clear() {
+    MclJNI.Fp_clear(swigCPtr, this);
+  }
+
+  public void setByCSPRNG() {
+    MclJNI.Fp_setByCSPRNG(swigCPtr, this);
+  }
+
+  public String toString(int base) {
+    return MclJNI.Fp_toString__SWIG_0(swigCPtr, this, base);
+  }
+
+  public String toString() {
+    return MclJNI.Fp_toString__SWIG_1(swigCPtr, this);
+  }
+
+  public void deserialize(byte[] cbuf) {
+    MclJNI.Fp_deserialize(swigCPtr, this, cbuf);
+  }
+
+  public byte[] serialize() { return MclJNI.Fp_serialize(swigCPtr, this); }
+
+}

ffi/java/com/herumi/mcl/Fr.java

@@ -0,0 +1,113 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class Fr {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected Fr(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(Fr obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        MclJNI.delete_Fr(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public Fr() {
+    this(MclJNI.new_Fr__SWIG_0(), true);
+  }
+
+  public Fr(Fr rhs) {
+    this(MclJNI.new_Fr__SWIG_1(Fr.getCPtr(rhs), rhs), true);
+  }
+
+  public Fr(int x) {
+    this(MclJNI.new_Fr__SWIG_2(x), true);
+  }
+
+  public Fr(String str, int base) {
+    this(MclJNI.new_Fr__SWIG_3(str, base), true);
+  }
+
+  public Fr(String str) {
+    this(MclJNI.new_Fr__SWIG_4(str), true);
+  }
+
+  public boolean equals(Fr rhs) {
+    return MclJNI.Fr_equals(swigCPtr, this, Fr.getCPtr(rhs), rhs);
+  }
+
+  public boolean isZero() {
+    return MclJNI.Fr_isZero(swigCPtr, this);
+  }
+
+  public boolean isOne() {
+    return MclJNI.Fr_isOne(swigCPtr, this);
+  }
+
+  public void setStr(String str, int base) {
+    MclJNI.Fr_setStr__SWIG_0(swigCPtr, this, str, base);
+  }
+
+  public void setStr(String str) {
+    MclJNI.Fr_setStr__SWIG_1(swigCPtr, this, str);
+  }
+
+  public void setInt(int x) {
+    MclJNI.Fr_setInt(swigCPtr, this, x);
+  }
+
+  public void clear() {
+    MclJNI.Fr_clear(swigCPtr, this);
+  }
+
+  public void setByCSPRNG() {
+    MclJNI.Fr_setByCSPRNG(swigCPtr, this);
+  }
+
+  public String toString(int base) {
+    return MclJNI.Fr_toString__SWIG_0(swigCPtr, this, base);
+  }
+
+  public String toString() {
+    return MclJNI.Fr_toString__SWIG_1(swigCPtr, this);
+  }
+
+  public void deserialize(byte[] cbuf) {
+    MclJNI.Fr_deserialize(swigCPtr, this, cbuf);
+  }
+
+  public void setLittleEndianMod(byte[] cbuf) {
+    MclJNI.Fr_setLittleEndianMod(swigCPtr, this, cbuf);
+  }
+
+  public void setHashOf(byte[] cbuf) {
+    MclJNI.Fr_setHashOf(swigCPtr, this, cbuf);
+  }
+
+  public byte[] serialize() { return MclJNI.Fr_serialize(swigCPtr, this); }
+
+}

ffi/java/com/herumi/mcl/G1.java

@@ -0,0 +1,93 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class G1 {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected G1(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(G1 obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        MclJNI.delete_G1(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public G1() {
+    this(MclJNI.new_G1__SWIG_0(), true);
+  }
+
+  public G1(G1 rhs) {
+    this(MclJNI.new_G1__SWIG_1(G1.getCPtr(rhs), rhs), true);
+  }
+
+  public G1(Fp x, Fp y) {
+    this(MclJNI.new_G1__SWIG_2(Fp.getCPtr(x), x, Fp.getCPtr(y), y), true);
+  }
+
+  public boolean equals(G1 rhs) {
+    return MclJNI.G1_equals(swigCPtr, this, G1.getCPtr(rhs), rhs);
+  }
+
+  public boolean isZero() {
+    return MclJNI.G1_isZero(swigCPtr, this);
+  }
+
+  public boolean isValidOrder() {
+    return MclJNI.G1_isValidOrder(swigCPtr, this);
+  }
+
+  public void set(Fp x, Fp y) {
+    MclJNI.G1_set(swigCPtr, this, Fp.getCPtr(x), x, Fp.getCPtr(y), y);
+  }
+
+  public void clear() {
+    MclJNI.G1_clear(swigCPtr, this);
+  }
+
+  public void setStr(String str, int base) {
+    MclJNI.G1_setStr__SWIG_0(swigCPtr, this, str, base);
+  }
+
+  public void setStr(String str) {
+    MclJNI.G1_setStr__SWIG_1(swigCPtr, this, str);
+  }
+
+  public String toString(int base) {
+    return MclJNI.G1_toString__SWIG_0(swigCPtr, this, base);
+  }
+
+  public String toString() {
+    return MclJNI.G1_toString__SWIG_1(swigCPtr, this);
+  }
+
+  public void deserialize(byte[] cbuf) {
+    MclJNI.G1_deserialize(swigCPtr, this, cbuf);
+  }
+
+  public byte[] serialize() { return MclJNI.G1_serialize(swigCPtr, this); }
+
+}

ffi/java/com/herumi/mcl/G2.java

@@ -0,0 +1,89 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class G2 {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected G2(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(G2 obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        MclJNI.delete_G2(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public G2() {
+    this(MclJNI.new_G2__SWIG_0(), true);
+  }
+
+  public G2(G2 rhs) {
+    this(MclJNI.new_G2__SWIG_1(G2.getCPtr(rhs), rhs), true);
+  }
+
+  public G2(Fp ax, Fp ay, Fp bx, Fp by) {
+    this(MclJNI.new_G2__SWIG_2(Fp.getCPtr(ax), ax, Fp.getCPtr(ay), ay, Fp.getCPtr(bx), bx, Fp.getCPtr(by), by), true);
+  }
+
+  public boolean equals(G2 rhs) {
+    return MclJNI.G2_equals(swigCPtr, this, G2.getCPtr(rhs), rhs);
+  }
+
+  public boolean isZero() {
+    return MclJNI.G2_isZero(swigCPtr, this);
+  }
+
+  public void set(Fp ax, Fp ay, Fp bx, Fp by) {
+    MclJNI.G2_set(swigCPtr, this, Fp.getCPtr(ax), ax, Fp.getCPtr(ay), ay, Fp.getCPtr(bx), bx, Fp.getCPtr(by), by);
+  }
+
+  public void clear() {
+    MclJNI.G2_clear(swigCPtr, this);
+  }
+
+  public void setStr(String str, int base) {
+    MclJNI.G2_setStr__SWIG_0(swigCPtr, this, str, base);
+  }
+
+  public void setStr(String str) {
+    MclJNI.G2_setStr__SWIG_1(swigCPtr, this, str);
+  }
+
+  public String toString(int base) {
+    return MclJNI.G2_toString__SWIG_0(swigCPtr, this, base);
+  }
+
+  public String toString() {
+    return MclJNI.G2_toString__SWIG_1(swigCPtr, this);
+  }
+
+  public void deserialize(byte[] cbuf) {
+    MclJNI.G2_deserialize(swigCPtr, this, cbuf);
+  }
+
+  public byte[] serialize() { return MclJNI.G2_serialize(swigCPtr, this); }
+
+}

ffi/java/com/herumi/mcl/GT.java

@@ -0,0 +1,81 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class GT {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected GT(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(GT obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        MclJNI.delete_GT(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public GT() {
+    this(MclJNI.new_GT__SWIG_0(), true);
+  }
+
+  public GT(GT rhs) {
+    this(MclJNI.new_GT__SWIG_1(GT.getCPtr(rhs), rhs), true);
+  }
+
+  public boolean equals(GT rhs) {
+    return MclJNI.GT_equals(swigCPtr, this, GT.getCPtr(rhs), rhs);
+  }
+
+  public boolean isOne() {
+    return MclJNI.GT_isOne(swigCPtr, this);
+  }
+
+  public void clear() {
+    MclJNI.GT_clear(swigCPtr, this);
+  }
+
+  public void setStr(String str, int base) {
+    MclJNI.GT_setStr__SWIG_0(swigCPtr, this, str, base);
+  }
+
+  public void setStr(String str) {
+    MclJNI.GT_setStr__SWIG_1(swigCPtr, this, str);
+  }
+
+  public String toString(int base) {
+    return MclJNI.GT_toString__SWIG_0(swigCPtr, this, base);
+  }
+
+  public String toString() {
+    return MclJNI.GT_toString__SWIG_1(swigCPtr, this);
+  }
+
+  public void deserialize(byte[] cbuf) {
+    MclJNI.GT_deserialize(swigCPtr, this, cbuf);
+  }
+
+  public byte[] serialize() { return MclJNI.GT_serialize(swigCPtr, this); }
+
+}

ffi/java/com/herumi/mcl/Mcl.java

@@ -0,0 +1,128 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class Mcl implements MclConstants {
+  public static void SystemInit(int curveType) {
+    MclJNI.SystemInit(curveType);
+  }
+
+  public static void neg(Fr y, Fr x) {
+    MclJNI.neg__SWIG_0(Fr.getCPtr(y), y, Fr.getCPtr(x), x);
+  }
+
+  public static void inv(Fr y, Fr x) {
+    MclJNI.inv__SWIG_0(Fr.getCPtr(y), y, Fr.getCPtr(x), x);
+  }
+
+  public static void add(Fr z, Fr x, Fr y) {
+    MclJNI.add__SWIG_0(Fr.getCPtr(z), z, Fr.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void sub(Fr z, Fr x, Fr y) {
+    MclJNI.sub__SWIG_0(Fr.getCPtr(z), z, Fr.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void mul(Fr z, Fr x, Fr y) {
+    MclJNI.mul__SWIG_0(Fr.getCPtr(z), z, Fr.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void mul(G1 z, G1 x, Fr y) {
+    MclJNI.mul__SWIG_1(G1.getCPtr(z), z, G1.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void mul(G2 z, G2 x, Fr y) {
+    MclJNI.mul__SWIG_2(G2.getCPtr(z), z, G2.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void div(Fr z, Fr x, Fr y) {
+    MclJNI.div__SWIG_0(Fr.getCPtr(z), z, Fr.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void pow(GT z, GT x, Fr y) {
+    MclJNI.pow(GT.getCPtr(z), z, GT.getCPtr(x), x, Fr.getCPtr(y), y);
+  }
+
+  public static void neg(Fp y, Fp x) {
+    MclJNI.neg__SWIG_1(Fp.getCPtr(y), y, Fp.getCPtr(x), x);
+  }
+
+  public static void inv(Fp y, Fp x) {
+    MclJNI.inv__SWIG_1(Fp.getCPtr(y), y, Fp.getCPtr(x), x);
+  }
+
+  public static void add(Fp z, Fp x, Fp y) {
+    MclJNI.add__SWIG_1(Fp.getCPtr(z), z, Fp.getCPtr(x), x, Fp.getCPtr(y), y);
+  }
+
+  public static void sub(Fp z, Fp x, Fp y) {
+    MclJNI.sub__SWIG_1(Fp.getCPtr(z), z, Fp.getCPtr(x), x, Fp.getCPtr(y), y);
+  }
+
+  public static void mul(Fp z, Fp x, Fp y) {
+    MclJNI.mul__SWIG_3(Fp.getCPtr(z), z, Fp.getCPtr(x), x, Fp.getCPtr(y), y);
+  }
+
+  public static void div(Fp z, Fp x, Fp y) {
+    MclJNI.div__SWIG_1(Fp.getCPtr(z), z, Fp.getCPtr(x), x, Fp.getCPtr(y), y);
+  }
+
+  public static void neg(G1 y, G1 x) {
+    MclJNI.neg__SWIG_2(G1.getCPtr(y), y, G1.getCPtr(x), x);
+  }
+
+  public static void dbl(G1 y, G1 x) {
+    MclJNI.dbl__SWIG_0(G1.getCPtr(y), y, G1.getCPtr(x), x);
+  }
+
+  public static void add(G1 z, G1 x, G1 y) {
+    MclJNI.add__SWIG_2(G1.getCPtr(z), z, G1.getCPtr(x), x, G1.getCPtr(y), y);
+  }
+
+  public static void sub(G1 z, G1 x, G1 y) {
+    MclJNI.sub__SWIG_2(G1.getCPtr(z), z, G1.getCPtr(x), x, G1.getCPtr(y), y);
+  }
+
+  public static void pairing(GT e, G1 P, G2 Q) {
+    MclJNI.pairing(GT.getCPtr(e), e, G1.getCPtr(P), P, G2.getCPtr(Q), Q);
+  }
+
+  public static void hashAndMapToG1(G1 P, byte[] cbuf) {
+    MclJNI.hashAndMapToG1(G1.getCPtr(P), P, cbuf);
+  }
+
+  public static void neg(G2 y, G2 x) {
+    MclJNI.neg__SWIG_3(G2.getCPtr(y), y, G2.getCPtr(x), x);
+  }
+
+  public static void dbl(G2 y, G2 x) {
+    MclJNI.dbl__SWIG_1(G2.getCPtr(y), y, G2.getCPtr(x), x);
+  }
+
+  public static void add(G2 z, G2 x, G2 y) {
+    MclJNI.add__SWIG_3(G2.getCPtr(z), z, G2.getCPtr(x), x, G2.getCPtr(y), y);
+  }
+
+  public static void sub(G2 z, G2 x, G2 y) {
+    MclJNI.sub__SWIG_3(G2.getCPtr(z), z, G2.getCPtr(x), x, G2.getCPtr(y), y);
+  }
+
+  public static void hashAndMapToG2(G2 P, byte[] cbuf) {
+    MclJNI.hashAndMapToG2(G2.getCPtr(P), P, cbuf);
+  }
+
+  public static void mul(GT z, GT x, GT y) {
+    MclJNI.mul__SWIG_4(GT.getCPtr(z), z, GT.getCPtr(x), x, GT.getCPtr(y), y);
+  }
+
+  public static void inv(GT y, GT x) {
+    MclJNI.inv__SWIG_2(GT.getCPtr(y), y, GT.getCPtr(x), x);
+  }
+
+}

ffi/java/com/herumi/mcl/MclConstants.java

@@ -0,0 +1,14 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public interface MclConstants {
+  public final static int BN254 = 0;
+  public final static int BLS12_381 = 5;
+}

ffi/java/com/herumi/mcl/MclJNI.java

@@ -0,0 +1,120 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class MclJNI {
+  public final static native void SystemInit(int jarg1);
+  public final static native void neg__SWIG_0(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_);
+  public final static native void inv__SWIG_0(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_);
+  public final static native void add__SWIG_0(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_, long jarg3, Fr jarg3_);
+  public final static native void sub__SWIG_0(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_, long jarg3, Fr jarg3_);
+  public final static native void mul__SWIG_0(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_, long jarg3, Fr jarg3_);
+  public final static native void mul__SWIG_1(long jarg1, G1 jarg1_, long jarg2, G1 jarg2_, long jarg3, Fr jarg3_);
+  public final static native void mul__SWIG_2(long jarg1, G2 jarg1_, long jarg2, G2 jarg2_, long jarg3, Fr jarg3_);
+  public final static native void div__SWIG_0(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_, long jarg3, Fr jarg3_);
+  public final static native void pow(long jarg1, GT jarg1_, long jarg2, GT jarg2_, long jarg3, Fr jarg3_);
+  public final static native long new_Fr__SWIG_0();
+  public final static native long new_Fr__SWIG_1(long jarg1, Fr jarg1_);
+  public final static native long new_Fr__SWIG_2(int jarg1);
+  public final static native long new_Fr__SWIG_3(String jarg1, int jarg2);
+  public final static native long new_Fr__SWIG_4(String jarg1);
+  public final static native boolean Fr_equals(long jarg1, Fr jarg1_, long jarg2, Fr jarg2_);
+  public final static native boolean Fr_isZero(long jarg1, Fr jarg1_);
+  public final static native boolean Fr_isOne(long jarg1, Fr jarg1_);
+  public final static native void Fr_setStr__SWIG_0(long jarg1, Fr jarg1_, String jarg2, int jarg3);
+  public final static native void Fr_setStr__SWIG_1(long jarg1, Fr jarg1_, String jarg2);
+  public final static native void Fr_setInt(long jarg1, Fr jarg1_, int jarg2);
+  public final static native void Fr_clear(long jarg1, Fr jarg1_);
+  public final static native void Fr_setByCSPRNG(long jarg1, Fr jarg1_);
+  public final static native String Fr_toString__SWIG_0(long jarg1, Fr jarg1_, int jarg2);
+  public final static native String Fr_toString__SWIG_1(long jarg1, Fr jarg1_);
+  public final static native void Fr_deserialize(long jarg1, Fr jarg1_, byte[] jarg2);
+  public final static native void Fr_setLittleEndianMod(long jarg1, Fr jarg1_, byte[] jarg2);
+  public final static native void Fr_setHashOf(long jarg1, Fr jarg1_, byte[] jarg2);
+  public final static native byte[] Fr_serialize(long jarg1, Fr jarg1_);
+  public final static native void delete_Fr(long jarg1);
+  public final static native void neg__SWIG_1(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_);
+  public final static native void inv__SWIG_1(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_);
+  public final static native void add__SWIG_1(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_);
+  public final static native void sub__SWIG_1(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_);
+  public final static native void mul__SWIG_3(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_);
+  public final static native void div__SWIG_1(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_);
+  public final static native long new_Fp__SWIG_0();
+  public final static native long new_Fp__SWIG_1(long jarg1, Fp jarg1_);
+  public final static native long new_Fp__SWIG_2(int jarg1);
+  public final static native long new_Fp__SWIG_3(String jarg1, int jarg2);
+  public final static native long new_Fp__SWIG_4(String jarg1);
+  public final static native boolean Fp_equals(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_);
+  public final static native boolean Fp_isZero(long jarg1, Fp jarg1_);
+  public final static native boolean Fp_isOne(long jarg1, Fp jarg1_);
+  public final static native void Fp_setStr__SWIG_0(long jarg1, Fp jarg1_, String jarg2, int jarg3);
+  public final static native void Fp_setStr__SWIG_1(long jarg1, Fp jarg1_, String jarg2);
+  public final static native void Fp_setInt(long jarg1, Fp jarg1_, int jarg2);
+  public final static native void Fp_clear(long jarg1, Fp jarg1_);
+  public final static native void Fp_setByCSPRNG(long jarg1, Fp jarg1_);
+  public final static native String Fp_toString__SWIG_0(long jarg1, Fp jarg1_, int jarg2);
+  public final static native String Fp_toString__SWIG_1(long jarg1, Fp jarg1_);
+  public final static native void Fp_deserialize(long jarg1, Fp jarg1_, byte[] jarg2);
+  public final static native byte[] Fp_serialize(long jarg1, Fp jarg1_);
+  public final static native void delete_Fp(long jarg1);
+  public final static native void neg__SWIG_2(long jarg1, G1 jarg1_, long jarg2, G1 jarg2_);
+  public final static native void dbl__SWIG_0(long jarg1, G1 jarg1_, long jarg2, G1 jarg2_);
+  public final static native void add__SWIG_2(long jarg1, G1 jarg1_, long jarg2, G1 jarg2_, long jarg3, G1 jarg3_);
+  public final static native void sub__SWIG_2(long jarg1, G1 jarg1_, long jarg2, G1 jarg2_, long jarg3, G1 jarg3_);
+  public final static native void pairing(long jarg1, GT jarg1_, long jarg2, G1 jarg2_, long jarg3, G2 jarg3_);
+  public final static native void hashAndMapToG1(long jarg1, G1 jarg1_, byte[] jarg2);
+  public final static native long new_G1__SWIG_0();
+  public final static native long new_G1__SWIG_1(long jarg1, G1 jarg1_);
+  public final static native long new_G1__SWIG_2(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_);
+  public final static native boolean G1_equals(long jarg1, G1 jarg1_, long jarg2, G1 jarg2_);
+  public final static native boolean G1_isZero(long jarg1, G1 jarg1_);
+  public final static native boolean G1_isValidOrder(long jarg1, G1 jarg1_);
+  public final static native void G1_set(long jarg1, G1 jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_);
+  public final static native void G1_clear(long jarg1, G1 jarg1_);
+  public final static native void G1_setStr__SWIG_0(long jarg1, G1 jarg1_, String jarg2, int jarg3);
+  public final static native void G1_setStr__SWIG_1(long jarg1, G1 jarg1_, String jarg2);
+  public final static native String G1_toString__SWIG_0(long jarg1, G1 jarg1_, int jarg2);
+  public final static native String G1_toString__SWIG_1(long jarg1, G1 jarg1_);
+  public final static native void G1_deserialize(long jarg1, G1 jarg1_, byte[] jarg2);
+  public final static native byte[] G1_serialize(long jarg1, G1 jarg1_);
+  public final static native void delete_G1(long jarg1);
+  public final static native void neg__SWIG_3(long jarg1, G2 jarg1_, long jarg2, G2 jarg2_);
+  public final static native void dbl__SWIG_1(long jarg1, G2 jarg1_, long jarg2, G2 jarg2_);
+  public final static native void add__SWIG_3(long jarg1, G2 jarg1_, long jarg2, G2 jarg2_, long jarg3, G2 jarg3_);
+  public final static native void sub__SWIG_3(long jarg1, G2 jarg1_, long jarg2, G2 jarg2_, long jarg3, G2 jarg3_);
+  public final static native void hashAndMapToG2(long jarg1, G2 jarg1_, byte[] jarg2);
+  public final static native long new_G2__SWIG_0();
+  public final static native long new_G2__SWIG_1(long jarg1, G2 jarg1_);
+  public final static native long new_G2__SWIG_2(long jarg1, Fp jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_, long jarg4, Fp jarg4_);
+  public final static native boolean G2_equals(long jarg1, G2 jarg1_, long jarg2, G2 jarg2_);
+  public final static native boolean G2_isZero(long jarg1, G2 jarg1_);
+  public final static native void G2_set(long jarg1, G2 jarg1_, long jarg2, Fp jarg2_, long jarg3, Fp jarg3_, long jarg4, Fp jarg4_, long jarg5, Fp jarg5_);
+  public final static native void G2_clear(long jarg1, G2 jarg1_);
+  public final static native void G2_setStr__SWIG_0(long jarg1, G2 jarg1_, String jarg2, int jarg3);
+  public final static native void G2_setStr__SWIG_1(long jarg1, G2 jarg1_, String jarg2);
+  public final static native String G2_toString__SWIG_0(long jarg1, G2 jarg1_, int jarg2);
+  public final static native String G2_toString__SWIG_1(long jarg1, G2 jarg1_);
+  public final static native void G2_deserialize(long jarg1, G2 jarg1_, byte[] jarg2);
+  public final static native byte[] G2_serialize(long jarg1, G2 jarg1_);
+  public final static native void delete_G2(long jarg1);
+  public final static native void mul__SWIG_4(long jarg1, GT jarg1_, long jarg2, GT jarg2_, long jarg3, GT jarg3_);
+  public final static native void inv__SWIG_2(long jarg1, GT jarg1_, long jarg2, GT jarg2_);
+  public final static native long new_GT__SWIG_0();
+  public final static native long new_GT__SWIG_1(long jarg1, GT jarg1_);
+  public final static native boolean GT_equals(long jarg1, GT jarg1_, long jarg2, GT jarg2_);
+  public final static native boolean GT_isOne(long jarg1, GT jarg1_);
+  public final static native void GT_clear(long jarg1, GT jarg1_);
+  public final static native void GT_setStr__SWIG_0(long jarg1, GT jarg1_, String jarg2, int jarg3);
+  public final static native void GT_setStr__SWIG_1(long jarg1, GT jarg1_, String jarg2);
+  public final static native String GT_toString__SWIG_0(long jarg1, GT jarg1_, int jarg2);
+  public final static native String GT_toString__SWIG_1(long jarg1, GT jarg1_);
+  public final static native void GT_deserialize(long jarg1, GT jarg1_, byte[] jarg2);
+  public final static native byte[] GT_serialize(long jarg1, GT jarg1_);
+  public final static native void delete_GT(long jarg1);
+}

ffi/java/com/herumi/mcl/PrivateKey.java

@@ -0,0 +1,87 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class PrivateKey {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected PrivateKey(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(PrivateKey obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        ElgamalJNI.delete_PrivateKey(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public String toStr() {
+    return ElgamalJNI.PrivateKey_toStr(swigCPtr, this);
+  }
+
+  public String toString() {
+    return ElgamalJNI.PrivateKey_toString(swigCPtr, this);
+  }
+
+  public void fromStr(String str) {
+    ElgamalJNI.PrivateKey_fromStr(swigCPtr, this, str);
+  }
+
+  public void save(String fileName) {
+    ElgamalJNI.PrivateKey_save(swigCPtr, this, fileName);
+  }
+
+  public void load(String fileName) {
+    ElgamalJNI.PrivateKey_load(swigCPtr, this, fileName);
+  }
+
+  public void init() {
+    ElgamalJNI.PrivateKey_init(swigCPtr, this);
+  }
+
+  public PublicKey getPublicKey() {
+    return new PublicKey(ElgamalJNI.PrivateKey_getPublicKey(swigCPtr, this), true);
+  }
+
+  public int dec(CipherText c, SWIGTYPE_p_bool b) {
+    return ElgamalJNI.PrivateKey_dec__SWIG_0(swigCPtr, this, CipherText.getCPtr(c), c, SWIGTYPE_p_bool.getCPtr(b));
+  }
+
+  public int dec(CipherText c) {
+    return ElgamalJNI.PrivateKey_dec__SWIG_1(swigCPtr, this, CipherText.getCPtr(c), c);
+  }
+
+  public void setCache(int rangeMin, int rangeMax) {
+    ElgamalJNI.PrivateKey_setCache(swigCPtr, this, rangeMin, rangeMax);
+  }
+
+  public void clearCache() {
+    ElgamalJNI.PrivateKey_clearCache(swigCPtr, this);
+  }
+
+  public PrivateKey() {
+    this(ElgamalJNI.new_PrivateKey(), true);
+  }
+
+}

ffi/java/com/herumi/mcl/PublicKey.java

@@ -0,0 +1,83 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class PublicKey {
+  private transient long swigCPtr;
+  protected transient boolean swigCMemOwn;
+
+  protected PublicKey(long cPtr, boolean cMemoryOwn) {
+    swigCMemOwn = cMemoryOwn;
+    swigCPtr = cPtr;
+  }
+
+  protected static long getCPtr(PublicKey obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+
+  @SuppressWarnings("deprecation")
+  protected void finalize() {
+    delete();
+  }
+
+  public synchronized void delete() {
+    if (swigCPtr != 0) {
+      if (swigCMemOwn) {
+        swigCMemOwn = false;
+        ElgamalJNI.delete_PublicKey(swigCPtr);
+      }
+      swigCPtr = 0;
+    }
+  }
+
+  public String toStr() {
+    return ElgamalJNI.PublicKey_toStr(swigCPtr, this);
+  }
+
+  public String toString() {
+    return ElgamalJNI.PublicKey_toString(swigCPtr, this);
+  }
+
+  public void fromStr(String str) {
+    ElgamalJNI.PublicKey_fromStr(swigCPtr, this, str);
+  }
+
+  public void save(String fileName) {
+    ElgamalJNI.PublicKey_save(swigCPtr, this, fileName);
+  }
+
+  public void load(String fileName) {
+    ElgamalJNI.PublicKey_load(swigCPtr, this, fileName);
+  }
+
+  public void enc(CipherText c, int m) {
+    ElgamalJNI.PublicKey_enc__SWIG_0(swigCPtr, this, CipherText.getCPtr(c), c, m);
+  }
+
+  public void enc(CipherText c, String str) {
+    ElgamalJNI.PublicKey_enc__SWIG_1(swigCPtr, this, CipherText.getCPtr(c), c, str);
+  }
+
+  public void rerandomize(CipherText c) {
+    ElgamalJNI.PublicKey_rerandomize(swigCPtr, this, CipherText.getCPtr(c), c);
+  }
+
+  public void add(CipherText c, int m) {
+    ElgamalJNI.PublicKey_add__SWIG_0(swigCPtr, this, CipherText.getCPtr(c), c, m);
+  }
+
+  public void add(CipherText c, String str) {
+    ElgamalJNI.PublicKey_add__SWIG_1(swigCPtr, this, CipherText.getCPtr(c), c, str);
+  }
+
+  public PublicKey() {
+    this(ElgamalJNI.new_PublicKey(), true);
+  }
+
+}

ffi/java/com/herumi/mcl/SWIGTYPE_p_bool.java

@@ -0,0 +1,26 @@
+/* ----------------------------------------------------------------------------
+ * This file was automatically generated by SWIG (http://www.swig.org).
+ * Version 4.0.2
+ *
+ * Do not make changes to this file unless you know what you are doing--modify
+ * the SWIG interface file instead.
+ * ----------------------------------------------------------------------------- */
+
+package com.herumi.mcl;
+
+public class SWIGTYPE_p_bool {
+  private transient long swigCPtr;
+
+  protected SWIGTYPE_p_bool(long cPtr, @SuppressWarnings("unused") boolean futureUse) {
+    swigCPtr = cPtr;
+  }
+
+  protected SWIGTYPE_p_bool() {
+    swigCPtr = 0;
+  }
+
+  protected static long getCPtr(SWIGTYPE_p_bool obj) {
+    return (obj == null) ? 0 : obj.swigCPtr;
+  }
+}
+

ffi/java/elgamal_impl.hpp

@@ -13,6 +13,11 @@ typedef mcl::FpT<mcl::ZnTag, 521> Zn;
 typedef mcl::EcT<Fp> Ec;
 typedef mcl::ElgamalT<Ec, Zn> Elgamal;
 
+#if defined(__GNUC__) && !defined(__EMSCRIPTEN__) && !defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdeprecated"
+#endif
+
 /*
 	init system
 	@param param [in] string such as "ecParamName hashName"
@@ -29,11 +34,11 @@ void SystemInit(const std::string& param) throw(std::exception)
 	if (iss >> ecParamStr >> hashNameStr) {
 		Param& p = Param::getParam();
 		p.ecParam = mcl::getEcParam(ecParamStr);
-		Zn::init(p.ecParam->n);
+		if (p.ecParam) {
-		Fp::init(p.ecParam->p);
+			mcl::initCurve<Ec, Zn>(p.ecParam->curveType);
-		Ec::init(p.ecParam->a, p.ecParam->b);
+			p.hashName = cybozu::crypto::Hash::getName(hashNameStr);
-		p.hashName = cybozu::crypto::Hash::getName(hashNameStr);
+			return;
-		return;
+		}
 	}
 	throw cybozu::Exception("SystemInit:bad param") << param;
 }
@@ -145,3 +150,7 @@ public:
 		self_.clearCache();
 	}
 };
+
+#if defined(__GNUC__) && !defined(__EMSCRIPTEN__) && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif

ffi/java/elgamal_wrap.cxx

@@ -1,6 +1,6 @@
 /* ----------------------------------------------------------------------------
  * This file was automatically generated by SWIG (http://www.swig.org).
- * Version 3.0.12
+ * Version 4.0.2
  *
  * This file is not intended to be easily readable and contains a number of
  * coding conventions designed to improve portability and efficiency. Do not make
@@ -182,15 +182,16 @@ template <typename T> T SwigValueInit() {
 
 /* Support for throwing Java exceptions */
 typedef enum {
-  SWIG_JavaOutOfMemoryError = 1, 
+  SWIG_JavaOutOfMemoryError = 1,
-  SWIG_JavaIOException, 
+  SWIG_JavaIOException,
-  SWIG_JavaRuntimeException, 
+  SWIG_JavaRuntimeException,
   SWIG_JavaIndexOutOfBoundsException,
   SWIG_JavaArithmeticException,
   SWIG_JavaIllegalArgumentException,
   SWIG_JavaNullPointerException,
   SWIG_JavaDirectorPureVirtual,
-  SWIG_JavaUnknownError
+  SWIG_JavaUnknownError,
+  SWIG_JavaIllegalStateException,
 } SWIG_JavaExceptionCodes;
 
 typedef struct {
@@ -211,6 +212,7 @@ static void SWIGUNUSED SWIG_JavaThrowException(JNIEnv *jenv, SWIG_JavaExceptionC
     { SWIG_JavaNullPointerException, "java/lang/NullPointerException" },
     { SWIG_JavaDirectorPureVirtual, "java/lang/RuntimeException" },
     { SWIG_JavaUnknownError,  "java/lang/UnknownError" },
+    { SWIG_JavaIllegalStateException, "java/lang/IllegalStateException" },
     { (SWIG_JavaExceptionCodes)0,  "java/lang/UnknownError" }
   };
   const SWIG_JavaExceptions_t *except_ptr = java_exceptions;
@@ -358,12 +360,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_SystemInit(JNIEnv *jenv,
   jenv->ReleaseStringUTFChars(jarg1, arg1_pstr); 
   try {
     SystemInit((std::string const &)*arg1);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -378,12 +378,10 @@ SWIGEXPORT jstring JNICALL Java_com_herumi_mcl_ElgamalJNI_CipherText_1toStr(JNIE
   arg1 = *(CipherText **)&jarg1; 
   try {
     result = ((CipherText const *)arg1)->toStr();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = jenv->NewStringUTF((&result)->c_str()); 
   return jresult;
 }
@@ -400,12 +398,10 @@ SWIGEXPORT jstring JNICALL Java_com_herumi_mcl_ElgamalJNI_CipherText_1toString(J
   arg1 = *(CipherText **)&jarg1; 
   try {
     result = ((CipherText const *)arg1)->toString();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = jenv->NewStringUTF((&result)->c_str()); 
   return jresult;
 }
@@ -430,12 +426,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_CipherText_1fromStr(JNIEn
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     (arg1)->fromStr((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -455,12 +449,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_CipherText_1add(JNIEnv *j
   } 
   try {
     (arg1)->add((CipherText const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -475,12 +467,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_CipherText_1mul_1_1SWIG_1
   arg2 = (int)jarg2; 
   try {
     (arg1)->mul(arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -503,12 +493,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_CipherText_1mul_1_1SWIG_1
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     (arg1)->mul((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -545,12 +533,10 @@ SWIGEXPORT jstring JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1toStr(JNIEn
   arg1 = *(PublicKey **)&jarg1; 
   try {
     result = ((PublicKey const *)arg1)->toStr();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = jenv->NewStringUTF((&result)->c_str()); 
   return jresult;
 }
@@ -567,12 +553,10 @@ SWIGEXPORT jstring JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1toString(JN
   arg1 = *(PublicKey **)&jarg1; 
   try {
     result = ((PublicKey const *)arg1)->toString();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = jenv->NewStringUTF((&result)->c_str()); 
   return jresult;
 }
@@ -597,12 +581,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1fromStr(JNIEnv
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     (arg1)->fromStr((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -625,12 +607,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1save(JNIEnv *j
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     ((PublicKey const *)arg1)->save((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -653,12 +633,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1load(JNIEnv *j
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     (arg1)->load((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -680,12 +658,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1enc_1_1SWIG_10
   arg3 = (int)jarg3; 
   try {
     ((PublicKey const *)arg1)->enc(*arg2,arg3);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -715,12 +691,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1enc_1_1SWIG_11
   jenv->ReleaseStringUTFChars(jarg3, arg3_pstr); 
   try {
     ((PublicKey const *)arg1)->enc(*arg2,(std::string const &)*arg3);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -740,12 +714,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1rerandomize(JN
   } 
   try {
     ((PublicKey const *)arg1)->rerandomize(*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -767,12 +739,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1add_1_1SWIG_10
   arg3 = (int)jarg3; 
   try {
     ((PublicKey const *)arg1)->add(*arg2,arg3);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -802,12 +772,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PublicKey_1add_1_1SWIG_11
   jenv->ReleaseStringUTFChars(jarg3, arg3_pstr); 
   try {
     ((PublicKey const *)arg1)->add(*arg2,(std::string const &)*arg3);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -844,12 +812,10 @@ SWIGEXPORT jstring JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1toStr(JNIE
   arg1 = *(PrivateKey **)&jarg1; 
   try {
     result = ((PrivateKey const *)arg1)->toStr();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = jenv->NewStringUTF((&result)->c_str()); 
   return jresult;
 }
@@ -866,12 +832,10 @@ SWIGEXPORT jstring JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1toString(J
   arg1 = *(PrivateKey **)&jarg1; 
   try {
     result = ((PrivateKey const *)arg1)->toString();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = jenv->NewStringUTF((&result)->c_str()); 
   return jresult;
 }
@@ -896,12 +860,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1fromStr(JNIEn
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     (arg1)->fromStr((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -924,12 +886,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1save(JNIEnv *
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     ((PrivateKey const *)arg1)->save((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -952,12 +912,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1load(JNIEnv *
   jenv->ReleaseStringUTFChars(jarg2, arg2_pstr); 
   try {
     (arg1)->load((std::string const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -970,12 +928,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1init(JNIEnv *
   arg1 = *(PrivateKey **)&jarg1; 
   try {
     (arg1)->init();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -990,12 +946,10 @@ SWIGEXPORT jlong JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1getPublicKey
   arg1 = *(PrivateKey **)&jarg1; 
   try {
     result = ((PrivateKey const *)arg1)->getPublicKey();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   *(PublicKey **)&jresult = new PublicKey((const PublicKey &)result); 
   return jresult;
 }
@@ -1021,12 +975,10 @@ SWIGEXPORT jint JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1dec_1_1SWIG_1
   arg3 = *(bool **)&jarg3; 
   try {
     result = (int)((PrivateKey const *)arg1)->dec((CipherText const &)*arg2,arg3);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = (jint)result; 
   return jresult;
 }
@@ -1050,12 +1002,10 @@ SWIGEXPORT jint JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1dec_1_1SWIG_1
   } 
   try {
     result = (int)((PrivateKey const *)arg1)->dec((CipherText const &)*arg2);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return 0;
   }
-  
   jresult = (jint)result; 
   return jresult;
 }
@@ -1074,12 +1024,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1setCache(JNIE
   arg3 = (int)jarg3; 
   try {
     (arg1)->setCache(arg2,arg3);
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 
@@ -1092,12 +1040,10 @@ SWIGEXPORT void JNICALL Java_com_herumi_mcl_ElgamalJNI_PrivateKey_1clearCache(JN
   arg1 = *(PrivateKey **)&jarg1; 
   try {
     (arg1)->clearCache();
-  }
+  } catch(std::exception &_e) {
-  catch(std::exception &_e) {
     SWIG_JavaThrowException(jenv, SWIG_JavaRuntimeException, (&_e)->what());
     return ;
   }
-  
 }
 
 

ffi/java/java.md

@@ -1,13 +1,14 @@
 # JNI for mcl (experimental)
 This library provides functionality to compute the optimal ate pairing
-over Barreto-Naehrig (BN) curves.
+over Barreto-Naehrig (BN) or BLS12-381 curves.
 
 # Initialization
-Load the library `mcl_bn256`.
+Load the library `mcljava`.
 ```
 import com.herumi.mcl.*;
 
-System.loadLibrary("mcl_bn256");
+System.loadLibrary(System.mapLibraryName("mcljava"));
+Mcl.SystemInit(curveType); // curveType = Mcl.BN254 or Mcl.BLS12_381
 ```
 
 # Classes
@@ -21,75 +22,67 @@ System.loadLibrary("mcl_bn256");
 ## Fr
 * `Fr::setInt(int x)` ; set by x
 * `Fr::setStr(String str)` ; set by str such as "123", "0xfff", etc.
-* `Fr::setRand()` ; randomly set
+* `Fr::setByCSPRNG()` ; randomly set
-* `Bn256.neg(Fr y, Fr x)` ; `y = -x`
+* `Mcl.neg(Fr y, Fr x)` ; `y = -x`
-* `Bn256.add(Fr z, Fr x, Fr y)` ; `z = x + y`
+* `Mcl.add(Fr z, Fr x, Fr y)` ; `z = x + y`
-* `Bn256.sub(Fr z, Fr x, Fr y)` ; `z = x - y`
+* `Mcl.sub(Fr z, Fr x, Fr y)` ; `z = x - y`
-* `Bn256.mul(Fr z, Fr x, Fr y)` ; `z = x * y`
+* `Mcl.mul(Fr z, Fr x, Fr y)` ; `z = x * y`
-* `Bn256.div(Fr z, Fr x, Fr y)` ; `z = x / y`
+* `Mcl.div(Fr z, Fr x, Fr y)` ; `z = x / y`
 
 ## G1
 
-* `G1::set(String x, String y)` ; set by (x, y)
+* `Mcl.neg(G1 y, G1 x)` ; `y = -x`
-* `G1::hashAndMapToG1(String m)` ; take SHA-256 of m and map it to an element of G1
+* `Mcl.dbl(G1 y, G1 x)` ; `y = 2x`
-* `G1::setStr(String str)` ; set by the result of `toString()` method
+* `Mcl.add(G1 z, G1 x, G1 y)` ; `z = x + y`
-* `Bn256.neg(G1 y, G1 x)` ; `y = -x`
+* `Mcl.sub(G1 z, G1 x, G1 y)` ; `z = x - y`
-* `Bn256.dbl(G1 y, G1 x)` ; `y = 2x`
+* `Mcl.mul(G1 z, G1 x, Fr y)` ; `z = x * y`
-* `Bn256.add(G1 z, G1 x, G1 y)` ; `z = x + y`
-* `Bn256.sub(G1 z, G1 x, G1 y)` ; `z = x - y`
-* `Bn256.mul(G1 z, G1 x, Fr y)` ; `z = x * y`
 
 ## G2
 
-* `G2::set(String xa, String xb, String ya, String yb)` ; set by ((xa, xb), (ya, yb))
+* `Mcl.neg(G2 y, G2 x)` ; `y = -x`
-* `G2::setStr(String str)` ; set by the result of `toString()` method
+* `Mcl.dbl(G2 y, G2 x)` ; `y = 2x`
-* `Bn256.neg(G2 y, G2 x)` ; `y = -x`
+* `Mcl.add(G2 z, G2 x, G2 y)` ; `z = x + y`
-* `Bn256.dbl(G2 y, G2 x)` ; `y = 2x`
+* `Mcl.sub(G2 z, G2 x, G2 y)` ; `z = x - y`
-* `Bn256.add(G2 z, G2 x, G2 y)` ; `z = x + y`
+* `Mcl.mul(G2 z, G2 x, Fr y)` ; `z = x * y`
-* `Bn256.sub(G2 z, G2 x, G2 y)` ; `z = x - y`
-* `Bn256.mul(G2 z, G2 x, Fr y)` ; `z = x * y`
 
 ## GT
 
 * `GT::setStr(String str)` ; set by the result of `toString()` method
-* `Bn256.mul(GT z, GT x, GT y)` ; `z = x * y`
+* `Mcl.mul(GT z, GT x, GT y)` ; `z = x * y`
-* `Bn256.pow(GT z, GT x, Fr y)` ; `z = x ^ y`
+* `Mcl.pow(GT z, GT x, Fr y)` ; `z = x ^ y`
+* `Mcl.inv(GT y, GT x)` ; `y = x^{-1}`
 
 ## pairing
-* `Bn256.pairing(GT e, G1 P, G2 Q)` ; e = e(P, Q)
+* `Mcl.pairing(GT e, G1 P, G2 Q)` ; e = e(P, Q)
 
 # BLS signature sample
 ```
-String xa = "12723517038133731887338407189719511622662176727675373276651903807414909099441";
+G2 Q = new G2();
-String xb = "4168783608814932154536427934509895782246573715297911553964171371032945126671";
+Mcl.hashAndMapToG2(Q, "abc".getBytes());
-String ya = "13891744915211034074451795021214165905772212241412891944830863846330766296736";
-String yb = "7937318970632701341203597196594272556916396164729705624521405069090520231616";
-
-G2 Q = new G2(xa, xb, ya, yb); // fixed point of G2
 
 Fr s = new Fr();
-s.setRand(); // secret key
+s.setByCSPRNG(); // secret key
 G2 pub = new G2();
-Bn256.mul(pub, Q, s); // public key = sQ
+Mcl.mul(pub, Q, s); // public key = sQ
 
 String m = "signature test";
 G1 H = new G1();
-H.hashAndMapToG1(m); // H = Hash(m)
+Mcl.hashAndMapToG1(H, m.getBytes()); // H = Hash(m)
 G1 sign = new G1();
-Bn256.mul(sign, H, s); // signature of m = s H
+Mcl.mul(sign, H, s); // signature of m = s H
 
 GT e1 = new GT();
 GT e2 = new GT();
-Bn256.pairing(e1, H, pub); // e1 = e(H, s Q)
+Mcl.pairing(e1, H, pub); // e1 = e(H, s Q)
-Bn256.pairing(e2, sign, Q); // e2 = e(s H, Q);
+Mcl.pairing(e2, sign, Q); // e2 = e(s H, Q);
 assertBool("verify signature", e1.equals(e2));
 ```
 
 # Make test
 ```
-cd java
+cd ffi/java
-make test_bn256
+make test
 ```
 
 # Sample code
-[Bn256Test.java](https://github.com/herumi/mcl/blob/master/java/Bn256Test.java)
+[MclTest.java](https://github.com/herumi/mcl/blob/master/ffi/java/MclTest.java)

ffi/java/make_wrap.bat

@@ -5,9 +5,11 @@ set SWIG=..\..\..\..\p\swig\swig.exe
 set PACKAGE_NAME=com.herumi.mcl
 set PACKAGE_DIR=%PACKAGE_NAME:.=\%
 if /i "%1"=="" (
-	set NAME=elgamal
+	set NAME=mcl
+	set LIBNAME=
 ) else (
 	set NAME=%1
+	set LIBNAME=%NAME%
 )
 
 echo [[run swig]]
@@ -15,7 +17,7 @@ mkdir %PACKAGE_DIR%
 set TOP_DIR=../..
 %SWIG% -java -package %PACKAGE_NAME% -outdir %PACKAGE_DIR% -c++ -Wall %NAME%.i
 echo [[make dll]]
-cl /MT /DNOMINMAX /LD /Ox /DNDEBUG /EHsc %NAME%_wrap.cxx %TOP_DIR%/src/fp.cpp -DMCL_NO_AUTOLINK -I%JAVA_INCLUDE% -I%JAVA_INCLUDE%\win32 -I%TOP_DIR%/include -I%TOP_DIR%/../cybozulib/include -I%TOP_DIR%/../cybozulib_ext/include -I%TOP_DIR%/../xbyak /link /LIBPATH:%TOP_DIR%/../cybozulib_ext/lib /OUT:%TOP_DIR%/bin/mcl_%NAME%.dll
+cl /MT /DNOMINMAX /LD /Ox /DNDEBUG /EHsc %NAME%_wrap.cxx %TOP_DIR%/src/fp.cpp -DMCL_NO_AUTOLINK -I%JAVA_INCLUDE% -I%JAVA_INCLUDE%\win32 -I%TOP_DIR%/include -I%TOP_DIR%/../cybozulib/include -I%TOP_DIR%/../cybozulib_ext/include -I%TOP_DIR%/../xbyak /link /LIBPATH:%TOP_DIR%/../cybozulib_ext/lib /OUT:%TOP_DIR%/bin/mcl%LIBNAME%java.dll
 
 call run-%NAME%.bat
 

ffi/java/mcl.i

@@ -0,0 +1,29 @@
+%module Mcl
+
+%include "std_string.i"
+%include "std_except.i"
+
+%apply(char *STRING, size_t LENGTH) { (const char *cbuf, size_t bufSize) };
+%{
+#include <mcl/bls12_381.hpp>
+
+#include "mcl_impl.hpp"
+
+%}
+
+%typemap(jtype) void serialize "byte[]"
+%typemap(jstype) void serialize "byte[]"
+%typemap(jni) void serialize "jbyteArray"
+%typemap(javaout) void serialize { return $jnicall; }
+%typemap(in, numinputs=0) std::string& out (std::string buf) "$1=&buf;"
+%typemap(argout) std::string& out {
+  $result = JCALL1(NewByteArray, jenv, $1->size());
+  JCALL4(SetByteArrayRegion, jenv, $result, 0, $1->size(), (const jbyte*)$1->c_str());
+}
+
+%include "mcl_impl.hpp"
+
+%javaconst(1);
+#define BN254 0
+#define BLS12_381 5
+

ffi/java/mcl_impl.hpp

@@ -0,0 +1,436 @@
+#include <mcl/bls12_381.hpp>
+#include <stdint.h>
+#include <sstream>
+
+#if defined(__GNUC__) && !defined(__EMSCRIPTEN__) && !defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdeprecated"
+#endif
+
+void SystemInit(int curveType) throw(std::exception)
+{
+	mcl::CurveParam cp;
+	switch (curveType) {
+	case MCL_BN254: cp = mcl::BN254; break;
+	case MCL_BN_SNARK1: cp = mcl::BN_SNARK1; break;
+	case MCL_BLS12_381: cp = mcl::BLS12_381; break;
+	default:
+		throw std::runtime_error("bad curveType");
+	}
+	mcl::bn::initPairing(cp);
+}
+
+template<class T>
+void deserializeT(T& x, const char *cbuf, size_t bufSize)
+{
+	if (x.deserialize(cbuf, bufSize) == 0) {
+		throw std::runtime_error("deserialize");
+	}
+}
+
+template<class T>
+void setLittleEndianModT(T& x, const char *cbuf, size_t bufSize)
+{
+	x.setLittleEndianMod(cbuf, bufSize);
+}
+
+template<class T>
+void setHashOfT(T& x, const char *cbuf, size_t bufSize)
+{
+	x.setHashOf(cbuf, bufSize);
+}
+
+template<class T>
+void serializeT(std::string& out, const T& x)
+{
+	out.resize(48 * 12);
+	size_t n = x.serialize(&out[0], out.size());
+	if (n == 0) throw std::runtime_error("serializeT");
+	out.resize(n);
+}
+
+class G1;
+class G2;
+class GT;
+/*
+	Fr = Z / rZ
+*/
+class Fr {
+	mcl::bn::Fr self_;
+	friend class G1;
+	friend class G2;
+	friend class GT;
+	friend void neg(Fr& y, const Fr& x);
+	friend void inv(Fr& y, const Fr& x);
+	friend void add(Fr& z, const Fr& x, const Fr& y);
+	friend void sub(Fr& z, const Fr& x, const Fr& y);
+	friend void mul(Fr& z, const Fr& x, const Fr& y);
+	friend void mul(G1& z, const G1& x, const Fr& y);
+	friend void mul(G2& z, const G2& x, const Fr& y);
+	friend void div(Fr& z, const Fr& x, const Fr& y);
+	friend void pow(GT& z, const GT& x, const Fr& y);
+public:
+	Fr() {}
+	Fr(const Fr& rhs) : self_(rhs.self_) {}
+	Fr(int x) : self_(x) {}
+	Fr(const std::string& str, int base = 0) throw(std::exception)
+		: self_(str, base) {}
+	bool equals(const Fr& rhs) const { return self_ == rhs.self_; }
+	bool isZero() const { return self_.isZero(); }
+	bool isOne() const { return self_.isOne(); }
+	void setStr(const std::string& str, int base = 0) throw(std::exception)
+	{
+		self_.setStr(str, base);
+	}
+	void setInt(int x)
+	{
+		self_ = x;
+	}
+	void clear()
+	{
+		self_.clear();
+	}
+	void setByCSPRNG()
+	{
+		self_.setByCSPRNG();
+	}
+	std::string toString(int base = 0) const throw(std::exception)
+	{
+		return self_.getStr(base);
+	}
+	void deserialize(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		deserializeT(self_, cbuf, bufSize);
+	}
+	void setLittleEndianMod(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		setLittleEndianModT(self_, cbuf, bufSize);
+	}
+	void setHashOf(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		setHashOfT(self_, cbuf, bufSize);
+	}
+	void serialize(std::string& out) const throw(std::exception)
+	{
+		serializeT(out, self_);
+	}
+};
+
+void neg(Fr& y, const Fr& x)
+{
+	mcl::bn::Fr::neg(y.self_, x.self_);
+}
+
+void inv(Fr& y, const Fr& x)
+{
+	mcl::bn::Fr::inv(y.self_, x.self_);
+}
+
+void add(Fr& z, const Fr& x, const Fr& y)
+{
+	mcl::bn::Fr::add(z.self_, x.self_, y.self_);
+}
+
+void sub(Fr& z, const Fr& x, const Fr& y)
+{
+	mcl::bn::Fr::sub(z.self_, x.self_, y.self_);
+}
+
+void mul(Fr& z, const Fr& x, const Fr& y)
+{
+	mcl::bn::Fr::mul(z.self_, x.self_, y.self_);
+}
+
+void div(Fr& z, const Fr& x, const Fr& y)
+{
+	mcl::bn::Fr::div(z.self_, x.self_, y.self_);
+}
+
+class Fp {
+	mcl::bn::Fp self_;
+	friend class G1;
+	friend class G2;
+	friend class GT;
+	friend void neg(Fp& y, const Fp& x);
+	friend void inv(Fp& y, const Fp& x);
+	friend void add(Fp& z, const Fp& x, const Fp& y);
+	friend void sub(Fp& z, const Fp& x, const Fp& y);
+	friend void mul(Fp& z, const Fp& x, const Fp& y);
+	friend void div(Fp& z, const Fp& x, const Fp& y);
+public:
+	Fp() {}
+	Fp(const Fp& rhs) : self_(rhs.self_) {}
+	Fp(int x) : self_(x) {}
+	Fp(const std::string& str, int base = 0) throw(std::exception)
+		: self_(str, base) {}
+	bool equals(const Fp& rhs) const { return self_ == rhs.self_; }
+	bool isZero() const { return self_.isZero(); }
+	bool isOne() const { return self_.isOne(); }
+	void setStr(const std::string& str, int base = 0) throw(std::exception)
+	{
+		self_.setStr(str, base);
+	}
+	void setInt(int x)
+	{
+		self_ = x;
+	}
+	void clear()
+	{
+		self_.clear();
+	}
+	void setByCSPRNG()
+	{
+		self_.setByCSPRNG();
+	}
+	std::string toString(int base = 0) const throw(std::exception)
+	{
+		return self_.getStr(base);
+	}
+	void deserialize(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		deserializeT(self_, cbuf, bufSize);
+	}
+	void serialize(std::string& out) const throw(std::exception)
+	{
+		serializeT(out, self_);
+	}
+};
+
+void neg(Fp& y, const Fp& x)
+{
+	mcl::bn::Fp::neg(y.self_, x.self_);
+}
+
+void inv(Fp& y, const Fp& x)
+{
+	mcl::bn::Fp::inv(y.self_, x.self_);
+}
+
+void add(Fp& z, const Fp& x, const Fp& y)
+{
+	mcl::bn::Fp::add(z.self_, x.self_, y.self_);
+}
+
+void sub(Fp& z, const Fp& x, const Fp& y)
+{
+	mcl::bn::Fp::sub(z.self_, x.self_, y.self_);
+}
+
+void mul(Fp& z, const Fp& x, const Fp& y)
+{
+	mcl::bn::Fp::mul(z.self_, x.self_, y.self_);
+}
+
+void div(Fp& z, const Fp& x, const Fp& y)
+{
+	mcl::bn::Fp::div(z.self_, x.self_, y.self_);
+}
+
+
+/*
+	#G1 = r
+*/
+class G1 {
+	mcl::bn::G1 self_;
+	friend void neg(G1& y, const G1& x);
+	friend void dbl(G1& y, const G1& x);
+	friend void add(G1& z, const G1& x, const G1& y);
+	friend void sub(G1& z, const G1& x, const G1& y);
+	friend void mul(G1& z, const G1& x, const Fr& y);
+	friend void pairing(GT& e, const G1& P, const G2& Q);
+	friend void hashAndMapToG1(G1& P, const char *cbuf, size_t bufSize) throw(std::exception);
+public:
+	G1() {}
+	G1(const G1& rhs) : self_(rhs.self_) {}
+	G1(const Fp& x, const Fp& y) throw(std::exception)
+		: self_(x.self_, y.self_) { }
+	bool equals(const G1& rhs) const { return self_ == rhs.self_; }
+	bool isZero() const { return self_.isZero(); }
+	bool isValidOrder() const { return self_.isValidOrder(); }
+	void set(const Fp& x, const Fp& y) throw(std::exception)
+	{
+		self_.set(x.self_, y.self_);
+	}
+	void clear()
+	{
+		self_.clear();
+	}
+	/*
+		compressed format
+	*/
+	void setStr(const std::string& str, int base = 0) throw(std::exception)
+	{
+		self_.setStr(str, base);
+	}
+	std::string toString(int base = 0) const throw(std::exception)
+	{
+		return self_.getStr(base);
+	}
+	void deserialize(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		deserializeT(self_, cbuf, bufSize);
+	}
+	void serialize(std::string& out) const throw(std::exception)
+	{
+		serializeT(out, self_);
+	}
+};
+
+void neg(G1& y, const G1& x)
+{
+	mcl::bn::G1::neg(y.self_, x.self_);
+}
+void dbl(G1& y, const G1& x)
+{
+	mcl::bn::G1::dbl(y.self_, x.self_);
+}
+void add(G1& z, const G1& x, const G1& y)
+{
+	mcl::bn::G1::add(z.self_, x.self_, y.self_);
+}
+void sub(G1& z, const G1& x, const G1& y)
+{
+	mcl::bn::G1::sub(z.self_, x.self_, y.self_);
+}
+void mul(G1& z, const G1& x, const Fr& y)
+{
+	mcl::bn::G1::mul(z.self_, x.self_, y.self_);
+}
+
+/*
+	#G2 = r
+*/
+class G2 {
+	mcl::bn::G2 self_;
+	friend void neg(G2& y, const G2& x);
+	friend void dbl(G2& y, const G2& x);
+	friend void add(G2& z, const G2& x, const G2& y);
+	friend void sub(G2& z, const G2& x, const G2& y);
+	friend void mul(G2& z, const G2& x, const Fr& y);
+	friend void pairing(GT& e, const G1& P, const G2& Q);
+	friend void hashAndMapToG2(G2& P, const char *cbuf, size_t bufSize) throw(std::exception);
+public:
+	G2() {}
+	G2(const G2& rhs) : self_(rhs.self_) {}
+	G2(const Fp& ax, const Fp& ay, const Fp& bx, const Fp& by) throw(std::exception)
+		: self_(mcl::bn::Fp2(ax.self_, ay.self_), mcl::bn::Fp2(bx.self_, by.self_))
+	{
+	}
+	bool equals(const G2& rhs) const { return self_ == rhs.self_; }
+	bool isZero() const { return self_.isZero(); }
+	void set(const Fp& ax, const Fp& ay, const Fp& bx, const Fp& by) throw(std::exception)
+	{
+		self_.set(mcl::bn::Fp2(ax.self_, ay.self_), mcl::bn::Fp2(bx.self_, by.self_));
+	}
+	void clear()
+	{
+		self_.clear();
+	}
+	/*
+		compressed format
+	*/
+	void setStr(const std::string& str, int base = 0) throw(std::exception)
+	{
+		self_.setStr(str, base);
+	}
+	std::string toString(int base = 0) const throw(std::exception)
+	{
+		return self_.getStr(base);
+	}
+	void deserialize(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		deserializeT(self_, cbuf, bufSize);
+	}
+	void serialize(std::string& out) const throw(std::exception)
+	{
+		serializeT(out, self_);
+	}
+};
+
+void neg(G2& y, const G2& x)
+{
+	mcl::bn::G2::neg(y.self_, x.self_);
+}
+void dbl(G2& y, const G2& x)
+{
+	mcl::bn::G2::dbl(y.self_, x.self_);
+}
+void add(G2& z, const G2& x, const G2& y)
+{
+	mcl::bn::G2::add(z.self_, x.self_, y.self_);
+}
+void sub(G2& z, const G2& x, const G2& y)
+{
+	mcl::bn::G2::sub(z.self_, x.self_, y.self_);
+}
+void mul(G2& z, const G2& x, const Fr& y)
+{
+	mcl::bn::G2::mul(z.self_, x.self_, y.self_);
+}
+
+/*
+	#GT = r
+*/
+class GT {
+	mcl::bn::Fp12 self_;
+	friend void mul(GT& z, const GT& x, const GT& y);
+	friend void inv(GT& y, GT& x);
+	friend void pow(GT& z, const GT& x, const Fr& y);
+	friend void pairing(GT& e, const G1& P, const G2& Q);
+public:
+	GT() {}
+	GT(const GT& rhs) : self_(rhs.self_) {}
+	bool equals(const GT& rhs) const { return self_ == rhs.self_; }
+	bool isOne() const { return self_.isOne(); }
+	void clear()
+	{
+		self_.clear();
+	}
+	void setStr(const std::string& str, int base = 0) throw(std::exception)
+	{
+		self_.setStr(str, base);
+	}
+	std::string toString(int base = 0) const throw(std::exception)
+	{
+		return self_.getStr(base);
+	}
+	void deserialize(const char *cbuf, size_t bufSize) throw(std::exception)
+	{
+		deserializeT(self_, cbuf, bufSize);
+	}
+	void serialize(std::string& out) const throw(std::exception)
+	{
+		serializeT(out, self_);
+	}
+};
+
+void mul(GT& z, const GT& x, const GT& y)
+{
+	mcl::bn::Fp12::mul(z.self_, x.self_, y.self_);
+}
+void inv(GT& y, GT& x)
+{
+	mcl::bn::Fp12::unitaryInv(y.self_, x.self_);
+}
+void pow(GT& z, const GT& x, const Fr& y)
+{
+	mcl::bn::Fp12::pow(z.self_, x.self_, y.self_);
+}
+void pairing(GT& e, const G1& P, const G2& Q)
+{
+	mcl::bn::pairing(e.self_, P.self_, Q.self_);
+}
+
+void hashAndMapToG1(G1& P, const char *cbuf, size_t bufSize) throw(std::exception)
+{
+	mcl::bn::hashAndMapToG1(P.self_, cbuf, bufSize);
+}
+
+void hashAndMapToG2(G2& P, const char *cbuf, size_t bufSize) throw(std::exception)
+{
+	mcl::bn::hashAndMapToG2(P.self_, cbuf, bufSize);
+}
+
+#if defined(__GNUC__) && !defined(__EMSCRIPTEN__) && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif

ffi/java/run-bn256.bat

@@ -1,9 +0,0 @@
-@echo off
-echo [[compile Bn256Test.java]]
-%JAVA_DIR%\bin\javac Bn256Test.java
-
-echo [[run Bn256Test]]
-set TOP_DIR=..\..
-pushd %TOP_DIR%\bin
-%JAVA_DIR%\bin\java -classpath ../ffi/java Bn256Test %1 %2 %3 %4 %5 %6
-popd

ffi/java/run-mcl.bat

@@ -0,0 +1,9 @@
+@echo off
+echo [[compile MclTest.java]]
+%JAVA_DIR%\bin\javac MclTest.java
+
+echo [[run MclTest]]
+set TOP_DIR=..\..
+pushd %TOP_DIR%\bin
+%JAVA_DIR%\bin\java -classpath ../ffi/java MclTest %1 %2 %3 %4 %5 %6
+popd

ffi/python/Makefile

@@ -0,0 +1,9 @@
+include ../../common.mk
+
+SHE384_256_SLIB=libmclshe384_256.$(LIB_SUF)
+
+she_test: ../../lib/$(SHE384_256_SLIB)
+	cd ../../lib && env LD_LIBRARY_PATH=./ python3 ../ffi/python/she.py
+
+../../lib/$(SHE384_256_SLIB):
+	make -C ../../ lib/$(SHE384_256_SLIB)

ffi/python/readme.md

@@ -0,0 +1,7 @@
+# sample for Python
+
+## SHE
+
+```
+make she_test
+```

ffi/python/she.py

@@ -1,6 +1,7 @@
 import os
 import platform
 from ctypes import *
+#from ctypes.util import find_library
 
 BN254 = 0
 BLS12_381 = 5
@@ -34,6 +35,7 @@ def init(curveType=BN254):
 		libName = 'mclshe384_256.dll'
 	else:
 		raise RuntimeError("not support yet", name)
+#	lib = cdll.LoadLibrary(find_library(libName))
 	lib = cdll.LoadLibrary(libName)
 	ret = lib.sheInit(curveType, MCLBN_COMPILED_TIME_VAR)
 	if ret != 0:

include/cybozu/atoi.hpp

@@ -7,7 +7,6 @@
 */
 
 #include <memory.h>
-#include <limits.h>
 #include <limits>
 #include <cybozu/exception.hpp>
 
@@ -173,7 +172,7 @@ public:
 	}
 	inline operator int() const
 	{
-		return atoi_local::convertToInt<int>(b_, p_, size_, "2147483648", INT_MIN, 214748364, '8');
+		return atoi_local::convertToInt<int>(b_, p_, size_, "2147483648", /*INT_MIN*/-2147483648, 214748364, '8');
 	}
 	inline operator unsigned int() const
 	{
@@ -181,7 +180,7 @@ public:
 	}
 	inline operator long long() const
 	{
-		return atoi_local::convertToInt<long long>(b_, p_, size_, "9223372036854775808", LLONG_MIN, 922337203685477580LL, '8');
+		return atoi_local::convertToInt<long long>(b_, p_, size_, "9223372036854775808", CYBOZU_LLONG_MIN, 922337203685477580LL, '8');
 	}
 	inline operator unsigned long long() const
 	{

include/cybozu/atomic.hpp

@@ -0,0 +1,146 @@
+#pragma once
+/**
+	@file
+	@brief atomic operation
+
+	@author MITSUNARI Shigeo(@herumi)
+	@author MITSUNARI Shigeo
+*/
+#include <cybozu/inttype.hpp>
+#ifdef _WIN32
+#ifndef WIN32_LEAN_AND_MEAN
+	#define WIN32_LEAN_AND_MEAN
+#endif
+#include <windows.h>
+#include <intrin.h>
+#else
+#include <emmintrin.h>
+#endif
+
+namespace cybozu {
+
+namespace atomic_local {
+
+template<size_t S>
+struct Tag {};
+
+template<>
+struct Tag<4> {
+	template<class T>
+	static inline T AtomicAddSub(T *p, T y)
+	{
+#ifdef _WIN32
+		return (T)_InterlockedExchangeAdd((long*)p, (long)y);
+#else
+		return static_cast<T>(__sync_fetch_and_add(p, y));
+#endif
+	}
+
+	template<class T>
+	static inline T AtomicCompareExchangeSub(T *p, T newValue, T oldValue)
+	{
+#ifdef _WIN32
+		return (T)_InterlockedCompareExchange((long*)p, (long)newValue, (long)oldValue);
+#else
+		return static_cast<T>(__sync_val_compare_and_swap(p, oldValue, newValue));
+#endif
+	}
+
+	template<class T>
+	static inline T AtomicExchangeSub(T *p, T newValue)
+	{
+#ifdef _WIN32
+		return (T)_InterlockedExchange((long*)p, (long)newValue);
+#else
+		return static_cast<T>(__sync_lock_test_and_set(p, newValue));
+#endif
+	}
+};
+
+template<>
+struct Tag<8> {
+#if (CYBOZU_OS_BIT == 64)
+	template<class T>
+	static inline T AtomicAddSub(T *p, T y)
+	{
+#ifdef _WIN32
+		return (T)_InterlockedExchangeAdd64((int64_t*)p, (int64_t)y);
+#else
+		return static_cast<T>(__sync_fetch_and_add(p, y));
+#endif
+	}
+#endif
+
+	template<class T>
+	static inline T AtomicCompareExchangeSub(T *p, T newValue, T oldValue)
+	{
+#ifdef _WIN32
+		return (T)_InterlockedCompareExchange64((int64_t*)p, (int64_t)newValue, (int64_t)oldValue);
+#else
+		return static_cast<T>(__sync_val_compare_and_swap(p, oldValue, newValue));
+#endif
+	}
+
+#if (CYBOZU_OS_BIT == 64)
+	template<class T>
+	static inline T AtomicExchangeSub(T *p, T newValue)
+	{
+#ifdef _WIN32
+		return (T)_InterlockedExchange64((int64_t*)p, (int64_t)newValue);
+#else
+		return static_cast<T>(__sync_lock_test_and_set(p, newValue));
+#endif
+	}
+#endif
+};
+
+} // atomic_local
+
+/**
+	atomic operation
+	see http://gcc.gnu.org/onlinedocs/gcc-4.4.0/gcc/Atomic-Builtins.html
+	http://msdn.microsoft.com/en-us/library/ms683504(VS.85).aspx
+*/
+/**
+	tmp = *p;
+	*p += y;
+	return tmp;
+*/
+template<class T>
+T AtomicAdd(T *p, T y)
+{
+	return atomic_local::Tag<sizeof(T)>::AtomicAddSub(p, y);
+}
+
+/**
+	tmp = *p;
+	if (*p == oldValue) *p = newValue;
+	return tmp;
+*/
+template<class T>
+T AtomicCompareExchange(T *p, T newValue, T oldValue)
+{
+	return atomic_local::Tag<sizeof(T)>::AtomicCompareExchangeSub(p, newValue, oldValue);
+}
+
+/**
+	tmp = *p;
+	*p = newValue;
+	return tmp;
+*/
+template<class T>
+T AtomicExchange(T *p, T newValue)
+{
+	return atomic_local::Tag<sizeof(T)>::AtomicExchangeSub(p, newValue);
+}
+
+inline void mfence()
+{
+#ifdef _MSC_VER
+	MemoryBarrier();
+#else
+	_mm_mfence();
+#endif
+}
+
+} // cybozu

include/cybozu/exception.hpp

@@ -42,7 +42,9 @@ public:
 #include <errno.h>
 #include <stdio.h>
 #ifdef _WIN32
-	#include <winsock2.h>
+	#ifndef WIN32_LEAN_AND_MEAN
+		#define WIN32_LEAN_AND_MEAN
+	#endif
 	#include <windows.h>
 #else
 	#include <string.h> // for strerror_r
@@ -201,7 +203,6 @@ public:
 	}
 	/**
 		convert NativeErrNo to string(maybe UTF8)
-		@param err [in] errno
 		@note Linux   : same as ConvertErrorNoToString
 			  Windows : for Win32 API(use en-us)
 	*/

include/cybozu/file.hpp

@@ -0,0 +1,626 @@
+#pragma once
+/**
+	@file
+	@brief file class and operations
+
+	@author MITSUNARI Shigeo(@herumi)
+	@remark mingw requires -lshlwapi option
+*/
+
+#include <assert.h>
+#include <sys/stat.h> // for stat
+#include <cybozu/exception.hpp>
+#include <vector>
+#include <ios>
+#ifdef _WIN32
+	#include <shlwapi.h>
+	#include <io.h>
+	#include <fcntl.h>
+	#include <shlobj.h>
+	#include <direct.h>
+	#ifndef WIN32_LEAN_AND_MEAN
+		#define WIN32_LEAN_AND_MEAN
+	#endif
+	#include <windows.h>
+	#ifdef _MSC_VER
+		#pragma comment(lib, "shlwapi.lib")
+		#pragma comment(lib, "shell32.lib")
+		#pragma comment(lib, "User32.lib")
+	#endif
+#else
+	#include <stdio.h>
+	#include <unistd.h>
+	#include <sys/types.h>
+	#include <fcntl.h>
+	#include <dirent.h>
+#endif
+#ifdef __APPLE__
+	#include <mach-o/dyld.h>
+#endif
+
+namespace cybozu {
+
+class File {
+	std::string name_; // used for only errmsg
+#ifdef _WIN32
+	typedef HANDLE handleType;
+#else
+	typedef int handleType;
+	enum {
+		INVALID_HANDLE_VALUE = -1
+	};
+#endif
+	handleType hdl_;
+	bool doClose_;
+	bool isReadOnly_;
+	File(const File&);
+	void operator=(const File&);
+#ifdef _WIN32
+	void openSub()
+	{
+	}
+#endif
+	void verifyMode(std::ios::openmode mode)
+	{
+		doClose_ = true;
+		bool isCorrectMode = true;
+		if (!!(mode & std::ios::in) == !!(mode & std::ios::out)) {
+			isCorrectMode = false;
+		} else {
+			if (mode & std::ios::in) {
+				isReadOnly_ = true;
+				if ((mode & std::ios::app) || (mode & std::ios::trunc)) isCorrectMode = false;
+			} else {
+				isReadOnly_ = false;
+				if ((mode & std::ios::app) && (mode & std::ios::trunc)) isCorrectMode = false;
+			}
+		}
+		if (!isCorrectMode) {
+			throw cybozu::Exception("File:open:bad mode") << name_ << mode;
+		}
+	}
+#ifdef _WIN32
+	HANDLE createFile(const std::string& name, DWORD access, DWORD share, DWORD disposition)
+	{
+		return  ::CreateFileA(name.c_str(), access, share, NULL, disposition, FILE_ATTRIBUTE_NORMAL, NULL);
+	}
+	HANDLE createFile(const std::wstring& name, DWORD access, DWORD share, DWORD disposition)
+	{
+		return  ::CreateFileW(name.c_str(), access, share, NULL, disposition, FILE_ATTRIBUTE_NORMAL, NULL);
+	}
+	template<class T>
+	void setHandle(const T& name, std::ios::openmode mode)
+	{
+		DWORD access = GENERIC_READ;
+		DWORD disposition = OPEN_EXISTING;
+		DWORD share = FILE_SHARE_READ | FILE_SHARE_WRITE;
+		if (mode & std::ios::out) {
+			access = GENERIC_WRITE;
+			disposition = CREATE_ALWAYS;
+			if (mode & std::ios::app) {
+				disposition = OPEN_ALWAYS;
+			}
+		}
+		hdl_ = createFile(name, access, share, disposition);
+	}
+#else
+	void setHandle(const std::string& name, std::ios::openmode mode)
+	{
+		int flags = O_RDONLY; // | O_NOATIME; /* can't use on NFS */
+		mode_t access = 0644;
+		if (mode & std::ios::out) {
+			flags = O_WRONLY | O_CREAT;
+			if (mode & std::ios::app) {
+				flags |= O_APPEND;
+			} else {
+				flags |= O_TRUNC;
+			}
+		}
+		hdl_ = ::open(name.c_str(), flags, access);
+	}
+#endif
+	template<class T>
+	void openSub(const T& name, std::ios::openmode mode)
+	{
+		if (isOpen()) throw cybozu::Exception("File:open:alread opened") << name_;
+		verifyMode(mode);
+		setHandle(name, mode);
+		if (isOpen()) {
+			if (mode & std::ios::app) {
+				seek(getSize(), std::ios::beg);
+			}
+			return;
+		}
+		throw cybozu::Exception("File:open") << name_ << cybozu::ErrorNo() << static_cast<int>(mode);
+	}
+public:
+	File()
+		: hdl_(INVALID_HANDLE_VALUE)
+		, doClose_(true)
+		, isReadOnly_(false)
+	{
+	}
+	File(const std::string& name, std::ios::openmode mode)
+		: hdl_(INVALID_HANDLE_VALUE)
+		, doClose_(true)
+		, isReadOnly_(false)
+	{
+		open(name, mode);
+	}
+	/*
+		construct with file handle
+		@param hdl [in] file handle
+	*/
+	explicit File(handleType hdl)
+		: hdl_(hdl)
+		, doClose_(false)
+		, isReadOnly_(false)
+
+	{
+	}
+	~File() CYBOZU_NOEXCEPT
+	{
+		if (!doClose_) return;
+		try {
+			sync();
+			close();
+		} catch (std::exception& e) {
+			fprintf(stderr, "File:dstr:%s\n", e.what());
+		} catch (...) {
+			fprintf(stderr, "File:dstr:unknown\n");
+		}
+	}
+	bool isOpen() const CYBOZU_NOEXCEPT { return hdl_ != INVALID_HANDLE_VALUE; }
+	/**
+		support mode
+		always binary mode
+		ios::in : read only
+		ios::out : write only(+truncate)
+		ios::out + ios::app(append)
+	*/
+	void open(const std::string& name, std::ios::openmode mode)
+	{
+		name_ = name;
+		openSub(name, mode);
+	}
+	void openW(const std::string& name)
+	{
+		open(name, std::ios::out | std::ios::binary | std::ios::trunc);
+	}
+	void openR(const std::string& name)
+	{
+		open(name, std::ios::in | std::ios::binary);
+	}
+#ifdef _WIN32
+	void open(const std::wstring& name, std::ios::openmode mode)
+	{
+		name_ = cybozu::exception::wstr2str(name);
+		openSub(name, mode);
+	}
+	File(const std::wstring& name, std::ios::openmode mode)
+		: hdl_(INVALID_HANDLE_VALUE)
+	{
+		open(name, mode);
+	}
+	void openW(const std::wstring& name)
+	{
+		open(name, std::ios::out | std::ios::binary | std::ios::trunc);
+	}
+	void openR(const std::wstring& name)
+	{
+		open(name, std::ios::in | std::ios::binary);
+	}
+#endif
+	void close()
+	{
+		if (!isOpen()) return;
+#ifdef _WIN32
+		bool isOK = ::CloseHandle(hdl_) != 0;
+#else
+		bool isOK = ::close(hdl_) == 0;
+#endif
+		hdl_ = INVALID_HANDLE_VALUE;
+		if (isOK) return;
+		throw cybozu::Exception("File:close") << name_ << cybozu::ErrorNo();
+	}
+	/*
+		sync
+		@param doFullSync [in] call sync(for only Linux)
+	*/
+	void sync(bool doFullSync = false)
+	{
+		cybozu::disable_warning_unused_variable(doFullSync);
+		if (!isOpen()) return;
+		if (isReadOnly_) return;
+#ifdef _WIN32
+		/* fail if isReadOnly_ */
+		if (!::FlushFileBuffers(hdl_)) goto ERR_EXIT;
+#elif defined(__linux__) || defined(__CYGWIN__)
+		if (doFullSync) {
+			if (::fsync(hdl_)) goto ERR_EXIT;
+		} else {
+			if (::fdatasync(hdl_)) goto ERR_EXIT;
+		}
+#else
+		if (::fcntl(hdl_, F_FULLFSYNC)) goto ERR_EXIT;
+#endif
+		return;
+	ERR_EXIT:
+		throw cybozu::Exception("File:sync") << name_ << cybozu::ErrorNo();
+	}
+	void write(const void *buf, size_t bufSize)
+	{
+		const char *p = static_cast<const char *>(buf);
+		while (bufSize > 0) {
+			uint32_t size = static_cast<uint32_t>(std::min<size_t>(0x7fffffff, bufSize));
+#ifdef _WIN32
+			DWORD writeSize;
+			if (!::WriteFile(hdl_, p, size, &writeSize, NULL)) goto ERR_EXIT;
+#else
+			ssize_t writeSize = ::write(hdl_, p, size);
+			if (writeSize < 0) {
+				if (errno == EINTR) continue;
+				goto ERR_EXIT;
+			}
+#endif
+			p += writeSize;
+			bufSize -= writeSize;
+		}
+		return;
+	ERR_EXIT:
+		throw cybozu::Exception("File:write") << name_ << cybozu::ErrorNo();
+	}
+	size_t readSome(void *buf, size_t bufSize)
+	{
+		uint32_t size = static_cast<uint32_t>(std::min<size_t>(0x7fffffff, bufSize));
+#ifdef _WIN32
+		DWORD readSize;
+		if (!::ReadFile(hdl_, buf, size, &readSize, NULL)) goto ERR_EXIT;
+#else
+	RETRY:
+		ssize_t readSize = ::read(hdl_, buf, size);
+		if (readSize < 0) {
+			if (errno == EINTR) goto RETRY;
+			goto ERR_EXIT;
+		}
+#endif
+		return readSize;
+	ERR_EXIT:
+		throw cybozu::Exception("File:read") << name_ << cybozu::ErrorNo();
+	}
+	void read(void *buf, size_t bufSize)
+	{
+		char *p = static_cast<char *>(buf);
+		while (bufSize > 0) {
+			size_t readSize = readSome(p, bufSize);
+			p += readSize;
+			bufSize -= readSize;
+		}
+	}
+	void seek(int64_t pos, std::ios::seek_dir dir)
+	{
+#ifdef _WIN32
+		LARGE_INTEGER largePos;
+		largePos.QuadPart = pos;
+		DWORD posMode = FILE_BEGIN;
+		switch (dir) {
+		case std::ios::beg:
+			posMode = FILE_BEGIN;
+			break;
+		case std::ios::cur:
+		default:
+			posMode = FILE_CURRENT;
+			break;
+		case std::ios::end:
+			posMode = FILE_END;
+			break;
+		}
+		bool isOK = SetFilePointerEx(hdl_, largePos, NULL, posMode) != 0;
+#else
+		int whence;
+		switch (dir) {
+		case std::ios::beg:
+			whence = SEEK_SET;
+			break;
+		case std::ios::cur:
+		default:
+			whence = SEEK_CUR;
+			break;
+		case std::ios::end:
+			whence = SEEK_END;
+			break;
+		}
+		bool isOK = lseek(hdl_, pos, whence) >= 0;
+#endif
+		if (isOK) return;
+		throw cybozu::Exception("File:seek") << name_ << cybozu::ErrorNo() << pos << static_cast<int>(dir);
+	}
+	uint64_t getSize() const
+	{
+		uint64_t fileSize;
+#ifdef _WIN32
+		LARGE_INTEGER size;
+		bool isOK = GetFileSizeEx(hdl_, &size) != 0;
+		fileSize = size.QuadPart;
+#else
+		struct stat stat;
+		bool isOK = fstat(hdl_, &stat) == 0;
+		fileSize = stat.st_size;
+#endif
+		if (isOK) return fileSize;
+		throw cybozu::Exception("File:getSize") << name_ << cybozu::ErrorNo();
+	}
+};
+
+/*
+	name has extension
+*/
+inline bool HasExtension(const std::string& name, const std::string& extension)
+{
+	const size_t extensionSize = extension.size();
+	if (extensionSize == 0) return true;
+	const size_t nameSize = name.size();
+	if (nameSize < extensionSize + 1) return false;
+	const char *p = &name[nameSize - extensionSize - 1];
+	return *p == '.' && memcmp(p + 1, &extension[0], extensionSize) == 0;
+}
+/*
+	split name as basename.suffix
+*/
+inline std::string GetBaseName(const std::string& name, std::string *suffix = 0)
+{
+	size_t pos = name.find_last_of('.');
+	if (pos == std::string::npos) {
+		if (suffix) suffix->clear();
+		return name;
+	}
+	if (suffix) {
+		*suffix = name.substr(pos + 1);
+	}
+	return name.substr(0, pos);
+}
+
+/**
+	replace \ with /
+*/
+inline void ReplaceBackSlash(std::string& str)
+{
+	for (size_t i = 0, n = str.size(); i < n; i++) {
+		if (str[i] == '\\') str[i] = '/';
+	}
+}
+
+/**
+	get exe path and baseNamme
+	@note file name is the form "xxx.exe" then baseName = xxx
+*/
+inline std::string GetExePath(std::string *baseName = 0)
+{
+	std::string path;
+	path.resize(4096);
+#ifdef _WIN32
+	if (!GetModuleFileNameA(NULL, &path[0], static_cast<int>(path.size()) - 2)) {
+		return "";
+	}
+	PathRemoveExtensionA(&path[0]);
+	if (baseName) {
+		*baseName = PathFindFileNameA(&path[0]);
+	}
+	if (::PathRemoveFileSpecA(&path[0])) {
+		::PathAddBackslashA(&path[0]);
+		path[0] = static_cast<char>(tolower(path[0]));
+		path.resize(strlen(&path[0]));
+		ReplaceBackSlash(path);
+	}
+#else
+#if defined(__APPLE__)
+	uint32_t size = (uint32_t)path.size();
+	if (_NSGetExecutablePath(&path[0], &size) != 0) {
+		return "";
+	}
+	path.resize(strlen(&path[0]));
+#else
+	int ret = readlink("/proc/self/exe", &path[0], path.size() - 2);
+	if (ret < 0) return "";
+	path.resize(ret);
+#endif
+	size_t pos = path.find_last_of('/');
+	if (pos != std::string::npos) {
+		if (baseName) {
+			const std::string name = path.substr(pos + 1);
+			std::string suffix;
+			std::string base = GetBaseName(name, &suffix);
+			if (suffix == "exe") {
+				*baseName = base;
+			} else {
+				*baseName = name;
+			}
+		}
+		path.resize(pos + 1);
+	}
+#endif
+	return path;
+}
+
+/**
+	get file size
+*/
+inline uint64_t GetFileSize(const std::string& name)
+{
+#ifdef _WIN32
+	struct __stat64 buf;
+	bool isOK = _stat64(name.c_str(), &buf) == 0;
+#else
+	struct stat buf;
+	bool isOK = stat(name.c_str(), &buf) == 0;
+#endif
+	if (isOK) return buf.st_size;
+	throw cybozu::Exception("GetFileSize") << name << cybozu::ErrorNo();
+}
+
+/**
+	verify whether path exists or not
+*/
+inline bool DoesFileExist(const std::string& path)
+{
+	if (path.empty()) return false;
+	std::string p = path;
+	char c = p[p.size() - 1];
+	if (c == '/' || c == '\\') {
+		p.resize(p.size() - 1);
+	}
+#ifdef _WIN32
+	struct _stat buf;
+	return _stat(p.c_str(), &buf) == 0;
+#else
+	struct stat buf;
+	return stat(p.c_str(), &buf) == 0;
+#endif
+}
+
+inline void RenameFile(const std::string& from, const std::string& to)
+{
+	if (DoesFileExist(to)) {
+		throw cybozu::Exception("RenameFile:file already exist") << from << to;
+	}
+#ifdef _WIN32
+	bool isOK = ::MoveFileExA(from.c_str(), to.c_str(), MOVEFILE_COPY_ALLOWED | MOVEFILE_WRITE_THROUGH) != 0;
+#else
+	bool isOK = ::rename(from.c_str(), to.c_str()) == 0;
+#endif
+	if (!isOK) {
+		throw cybozu::Exception("RenameFile") << from << to << cybozu::ErrorNo();
+	}
+}
+
+/**
+	remove file
+*/
+inline void RemoveFile(const std::string& name)
+{
+#ifdef _WIN32
+	bool isOK = DeleteFileA(name.c_str()) != 0;
+#else
+	bool isOK = unlink(name.c_str()) == 0;
+#endif
+	if (!isOK) {
+		throw cybozu::Exception("RemoveFile") << name << cybozu::ErrorNo();
+	}
+}
+
+/*
+	remark of isFile()
+	not directory on Windows
+	not contain symbolic link on Linux
+*/
+struct FileInfo {
+	std::string name;
+	uint32_t attr; // dwFileAttributes for Windows, d_type for Linux
+#ifdef _WIN32
+	bool isUnknown() const { return attr == 0; }
+	bool isDirectory() const { verify(); return (attr & FILE_ATTRIBUTE_DIRECTORY) != 0; }
+	bool isFile() const { verify(); return !isDirectory(); }
+#else
+	bool isUnknown() const { return attr == DT_UNKNOWN; }
+	bool isDirectory() const { verify(); return attr == DT_DIR; }
+	bool isFile() const { verify(); return attr == DT_REG; }
+#endif
+	FileInfo() : attr(0) {}
+	FileInfo(const std::string& name, uint32_t attr) : name(name), attr(attr) {}
+	void verify() const
+	{
+		if (isUnknown()) throw cybozu::Exception("FileInfo:unknown attr") << name;
+	}
+};
+
+typedef std::vector<FileInfo> FileList;
+
+
+namespace file_local {
+
+inline void filterAndPush(FileList& list, const FileInfo& fi, const std::string& extension, bool cond(const std::string&, const std::string&))
+{
+	if (fi.name == "." || fi.name == "..") {
+		return;
+	}
+	if (cond(fi.name, extension)) {
+		list.push_back(fi);
+	}
+}
+
+} // cybozu::file_local
+
+/**
+	get file name in dir
+	@param list [out] FileList
+	@param dir [in] directory
+	@param extension [in] select files(including directory) having extension such as "cpp" ; select all if suffix is empty
+	@param cond [in] filter function (select if cond(targetFile, suffix) is true)
+	@note "." and ".." are excluded
+*/
+inline bool GetFileList(FileList &list, const std::string& dir, const std::string& extension = "", bool (*cond)(const std::string&, const std::string&) = cybozu::HasExtension)
+{
+#ifdef _WIN32
+	std::string path = dir + "/*";
+	WIN32_FIND_DATAA fd;
+	struct Handle {
+		Handle(HANDLE hdl)
+			: hdl_(hdl)
+		{
+		}
+		~Handle()
+		{
+			if (hdl_ != INVALID_HANDLE_VALUE) {
+				FindClose(hdl_);
+			}
+		}
+		HANDLE hdl_;
+	};
+	Handle hdl(FindFirstFileA(path.c_str(), &fd));
+	if (hdl.hdl_ == INVALID_HANDLE_VALUE) {
+		return false;
+	}
+	do {
+		FileInfo fi(fd.cFileName, fd.dwFileAttributes);
+		file_local::filterAndPush(list, fi, extension, cond);
+	} while (FindNextFileA(hdl.hdl_, &fd) != 0);
+	return true;
+#else
+	struct Handle {
+		DIR *dir_;
+		Handle(DIR *dir)
+			: dir_(dir)
+		{
+			if (dir_ == 0) {
+				perror("opendir");
+			}
+		}
+		~Handle()
+		{
+			if (dir_) {
+				if (::closedir(dir_)) {
+					perror("closedir");
+				}
+			}
+		}
+		bool isValid() const { return dir_ != 0; }
+	};
+	Handle hdl(::opendir(dir.c_str()));
+	if (!hdl.isValid()) return false;
+	for (;;) {
+		struct dirent *dp = ::readdir(hdl.dir_);
+		if (dp == 0) return true;
+		FileInfo fi(dp->d_name, (uint8_t)dp->d_type);
+		file_local::filterAndPush(list, fi, extension, cond);
+	}
+#endif
+}
+
+inline FileList GetFileList(const std::string& dir, const std::string& extension = "", bool (*cond)(const std::string&, const std::string&) = cybozu::HasExtension)
+{
+	FileList fl;
+	if (GetFileList(fl, dir, extension, cond)) return fl;
+	throw cybozu::Exception("cybozu:GetFileList") << dir << cybozu::ErrorNo();
+}
+
+} // cybozu

include/cybozu/hash.hpp

@@ -54,7 +54,7 @@ namespace std { CYBOZU_NAMESPACE_TR1_BEGIN
 	#pragma warning(push)
 	#pragma warning(disable : 4099) // missmatch class and struct
 #endif
-#ifndef __APPLE__
+#if !(defined(__APPLE__) && defined(__clang__))
 template<class T>
 struct hash;
 #endif

include/cybozu/inttype.hpp

@@ -72,6 +72,9 @@
 	#endif
 #endif
 
+// LLONG_MIN in limits.h is not defined in some env.
+#define CYBOZU_LLONG_MIN (-9223372036854775807ll-1)
+
 #define CYBOZU_CPP_VERSION_CPP03 0
 #define CYBOZU_CPP_VERSION_TR1 1
 #define CYBOZU_CPP_VERSION_CPP11 2
@@ -88,13 +91,13 @@
 	#define CYBOZU_CPP_VERSION CYBOZU_CPP_VERSION_CPP17
 #elif (__cplusplus >= 201402)
 	#define CYBOZU_CPP_VERSION CYBOZU_CPP_VERSION_CPP14
-#elif (__cplusplus >= 201103) || (_MSC_VER >= 1500) || defined(__GXX_EXPERIMENTAL_CXX0X__)
+#elif (__cplusplus >= 201103) || (defined(_MSC_VER) && _MSC_VER >= 1500) || defined(__GXX_EXPERIMENTAL_CXX0X__)
 	#if defined(_MSC_VER) && (_MSC_VER <= 1600)
 		#define CYBOZU_CPP_VERSION CYBOZU_CPP_VERSION_TR1
 	#else
 		#define CYBOZU_CPP_VERSION CYBOZU_CPP_VERSION_CPP11
 	#endif
-#elif CYBOZU_GNUC_PREREQ(4, 5) || (CYBOZU_GNUC_PREREQ(4, 2) && __GLIBCXX__ >= 20070719) || defined(__INTEL_COMPILER) || (__clang_major__ >= 3)
+#elif CYBOZU_GNUC_PREREQ(4, 5) || (CYBOZU_GNUC_PREREQ(4, 2) && (defined(__GLIBCXX__) &&__GLIBCXX__ >= 20070719)) || defined(__INTEL_COMPILER) || (__clang_major__ >= 3)
 	#define CYBOZU_CPP_VERSION CYBOZU_CPP_VERSION_TR1
 #else
 	#define CYBOZU_CPP_VERSION CYBOZU_CPP_VERSION_CPP03
@@ -115,7 +118,7 @@
 #endif
 
 #ifndef CYBOZU_OS_BIT
-	#if defined(_WIN64) || defined(__x86_64__) || defined(__AARCH64EL__) || defined(__EMSCRIPTEN__)
+	#if defined(_WIN64) || defined(__x86_64__) || defined(__AARCH64EL__) || defined(__EMSCRIPTEN__) || defined(__LP64__)
 		#define CYBOZU_OS_BIT 64
 	#else
 		#define CYBOZU_OS_BIT 32
@@ -143,6 +146,8 @@
 		#define CYBOZU_ENDIAN CYBOZU_ENDIAN_LITTLE
 	#elif (CYBOZU_HOST == CYBOZU_HOST_ARM) && (defined(__ARM_EABI__) || defined(__AARCH64EL__))
 		#define CYBOZU_ENDIAN CYBOZU_ENDIAN_LITTLE
+	#elif defined(__s390x__)
+		#define CYBOZU_ENDIAN CYBOZU_ENDIAN_BIG
 	#else
 		#define CYBOZU_ENDIAN CYBOZU_ENDIAN_UNKNOWN
 	#endif
@@ -150,8 +155,10 @@
 
 #if CYBOZU_CPP_VERSION >= CYBOZU_CPP_VERSION_CPP11
 	#define CYBOZU_NOEXCEPT noexcept
+	#define CYBOZU_NULLPTR nullptr
 #else
 	#define CYBOZU_NOEXCEPT throw()
+	#define CYBOZU_NULLPTR 0
 #endif
 namespace cybozu {
 template<class T>

include/cybozu/itoa.hpp

@@ -5,7 +5,6 @@
 
 	@author MITSUNARI Shigeo(@herumi)
 */
-#include <limits.h>
 #ifndef CYBOZU_DONT_USE_STRING
 #include <string>
 #endif
@@ -122,7 +121,7 @@ size_t uintToBin(char *buf, size_t bufSize, UT x)
 template<class T>
 size_t intToDec(char *buf, size_t bufSize, T x)
 {
-	if (x == LLONG_MIN) {
+	if (x == CYBOZU_LLONG_MIN) {
 		const char minStr[] = "-9223372036854775808";
 		const size_t minStrLen = sizeof(minStr) - 1;
 		if (bufSize < minStrLen) {

include/cybozu/mutex.hpp

@@ -8,6 +8,9 @@
 */
 
 #ifdef _WIN32
+	#ifndef WIN32_LEAN_AND_MEAN
+		#define WIN32_LEAN_AND_MEAN
+	#endif
 	#include <windows.h>
 #else
 	#include <pthread.h>

include/cybozu/option.hpp

@@ -145,7 +145,7 @@ bool convertInt(T* x, const char *str)
 	if (factor > 1) {
 		if ((std::numeric_limits<T>::min)() / factor <= y
 			&& y <= (std::numeric_limits<T>::max)() / factor) {
-			*x = y * factor;
+			*x = static_cast<T>(y * factor);
 		} else {
 			return false;
 		}
@@ -155,9 +155,29 @@ bool convertInt(T* x, const char *str)
 	return true;
 }
 
+template<class T>
+void convertToStr(std::ostream& os, const T* p)
+{
+	os << *p;
+}
+template<>inline void convertToStr(std::ostream& os, const int8_t* p)
+{
+	os << static_cast<int>(*p);
+}
+template<>inline void convertToStr(std::ostream& os, const uint8_t* p)
+{
+	os << static_cast<int>(*p);
+}
+
 #define CYBOZU_OPTION_DEFINE_CONVERT_INT(type) \
 template<>inline bool convert(type* x, const char *str) { return convertInt(x, str); }
 
+CYBOZU_OPTION_DEFINE_CONVERT_INT(int8_t)
+CYBOZU_OPTION_DEFINE_CONVERT_INT(uint8_t)
+
+CYBOZU_OPTION_DEFINE_CONVERT_INT(int16_t)
+CYBOZU_OPTION_DEFINE_CONVERT_INT(uint16_t)
+
 CYBOZU_OPTION_DEFINE_CONVERT_INT(int)
 CYBOZU_OPTION_DEFINE_CONVERT_INT(long)
 CYBOZU_OPTION_DEFINE_CONVERT_INT(long long)
@@ -185,7 +205,7 @@ struct Holder : public HolderBase {
 	std::string toStr() const
 	{
 		std::ostringstream os;
-		os << *p_;
+		convertToStr(os, p_);
 		return os.str();
 	}
 	const void *get() const { return (void*)p_; }
@@ -345,7 +365,7 @@ class Option {
 	template<class T, class U>
 	void append(T *pvar, const U& defaultVal, bool isMust, const char *opt, const std::string& help = "")
 	{
-		*pvar = defaultVal;
+		*pvar = static_cast<const T&>(defaultVal);
 		appendSub(pvar, N_is1, isMust, opt, help);
 	}
 	/*

include/cybozu/random_generator.hpp

@@ -11,13 +11,14 @@
 #include <cybozu/exception.hpp>
 #endif
 #ifdef _WIN32
-#include <winsock2.h>
+#ifndef WIN32_LEAN_AND_MEAN
+	#define WIN32_LEAN_AND_MEAN
+#endif
 #include <windows.h>
 #include <wincrypt.h>
 #ifdef _MSC_VER
 #pragma comment (lib, "advapi32.lib")
 #endif
-#include <cybozu/critical_section.hpp>
 #else
 #include <sys/types.h>
 #include <fcntl.h>
@@ -32,10 +33,9 @@ public:
 #ifdef _WIN32
 	RandomGenerator()
 		: prov_(0)
-		, pos_(bufSize)
 	{
-		DWORD flagTbl[] = { 0, CRYPT_NEWKEYSET };
+		DWORD flagTbl[] = { CRYPT_VERIFYCONTEXT | CRYPT_SILENT, 0, CRYPT_MACHINE_KEYSET };
-		for (int i = 0; i < 2; i++) {
+		for (size_t i = 0; i < CYBOZU_NUM_OF_ARRAY(flagTbl); i++) {
 			if (CryptAcquireContext(&prov_, NULL, NULL, PROV_RSA_FULL, flagTbl[i]) != 0) return;
 		}
 #ifdef CYBOZU_DONT_USE_EXCEPTION
@@ -62,29 +62,11 @@ public:
 	template<class T>
 	void read(bool *pb, T *buf, size_t bufNum)
 	{
-		cybozu::AutoLockCs al(cs_);
 		const size_t byteSize = sizeof(T) * bufNum;
-		if (byteSize > bufSize) {
+		*pb = read_inner(buf, byteSize);
-			if (!read_inner(buf, byteSize)) {
-				*pb = false;
-				return;
-			}
-		} else {
-			if (pos_ + byteSize > bufSize) {
-				read_inner(buf_, bufSize);
-				pos_ = 0;
-			}
-			memcpy(buf, buf_ + pos_, byteSize);
-			pos_ += byteSize;
-		}
-		*pb = true;
 	}
 private:
 	HCRYPTPROV prov_;
-	static const size_t bufSize = 1024;
-	char buf_[bufSize];
-	size_t pos_;
-	cybozu::CriticalSection cs_;
 #else
 	RandomGenerator()
 		: fp_(::fopen("/dev/urandom", "rb"))

include/cybozu/sha2.hpp

@@ -13,6 +13,7 @@
 #ifndef CYBOZU_DONT_USE_STRING
 #include <string>
 #endif
+#include <memory.h>
 
 #ifdef CYBOZU_USE_OPENSSL_SHA
 #ifdef __APPLE__
@@ -468,3 +469,61 @@ public:
 } // cybozu
 
 #endif
+
+namespace cybozu {
+
+namespace sha2_local {
+
+inline void hmac256_inner(void *hmac, const void *key, size_t keySize, const void *msg, size_t msgSize, bool addZeroByte)
+{
+	const uint8_t ipad = 0x36;
+	const uint8_t opad = 0x5c;
+	uint8_t k[64];
+	Sha256 hash;
+	if (keySize > 64) {
+		hash.digest(k, 32, key, keySize);
+		hash.clear();
+		keySize = 32;
+	} else {
+		memcpy(k, key, keySize);
+	}
+	for (size_t i = 0; i < keySize; i++) {
+		k[i] = k[i] ^ ipad;
+	}
+	memset(k + keySize, ipad, 64 - keySize);
+	hash.update(k, 64);
+	if (addZeroByte) {
+		hash.update(msg, msgSize);
+		const char zero = '\x00';
+		hash.digest(hmac, 32, &zero, 1);
+	} else {
+		hash.digest(hmac, 32, msg, msgSize);
+	}
+	hash.clear();
+	for (size_t i = 0; i < 64; i++) {
+		k[i] = k[i] ^ (ipad ^ opad);
+	}
+	hash.update(k, 64);
+	hash.digest(hmac, 32, hmac, 32);
+}
+
+} // cybozu::sha2_local
+
+/*
+	HMAC-SHA-256
+	hmac must have 32 bytes buffer
+*/
+inline void hmac256(void *hmac, const void *key, size_t keySize, const void *msg, size_t msgSize)
+{
+	sha2_local::hmac256_inner(hmac, key, keySize, msg, msgSize, false);
+}
+
+/*
+	hmac256 for [msg] + [\x00]
+*/
+inline void hmac256addZeroByte(void *hmac, const void *key, size_t keySize, const void *msg, size_t msgSize)
+{
+	sha2_local::hmac256_inner(hmac, key, keySize, msg, msgSize, true);
+}
+
+} // cybozu

include/cybozu/socket.hpp

@@ -0,0 +1,785 @@
+#pragma once
+/**
+	@file
+	@brief tiny socket class
+
+	@author MITSUNARI Shigeo(@herumi)
+	@author MITSUNARI Shigeo
+	@remark mingw requires -lws2_32 option
+*/
+#include <errno.h>
+#include <assert.h>
+#include <stdio.h>
+#ifdef _WIN32
+	#ifndef WIN32_LEAN_AND_MEAN
+		#define WIN32_LEAN_AND_MEAN
+	#endif
+	#include <windows.h>
+	#include <winsock2.h>
+	#include <ws2tcpip.h> // for socklen_t
+	#ifdef _MSC_VER
+		#pragma comment(lib, "ws2_32.lib")
+		#pragma comment(lib, "iphlpapi.lib")
+		#pragma warning(push)
+		#pragma warning(disable : 4127) // constant condition
+	#endif
+#else
+	#include <unistd.h>
+	#include <sys/socket.h>
+	#include <sys/ioctl.h>
+	#include <netinet/tcp.h>
+	#include <arpa/inet.h>
+	#include <netdb.h>
+	#include <memory.h>
+	#include <signal.h>
+#endif
+#ifndef NDEBUG
+	#include <stdio.h>
+#endif
+
+#include <cybozu/atomic.hpp>
+#include <cybozu/exception.hpp>
+#include <cybozu/itoa.hpp>
+#include <string>
+
+#ifdef __linux__
+//	#define CYBOZU_SOCKET_USE_EPOLL
+	#include <sys/epoll.h>
+#endif
+
+namespace cybozu {
+
+#ifdef _MSC_VER
+struct NetErrorNo : public cybozu::ErrorNo {
+	NetErrorNo(NativeErrorNo err)
+		: cybozu::ErrorNo(err)
+	{
+	}
+	NetErrorNo()
+		: cybozu::ErrorNo(WSAGetLastError())
+	{
+	}
+};
+#else
+typedef cybozu::ErrorNo NetErrorNo;
+#endif
+
+#ifdef CYBOZU_SOCKET_USE_EPOLL
+
+namespace experimental {
+
+struct EpollEvent {
+	struct epoll_event ev_;
+	EpollEvent()
+	{
+		memset(&ev_, 0, sizeof(ev_));
+	}
+	void set(int fd, uint32_t events = EPOLLIN)
+	{
+		ev_.events = events;
+		ev_.data.fd = fd;
+	}
+	int getFd() const { return ev_.data.fd; }
+};
+
+class Epoll {
+	int efd_;
+	bool verify(const char *msg, int ret, int *err) const {
+		if (ret >= 0) return true;
+		if (err == 0) throw cybozu::Exception(msg) << cybozu::NetErrorNo();
+		*err = errno;
+		return false;
+	}
+public:
+	Epoll() : efd_(-1) {}
+	bool init(int *err = 0)
+	{
+		efd_ = ::epoll_create1(0);
+		return verify("Epoll:init", efd_, err);
+	}
+	~Epoll()
+	{
+		if (efd_ >= 0) ::close(efd_);
+	}
+	/*
+		throw if err == NULL
+	*/
+	bool ctrl(int op, int fd, EpollEvent *ev, int *err = 0) {
+		int ret = ::epoll_ctl(efd_, op, fd, &ev->ev_);
+		return verify("Epoll:ctrl", ret, err);
+	}
+	bool add(int fd, uint32_t events = EPOLLIN, int *err = 0) {
+		EpollEvent ev;
+		ev.set(fd, events);
+		return ctrl(EPOLL_CTL_ADD, fd, &ev, err);
+	}
+	bool del(int fd, int *err = 0) {
+		return ctrl(EPOLL_CTL_DEL, fd, NULL, err);
+	}
+	/*
+		msec : 0 : block
+	*/
+	int wait(EpollEvent *ev, int maxEv, int msec = 0)
+	{
+		/*
+		 0 : return immediately
+		-1 : block indefinitely
+		*/
+		if (msec == 0) {
+			msec = -1;
+		} else if (msec == -1) {
+			msec = 0;
+		}
+		int ret = ::epoll_wait(efd_, &ev->ev_, maxEv, msec);
+		if (ret == 0) return 0; // timeout
+		if (ret < 0) return -errno;
+		return ret;
+	}
+};
+
+struct AutoLock {
+	Epoll& ep_;
+	int fd_;
+	AutoLock(Epoll& ep, int fd, int events = EPOLLIN)
+		: ep_(ep)
+		, fd_(fd)
+	{
+		ep_.add(fd, events);
+	}
+	~AutoLock()
+	{
+		int err;
+		ep_.del(fd_, &err);
+	}
+};
+
+} // cybozu::experimental
+#endif
+
+namespace ssl {
+class ClientSocket;
+};
+
+namespace socket_local {
+
+#ifdef _WIN32
+	typedef SOCKET SocketHandle;
+#else
+	typedef int SocketHandle;
+#endif
+
+struct InitTerm {
+	/** call once for init */
+	InitTerm()
+	{
+#ifdef _WIN32
+		WSADATA data;
+		int err = ::WSAStartup(MAKEWORD(2, 2), &data);
+		if (err) {
+			fprintf(stderr, "WSAStartup failed : %d\n", err);
+			exit(1);
+		}
+#else
+		::signal(SIGPIPE, SIG_IGN);
+#endif
+	}
+	/** call once for term */
+	~InitTerm()
+	{
+#ifdef _WIN32
+		::WSACleanup();
+#endif
+	}
+	void dummyCall() { }
+};
+
+template<int dummy = 0>
+struct InstanceIsHere { static InitTerm it_; };
+
+template<int dummy>
+InitTerm InstanceIsHere<dummy>::it_;
+
+struct DummyCall {
+	DummyCall() { InstanceIsHere<>::it_.dummyCall(); }
+};
+
+} // cybozu::socket_local
+
+class SocketAddr {
+	union {
+		// http://www.coins.tsukuba.ac.jp/~syspro/2010/No6_files/sockaddr.html
+		struct sockaddr sa; /* 16byte */
+		struct sockaddr_in v4; /* 16byte */
+		struct sockaddr_in6 v6;
+	} addr_;
+	socklen_t addrlen_;
+	int family_;
+	friend class Socket;
+	void verify() // call in only Socket::accept
+	{
+		if (addrlen_ == sizeof(addr_.v4) && addr_.sa.sa_family == AF_INET) {
+			family_ = AF_INET;
+			return;
+		}
+		if (addrlen_ == sizeof(addr_.v6) && addr_.sa.sa_family == AF_INET6) {
+			family_ = AF_INET6;
+			return;
+		}
+		throw cybozu::Exception("cybozu:SocketAddr:verify") << addrlen_;
+	}
+public:
+	SocketAddr()
+		: addrlen_(0)
+		, family_(0)
+	{
+	}
+	SocketAddr(const std::string& address, uint16_t port, bool forceIpV6 = false)
+	{
+		set(address, port, forceIpV6);
+	}
+	void set(const std::string& address, uint16_t port, bool forceIpV6 = false)
+	{
+		char portStr[16];
+		CYBOZU_SNPRINTF(portStr, sizeof(portStr), "%d", port);
+		memset(&addr_, 0, sizeof(addr_));
+		addrlen_ = 0;
+		family_ = 0;
+
+		struct addrinfo *result = 0;
+		struct addrinfo hints;
+		memset(&hints, 0, sizeof(struct addrinfo));
+		hints.ai_family = AF_INET;
+		hints.ai_socktype = SOCK_STREAM;
+		hints.ai_protocol = IPPROTO_TCP;
+		hints.ai_flags = AI_NUMERICSERV; // AI_PASSIVE;
+		const int s = getaddrinfo(address.c_str(), portStr, &hints, &result);
+		// s == EAI_AGAIN
+		if (s || forceIpV6) {
+			hints.ai_family = AF_INET6;
+			hints.ai_flags |= AI_V4MAPPED;
+			if (getaddrinfo(address.c_str(), portStr, &hints, &result)) {
+				goto ERR_EXIT;
+			}
+		}
+		{
+			bool found = false;
+			for (const struct addrinfo *p = result; p; p = p->ai_next) {
+				const int family = p->ai_family;
+				if (family == hints.ai_family) {
+					if (p->ai_addrlen > sizeof(addr_)) {
+						break;
+					}
+					memcpy(&addr_, p->ai_addr, p->ai_addrlen);
+					addrlen_ = (socklen_t)p->ai_addrlen;
+					family_ = family;
+					found = true;
+					break;
+				}
+			}
+			freeaddrinfo(result);
+			if (found) return;
+		}
+	ERR_EXIT:
+		throw cybozu::Exception("SocketAddr:set") << address << port << cybozu::NetErrorNo();
+	}
+	socklen_t getSize() const { return addrlen_; }
+	int getFamily() const { return family_; }
+	const struct sockaddr *get() const { return &addr_.sa; }
+	uint16_t getPort() const {
+		if (family_ == AF_INET) {
+			return ntohs(addr_.v4.sin_port);
+		} else if (family_ == AF_INET6) {
+			return ntohs(addr_.v6.sin6_port);
+		}
+		throw cybozu::Exception("SocketAddr:getPort:bad family") << family_;
+	}
+	// compare addr without port
+	bool hasSameAddr(const SocketAddr& rhs) const
+	{
+		const uint8_t *v4 = 0;
+		const uint8_t *v6 = 0;
+		if (family_ == AF_INET) {
+			if (rhs.family_ == AF_INET) return memcmp(&addr_.v4.sin_addr, &rhs.addr_.v4.sin_addr, sizeof(in_addr)) == 0;
+			if (rhs.family_ != AF_INET6) return false;
+			v4 = (const uint8_t*)&addr_.v4.sin_addr;
+			v6 = (const uint8_t*)&rhs.addr_.v6.sin6_addr;
+		} else if (family_ != AF_INET6) {
+			return false;
+		} else {
+			if (rhs.family_ == AF_INET6) return memcmp(&addr_.v6.sin6_addr, &rhs.addr_.v6.sin6_addr, sizeof(in6_addr)) == 0;
+			if (rhs.family_ != AF_INET) return false;
+			v4 = (const uint8_t*)&rhs.addr_.v4.sin_addr;
+			v6 = (const uint8_t*)&addr_.v6.sin6_addr;
+		}
+		// Ipv6-mapped?
+		const uint8_t header[12] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xff, 0xff };
+		return memcmp(v6, header, 12) == 0 && memcmp(v6 + 12, v4, 4) == 0;
+	}
+	std::string toStr() const
+	{
+		if (family_ == AF_INET || family_ == AF_INET6) {
+			char buf[INET6_ADDRSTRLEN];
+			assert(INET_ADDRSTRLEN <= INET6_ADDRSTRLEN);
+			const bool isIPv4 = family_ == AF_INET;
+			const void *pa = isIPv4 ? (const void*)&addr_.v4.sin_addr : (const void*)&addr_.v6.sin6_addr;
+			// not "const void*" because of vc
+			const char *p = inet_ntop(family_, const_cast<void*>(pa), buf, sizeof(buf));
+			if (!p) throw cybozu::Exception("cybozu:SocketAddr:toStr") << cybozu::NetErrorNo();
+			if (isIPv4) return std::string(p) + ':' + cybozu::itoa(getPort());
+			return std::string("[") + p + "]:" + cybozu::itoa(getPort());
+		}
+		throw cybozu::Exception("cybozu:SocketAddr:toStr:bad family_") << family_;
+	}
+};
+/*
+	socket class
+	@note ower is moved if copied
+*/
+class Socket {
+	friend class cybozu::ssl::ClientSocket;
+private:
+	cybozu::socket_local::SocketHandle sd_;
+	Socket(const Socket&);
+	void operator=(const Socket&);
+#ifdef WIN32
+	void setTimeout(int type, int msec)
+	{
+		setSocketOption(type, msec);
+	}
+	/* return msec */
+	int getTimeout(int type) const
+	{
+		return getSocketOption(type);
+	}
+#else
+	void setTimeout(int type, int msec)
+	{
+		struct timeval t;
+		t.tv_sec = msec / 1000;
+		t.tv_usec = (msec % 1000) * 1000;
+		setSocketOption(type, t);
+	}
+	/* return msec */
+	int getTimeout(int type) const
+	{
+		struct timeval t;
+		getSocketOption(type, &t);
+		return t.tv_sec * 1000 + t.tv_usec / 1000; /* msec */
+	}
+#endif
+	void setBlocking(bool isBlocking)
+	{
+#ifdef _WIN32
+		u_long val = isBlocking ? 0 : 1;
+		int ret = ::ioctlsocket(sd_, FIONBIO, &val);
+#else
+		int val = isBlocking ? 0 : 1;
+		int ret = ::ioctl(sd_, FIONBIO, &val);
+#endif
+		if (ret < 0) throw cybozu::Exception("Socket:setBlocking") << cybozu::NetErrorNo() << isBlocking;
+	}
+public:
+#ifndef _WIN32
+	static const int INVALID_SOCKET = -1;
+#endif
+	Socket()
+		: sd_(INVALID_SOCKET)
+	{
+	}
+
+	bool isValid() const { return sd_ != INVALID_SOCKET; }
+
+	// move
+#if CYBOZU_CPP_VERSION >= CYBOZU_CPP_VERSION_CPP11
+	Socket(Socket&& rhs)
+		: sd_(INVALID_SOCKET)
+	{
+		sd_ = cybozu::AtomicExchange(&rhs.sd_, sd_);
+	}
+#endif
+	// close and move
+	void moveFrom(Socket& rhs)
+	{
+		close();
+		sd_ = cybozu::AtomicExchange(&rhs.sd_, INVALID_SOCKET);
+	}
+#if CYBOZU_CPP_VERSION >= CYBOZU_CPP_VERSION_CPP11
+	void operator=(Socket&& rhs)
+#else
+	void operator=(Socket& rhs)
+#endif
+	{
+		moveFrom(rhs);
+	}
+
+	~Socket()
+	{
+		close(cybozu::DontThrow);
+	}
+
+	bool close(bool dontThrow = false)
+	{
+		cybozu::socket_local::SocketHandle sd = cybozu::AtomicExchange(&sd_, INVALID_SOCKET);
+		if (sd == INVALID_SOCKET) return true;
+#ifdef _WIN32
+		// ::shutdown(sd, SD_SEND);
+		// shutdown is called in closesocket
+		bool isOK = ::closesocket(sd) == 0;
+#else
+		bool isOK = ::close(sd) == 0;
+#endif
+		if (!dontThrow && !isOK) throw cybozu::Exception("Socket:close") << cybozu::NetErrorNo();
+		return isOK;
+	}
+	/*
+		how 0 : SHUTRD ; disallow read
+		    1 : SHUT_WR ; disallow write
+		    2 : SHUT_RDWR ; disallow read/write
+	*/
+	bool shutdown(int how, bool dontThrow = false)
+	{
+		bool isOK = ::shutdown(sd_, how) == 0;
+		if (!dontThrow && !isOK) throw cybozu::Exception("Socket:waitForClose:shutdown") << cybozu::NetErrorNo();
+		return isOK;
+	}
+	/*
+		send FIN and wait for remote's close().
+		this function is used for the following situation.
+		sock.write()
+		sock.waitForClose()
+		sock.close()
+	*/
+	void waitForClose()
+	{
+		if (sd_ == INVALID_SOCKET) return;
+		//	send FIN and this socket can't write any data.
+		shutdown(1);
+		// wait for FIN from the peer.
+		char buf[1];
+		ssize_t readSize = readSome(buf, sizeof(buf));
+		if (readSize != 0) {
+			throw cybozu::Exception("Socket:waitForClose:readSome:bad size") << readSize;
+		}
+	}
+
+	/*!
+		receive data
+		@param buf [out] receive buffer
+		@param bufSize [in] receive buffer size(byte)
+		@note return read size
+	*/
+	size_t readSome(void *buf, size_t bufSize)
+	{
+		int size = (int)(std::min)((size_t)0x7fffffff, bufSize);
+#ifdef _WIN32
+		int readSize = ::recv(sd_, (char *)buf, size, 0);
+#else
+	RETRY:
+		ssize_t readSize = ::read(sd_, buf, size);
+		if (readSize < 0 && errno == EINTR) goto RETRY;
+#endif
+		if (readSize < 0) throw cybozu::Exception("Socket:readSome") << cybozu::NetErrorNo() << bufSize;
+		return readSize;
+	}
+
+	/*!
+		receive all data unless timeout
+		@param buf [out] receive buffer
+		@param bufSize [in] receive buffer size(byte)
+	*/
+	void read(void *buf, size_t bufSize)
+	{
+		char *p = (char *)buf;
+		while (bufSize > 0) {
+			size_t readSize = readSome(p, bufSize);
+			if (readSize == 0) throw cybozu::Exception("Socket:read:readSize is zero");
+			p += readSize;
+			bufSize -= readSize;
+		}
+	}
+	/*!
+		write all data
+		@param buf [out] send buffer
+		@param bufSize [in] send buffer size(byte)
+	*/
+	void write(bool *pb, const void *buf, size_t bufSize)
+	{
+		const char *p = (const char *)buf;
+		while (bufSize > 0) {
+			int size = (int)(std::min)(size_t(0x7fffffff), bufSize);
+#ifdef _WIN32
+			int writeSize = ::send(sd_, p, size, 0);
+#else
+			int writeSize = ::write(sd_, p, size);
+			if (writeSize < 0 && errno == EINTR) continue;
+#endif
+			if (writeSize < 0) {
+				*pb = false;
+				return;
+			}
+			p += writeSize;
+			bufSize -= writeSize;
+		}
+		*pb = true;
+	}
+	void write(const void *buf, size_t bufSize)
+	{
+		bool b;
+		write(&b, buf, bufSize);
+		if (!b) throw cybozu::Exception("Socket:write") << cybozu::NetErrorNo() << bufSize;
+	}
+	/**
+		connect to address:port
+		@param address [in] address
+		@param port [in] port
+		@param msec: 0 : block
+	*/
+	void connect(const std::string& address, uint16_t port, int msec = 0)
+	{
+		SocketAddr addr;
+		addr.set(address, port);
+		connect(addr, msec);
+	}
+	/**
+		connect to resolved socket addr
+	*/
+	void connect(const cybozu::SocketAddr& addr, int msec = 0)
+	{
+		if (isValid()) throw cybozu::Exception("Socket:connect:already connect");
+		sd_ = ::socket(addr.getFamily(), SOCK_STREAM, IPPROTO_TCP);
+		if (!isValid()) {
+			throw cybozu::Exception("Socket:connect:socket") << cybozu::NetErrorNo();
+		}
+		if (msec == 0) {
+			if (::connect(sd_, addr.get(), addr.getSize()) < 0) {
+				throw cybozu::Exception("Socket:connect") << cybozu::NetErrorNo() << addr.toStr();
+			}
+		} else {
+			setBlocking(false);
+			if (::connect(sd_, addr.get(), addr.getSize()) < 0) {
+#ifdef _WIN32
+				bool inProgress = WSAGetLastError() == WSAEWOULDBLOCK;
+#else
+				bool inProgress = errno == EINPROGRESS;
+#endif
+				if (!inProgress) throw cybozu::Exception("Socket:connect:not in progress") << cybozu::NetErrorNo() << addr.toStr();
+				if (!queryAccept(msec, false)) throw cybozu::Exception("Socket:connect:timeout") << addr.toStr();
+				int err = getSocketOption(SO_ERROR);
+				if (err != 0) throw cybozu::Exception("Socket::connect:bad socket") << cybozu::NetErrorNo(err);
+			}
+			setBlocking(true);
+		}
+	}
+
+	static const int allowIPv4 = 1;
+	static const int allowIPv6 = 2;
+	/**
+		init for server
+		@param port [in] port number
+	*/
+	void bind(uint16_t port, int mode = allowIPv4 | allowIPv6)
+	{
+		const int family = (mode & allowIPv6) ? AF_INET6 : AF_INET;
+		sd_ = ::socket(family, SOCK_STREAM, IPPROTO_TCP);
+		if (!isValid()) {
+			throw cybozu::Exception("Socket:bind:socket") << cybozu::NetErrorNo();
+		}
+		setSocketOption(SO_REUSEADDR, 1);
+		struct sockaddr_in6 addr6;
+		struct sockaddr_in addr4;
+		struct sockaddr *addr;
+		socklen_t addrLen;
+		if (mode & allowIPv6) {
+			setSocketOption(IPV6_V6ONLY, (mode & allowIPv4) ? 0 : 1, IPPROTO_IPV6);
+			memset(&addr6, 0, sizeof(addr6));
+			addr6.sin6_family = AF_INET6;
+			addr6.sin6_port = htons(port);
+			addr = (struct sockaddr*)&addr6;
+			addrLen = sizeof(addr6);
+		} else {
+			memset(&addr4, 0, sizeof(addr4));
+			addr4.sin_family = AF_INET;
+			addr4.sin_port = htons(port);
+			addr = (struct sockaddr*)&addr4;
+			addrLen = sizeof(addr4);
+		}
+		if (::bind(sd_, addr, addrLen) == 0) {
+			if (::listen(sd_, SOMAXCONN) == 0) {
+				return;
+			}
+		}
+		cybozu::NetErrorNo keep;
+		close(cybozu::DontThrow);
+		throw cybozu::Exception("Socket:bind") << keep;
+	}
+
+	/**
+		return positive if accepted
+		return zero if timeout
+		return negative(-errno) if error
+	*/
+	int queryAcceptNoThrow(int msec = 1000, bool checkWrite = true)
+	{
+		if (sd_ == INVALID_SOCKET) return -EBADF;
+#ifdef CYBOZU_SOCKET_USE_EPOLL
+		int err;
+		experimental::Epoll ep;
+		if (!ep.init(&err)) return -err;
+		uint32_t events = checkWrite ? EPOLLIN : EPOLLOUT;
+		experimental::AutoLock al(ep, sd_, events);
+		experimental::EpollEvent ev;
+		int ret = ep.wait(&ev, 1, msec);
+		if (ret != 1) return ret;
+		assert(ev.getFd() == sd_);
+		return ret;
+#else
+#ifndef _WIN32
+		// https://msdn.microsoft.com/en-us/library/windows/desktop/ms739169.aspx
+		if (sd_ >= FD_SETSIZE) return -EMFILE;
+#endif
+		struct timeval timeout;
+		timeout.tv_sec = msec / 1000;
+		timeout.tv_usec = (msec % 1000) * 1000;
+		fd_set fds;
+		FD_ZERO(&fds);
+		FD_SET((unsigned)sd_, &fds);
+		int fdNum;
+		if (checkWrite) {
+			fdNum = ::select((int)sd_ + 1, &fds, 0, 0, &timeout);
+		} else {
+			fdNum = ::select((int)sd_ + 1, 0, &fds, 0, &timeout);
+		}
+		if (fdNum < 0) return -errno;
+		return fdNum;
+#endif
+	}
+	/**
+		return true if acceptable, otherwise false
+		return false if one second passed
+		while (!server.queryAccept()) {
+		}
+		client.accept(server);
+	*/
+	bool queryAccept(int msec = 1000, bool checkWrite = true)
+	{
+		int ret = queryAcceptNoThrow(msec, checkWrite);
+		if (ret < 0) throw cybozu::Exception("Socket:queryAccept") << cybozu::NetErrorNo(-ret);
+		return ret > 0;
+	}
+
+	/**
+		accept for server
+	*/
+	void accept(Socket& client, SocketAddr *paddr = 0) const
+	{
+		if (paddr) {
+			struct sockaddr *psa = &paddr->addr_.sa;
+			paddr->addrlen_ = sizeof(paddr->addr_);
+			client.sd_ = ::accept(sd_, psa, &paddr->addrlen_);
+			paddr->verify();
+		} else {
+			client.sd_ = ::accept(sd_, 0, 0);
+		}
+		if (!client.isValid()) throw cybozu::Exception("Socket:accept") << cybozu::NetErrorNo();
+	}
+
+	template<typename T>
+	void setSocketOption(int optname, const T& value, int level = SOL_SOCKET)
+	{
+		bool isOK = setsockopt(sd_, level, optname, cybozu::cast<const char*>(&value), sizeof(T)) == 0;
+		if (!isOK) throw cybozu::Exception("Socket:setSocketOption") << cybozu::NetErrorNo();
+	}
+	template<typename T>
+	void getSocketOption(int optname, T* value, int level = SOL_SOCKET) const
+	{
+		socklen_t len = (socklen_t)sizeof(T);
+		bool isOK = getsockopt(sd_, level, optname, cybozu::cast<char*>(value), &len) == 0;
+		if (!isOK) throw cybozu::Exception("Socket:getSocketOption") << cybozu::NetErrorNo();
+	}
+	int getSocketOption(int optname) const
+	{
+		int ret;
+		getSocketOption(optname, &ret);
+		return ret;
+	}
+	/**
+		setup linger
+	*/
+	void setLinger(uint16_t l_onoff, uint16_t l_linger)
+	{
+		struct linger linger;
+		linger.l_onoff = l_onoff;
+		linger.l_linger = l_linger;
+		setSocketOption(SO_LINGER, &linger);
+	}
+	/**
+		get receive buffer size
+		@retval positive buffer size(byte)
+		@retval -1 error
+	*/
+	int getReceiveBufferSize() const
+	{
+		return getSocketOption(SO_RCVBUF);
+	}
+	/**
+		set receive buffer size
+		@param size [in] buffer size(byte)
+	*/
+	void setReceiveBufferSize(int size)
+	{
+		setSocketOption(SO_RCVBUF, size);
+	}
+	/**
+		get send buffer size
+		@retval positive buffer size(byte)
+		@retval -1 error
+	*/
+	int getSendBufferSize() const
+	{
+		return getSocketOption(SO_SNDBUF);
+	}
+	/**
+		sed send buffer size
+		@param size [in] buffer size(byte)
+	*/
+	void setSendBufferSize(int size)
+	{
+		setSocketOption(SO_SNDBUF, size);
+	}
+	/**
+		set send timeout
+		@param msec [in] msec
+	*/
+	void setSendTimeout(int msec)
+	{
+		setTimeout(SO_SNDTIMEO, msec);
+	}
+	/**
+		set receive timeout
+		@param msec [in] msec
+	*/
+	void setReceiveTimeout(int msec)
+	{
+		setTimeout(SO_RCVTIMEO, msec);
+	}
+	/**
+		get send timeout(msec)
+	*/
+	int getSendTimeout() const
+	{
+		return getTimeout(SO_SNDTIMEO);
+	}
+	/**
+		get receive timeout(msec)
+	*/
+	int getReceiveTimeout() const
+	{
+		return getTimeout(SO_RCVTIMEO);
+	}
+};
+
+} // cybozu
+
+#ifdef _MSC_VER
+	#pragma warning(pop)
+#endif

include/cybozu/stream.hpp

@@ -9,7 +9,9 @@
 #include <string>
 #include <iosfwd>
 #endif
+#ifndef CYBOZU_DONT_USE_EXCEPTION
 #include <cybozu/exception.hpp>
+#endif
 #include <memory.h>
 
 namespace cybozu {

include/cybozu/time.hpp

@@ -0,0 +1,281 @@
+#pragma once
+/**
+	@file
+	@brief tiny time class
+
+	@author MITSUNARI Shigeo(@herumi)
+*/
+#include <ctime>
+#include <cybozu/exception.hpp>
+#include <cybozu/atoi.hpp>
+#include <cybozu/itoa.hpp>
+#ifdef _WIN32
+	#include <sys/timeb.h>
+#else
+	#include <sys/time.h>
+#endif
+
+namespace cybozu {
+
+/**
+	time struct with time_t and msec
+	@note time MUST be latesr than 1970/1/1
+*/
+class Time {
+	static const uint64_t epochBias = 116444736000000000ull;
+	std::time_t time_;
+	int msec_;
+public:
+	explicit Time(std::time_t time = 0, int msec = 0)
+		: time_(time)
+		, msec_(msec)
+	{
+	}
+	explicit Time(bool doSet)
+	{
+		if (doSet) setCurrentTime();
+	}
+	Time& setTime(std::time_t time, int msec = 0)
+	{
+		time_ = time;
+		msec_ = msec;
+		return *this;
+	}
+	/*
+		Windows FILETIME is defined as
+		struct FILETILME {
+			DWORD dwLowDateTime;
+			DWORD dwHighDateTime;
+		};
+		the value represents the number of 100-nanosecond intervals since January 1, 1601 (UTC).
+	*/
+	void setByFILETIME(uint32_t low, uint32_t high)
+	{
+		const uint64_t fileTime = (((uint64_t(high) << 32) | low) - epochBias) / 10000;
+		time_ = fileTime / 1000;
+		msec_ = fileTime % 1000;
+	}
+	/*
+		DWORD is defined as unsigned long in windows
+	*/
+	template<class dword>
+	void getFILETIME(dword& low, dword& high) const
+	{
+		const uint64_t fileTime = (time_ * 1000 + msec_) * 10000 + epochBias;
+		low = dword(fileTime);
+		high = dword(fileTime >> 32);
+	}
+	explicit Time(const std::string& in)
+	{
+		fromString(in);
+	}
+	explicit Time(const char *in)
+	{
+		fromString(in, in + strlen(in));
+	}
+	const std::time_t& getTime() const { return time_; }
+	int getMsec() const { return msec_; }
+	double getTimeSec() const { return time_ + msec_ * 1e-3; }
+	void addSec(int sec) { time_ += sec; }
+	bool operator<(const Time& rhs) const { return (time_ < rhs.time_) || (time_ == rhs.time_ && msec_ < rhs.msec_); }
+//	bool operator<=(const Time& rhs) const { return (*this < rhs) || (*this == rhs); }
+//	bool operator>(const Time& rhs) const { return rhs < *this; }
+	bool operator==(const Time& rhs) const { return (time_ == rhs.time_) && (msec_ == rhs.msec_); }
+	bool operator!=(const Time& rhs) const { return !(*this == rhs); }
+	/**
+		set time from string such as
+		2009-Jan-23T02:53:44Z
+		2009-Jan-23T02:53:44.078Z
+		2009-01-23T02:53:44Z
+		2009-01-23T02:53:44.078Z
+		@note 'T' may be ' '. '-' may be '/'. last char 'Z' is omissible
+	*/
+	void fromString(bool *pb, const std::string& in) { fromString(pb, &in[0], &in[0] + in.size()); }
+	void fromString(const std::string& in) { fromString(0, in); }
+
+	void fromString(bool *pb, const char *begin, const char *end)
+	{
+		const size_t len = end - begin;
+		if (len >= 19) {
+			const char *p = begin;
+			struct tm tm;
+			int num;
+			bool b;
+			tm.tm_year = getNum(&b, p, 4, 1970, 3000) - 1900;
+			if (!b) goto ERR;
+			p += 4;
+			char sep = *p++;
+			if (sep != '-' && sep != '/') goto ERR;
+
+			p = getMonth(&num, p);
+			if (p == 0) goto ERR;
+			tm.tm_mon = num;
+			if (*p++ != sep) goto ERR;
+
+			tm.tm_mday = getNum(&b, p, 2, 1, 31);
+			if (!b) goto ERR;
+			p += 2;
+			if (*p != ' ' && *p != 'T') goto ERR;
+			p++;
+
+			tm.tm_hour = getNum(&b, p, 2, 0, 23);
+			if (!b) goto ERR;
+			p += 2;
+			if (*p++ != ':') goto ERR;
+
+			tm.tm_min = getNum(&b, p, 2, 0, 59);
+			if (!b) goto ERR;
+			p += 2;
+			if (*p++ != ':') goto ERR;
+
+			tm.tm_sec = getNum(&b, p, 2, 0, 59);
+			if (!b) goto ERR;
+			p += 2;
+
+			if (p == end) {
+				msec_ = 0;
+			} else if (p + 1 == end && *p == 'Z') {
+				msec_ = 0;
+				p++;
+			} else if (*p == '.' && (p + 4 == end || (p + 5 == end && *(p + 4) == 'Z'))) {
+				msec_ = getNum(&b, p + 1, 3, 0, 999);
+				if (!b) goto ERR;
+//				p += 4;
+			} else {
+				goto ERR;
+			}
+#ifdef _WIN32
+			time_ = _mkgmtime64(&tm);
+			if (time_ == -1) goto ERR;
+#else
+			time_ = timegm(&tm);
+#endif
+			if (pb) {
+				*pb = true;
+			}
+			return;
+		}
+	ERR:
+		if (pb) {
+			*pb = false;
+			return;
+		}
+		throw cybozu::Exception("time::fromString") << std::string(begin, 24);
+	}
+	void fromString(const char *begin, const char *end) { fromString(0, begin, end); }
+
+	/**
+		get current time with format
+		@param out [out] output string
+		@param format [in] foramt for strftime and append three digits for msec
+		@param appendMsec [in] appemd <mmm>
+		@param doClear (append to out if false)
+		@note ex. "%Y-%b-%d %H:%M:%S." to get 2009-Jan-23 02:53:44.078
+	*/
+	void toString(std::string& out, const char *format, bool appendMsec = true, bool doClear = true) const
+	{
+		if (doClear) out.clear();
+		char buf[128];
+		struct tm tm;
+#ifdef _WIN32
+		bool isOK = _gmtime64_s(&tm, &time_) == 0;
+#else
+		bool isOK = gmtime_r(&time_, &tm) != 0;
+#endif
+		if (!isOK) throw cybozu::Exception("time::toString") << time_;
+#ifdef __GNUC__
+	#pragma GCC diagnostic push
+	#pragma GCC diagnostic ignored "-Wformat-nonliteral"
+#endif
+		if (std::strftime(buf, sizeof(buf), format, &tm) == 0) {
+			throw cybozu::Exception("time::toString::too long") << format << time_;
+		}
+#ifdef __GNUC__
+	#pragma GCC diagnostic pop
+#endif
+		out += buf;
+		if (appendMsec) {
+			out += cybozu::itoaWithZero(msec_, 3);
+		}
+	}
+
+	/**
+		get current time such as 2009-01-23 02:53:44.078
+		@param out [out] sink string
+	*/
+	void toString(std::string& out, bool appendMsec = true, bool doClear = true) const
+	{
+		const char *format = appendMsec ? "%Y-%m-%d %H:%M:%S." : "%Y-%m-%d %H:%M:%S";
+		toString(out, format, appendMsec, doClear);
+	}
+	std::string toString(bool appendMsec = true) const { std::string out; toString(out, appendMsec); return out; }
+	/**
+		get current time
+	*/
+	Time& setCurrentTime()
+	{
+#ifdef _WIN32
+		struct _timeb timeb;
+		_ftime_s(&timeb);
+		time_ = timeb.time;
+		msec_ = timeb.millitm;
+#else
+		struct timeval tv;
+		gettimeofday(&tv, 0);
+		time_ = tv.tv_sec;
+		msec_ = tv.tv_usec / 1000;
+#endif
+		return *this;
+	}
+private:
+
+	int getNum(bool *b, const char *in, size_t len, int min, int max) const
+	{
+		int ret = cybozu::atoi(b, in, len);
+		if (min <= ret && ret <= max) {
+			return ret;
+		} else {
+			*b = false;
+			return 0;
+		}
+	}
+
+	/*
+		convert month-str to [0, 11]
+		@param ret [out] return idx
+		@param p [in] month-str
+		@retval next pointer or null
+	*/
+	const char *getMonth(int *ret, const char *p) const
+	{
+		static const char monthTbl[12][4] = {
+			"Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"
+		};
+
+		for (size_t i = 0; i < CYBOZU_NUM_OF_ARRAY(monthTbl); i++) {
+			if (memcmp(p, monthTbl[i], 3) == 0) {
+				*ret = (int)i;
+				return p + 3;
+			}
+		}
+		bool b;
+		*ret = getNum(&b, p, 2, 1, 12) - 1;
+		if (b) {
+			return p + 2;
+		} else {
+			return 0;
+		}
+	}
+};
+
+inline std::ostream& operator<<(std::ostream& os, const cybozu::Time& time)
+{
+	return os << time.toString();
+}
+
+inline double GetCurrentTimeSec()
+{
+	return cybozu::Time(true).getTimeSec();
+}
+
+} // cybozu

include/mcl/aggregate_sig.hpp

@@ -62,16 +62,32 @@ public:
 		friend class SecretKey;
 		friend class PublicKey;
 	public:
+		template<class InputStream>
+		void load(bool *pb, InputStream& is, int ioMode = IoSerialize)
+		{
+			S_.load(pb, is, ioMode);
+		}
+		template<class OutputStream>
+		void save(bool *pb, OutputStream& os, int ioMode = IoSerialize) const
+		{
+			S_.save(pb, os, ioMode);
+		}
+#ifndef CYBOZU_DONT_USE_EXCEPTION
 		template<class InputStream>
 		void load(InputStream& is, int ioMode = IoSerialize)
 		{
-			S_.load(is, ioMode);
+			bool b;
+			load(&b, is, ioMode);
+			if (!b) throw cybozu::Exception("Signature:load");
 		}
 		template<class OutputStream>
 		void save(OutputStream& os, int ioMode = IoSerialize) const
 		{
-			S_.save(os, ioMode);
+			bool b;
+			save(&b, os, ioMode);
+			if (!b) throw cybozu::Exception("Signature:save");
 		}
+#endif
 		friend std::istream& operator>>(std::istream& is, Signature& self)
 		{
 			self.load(is, fp::detectIoMode(G1::getIoMode(), is));
@@ -155,16 +171,32 @@ public:
 		friend class SecretKey;
 		friend class Signature;
 	public:
+		template<class InputStream>
+		void load(bool *pb, InputStream& is, int ioMode = IoSerialize)
+		{
+			xQ_.load(pb, is, ioMode);
+		}
+		template<class OutputStream>
+		void save(bool *pb, OutputStream& os, int ioMode = IoSerialize) const
+		{
+			xQ_.save(pb, os, ioMode);
+		}
+#ifndef CYBOZU_DONT_USE_EXCEPTION
 		template<class InputStream>
 		void load(InputStream& is, int ioMode = IoSerialize)
 		{
-			xQ_.load(is, ioMode);
+			bool b;
+			load(&b, is, ioMode);
+			if (!b) throw cybozu::Exception("PublicKey:load");
 		}
 		template<class OutputStream>
 		void save(OutputStream& os, int ioMode = IoSerialize) const
 		{
-			xQ_.save(os, ioMode);
+			bool b;
+			save(&b, os, ioMode);
+			if (!b) throw cybozu::Exception("PublicKey:save");
 		}
+#endif
 		friend std::istream& operator>>(std::istream& is, PublicKey& self)
 		{
 			self.load(is, fp::detectIoMode(G2::getIoMode(), is));
@@ -208,16 +240,32 @@ public:
 		friend class PublicKey;
 		friend class Signature;
 	public:
+		template<class InputStream>
+		void load(bool *pb, InputStream& is, int ioMode = IoSerialize)
+		{
+			x_.load(pb, is, ioMode);
+		}
+		template<class OutputStream>
+		void save(bool *pb, OutputStream& os, int ioMode = IoSerialize) const
+		{
+			x_.save(pb, os, ioMode);
+		}
+#ifndef CYBOZU_DONT_USE_EXCEPTION
 		template<class InputStream>
 		void load(InputStream& is, int ioMode = IoSerialize)
 		{
-			x_.load(is, ioMode);
+			bool b;
+			load(&b, is, ioMode);
+			if (!b) throw cybozu::Exception("SecretKey:load");
 		}
 		template<class OutputStream>
 		void save(OutputStream& os, int ioMode = IoSerialize) const
 		{
-			x_.save(os, ioMode);
+			bool b;
+			save(&b, os, ioMode);
+			if (!b) throw cybozu::Exception("SecretKey:save");
 		}
+#endif
 		friend std::istream& operator>>(std::istream& is, SecretKey& self)
 		{
 			self.load(is, fp::detectIoMode(Fr::getIoMode(), is));

include/mcl/ahe.hpp

@@ -1,76 +0,0 @@
-#pragma once
-/**
-	@file
-	@brief 192/256-bit additive homomorphic encryption by lifted-ElGamal
-	@author MITSUNARI Shigeo(@herumi)
-	@license modified new BSD license
-	http://opensource.org/licenses/BSD-3-Clause
-*/
-#include <mcl/elgamal.hpp>
-#include <mcl/ecparam.hpp>
-
-namespace mcl {
-
-#ifdef MCL_USE_AHE192
-namespace ahe192 {
-
-const mcl::EcParam& para = mcl::ecparam::NIST_P192;
-
-typedef mcl::FpT<mcl::FpTag, 192> Fp;
-typedef mcl::FpT<mcl::ZnTag, 192> Zn;
-typedef mcl::EcT<Fp> Ec;
-typedef mcl::ElgamalT<Ec, Zn> ElgamalEc;
-typedef ElgamalEc::PrivateKey SecretKey;
-typedef ElgamalEc::PublicKey PublicKey;
-typedef ElgamalEc::CipherText CipherText;
-
-static inline void initAhe()
-{
-	Fp::init(para.p);
-	Zn::init(para.n);
-	Ec::init(para.a, para.b);
-	Ec::setIoMode(16);
-	Zn::setIoMode(16);
-}
-
-static inline void initSecretKey(SecretKey& sec)
-{
-	const Ec P(Fp(para.gx), Fp(para.gy));
-	sec.init(P, Zn::getBitSize());
-}
-
-} //mcl::ahe192
-#endif
-
-#ifdef MCL_USE_AHE256
-namespace ahe256 {
-
-const mcl::EcParam& para = mcl::ecparam::NIST_P256;
-
-typedef mcl::FpT<mcl::FpTag, 256> Fp;
-typedef mcl::FpT<mcl::ZnTag, 256> Zn;
-typedef mcl::EcT<Fp> Ec;
-typedef mcl::ElgamalT<Ec, Zn> ElgamalEc;
-typedef ElgamalEc::PrivateKey SecretKey;
-typedef ElgamalEc::PublicKey PublicKey;
-typedef ElgamalEc::CipherText CipherText;
-
-static inline void initAhe()
-{
-	Fp::init(para.p);
-	Zn::init(para.n);
-	Ec::init(para.a, para.b);
-	Ec::setIoMode(16);
-	Zn::setIoMode(16);
-}
-
-static inline void initSecretKey(SecretKey& sec)
-{
-	const Ec P(Fp(para.gx), Fp(para.gy));
-	sec.init(P, Zn::getBitSize());
-}
-
-} //mcl::ahe256
-#endif
-
-} //  mcl

include/mcl/array.hpp

@@ -118,6 +118,7 @@ class FixedArray {
 		y = t;
 	}
 public:
+	typedef T value_type;
 	FixedArray() : n_(0) {}
 	bool resize(size_t n)
 	{
@@ -125,6 +126,15 @@ public:
 		n_ = n;
 		return true;
 	}
+	void push(bool *pb, const T& x)
+	{
+		if (n_ == maxSize) {
+			*pb = false;
+			return;
+		}
+		p_[n_++] = x;
+		*pb = true;
+	}
 	bool copy(const FixedArray<T, maxSize>& rhs)
 	{
 		if (this == &rhs) return true;

include/mcl/bn.h

@@ -75,35 +75,47 @@ typedef struct mclBnFp2 mclBnFp2;
 
 #else
 
+typedef struct {
+	uint64_t d[MCLBN_FP_UNIT_SIZE];
+} mclBnFp;
+
+/*
+	x = d[0] + d[1] i where i^2 = -1
+*/
+typedef struct {
+	mclBnFp d[2];
+} mclBnFp2;
+
+/*
+	G1 and G2 are isomorphism to Fr
+*/
 typedef struct {
 	uint64_t d[MCLBN_FR_UNIT_SIZE];
 } mclBnFr;
 
+/*
+	G1 is defined over Fp
+*/
 typedef struct {
-	uint64_t d[MCLBN_FP_UNIT_SIZE * 3];
+	mclBnFp x, y, z;
 } mclBnG1;
 
 typedef struct {
-	uint64_t d[MCLBN_FP_UNIT_SIZE * 2 * 3];
+	mclBnFp2 x, y, z;
 } mclBnG2;
 
 typedef struct {
-	uint64_t d[MCLBN_FP_UNIT_SIZE * 12];
+	mclBnFp d[12];
 } mclBnGT;
 
-typedef struct {
-	uint64_t d[MCLBN_FP_UNIT_SIZE];
-} mclBnFp;
-
-typedef struct {
-	mclBnFp d[2];
-} mclBnFp2;
-
 #endif
 
 #include <mcl/curve_type.h>
 
+#define MCLBN_IO_EC_AFFINE 0
+#define MCLBN_IO_EC_PROJ 1024
 #define MCLBN_IO_SERIALIZE_HEX_STR 2048
+
 // for backword compatibility
 enum {
 	mclBn_CurveFp254BNb = 0,
@@ -115,7 +127,7 @@ enum {
 };
 
 // return 0xABC which means A.BC
-MCLBN_DLL_API int mclBn_getVersion();
+MCLBN_DLL_API int mclBn_getVersion(void);
 /*
 	init library
 	@param curve [in] type of bn curve
@@ -131,6 +143,7 @@ MCLBN_DLL_API int mclBn_getVersion();
 */
 MCLBN_DLL_API int mclBn_init(int curve, int compiledTimeVar);
 
+MCLBN_DLL_API int mclBn_getCurveType(void);
 
 /*
 	pairing : G1 x G2 -> GT
@@ -179,10 +192,20 @@ MCLBN_DLL_API mclSize mclBn_getFieldOrder(char *buf, mclSize maxBufSize);
 
 /*
 	set ETH serialization mode for BLS12-381
-	@param ETHserialization [in] 1:enable,  0:disable
+	@param enable [in] 1:enable,  0:disable
 	@note ignore the flag if curve is not BLS12-381
 */
-MCLBN_DLL_API void mclBn_setETHserialization(int ETHserialization);
+MCLBN_DLL_API void mclBn_setETHserialization(int enable);
+
+// return 1 if ETH serialization mode else 0
+MCLBN_DLL_API int mclBn_getETHserialization(void);
+
+/*
+	set map-to-function to mode (only support MCL_MAP_TO_MODE_HASH_TO_CURVE_07)
+	return 0 if success else -1
+*/
+MCLBN_DLL_API int mclBn_setMapToMode(int mode);
+
 ////////////////////////////////////////////////
 /*
 	deserialize
@@ -238,6 +261,8 @@ MCLBN_DLL_API void mclBnFp2_clear(mclBnFp2 *x);
 // set x to y
 MCLBN_DLL_API void mclBnFr_setInt(mclBnFr *y, mclInt x);
 MCLBN_DLL_API void mclBnFr_setInt32(mclBnFr *y, int x);
+MCLBN_DLL_API void mclBnFp_setInt(mclBnFp *y, mclInt x);
+MCLBN_DLL_API void mclBnFp_setInt32(mclBnFp *y, int x);
 
 // x = buf & (1 << bitLen(r)) - 1
 // if (x >= r) x &= (1 << (bitLen(r) - 1)) - 1
@@ -245,25 +270,49 @@ MCLBN_DLL_API void mclBnFr_setInt32(mclBnFr *y, int x);
 MCLBN_DLL_API int mclBnFr_setLittleEndian(mclBnFr *x, const void *buf, mclSize bufSize);
 MCLBN_DLL_API int mclBnFp_setLittleEndian(mclBnFp *x, const void *buf, mclSize bufSize);
 
+/*
+	write a value as little endian
+	return written size if success else 0
+	@note buf[0] = 0 and return 1 if the value is zero
+*/
+MCLBN_DLL_API mclSize mclBnFr_getLittleEndian(void *buf, mclSize maxBufSize, const mclBnFr *x);
+MCLBN_DLL_API mclSize mclBnFp_getLittleEndian(void *buf, mclSize maxBufSize, const mclBnFp *x);
+
 // set (buf mod r) to x
 // return 0 if bufSize <= (byte size of Fr * 2) else -1
 MCLBN_DLL_API int mclBnFr_setLittleEndianMod(mclBnFr *x, const void *buf, mclSize bufSize);
+MCLBN_DLL_API int mclBnFr_setBigEndianMod(mclBnFr *x, const void *buf, mclSize bufSize);
 // set (buf mod p) to x
 // return 0 if bufSize <= (byte size of Fp * 2) else -1
 MCLBN_DLL_API int mclBnFp_setLittleEndianMod(mclBnFp *x, const void *buf, mclSize bufSize);
+MCLBN_DLL_API int mclBnFp_setBigEndianMod(mclBnFp *x, const void *buf, mclSize bufSize);
 
 // return 1 if true and 0 otherwise
 MCLBN_DLL_API int mclBnFr_isValid(const mclBnFr *x);
 MCLBN_DLL_API int mclBnFr_isEqual(const mclBnFr *x, const mclBnFr *y);
 MCLBN_DLL_API int mclBnFr_isZero(const mclBnFr *x);
 MCLBN_DLL_API int mclBnFr_isOne(const mclBnFr *x);
+MCLBN_DLL_API int mclBnFr_isOdd(const mclBnFr *x);
+// return 1 if half <= x < r, where half = (r + 1) / 2 else 0
+MCLBN_DLL_API int mclBnFr_isNegative(const mclBnFr *x);
 
+MCLBN_DLL_API int mclBnFp_isValid(const mclBnFp *x);
 MCLBN_DLL_API int mclBnFp_isEqual(const mclBnFp *x, const mclBnFp *y);
+MCLBN_DLL_API int mclBnFp_isZero(const mclBnFp *x);
+MCLBN_DLL_API int mclBnFp_isOne(const mclBnFp *x);
+MCLBN_DLL_API int mclBnFp_isOdd(const mclBnFp *x);
+// return 1 if half <= x < p, where half = (p + 1) / 2 else 0
+MCLBN_DLL_API int mclBnFp_isNegative(const mclBnFp *x);
+
 MCLBN_DLL_API int mclBnFp2_isEqual(const mclBnFp2 *x, const mclBnFp2 *y);
+MCLBN_DLL_API int mclBnFp2_isZero(const mclBnFp2 *x);
+MCLBN_DLL_API int mclBnFp2_isOne(const mclBnFp2 *x);
+
 
 #ifndef MCL_DONT_USE_CSRPNG
 // return 0 if success
 MCLBN_DLL_API int mclBnFr_setByCSPRNG(mclBnFr *x);
+MCLBN_DLL_API int mclBnFp_setByCSPRNG(mclBnFp *x);
 
 /*
 	set user-defined random function for setByCSPRNG
@@ -294,6 +343,28 @@ MCLBN_DLL_API void mclBnFr_sub(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
 MCLBN_DLL_API void mclBnFr_mul(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
 MCLBN_DLL_API void mclBnFr_div(mclBnFr *z, const mclBnFr *x, const mclBnFr *y);
 
+MCLBN_DLL_API void mclBnFp_neg(mclBnFp *y, const mclBnFp *x);
+MCLBN_DLL_API void mclBnFp_inv(mclBnFp *y, const mclBnFp *x);
+MCLBN_DLL_API void mclBnFp_sqr(mclBnFp *y, const mclBnFp *x);
+MCLBN_DLL_API void mclBnFp_add(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+MCLBN_DLL_API void mclBnFp_sub(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+MCLBN_DLL_API void mclBnFp_mul(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+MCLBN_DLL_API void mclBnFp_div(mclBnFp *z, const mclBnFp *x, const mclBnFp *y);
+
+MCLBN_DLL_API void mclBnFp2_neg(mclBnFp2 *y, const mclBnFp2 *x);
+MCLBN_DLL_API void mclBnFp2_inv(mclBnFp2 *y, const mclBnFp2 *x);
+MCLBN_DLL_API void mclBnFp2_sqr(mclBnFp2 *y, const mclBnFp2 *x);
+MCLBN_DLL_API void mclBnFp2_add(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+MCLBN_DLL_API void mclBnFp2_sub(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+MCLBN_DLL_API void mclBnFp2_mul(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+MCLBN_DLL_API void mclBnFp2_div(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y);
+
+// y is one of square root of x
+// return 0 if success else -1
+MCLBN_DLL_API int mclBnFr_squareRoot(mclBnFr *y, const mclBnFr *x);
+MCLBN_DLL_API int mclBnFp_squareRoot(mclBnFp *y, const mclBnFp *x);
+MCLBN_DLL_API int mclBnFp2_squareRoot(mclBnFp2 *y, const mclBnFp2 *x);
+
 ////////////////////////////////////////////////
 // set zero
 MCLBN_DLL_API void mclBnG1_clear(mclBnG1 *x);
@@ -365,26 +436,37 @@ MCLBN_DLL_API int mclBnGT_isZero(const mclBnGT *x);
 MCLBN_DLL_API int mclBnGT_isOne(const mclBnGT *x);
 
 MCLBN_DLL_API void mclBnGT_neg(mclBnGT *y, const mclBnGT *x);
-MCLBN_DLL_API void mclBnGT_inv(mclBnGT *y, const mclBnGT *x);
 MCLBN_DLL_API void mclBnGT_sqr(mclBnGT *y, const mclBnGT *x);
 MCLBN_DLL_API void mclBnGT_add(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
 MCLBN_DLL_API void mclBnGT_sub(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
 MCLBN_DLL_API void mclBnGT_mul(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
 MCLBN_DLL_API void mclBnGT_div(mclBnGT *z, const mclBnGT *x, const mclBnGT *y);
 
+// y = conjugate of x in Fp12, which is equal to the inverse of x if |x|^r = 1
+MCLBN_DLL_API void mclBnGT_inv(mclBnGT *y, const mclBnGT *x);
+// use invGeneric when x in Fp12 is not in GT
+MCLBN_DLL_API void mclBnGT_invGeneric(mclBnGT *y, const mclBnGT *x);
+
 /*
 	pow for all elements of Fp12
 */
 MCLBN_DLL_API void mclBnGT_powGeneric(mclBnGT *z, const mclBnGT *x, const mclBnFr *y);
 /*
-	pow for only {x|x^r = 1} in Fp12 by GLV method
+	pow for only {x|x^r = 1} in GT by GLV method
 	the value generated by pairing satisfies the condition
 */
 MCLBN_DLL_API void mclBnGT_pow(mclBnGT *z, const mclBnGT *x, const mclBnFr *y);
 
+// z = sum_{i=0}^{n-1} x[i] y[i]
+MCLBN_DLL_API void mclBnG1_mulVec(mclBnG1 *z, const mclBnG1 *x, const mclBnFr *y, mclSize n);
+MCLBN_DLL_API void mclBnG2_mulVec(mclBnG2 *z, const mclBnG2 *x, const mclBnFr *y, mclSize n);
+MCLBN_DLL_API void mclBnGT_powVec(mclBnGT *z, const mclBnGT *x, const mclBnFr *y, mclSize n);
+
 MCLBN_DLL_API void mclBn_pairing(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y);
 MCLBN_DLL_API void mclBn_finalExp(mclBnGT *y, const mclBnGT *x);
 MCLBN_DLL_API void mclBn_millerLoop(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y);
+// z = prod_{i=0}^{n-1} millerLoop(x[i], y[i])
+MCLBN_DLL_API void mclBn_millerLoopVec(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y, mclSize n);
 
 // return precomputedQcoeffSize * sizeof(Fp6) / sizeof(uint64_t)
 MCLBN_DLL_API int mclBn_getUint64NumToPrecompute(void);
@@ -410,7 +492,8 @@ MCLBN_DLL_API int mclBn_G2LagrangeInterpolation(mclBnG2 *out, const mclBnFr *xVe
 /*
 	evaluate polynomial
 	out = f(x) = c[0] + c[1] * x + c[2] * x^2 + ... + c[cSize - 1] * x^(cSize - 1)
-	@note cSize >= 2
+	return 0 if success else -1
+	@note cSize >= 1
 */
 MCLBN_DLL_API int mclBn_FrEvaluatePolynomial(mclBnFr *out, const mclBnFr *cVec, mclSize cSize, const mclBnFr *x);
 MCLBN_DLL_API int mclBn_G1EvaluatePolynomial(mclBnG1 *out, const mclBnG1 *cVec, mclSize cSize, const mclBnFr *x);

include/mcl/config.hpp

@@ -0,0 +1,34 @@
+#pragma once
+/**
+	@file
+	@brief constant macro
+	@author MITSUNARI Shigeo(@herumi)
+	@license modified new BSD license
+	http://opensource.org/licenses/BSD-3-Clause
+*/
+#include <cybozu/inttype.hpp>
+
+#ifndef MCL_SIZEOF_UNIT
+	#if defined(CYBOZU_OS_BIT) && (CYBOZU_OS_BIT == 32)
+		#define MCL_SIZEOF_UNIT 4
+	#else
+		#define MCL_SIZEOF_UNIT 8
+	#endif
+#endif
+
+#ifndef MCL_MAX_BIT_SIZE
+	#define MCL_MAX_BIT_SIZE 512
+#endif
+
+namespace mcl {
+
+namespace fp {
+
+#if MCL_SIZEOF_UNIT == 8
+typedef uint64_t Unit;
+#else
+typedef uint32_t Unit;
+#endif
+#define MCL_UNIT_BIT_SIZE (MCL_SIZEOF_UNIT * 8)
+
+} } // mcl::fp

include/mcl/conversion.hpp

@@ -1,6 +1,7 @@
 #pragma once
 #include <cybozu/itoa.hpp>
 #include <cybozu/stream.hpp>
+#include <mcl/config.hpp>
 /**
 	@file
 	@brief convertion bin/dec/hex <=> array
@@ -15,6 +16,44 @@
 
 namespace mcl { namespace fp {
 
+/*
+	treat src[] as a little endian and set dst[]
+	fill remain of dst if sizeof(D) * dstN > sizeof(S) * srcN
+	return false if sizeof(D) * dstN < sizeof(S) * srcN
+*/
+template<class D, class S>
+bool convertArrayAsLE(D *dst, size_t dstN, const S *src, size_t srcN)
+{
+	char assert_D_is_unsigned[D(-1) < 0 ? -1 : 1];
+	char assert_S_is_unsigned[S(-1) < 0 ? -1 : 1];
+	(void)assert_D_is_unsigned;
+	(void)assert_S_is_unsigned;
+	if (sizeof(D) * dstN < sizeof(S) * srcN) return false;
+	size_t pos = 0;
+	if (sizeof(D) < sizeof(S)) {
+		for (size_t i = 0; i < srcN; i++) {
+			S s = src[i];
+			for (size_t j = 0; j < sizeof(S); j += sizeof(D)) {
+				dst[pos++] = D(s);
+				s >>= sizeof(D) * 8;
+			}
+		}
+		for (; pos < dstN; pos++) {
+			dst[pos] = 0;
+		}
+	} else {
+		for (size_t i = 0; i < dstN; i++) {
+			D u = 0;
+			for (size_t j = 0; j < sizeof(D); j += sizeof(S)) {
+				S s = (pos < srcN) ? src[pos++] : 0;
+				u |= D(s) << (j * 8);
+			}
+			dst[i] = u;
+		}
+	}
+	return true;
+}
+
 namespace local {
 
 inline bool isSpace(char c)
@@ -208,8 +247,6 @@ inline bool parsePrefix(size_t *readSize, bool *isMinus, int *base, const char *
 			if (*base == 0 || *base == 2) {
 				*base = 2;
 				pos += 2;
-			} else {
-				return false;
 			}
 		}
 	}
@@ -337,31 +374,22 @@ inline size_t binToArray(UT *x, size_t maxN, const char *buf, size_t bufSize)
 }
 
 /*
-	little endian x[0, xn) to buf
+	little endian t[0, tn) to buf
 	return written size if success else 0
 	data is buf[bufSize - retval, bufSize)
 */
-template<class UT>
+inline size_t inner_arrayToDec(char *buf, size_t bufSize, uint32_t *t, size_t tn)
-inline size_t arrayToDec(char *buf, size_t bufSize, const UT *x, size_t xn)
 {
-	const size_t maxN = 64;
-	uint32_t t[maxN];
-	if (sizeof(UT) == 8) {
-		xn *= 2;
-	}
-	if (xn > maxN) return 0;
-	memcpy(t, x, xn * sizeof(t[0]));
-
 	const size_t width = 9;
 	const uint32_t i1e9 = 1000000000U;
 	size_t pos = 0;
 	for (;;) {
-		uint32_t r = local::divU32(t, t, xn, i1e9);
+		uint32_t r = local::divU32(t, t, tn, i1e9);
-		while (xn > 0 && t[xn - 1] == 0) xn--;
+		while (tn > 0 && t[tn - 1] == 0) tn--;
 		size_t len = cybozu::itoa_local::uintToDec(buf, bufSize - pos, r);
 		if (len == 0) return 0;
 		assert(0 < len && len <= width);
-		if (xn == 0) return pos + len;
+		if (tn == 0) return pos + len;
 		// fill (width - len) '0'
 		for (size_t j = 0; j < width - len; j++) {
 			buf[bufSize - pos - width + j] = '0';
@@ -370,21 +398,33 @@ inline size_t arrayToDec(char *buf, size_t bufSize, const UT *x, size_t xn)
 	}
 }
 
+inline size_t arrayToDec(char *buf, size_t bufSize, const uint32_t *x, size_t xn)
+{
+	uint32_t *t = (uint32_t*)CYBOZU_ALLOCA(sizeof(uint32_t) * xn);
+	memcpy(t, x, sizeof(uint32_t) * xn);
+	return inner_arrayToDec(buf, bufSize, t, xn);
+}
+
+inline size_t arrayToDec(char *buf, size_t bufSize, const uint64_t *x, size_t xn)
+{
+	uint32_t *t = (uint32_t*)CYBOZU_ALLOCA(sizeof(uint32_t) * xn * 2);
+	for (size_t i = 0; i < xn; i++) {
+		uint64_t v = x[i];
+		t[i * 2 + 0] = uint32_t(v);
+		t[i * 2 + 1] = uint32_t(v >> 32);
+	}
+	return inner_arrayToDec(buf, bufSize, t, xn * 2);
+}
+
 /*
 	convert buf[0, bufSize) to x[0, num)
 	return written num if success else 0
 */
-template<class UT>
+inline size_t decToArray(uint32_t *x, size_t maxN, const char *buf, size_t bufSize)
-inline size_t decToArray(UT *_x, size_t maxN, const char *buf, size_t bufSize)
 {
-	assert(sizeof(UT) == 4 || sizeof(UT) == 8);
 	const size_t width = 9;
 	const uint32_t i1e9 = 1000000000U;
 	if (maxN == 0) return 0;
-	if (sizeof(UT) == 8) {
-		maxN *= 2;
-	}
-	uint32_t *x = reinterpret_cast<uint32_t*>(_x);
 	size_t xn = 1;
 	x[0] = 0;
 	while (bufSize > 0) {
@@ -406,10 +446,20 @@ inline size_t decToArray(UT *_x, size_t maxN, const char *buf, size_t bufSize)
 		buf += n;
 		bufSize -= n;
 	}
-	if (sizeof(UT) == 8 && (xn & 1)) {
+	return xn;
-		x[xn++] = 0;
+}
+
+inline size_t decToArray(uint64_t *x, size_t maxN, const char *buf, size_t bufSize)
+{
+	uint32_t *t = (uint32_t*)CYBOZU_ALLOCA(sizeof(uint32_t) * maxN * 2);
+	size_t xn = decToArray(t, maxN * 2, buf, bufSize);
+	if (xn & 1) {
+		t[xn++] = 0;
+	}
+	for (size_t i = 0; i < xn; i += 2) {
+		x[i / 2] = (uint64_t(t[i + 1]) << 32) | t[i];
 	}
-	return xn / (sizeof(UT) / 4);
+	return xn / 2;
 }
 
 /*

include/mcl/curve_type.h

@@ -18,7 +18,7 @@ enum {
 
 	/*
 		for only G1
-		the size of curve must be less or equal to MCLBN_FP_UNIT_SIZE
+		the size of curve must be <= MCLBN_FP_UNIT_SIZE
 	*/
 	MCL_EC_BEGIN = 100,
 	MCL_SECP192K1 = MCL_EC_BEGIN,
@@ -29,7 +29,90 @@ enum {
 	MCL_NIST_P192 = 105,
 	MCL_NIST_P224 = 106,
 	MCL_NIST_P256 = 107,
-	MCL_EC_END = MCL_NIST_P256 + 1,
+	MCL_SECP160K1 = 108,
+	MCL_P160_1 = 109,
+	MCL_EC_END = MCL_P160_1 + 1,
 	MCL_NIST_P384 = MCL_SECP384R1,
 	MCL_NIST_P521 = MCL_SECP521R1
 };
+
+/*
+	remark : if irtf-cfrg-hash-to-curve is compeletely fixed, then
+	MCL_MAP_TO_MODE_WB19, MCL_MAP_TO_MODE_HASH_TO_CURVE_0? will be removed and
+	only MCL_MAP_TO_MODE_HASH_TO_CURVE will be available.
+*/
+enum {
+	MCL_MAP_TO_MODE_ORIGINAL, // see MapTo::calcBN
+	MCL_MAP_TO_MODE_TRY_AND_INC, // try-and-incremental-x
+	MCL_MAP_TO_MODE_ETH2, // (deprecated) old eth2.0 spec
+	MCL_MAP_TO_MODE_WB19, // (deprecated) used in new eth2.0 spec
+	MCL_MAP_TO_MODE_HASH_TO_CURVE_05 = MCL_MAP_TO_MODE_WB19, // (deprecated) draft-irtf-cfrg-hash-to-curve-05
+	MCL_MAP_TO_MODE_HASH_TO_CURVE_06, // (deprecated) draft-irtf-cfrg-hash-to-curve-06
+	MCL_MAP_TO_MODE_HASH_TO_CURVE_07, // draft-irtf-cfrg-hash-to-curve-07
+	MCL_MAP_TO_MODE_HASH_TO_CURVE = MCL_MAP_TO_MODE_HASH_TO_CURVE_07 // the latset version
+};
+
+#ifdef __cplusplus
+
+#include <string.h>
+#include <assert.h>
+
+namespace mcl {
+
+struct CurveParam {
+	/*
+		y^2 = x^3 + b
+		i^2 = -1
+		xi = xi_a + i
+		v^3 = xi
+		w^2 = v
+	*/
+	const char *z;
+	int b; // y^2 = x^3 + b
+	int xi_a; // xi = xi_a + i
+	/*
+		BN254, BN381 : Dtype
+		BLS12-381 : Mtype
+	*/
+	bool isMtype;
+	int curveType; // same in curve_type.h
+	bool operator==(const CurveParam& rhs) const
+	{
+		return strcmp(z, rhs.z) == 0 && b == rhs.b && xi_a == rhs.xi_a && isMtype == rhs.isMtype;
+	}
+	bool operator!=(const CurveParam& rhs) const { return !operator==(rhs); }
+};
+
+const CurveParam BN254 = { "-0x4080000000000001", 2, 1, false, MCL_BN254 }; // -(2^62 + 2^55 + 1)
+// provisional(experimental) param with maxBitSize = 384
+const CurveParam BN381_1 = { "-0x400011000000000000000001", 2, 1, false, MCL_BN381_1 }; // -(2^94 + 2^76 + 2^72 + 1) // A Family of Implementation-Friendly BN Elliptic Curves
+const CurveParam BN381_2 = { "-0x400040090001000000000001", 2, 1, false, MCL_BN381_2 }; // -(2^94 + 2^78 + 2^67 + 2^64 + 2^48 + 1) // used in relic-toolkit
+const CurveParam BN462 = { "0x4001fffffffffffffffffffffbfff", 5, 2, false, MCL_BN462 }; // 2^114 + 2^101 - 2^14 - 1 // https://eprint.iacr.org/2017/334
+const CurveParam BN_SNARK1 = { "4965661367192848881", 3, 9, false, MCL_BN_SNARK1 };
+const CurveParam BLS12_381 = { "-0xd201000000010000", 4, 1, true, MCL_BLS12_381 };
+const CurveParam BN160 = { "0x4000000031", 3, 4, false, MCL_BN160 };
+
+#ifdef __clang__
+	#pragma GCC diagnostic push
+	#pragma GCC diagnostic ignored "-Wreturn-type-c-linkage"
+#endif
+inline const CurveParam* getCurveParam(int type)
+{
+	switch (type) {
+	case MCL_BN254: return &mcl::BN254;
+	case MCL_BN381_1: return &mcl::BN381_1;
+	case MCL_BN381_2: return &mcl::BN381_2;
+	case MCL_BN462: return &mcl::BN462;
+	case MCL_BN_SNARK1: return &mcl::BN_SNARK1;
+	case MCL_BLS12_381: return &mcl::BLS12_381;
+	case MCL_BN160: return &mcl::BN160;
+	default:
+		return 0;
+	}
+}
+#ifdef __clang__
+	#pragma GCC diagnostic pop
+#endif
+
+} // mcl
+#endif

include/mcl/ecdsa.h

@@ -1,105 +0,0 @@
-#pragma once
-/**
-	@file
-	@brief C interface of ECDSA
-	@author MITSUNARI Shigeo(@herumi)
-	@license modified new BSD license
-	http://opensource.org/licenses/BSD-3-Clause
-*/
-#include <stdint.h> // for uint64_t, uint8_t
-#include <stdlib.h> // for size_t
-
-#if defined(_MSC_VER)
-	#ifdef ECDSA_DLL_EXPORT
-		#define ECDSA_DLL_API __declspec(dllexport)
-	#else
-		#define ECDSA_DLL_API __declspec(dllimport)
-		#ifndef ECDSA_NO_AUTOLINK
-			#pragma comment(lib, "mclecdsa.lib")
-		#endif
-	#endif
-#elif defined(__EMSCRIPTEN__)
-	#define ECDSA_DLL_API __attribute__((used))
-#else
-	#define ECDSA_DLL_API
-#endif
-
-#ifndef mclSize
-	#ifdef __EMSCRIPTEN__
-		// avoid 64-bit integer
-		#define mclSize unsigned int
-		#define mclInt int
-	#else
-		// use #define for cgo
-		#define mclSize size_t
-		#define mclInt int64_t
-	#endif
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifdef ECDSA_NOT_DEFINE_STRUCT
-
-typedef struct ecdsaSecretKey ecdsaSecretKey;
-typedef struct ecdsaPublicKey ecdsaPublicKey;
-typedef struct ecdsaSignature ecdsaSignature;
-
-#else
-
-typedef struct {
-	uint64_t d[4];
-} ecdsaSecretKey;
-
-typedef struct {
-	uint64_t d[4 * 3];
-} ecdsaPublicKey;
-
-typedef struct {
-	uint64_t d[4 * 2];
-} ecdsaSignature;
-
-#endif
-
-struct ecdsaPrecomputedPublicKey;
-
-/*
-	init library
-	return 0 if success
-	@note not threadsafe
-*/
-ECDSA_DLL_API int ecdsaInit(void);
-
-// return written byte size if success else 0
-ECDSA_DLL_API mclSize ecdsaSecretKeySerialize(void *buf, mclSize maxBufSize, const ecdsaSecretKey *sec);
-ECDSA_DLL_API mclSize ecdsaPublicKeySerialize(void *buf, mclSize maxBufSize, const ecdsaPublicKey *pub);
-ECDSA_DLL_API mclSize ecdsaSignatureSerialize(void *buf, mclSize maxBufSize, const ecdsaSignature *sig);
-
-// return read byte size if sucess else 0
-ECDSA_DLL_API mclSize ecdsaSecretKeyDeserialize(ecdsaSecretKey* sec, const void *buf, mclSize bufSize);
-ECDSA_DLL_API mclSize ecdsaPublicKeyDeserialize(ecdsaPublicKey* pub, const void *buf, mclSize bufSize);
-ECDSA_DLL_API mclSize ecdsaSignatureDeserialize(ecdsaSignature* sig, const void *buf, mclSize bufSize);
-
-//	return 0 if success
-ECDSA_DLL_API int ecdsaSecretKeySetByCSPRNG(ecdsaSecretKey *sec);
-
-ECDSA_DLL_API void ecdsaGetPublicKey(ecdsaPublicKey *pub, const ecdsaSecretKey *sec);
-
-ECDSA_DLL_API void ecdsaSign(ecdsaSignature *sig, const ecdsaSecretKey *sec, const void *m, mclSize size);
-
-// return 1 if valid
-ECDSA_DLL_API int ecdsaVerify(const ecdsaSignature *sig, const ecdsaPublicKey *pub, const void *m, mclSize size);
-ECDSA_DLL_API int ecdsaVerifyPrecomputed(const ecdsaSignature *sig, const ecdsaPrecomputedPublicKey *pub, const void *m, mclSize size);
-
-// return nonzero if success
-ECDSA_DLL_API ecdsaPrecomputedPublicKey *ecdsaPrecomputedPublicKeyCreate();
-// call this function to avoid memory leak
-ECDSA_DLL_API void ecdsaPrecomputedPublicKeyDestroy(ecdsaPrecomputedPublicKey *ppub);
-// return 0 if success
-ECDSA_DLL_API int ecdsaPrecomputedPublicKeyInit(ecdsaPrecomputedPublicKey *ppub, const ecdsaPublicKey *pub);
-
-#ifdef __cplusplus
-}
-#endif
-

include/mcl/ecdsa.hpp

@@ -1,257 +0,0 @@
-#pragma once
-/**
-	@file
-	@brief ECDSA
-	@author MITSUNARI Shigeo(@herumi)
-	@license modified new BSD license
-	http://opensource.org/licenses/BSD-3-Clause
-*/
-#include <mcl/fp.hpp>
-#include <mcl/ec.hpp>
-#include <mcl/ecparam.hpp>
-#include <mcl/window_method.hpp>
-
-namespace mcl { namespace ecdsa {
-
-namespace local {
-
-#ifndef MCLSHE_WIN_SIZE
-	#define MCLSHE_WIN_SIZE 10
-#endif
-static const size_t winSize = MCLSHE_WIN_SIZE;
-
-struct FpTag;
-struct ZnTag;
-
-} // mcl::ecdsa::local
-
-typedef mcl::FpT<local::FpTag, 256> Fp;
-typedef mcl::FpT<local::ZnTag, 256> Zn;
-typedef mcl::EcT<Fp> Ec;
-
-namespace local {
-
-struct Param {
-	mcl::EcParam ecParam;
-	Ec P;
-	mcl::fp::WindowMethod<Ec> Pbase;
-};
-
-inline Param& getParam()
-{
-	static Param p;
-	return p;
-}
-
-inline void be32toZn(Zn& x, const mcl::fp::Unit *buf)
-{
-	const size_t n = 32;
-	const unsigned char *p = (const unsigned char*)buf;
-	unsigned char be[n];
-	for (size_t i = 0; i < n; i++) {
-		be[i] = p[n - 1 - i];
-	}
-	x.setArrayMaskMod(be, n);
-}
-
-/*
-	y = x mod n
-*/
-inline void FpToZn(Zn& y, const Fp& x)
-{
-	fp::Block b;
-	x.getBlock(b);
-	y.setArrayMaskMod(b.p, b.n);
-}
-
-inline void setHashOf(Zn& x, const void *msg, size_t msgSize)
-{
-	mcl::fp::Unit xBuf[256 / 8 / sizeof(mcl::fp::Unit)];
-	uint32_t hashSize = mcl::fp::sha256(xBuf, sizeof(xBuf), msg, (uint32_t)msgSize);
-	assert(hashSize == sizeof(xBuf));
-	(void)hashSize;
-	be32toZn(x, xBuf);
-}
-
-} // mcl::ecdsa::local
-
-const local::Param& param = local::getParam();
-
-inline void init(bool *pb)
-{
-	const mcl::EcParam& ecParam = mcl::ecparam::secp256k1;
-	Zn::init(pb, ecParam.n);
-	if (!*pb) return;
-	Fp::init(pb, ecParam.p);
-	if (!*pb) return;
-	Ec::init(pb, ecParam.a, ecParam.b);
-	if (!*pb) return;
-	Zn::setIoMode(16);
-	Fp::setIoMode(16);
-	Ec::setIoMode(mcl::IoEcAffine);
-	local::Param& p = local::getParam();
-	p.ecParam = ecParam;
-	Fp x, y;
-	x.setStr(pb, ecParam.gx);
-	if (!*pb) return;
-	y.setStr(pb, ecParam.gy);
-	if (!*pb) return;
-	p.P.set(pb, x, y);
-	if (!*pb) return;
-	p.Pbase.init(pb, p.P, ecParam.bitSize, local::winSize);
-}
-
-#ifndef CYBOZU_DONT_USE_EXCEPTION
-inline void init()
-{
-	bool b;
-	init(&b);
-	if (!b) throw cybozu::Exception("ecdsa:init");
-}
-#endif
-
-typedef Zn SecretKey;
-typedef Ec PublicKey;
-
-struct PrecomputedPublicKey {
-	mcl::fp::WindowMethod<Ec> pubBase_;
-	void init(bool *pb, const PublicKey& pub)
-	{
-		pubBase_.init(pb, pub, param.ecParam.bitSize, local::winSize);
-	}
-#ifndef CYBOZU_DONT_USE_EXCEPTION
-	void init(const PublicKey& pub)
-	{
-		bool b;
-		init(&b, pub);
-		if (!b) throw cybozu::Exception("ecdsa:PrecomputedPublicKey:init");
-	}
-#endif
-};
-
-inline void getPublicKey(PublicKey& pub, const SecretKey& sec)
-{
-	Ec::mul(pub, param.P, sec);
-	pub.normalize();
-}
-
-struct Signature : public mcl::fp::Serializable<Signature> {
-	Zn r, s;
-	template<class InputStream>
-	void load(bool *pb, InputStream& is, int ioMode = IoSerialize)
-	{
-		r.load(pb, is, ioMode); if (!*pb) return;
-		s.load(pb, is, ioMode);
-	}
-	template<class OutputStream>
-	void save(bool *pb, OutputStream& os, int ioMode = IoSerialize) const
-	{
-		const char sep = *fp::getIoSeparator(ioMode);
-		r.save(pb, os, ioMode); if (!*pb) return;
-		if (sep) {
-			cybozu::writeChar(pb, os, sep);
-			if (!*pb) return;
-		}
-		s.save(pb, os, ioMode);
-	}
-#ifndef CYBOZU_DONT_USE_EXCEPTION
-	template<class InputStream>
-	void load(InputStream& is, int ioMode = IoSerialize)
-	{
-		bool b;
-		load(&b, is, ioMode);
-		if (!b) throw cybozu::Exception("ecdsa:Signature:load");
-	}
-	template<class OutputStream>
-	void save(OutputStream& os, int ioMode = IoSerialize) const
-	{
-		bool b;
-		save(&b, os, ioMode);
-		if (!b) throw cybozu::Exception("ecdsa:Signature:save");
-	}
-#endif
-#ifndef CYBOZU_DONT_USE_STRING
-	friend std::istream& operator>>(std::istream& is, Signature& self)
-	{
-		self.load(is, fp::detectIoMode(Ec::getIoMode(), is));
-		return is;
-	}
-	friend std::ostream& operator<<(std::ostream& os, const Signature& self)
-	{
-		self.save(os, fp::detectIoMode(Ec::getIoMode(), os));
-		return os;
-	}
-#endif
-};
-
-inline void sign(Signature& sig, const SecretKey& sec, const void *msg, size_t msgSize)
-{
-	Zn& r = sig.r;
-	Zn& s = sig.s;
-	Zn z, k;
-	local::setHashOf(z, msg, msgSize);
-	Ec Q;
-	for (;;) {
-		k.setByCSPRNG();
-		param.Pbase.mul(Q, k);
-		if (Q.isZero()) continue;
-		Q.normalize();
-		local::FpToZn(r, Q.x);
-		if (r.isZero()) continue;
-		Zn::mul(s, r, sec);
-		s += z;
-		if (s.isZero()) continue;
-		s /= k;
-		return;
-	}
-}
-
-namespace local {
-
-inline void mulDispatch(Ec& Q, const PublicKey& pub, const Zn& y)
-{
-	Ec::mul(Q, pub, y);
-}
-
-inline void mulDispatch(Ec& Q, const PrecomputedPublicKey& ppub, const Zn& y)
-{
-	ppub.pubBase_.mul(Q, y);
-}
-
-template<class Pub>
-inline bool verify(const Signature& sig, const Pub& pub, const void *msg, size_t msgSize)
-{
-	const Zn& r = sig.r;
-	const Zn& s = sig.s;
-	if (r.isZero() || s.isZero()) return false;
-	Zn z, w, u1, u2;
-	local::setHashOf(z, msg, msgSize);
-	Zn::inv(w, s);
-	Zn::mul(u1, z, w);
-	Zn::mul(u2, r, w);
-	Ec Q1, Q2;
-	param.Pbase.mul(Q1, u1);
-//	Ec::mul(Q2, pub, u2);
-	local::mulDispatch(Q2, pub, u2);
-	Q1 += Q2;
-	if (Q1.isZero()) return false;
-	Q1.normalize();
-	Zn x;
-	local::FpToZn(x, Q1.x);
-	return r == x;
-}
-
-} // mcl::ecdsa::local
-
-inline bool verify(const Signature& sig, const PublicKey& pub, const void *msg, size_t msgSize)
-{
-	return local::verify(sig, pub, msg, msgSize);
-}
-
-inline bool verify(const Signature& sig, const PrecomputedPublicKey& ppub, const void *msg, size_t msgSize)
-{
-	return local::verify(sig, ppub, msg, msgSize);
-}
-
-} } // mcl::ecdsa
-

include/mcl/ecparam.hpp

@@ -6,10 +6,23 @@
 	@license modified new BSD license
 	http://opensource.org/licenses/BSD-3-Clause
 */
-#include <mcl/ec.hpp>
 #include <mcl/curve_type.h>
 
-namespace mcl { namespace ecparam {
+namespace mcl {
+
+struct EcParam {
+	const char *name;
+	const char *p;
+	const char *a;
+	const char *b;
+	const char *gx;
+	const char *gy;
+	const char *n;
+	size_t bitSize; // bit length of p
+	int curveType;
+};
+
+namespace ecparam {
 
 const struct mcl::EcParam secp160k1 = {
 	"secp160k1",
@@ -20,9 +33,9 @@ const struct mcl::EcParam secp160k1 = {
 	"0x938cf935318fdced6bc28286531733c3f03c4fee",
 	"0x100000000000000000001b8fa16dfab9aca16b6b3",
 	160,
-	-1
+	MCL_SECP160K1
 };
-// p=2^160 + 7
+// p=2^160 + 7 (for test)
 const struct mcl::EcParam p160_1 = {
 	"p160_1",
 	"0x10000000000000000000000000000000000000007",
@@ -32,7 +45,7 @@ const struct mcl::EcParam p160_1 = {
 	"1236612389951462151661156731535316138439983579284",
 	"1461501637330902918203683518218126812711137002561",
 	161,
-	-1
+	MCL_P160_1
 };
 const struct mcl::EcParam secp192k1 = {
 	"secp192k1",
@@ -181,11 +194,16 @@ inline const mcl::EcParam* getEcParam(int curve)
 	case MCL_SECP224K1: return &ecparam::secp224k1;
 	case MCL_SECP256K1: return &ecparam::secp256k1;
 	case MCL_SECP384R1: return &ecparam::secp384r1;
+	case MCL_SECP521R1: return &ecparam::secp521r1;
 	case MCL_NIST_P192: return &ecparam::NIST_P192;
 	case MCL_NIST_P224: return &ecparam::NIST_P224;
 	case MCL_NIST_P256: return &ecparam::NIST_P256;
+	case MCL_SECP160K1: return &ecparam::secp160k1;
+	case MCL_P160_1: return &ecparam::p160_1;
 	default: return 0;
 	}
 }
 
 } // mcl
+
+#include <mcl/ec.hpp>

include/mcl/elgamal.hpp

@@ -386,7 +386,7 @@ struct ElgamalT {
 		create table g^i for i in [rangeMin, rangeMax]
 	*/
 	struct PowerCache {
-#if (CYBOZU_CPP_VERSION > CYBOZU_CPP_VERSION_CP03)
+#if (CYBOZU_CPP_VERSION > CYBOZU_CPP_VERSION_CPP03)
 		typedef CYBOZU_NAMESPACE_STD::unordered_map<Ec, int> Cache;
 #else
 		typedef std::map<Ec, int> Cache;

include/mcl/fp.hpp

@@ -38,13 +38,6 @@ struct ZnTag;
 
 namespace fp {
 
-// copy src to dst as little endian
-void copyUnitToByteAsLE(uint8_t *dst, const Unit *src, size_t byteSize);
-// copy src to dst as little endian
-void copyByteToUnitAsLE(Unit *dst, const uint8_t *src, size_t byteSize);
-
-bool copyAndMask(Unit *y, const void *x, size_t xByteSize, const Op& op, MaskMode maskMode);
-
 uint64_t getUint64(bool *pb, const fp::Block& b);
 int64_t getInt64(bool *pb, fp::Block& b, const fp::Op& op);
 
@@ -73,13 +66,15 @@ bool isEnableJIT(); // 1st call is not threadsafe
 uint32_t sha256(void *out, uint32_t maxOutSize, const void *msg, uint32_t msgSize);
 uint32_t sha512(void *out, uint32_t maxOutSize, const void *msg, uint32_t msgSize);
 
+// draft-07 outSize = 128 or 256
+void expand_message_xmd(uint8_t out[], size_t outSize, const void *msg, size_t msgSize, const void *dst, size_t dstSize);
+
 namespace local {
 
-inline void byteSwap(void *x, size_t n)
+inline void byteSwap(uint8_t *x, size_t n)
 {
-	char *p = (char *)x;
 	for (size_t i = 0; i < n / 2; i++) {
-		fp::swap_(p[i], p[n - 1 - i]);
+		fp::swap_(x[i], x[n - 1 - i]);
 	}
 }
 
@@ -105,6 +100,47 @@ private:
 	template<class Fp> friend class FpDblT;
 	template<class Fp> friend class Fp2T;
 	template<class Fp> friend struct Fp6T;
+#ifdef MCL_XBYAK_DIRECT_CALL
+	static inline void addA(Unit *z, const Unit *x, const Unit *y)
+	{
+		op_.fp_add(z, x, y, op_.p);
+	}
+	static inline void subA(Unit *z, const Unit *x, const Unit *y)
+	{
+		op_.fp_sub(z, x, y, op_.p);
+	}
+	static inline void negA(Unit *y, const Unit *x)
+	{
+		op_.fp_neg(y, x, op_.p);
+	}
+	static inline void mulA(Unit *z, const Unit *x, const Unit *y)
+	{
+		op_.fp_mul(z, x, y, op_.p);
+	}
+	static inline void sqrA(Unit *y, const Unit *x)
+	{
+		op_.fp_sqr(y, x, op_.p);
+	}
+	static inline void mul2A(Unit *y, const Unit *x)
+	{
+		op_.fp_mul2(y, x, op_.p);
+	}
+#endif
+	static inline void mul9A(Unit *y, const Unit *x)
+	{
+		mulSmall(y, x, 9);
+//		op_.fp_mul9(y, x, op_.p);
+	}
+	static inline void mulSmall(Unit *z, const Unit *x, const uint32_t y)
+	{
+		assert(y <= op_.smallModp.maxMulN);
+		Unit xy[maxSize + 1];
+		op_.fp_mulUnitPre(xy, x, y);
+		int v = op_.smallModp.approxMul(xy);
+		const Unit *pv = op_.smallModp.getPmul(v);
+		op_.fp_subPre(z, xy, pv);
+		op_.fp_sub(z, z, op_.p, op_.p);
+	}
 public:
 	typedef FpT<tag, maxBitSize> BaseFp;
 	// return pointer to array v_[]
@@ -115,6 +151,7 @@ public:
 	static inline size_t getBitSize() { return op_.bitSize; }
 	static inline size_t getByteSize() { return (op_.bitSize + 7) / 8; }
 	static inline const fp::Op& getOp() { return op_; }
+	static inline fp::Op& getOpNonConst() { return op_; }
 	void dump() const
 	{
 		const size_t N = op_.N;
@@ -131,6 +168,9 @@ public:
 	{
 		assert(maxBitSize <= MCL_MAX_BIT_SIZE);
 		*pb = op_.init(p, maxBitSize, xi_a, mode);
+#ifdef MCL_DUMP_JIT
+		return;
+#endif
 		if (!*pb) return;
 		{ // set oneRep
 			FpT& one = *reinterpret_cast<FpT*>(op_.oneRep);
@@ -147,16 +187,27 @@ public:
 		ioMode_ = 0;
 		isETHserialization_ = false;
 #ifdef MCL_XBYAK_DIRECT_CALL
-		add = fp::func_ptr_cast<void (*)(FpT& z, const FpT& x, const FpT& y)>(op_.fp_addA_);
+		if (op_.fp_addA_ == 0) {
-		if (add == 0) add = addC;
+			op_.fp_addA_ = addA;
-		sub = fp::func_ptr_cast<void (*)(FpT& z, const FpT& x, const FpT& y)>(op_.fp_subA_);
+		}
-		if (sub == 0) sub = subC;
+		if (op_.fp_subA_ == 0) {
-		neg = fp::func_ptr_cast<void (*)(FpT& y, const FpT& x)>(op_.fp_negA_);
+			op_.fp_subA_ = subA;
-		if (neg == 0) neg = negC;
+		}
-		mul = fp::func_ptr_cast<void (*)(FpT& z, const FpT& x, const FpT& y)>(op_.fp_mulA_);
+		if (op_.fp_negA_ == 0) {
-		if (mul == 0) mul = mulC;
+			op_.fp_negA_ = negA;
-		sqr = fp::func_ptr_cast<void (*)(FpT& y, const FpT& x)>(op_.fp_sqrA_);
+		}
-		if (sqr == 0) sqr = sqrC;
+		if (op_.fp_mulA_ == 0) {
+			op_.fp_mulA_ = mulA;
+		}
+		if (op_.fp_sqrA_ == 0) {
+			op_.fp_sqrA_ = sqrA;
+		}
+		if (op_.fp_mul2A_ == 0) {
+			op_.fp_mul2A_ = mul2A;
+		}
+		if (op_.fp_mul9A_ == 0) {
+			op_.fp_mul9A_ = mul9A;
+		}
 #endif
 		*pb = true;
 	}
@@ -224,7 +275,7 @@ public:
 		} else {
 			clear();
 			if (x) {
-				int64_t y = x < 0 ? -x : x;
+				uint64_t y = fp::abs_(x);
 				if (sizeof(Unit) == 8) {
 					v_[0] = y;
 				} else {
@@ -254,26 +305,28 @@ public:
 	{
 		if (isMont()) op_.fromMont(v_, v_);
 	}
+	// deny a string with large length even if the value is in Fp
 	template<class InputStream>
 	void load(bool *pb, InputStream& is, int ioMode)
 	{
 		bool isMinus = false;
 		*pb = false;
-		if (ioMode & (IoArray | IoArrayRaw | IoSerialize | IoSerializeHexStr)) {
+		if (fp::isIoSerializeMode(ioMode)) {
 			const size_t n = getByteSize();
-			v_[op_.N - 1] = 0;
+			uint8_t *buf = (uint8_t*)CYBOZU_ALLOCA(n);
 			size_t readSize;
 			if (ioMode & IoSerializeHexStr) {
-				readSize = mcl::fp::readHexStr(v_, n, is);
+				readSize = mcl::fp::readHexStr(buf, n, is);
 			} else {
-				readSize = cybozu::readSome(v_, n, is);
+				readSize = cybozu::readSome(buf, n, is);
 			}
+			if (readSize != n) return;
 			if (isETHserialization_ && ioMode & (IoSerialize | IoSerializeHexStr)) {
-				fp::local::byteSwap(v_, n);
+				fp::local::byteSwap(buf, n);
 			}
-			if (readSize != n) return;
+			fp::convertArrayAsLE(v_, op_.N, buf, n);
 		} else {
-			char buf[1024];
+			char buf[sizeof(*this) * 8 + 2]; // '0b' + max binary format length
 			size_t n = fp::local::loadWord(buf, sizeof(buf), is);
 			if (n == 0) return;
 			n = fp::strToArray(&isMinus, v_, op_.N, buf, n, ioMode);
@@ -295,24 +348,23 @@ public:
 	void save(bool *pb, OutputStream& os, int ioMode) const
 	{
 		const size_t n = getByteSize();
-		if (ioMode & (IoArray | IoArrayRaw | IoSerialize | IoSerializeHexStr)) {
+		if (fp::isIoSerializeMode(ioMode)) {
+			const size_t xn = sizeof(fp::Unit) * op_.N;
+			uint8_t *x = (uint8_t*)CYBOZU_ALLOCA(xn);
 			if (ioMode & IoArrayRaw) {
-				cybozu::write(pb, os, v_, n);
+				fp::convertArrayAsLE(x, xn, v_, op_.N);
+				cybozu::write(pb, os, x, n);
 			} else {
 				fp::Block b;
 				getBlock(b);
-				const char *src = (const char *)b.p;
+				fp::convertArrayAsLE(x, xn, b.p, b.n);
-				char rev[fp::maxUnitSize * sizeof(fp::Unit)];
 				if (isETHserialization_ && ioMode & (IoSerialize | IoSerializeHexStr)) {
-					for (size_t i = 0; i < n; i++) {
+					fp::local::byteSwap(x, n);
-						rev[i] = src[n - 1 - i];
-					}
-					src = rev;
 				}
 				if (ioMode & IoSerializeHexStr) {
-					mcl::fp::writeHexStr(pb, os, src, n);
+					mcl::fp::writeHexStr(pb, os, x, n);
 				} else {
-					cybozu::write(pb, os, src, n);
+					cybozu::write(pb, os, x, n);
 				}
 			}
 			return;
@@ -329,31 +381,65 @@ public:
 		cybozu::write(pb, os, buf + sizeof(buf) - len, len);
 	}
 	/*
-		mode = Mod : set x mod p if sizeof(S) * n <= 64 else error
+		treat x as little endian
+		if x >= p then error
 	*/
 	template<class S>
-	void setArray(bool *pb, const S *x, size_t n, mcl::fp::MaskMode mode = fp::NoMask)
+	void setArray(bool *pb, const S *x, size_t n)
 	{
-		*pb = fp::copyAndMask(v_, x, sizeof(S) * n, op_, mode);
+		if (!fp::convertArrayAsLE(v_, op_.N, x, n)) {
+			*pb = false;
+			return;
+		}
+		if (fp::isGreaterOrEqualArray(v_, op_.p, op_.N)) {
+			*pb = false;
+			return;
+		}
+		*pb = true;
 		toMont();
 	}
 	/*
-		mask x with (1 << bitLen) and subtract p if x >= p
+		treat x as little endian
+		x &= (1 << bitLen) = 1
+		x &= (1 << (bitLen - 1)) - 1 if x >= p
 	*/
 	template<class S>
-	void setArrayMaskMod(const S *x, size_t n)
+	void setArrayMask(const S *x, size_t n)
 	{
-		fp::copyAndMask(v_, x, sizeof(S) * n, op_, fp::MaskAndMod);
+		const size_t dstByte = sizeof(fp::Unit) * op_.N;
+		if (sizeof(S) * n > dstByte) {
+			n = dstByte / sizeof(S);
+		}
+		bool b = fp::convertArrayAsLE(v_, op_.N, x, n);
+		assert(b);
+		(void)b;
+		fp::maskArray(v_, op_.N, op_.bitSize);
+		if (fp::isGreaterOrEqualArray(v_, op_.p, op_.N)) {
+			fp::maskArray(v_, op_.N, op_.bitSize - 1);
+		}
 		toMont();
 	}
-
 	/*
-		mask x with (1 << (bitLen - 1)) - 1 if x >= p
+		set (x as little endian) % p
+		error if size of x >= sizeof(Fp) * 2
 	*/
 	template<class S>
-	void setArrayMask(const S *x, size_t n)
+	void setArrayMod(bool *pb, const S *x, size_t n)
 	{
-		fp::copyAndMask(v_, x, sizeof(S) * n, op_, fp::SmallMask);
+		if (sizeof(S) * n > sizeof(fp::Unit) * op_.N * 2) {
+			*pb = false;
+			return;
+		}
+		mpz_class mx;
+		gmp::setArray(pb, mx, x, n);
+		if (!*pb) return;
+#ifdef MCL_USE_VINT
+		op_.modp.modp(mx, mx);
+#else
+		mx %= op_.mp;
+#endif
+		gmp::getArray(pb, v_, op_.N, mx);
+		if (!*pb) return;
 		toMont();
 	}
 	void getBlock(fp::Block& b) const
@@ -366,14 +452,82 @@ public:
 			b.p = &v_[0];
 		}
 	}
+	/*
+		write a value with little endian
+		write buf[0] = 0 and return 1 if the value is 0
+		return written size if success else 0
+	*/
+	size_t getLittleEndian(uint8_t *buf, size_t maxN) const
+	{
+		fp::Block b;
+		getBlock(b);
+		size_t n = sizeof(fp::Unit) * b.n;
+		uint8_t *t = (uint8_t*)CYBOZU_ALLOCA(n);
+		if (!fp::convertArrayAsLE(t, n, b.p, b.n)) {
+			return 0;
+		}
+		while (n > 0) {
+			if (t[n - 1]) break;
+			n--;
+		}
+		if (n == 0) n = 1; // zero
+		if (maxN < n) return 0;
+		for (size_t i = 0; i < n; i++) {
+			buf[i] = t[i];
+		}
+		return n;
+	}
+	/*
+		set (little endian % p)
+		error if xn > 64
+	*/
+	void setLittleEndianMod(bool *pb, const uint8_t *x, size_t xn)
+	{
+		if (xn > 64) {
+			*pb = false;
+			return;
+		}
+		setArrayMod(pb, x, xn);
+	}
+	/*
+		set (big endian % p)
+		error if xn > 64
+	*/
+	void setBigEndianMod(bool *pb, const uint8_t *x, size_t xn)
+	{
+		if (xn > 64) {
+			*pb = false;
+			return;
+		}
+		uint8_t swapX[64];
+		for (size_t i = 0; i < xn; i++) {
+			swapX[xn - 1 - i] = x[i];
+		}
+		setArrayMod(pb, swapX, xn);
+	}
 	void setByCSPRNG(bool *pb, fp::RandGen rg = fp::RandGen())
 	{
 		if (rg.isZero()) rg = fp::RandGen::get();
-		rg.read(pb, v_, op_.N * sizeof(Unit)); // byte size
+		uint8_t x[sizeof(*this)];
+		const size_t n = op_.N * sizeof(Unit);
+		rg.read(pb, x, n); // byte size
 		if (!pb) return;
+		fp::convertArrayAsLE(v_, op_.N, x, n);
 		setArrayMask(v_, op_.N);
 	}
 #ifndef CYBOZU_DONT_USE_EXCEPTION
+	void setLittleEndianMod(const uint8_t *buf, size_t bufSize)
+	{
+		bool b;
+		setLittleEndianMod(&b, buf, bufSize);
+		if (!b) throw cybozu::Exception("setLittleEndianMod");
+	}
+	void setBigEndianMod(const uint8_t *buf, size_t bufSize)
+	{
+		bool b;
+		setBigEndianMod(&b, buf, bufSize);
+		if (!b) throw cybozu::Exception("setBigEndianMod");
+	}
 	void setByCSPRNG(fp::RandGen rg = fp::RandGen())
 	{
 		bool b;
@@ -386,11 +540,15 @@ public:
 		setByCSPRNG(rg);
 	}
 	/*
-		hash msg and mask with (1 << (bitLen - 1)) - 1
+		x = SHA-256(msg) as little endian
+		p = order of a finite field
+		L = bit size of p
+		x &= (1 << L) - 1
+		if (x >= p) x &= (1 << (L - 1)) - 1
 	*/
 	void setHashOf(const void *msg, size_t msgSize)
 	{
-		char buf[MCL_MAX_HASH_BIT_SIZE / 8];
+		uint8_t buf[MCL_MAX_HASH_BIT_SIZE / 8];
 		uint32_t size = op_.hash(buf, static_cast<uint32_t>(sizeof(buf)), msg, static_cast<uint32_t>(msgSize));
 		setArrayMask(buf, size);
 	}
@@ -408,32 +566,78 @@ public:
 		}
 		setArray(pb, gmp::getUnit(x), gmp::getUnitSize(x));
 	}
+	static void add(FpT& z, const FpT& x, const FpT& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		op_.fp_addA_(z.v_, x.v_, y.v_);
+#else
+		op_.fp_add(z.v_, x.v_, y.v_, op_.p);
+#endif
+	}
+	static void sub(FpT& z, const FpT& x, const FpT& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		op_.fp_subA_(z.v_, x.v_, y.v_);
+#else
+		op_.fp_sub(z.v_, x.v_, y.v_, op_.p);
+#endif
+	}
+	static void neg(FpT& y, const FpT& x)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		op_.fp_negA_(y.v_, x.v_);
+#else
+		op_.fp_neg(y.v_, x.v_, op_.p);
+#endif
+	}
+	static void mul(FpT& z, const FpT& x, const FpT& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		op_.fp_mulA_(z.v_, x.v_, y.v_);
+#else
+		op_.fp_mul(z.v_, x.v_, y.v_, op_.p);
+#endif
+	}
+	static void sqr(FpT& y, const FpT& x)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		op_.fp_sqrA_(y.v_, x.v_);
+#else
+		op_.fp_sqr(y.v_, x.v_, op_.p);
+#endif
+	}
+	static void mul2(FpT& y, const FpT& x)
+	{
 #ifdef MCL_XBYAK_DIRECT_CALL
-	static void (*add)(FpT& z, const FpT& x, const FpT& y);
+		op_.fp_mul2A_(y.v_, x.v_);
-	static inline void addC(FpT& z, const FpT& x, const FpT& y) { op_.fp_add(z.v_, x.v_, y.v_, op_.p); }
-	static void (*sub)(FpT& z, const FpT& x, const FpT& y);
-	static inline void subC(FpT& z, const FpT& x, const FpT& y) { op_.fp_sub(z.v_, x.v_, y.v_, op_.p); }
-	static void (*neg)(FpT& y, const FpT& x);
-	static inline void negC(FpT& y, const FpT& x) { op_.fp_neg(y.v_, x.v_, op_.p); }
-	static void (*mul)(FpT& z, const FpT& x, const FpT& y);
-	static inline void mulC(FpT& z, const FpT& x, const FpT& y) { op_.fp_mul(z.v_, x.v_, y.v_, op_.p); }
-	static void (*sqr)(FpT& y, const FpT& x);
-	static inline void sqrC(FpT& y, const FpT& x) { op_.fp_sqr(y.v_, x.v_, op_.p); }
 #else
-	static inline void add(FpT& z, const FpT& x, const FpT& y) { op_.fp_add(z.v_, x.v_, y.v_, op_.p); }
+		op_.fp_mul2(y.v_, x.v_, op_.p);
-	static inline void sub(FpT& z, const FpT& x, const FpT& y) { op_.fp_sub(z.v_, x.v_, y.v_, op_.p); }
-	static inline void neg(FpT& y, const FpT& x) { op_.fp_neg(y.v_, x.v_, op_.p); }
-	static inline void mul(FpT& z, const FpT& x, const FpT& y) { op_.fp_mul(z.v_, x.v_, y.v_, op_.p); }
-	static inline void sqr(FpT& y, const FpT& x) { op_.fp_sqr(y.v_, x.v_, op_.p); }
 #endif
+	}
+	static void mul9(FpT& y, const FpT& x)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		op_.fp_mul9A_(y.v_, x.v_);
+#else
+		mul9A(y.v_, x.v_);
+#endif
+	}
 	static inline void addPre(FpT& z, const FpT& x, const FpT& y) { op_.fp_addPre(z.v_, x.v_, y.v_); }
 	static inline void subPre(FpT& z, const FpT& x, const FpT& y) { op_.fp_subPre(z.v_, x.v_, y.v_); }
+	static inline void mulSmall(FpT& z, const FpT& x, const uint32_t y)
+	{
+		mulSmall(z.v_, x.v_, y);
+	}
 	static inline void mulUnit(FpT& z, const FpT& x, const Unit y)
 	{
 		if (mulSmallUnit(z, x, y)) return;
 		op_.fp_mulUnit(z.v_, x.v_, y, op_.p);
 	}
-	static inline void inv(FpT& y, const FpT& x) { op_.fp_invOp(y.v_, x.v_, op_); }
+	static inline void inv(FpT& y, const FpT& x)
+	{
+		assert(!x.isZero());
+		op_.fp_invOp(y.v_, x.v_, op_);
+	}
 	static inline void divBy2(FpT& y, const FpT& x)
 	{
 #if 0
@@ -526,9 +730,12 @@ public:
 	}
 	static void setETHserialization(bool ETHserialization)
 	{
-		if (getBitSize() != 381) return;
 		isETHserialization_ = ETHserialization;
 	}
+	static bool getETHserialization()
+	{
+		return isETHserialization_;
+	}
 	static inline bool isETHserialization() { return isETHserialization_; }
 	static inline int getIoMode() { return ioMode_; }
 	static inline size_t getModBitLen() { return getBitSize(); }
@@ -655,16 +862,10 @@ template<class tag, size_t maxBitSize> fp::Op FpT<tag, maxBitSize>::op_;
 template<class tag, size_t maxBitSize> FpT<tag, maxBitSize> FpT<tag, maxBitSize>::inv2_;
 template<class tag, size_t maxBitSize> int FpT<tag, maxBitSize>::ioMode_ = IoAuto;
 template<class tag, size_t maxBitSize> bool FpT<tag, maxBitSize>::isETHserialization_ = false;
-#ifdef MCL_XBYAK_DIRECT_CALL
-template<class tag, size_t maxBitSize> void (*FpT<tag, maxBitSize>::add)(FpT& z, const FpT& x, const FpT& y);
-template<class tag, size_t maxBitSize> void (*FpT<tag, maxBitSize>::sub)(FpT& z, const FpT& x, const FpT& y);
-template<class tag, size_t maxBitSize> void (*FpT<tag, maxBitSize>::neg)(FpT& y, const FpT& x);
-template<class tag, size_t maxBitSize> void (*FpT<tag, maxBitSize>::mul)(FpT& z, const FpT& x, const FpT& y);
-template<class tag, size_t maxBitSize> void (*FpT<tag, maxBitSize>::sqr)(FpT& y, const FpT& x);
-#endif
 
 } // mcl
 
+#ifndef CYBOZU_DONT_USE_EXCEPTION
 #ifdef CYBOZU_USE_BOOST
 namespace mcl {
 
@@ -688,6 +889,7 @@ struct hash<mcl::FpT<tag, maxBitSize> > {
 
 CYBOZU_NAMESPACE_TR1_END } // std::tr1
 #endif
+#endif
 
 #ifdef _MSC_VER
 	#pragma warning(pop)

include/mcl/fp_tower.hpp

@@ -10,22 +10,18 @@
 
 namespace mcl {
 
+template<class Fp> struct Fp12T;
+template<class Fp> class BNT;
+template<class Fp> struct Fp2DblT;
+
 template<class Fp>
 class FpDblT : public fp::Serializable<FpDblT<Fp> > {
 	typedef fp::Unit Unit;
 	Unit v_[Fp::maxSize * 2];
+	friend struct Fp2DblT<Fp>;
 public:
 	static size_t getUnitSize() { return Fp::op_.N * 2; }
-	FpDblT() : v_()
+	const fp::Unit *getUnit() const { return v_; }
-	{
-	}
-	FpDblT(const FpDblT& rhs)
-	{
-		const size_t n = getUnitSize();
-		for (size_t i = 0; i < n; i++) {
-			v_[i] = rhs.v_[i];
-		}
-	}
 	void dump() const
 	{
 		const size_t n = getUnitSize();
@@ -117,31 +113,42 @@ public:
 	{
 		gmp::setArray(pb, x, v_, Fp::op_.N * 2);
 	}
+	static inline void add(FpDblT& z, const FpDblT& x, const FpDblT& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fpDbl_addA_(z.v_, x.v_, y.v_);
+#else
+		Fp::op_.fpDbl_add(z.v_, x.v_, y.v_, Fp::op_.p);
+#endif
+	}
+	static inline void sub(FpDblT& z, const FpDblT& x, const FpDblT& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fpDbl_subA_(z.v_, x.v_, y.v_);
+#else
+		Fp::op_.fpDbl_sub(z.v_, x.v_, y.v_, Fp::op_.p);
+#endif
+	}
+	static inline void mod(Fp& z, const FpDblT& xy)
+	{
 #ifdef MCL_XBYAK_DIRECT_CALL
-	static void (*add)(FpDblT& z, const FpDblT& x, const FpDblT& y);
+		Fp::op_.fpDbl_modA_(z.v_, xy.v_);
-	static void (*sub)(FpDblT& z, const FpDblT& x, const FpDblT& y);
-	static void (*mod)(Fp& z, const FpDblT& xy);
-	static void (*addPre)(FpDblT& z, const FpDblT& x, const FpDblT& y);
-	static void (*subPre)(FpDblT& z, const FpDblT& x, const FpDblT& y);
-	static void addC(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_add(z.v_, x.v_, y.v_, Fp::op_.p); }
-	static void subC(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_sub(z.v_, x.v_, y.v_, Fp::op_.p); }
-	static void modC(Fp& z, const FpDblT& xy) { Fp::op_.fpDbl_mod(z.v_, xy.v_, Fp::op_.p); }
-	static void addPreC(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_addPre(z.v_, x.v_, y.v_); }
-	static void subPreC(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_subPre(z.v_, x.v_, y.v_); }
 #else
-	static void add(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_add(z.v_, x.v_, y.v_, Fp::op_.p); }
+		Fp::op_.fpDbl_mod(z.v_, xy.v_, Fp::op_.p);
-	static void sub(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_sub(z.v_, x.v_, y.v_, Fp::op_.p); }
+#endif
-	static void mod(Fp& z, const FpDblT& xy) { Fp::op_.fpDbl_mod(z.v_, xy.v_, Fp::op_.p); }
+	}
+#ifdef MCL_XBYAK_DIRECT_CALL
+	static void addA(Unit *z, const Unit *x, const Unit *y) { Fp::op_.fpDbl_add(z, x, y, Fp::op_.p); }
+	static void subA(Unit *z, const Unit *x, const Unit *y) { Fp::op_.fpDbl_sub(z, x, y, Fp::op_.p); }
+	static void modA(Unit *z, const Unit *xy) { Fp::op_.fpDbl_mod(z, xy, Fp::op_.p); }
+#endif
 	static void addPre(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_addPre(z.v_, x.v_, y.v_); }
 	static void subPre(FpDblT& z, const FpDblT& x, const FpDblT& y) { Fp::op_.fpDbl_subPre(z.v_, x.v_, y.v_); }
-#endif
-	static void mulPreC(FpDblT& xy, const Fp& x, const Fp& y) { Fp::op_.fpDbl_mulPre(xy.v_, x.v_, y.v_); }
-	static void sqrPreC(FpDblT& xx, const Fp& x) { Fp::op_.fpDbl_sqrPre(xx.v_, x.v_); }
 	/*
 		mul(z, x, y) = mulPre(xy, x, y) + mod(z, xy)
 	*/
-	static void (*mulPre)(FpDblT& xy, const Fp& x, const Fp& y);
+	static void mulPre(FpDblT& xy, const Fp& x, const Fp& y) { Fp::op_.fpDbl_mulPre(xy.v_, x.v_, y.v_); }
-	static void (*sqrPre)(FpDblT& xx, const Fp& x);
+	static void sqrPre(FpDblT& xx, const Fp& x) { Fp::op_.fpDbl_sqrPre(xx.v_, x.v_); }
 	static void mulUnit(FpDblT& z, const FpDblT& x, Unit y)
 	{
 		if (mulSmallUnit(z, x, y)) return;
@@ -149,47 +156,23 @@ public:
 	}
 	static void init()
 	{
-		const mcl::fp::Op& op = Fp::getOp();
 #ifdef MCL_XBYAK_DIRECT_CALL
-		add = fp::func_ptr_cast<void (*)(FpDblT&, const FpDblT&, const FpDblT&)>(op.fpDbl_addA_);
+		mcl::fp::Op& op = Fp::op_;
-		if (add == 0) add = addC;
+		if (op.fpDbl_addA_ == 0) {
-		sub = fp::func_ptr_cast<void (*)(FpDblT&, const FpDblT&, const FpDblT&)>(op.fpDbl_subA_);
+			op.fpDbl_addA_ = addA;
-		if (sub == 0) sub = subC;
-		mod = fp::func_ptr_cast<void (*)(Fp&, const FpDblT&)>(op.fpDbl_modA_);
-		if (mod == 0) mod = modC;
-		addPre = fp::func_ptr_cast<void (*)(FpDblT&, const FpDblT&, const FpDblT&)>(op.fpDbl_addPre);
-		if (addPre == 0) addPre = addPreC;
-		subPre = fp::func_ptr_cast<void (*)(FpDblT&, const FpDblT&, const FpDblT&)>(op.fpDbl_subPre);
-		if (subPre == 0) subPre = subPreC;
-#endif
-		if (op.fpDbl_mulPreA_) {
-			mulPre = fp::func_ptr_cast<void (*)(FpDblT&, const Fp&, const Fp&)>(op.fpDbl_mulPreA_);
-		} else {
-			mulPre = mulPreC;
 		}
-		if (op.fpDbl_sqrPreA_) {
+		if (op.fpDbl_subA_ == 0) {
-			sqrPre = fp::func_ptr_cast<void (*)(FpDblT&, const Fp&)>(op.fpDbl_sqrPreA_);
+			op.fpDbl_subA_ = subA;
-		} else {
-			sqrPre = sqrPreC;
 		}
+		if (op.fpDbl_modA_ == 0) {
+			op.fpDbl_modA_ = modA;
+		}
+#endif
 	}
 	void operator+=(const FpDblT& x) { add(*this, *this, x); }
 	void operator-=(const FpDblT& x) { sub(*this, *this, x); }
 };
 
-#ifdef MCL_XBYAK_DIRECT_CALL
-template<class Fp> void (*FpDblT<Fp>::add)(FpDblT&, const FpDblT&, const FpDblT&);
-template<class Fp> void (*FpDblT<Fp>::sub)(FpDblT&, const FpDblT&, const FpDblT&);
-template<class Fp> void (*FpDblT<Fp>::mod)(Fp&, const FpDblT&);
-template<class Fp> void (*FpDblT<Fp>::addPre)(FpDblT&, const FpDblT&, const FpDblT&);
-template<class Fp> void (*FpDblT<Fp>::subPre)(FpDblT&, const FpDblT&, const FpDblT&);
-#endif
-template<class Fp> void (*FpDblT<Fp>::mulPre)(FpDblT&, const Fp&, const Fp&);
-template<class Fp> void (*FpDblT<Fp>::sqrPre)(FpDblT&, const Fp&);
-
-template<class Fp> struct Fp12T;
-template<class Fp> class BNT;
-template<class Fp> struct Fp2DblT;
 /*
 	beta = -1
 	Fp2 = F[i] / (i^2 + 1)
@@ -240,22 +223,79 @@ public:
 		a = a_;
 		b = b_;
 	}
+	static void add(Fp2T& z, const Fp2T& x, const Fp2T& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fp2_addA_(z.a.v_, x.a.v_, y.a.v_);
+#else
+		addA(z.a.v_, x.a.v_, y.a.v_);
+#endif
+	}
+	static void sub(Fp2T& z, const Fp2T& x, const Fp2T& y)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fp2_subA_(z.a.v_, x.a.v_, y.a.v_);
+#else
+		subA(z.a.v_, x.a.v_, y.a.v_);
+#endif
+	}
+	static void neg(Fp2T& y, const Fp2T& x)
+	{
 #ifdef MCL_XBYAK_DIRECT_CALL
-	static void (*add)(Fp2T& z, const Fp2T& x, const Fp2T& y);
+		Fp::op_.fp2_negA_(y.a.v_, x.a.v_);
-	static void (*sub)(Fp2T& z, const Fp2T& x, const Fp2T& y);
-	static void (*neg)(Fp2T& y, const Fp2T& x);
-	static void (*mul)(Fp2T& z, const Fp2T& x, const Fp2T& y);
-	static void (*sqr)(Fp2T& y, const Fp2T& x);
 #else
-	static void add(Fp2T& z, const Fp2T& x, const Fp2T& y) { addC(z, x, y); }
+		negA(y.a.v_, x.a.v_);
-	static void sub(Fp2T& z, const Fp2T& x, const Fp2T& y) { subC(z, x, y); }
-	static void neg(Fp2T& y, const Fp2T& x) { negC(y, x); }
-	static void mul(Fp2T& z, const Fp2T& x, const Fp2T& y) { mulC(z, x, y); }
-	static void sqr(Fp2T& y, const Fp2T& x) { sqrC(y, x); }
 #endif
-	static void (*mul_xi)(Fp2T& y, const Fp2T& x);
+	}
-	static void addPre(Fp2T& z, const Fp2T& x, const Fp2T& y) { Fp::addPre(z.a, x.a, y.a); Fp::addPre(z.b, x.b, y.b); }
+	static void mul(Fp2T& z, const Fp2T& x, const Fp2T& y)
-	static void inv(Fp2T& y, const Fp2T& x) { Fp::op_.fp2_inv(y.a.v_, x.a.v_); }
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fp2_mulA_(z.a.v_, x.a.v_, y.a.v_);
+#else
+		mulA(z.a.v_, x.a.v_, y.a.v_);
+#endif
+	}
+	static void sqr(Fp2T& y, const Fp2T& x)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fp2_sqrA_(y.a.v_, x.a.v_);
+#else
+		sqrA(y.a.v_, x.a.v_);
+#endif
+	}
+	static void mul2(Fp2T& y, const Fp2T& x)
+	{
+#ifdef MCL_XBYAK_DIRECT_CALL
+		Fp::op_.fp2_mul2A_(y.a.v_, x.a.v_);
+#else
+		mul2A(y.a.v_, x.a.v_);
+#endif
+	}
+	static void mul_xi(Fp2T& y, const Fp2T& x)
+	{
+		Fp::op_.fp2_mul_xiA_(y.a.v_, x.a.v_);
+	}
+	/*
+		x = a + bi
+		1 / x = (a - bi) / (a^2 + b^2)
+	*/
+	static void inv(Fp2T& y, const Fp2T& x)
+	{
+		assert(!x.isZero());
+		const Fp& a = x.a;
+		const Fp& b = x.b;
+		Fp r;
+		norm(r, x);
+		Fp::inv(r, r); // r = 1 / (a^2 + b^2)
+		Fp::mul(y.a, a, r);
+		Fp::mul(y.b, b, r);
+		Fp::neg(y.b, y.b);
+	}
+	static void addPre(Fp2T& z, const Fp2T& x, const Fp2T& y)
+	{
+		Fp::addPre(z.a, x.a, y.a);
+		Fp::addPre(z.b, x.b, y.b);
+	}
 	static void divBy2(Fp2T& y, const Fp2T& x)
 	{
 		Fp::divBy2(y.a, x.a);
@@ -360,12 +400,14 @@ public:
 		Fp::mul(y.b, x.b, t2);
 		return true;
 	}
+	// y = a^2 + b^2
 	static void inline norm(Fp& y, const Fp2T& x)
 	{
-		Fp aa, bb;
+		FpDbl AA, BB;
-		Fp::sqr(aa, x.a);
+		FpDbl::sqrPre(AA, x.a);
-		Fp::sqr(bb, x.b);
+		FpDbl::sqrPre(BB, x.b);
-		Fp::add(y, aa, bb);
+		FpDbl::addPre(AA, AA, BB);
+		FpDbl::mod(y, AA);
 	}
 	/*
 		Frobenius
@@ -389,31 +431,40 @@ public:
 	}
 
 	static uint32_t get_xi_a() { return Fp::getOp().xi_a; }
-	static void init()
+	static void init(bool *pb)
 	{
-//		assert(Fp::maxSize <= 256);
 		mcl::fp::Op& op = Fp::op_;
 		assert(op.xi_a);
-		mul_xi = 0;
+		// assume p < W/4 where W = 1 << (N * sizeof(Unit) * 8)
+		if ((op.p[op.N - 1] >> (sizeof(fp::Unit) * 8 - 2)) != 0) {
+			*pb = false;
+			return;
+		}
 #ifdef MCL_XBYAK_DIRECT_CALL
-		add = fp::func_ptr_cast<void (*)(Fp2T& z, const Fp2T& x, const Fp2T& y)>(op.fp2_addA_);
+		if (op.fp2_addA_ == 0) {
-		if (add == 0) add = addC;
+			op.fp2_addA_ = addA;
-		sub = fp::func_ptr_cast<void (*)(Fp2T& z, const Fp2T& x, const Fp2T& y)>(op.fp2_subA_);
+		}
-		if (sub == 0) sub = subC;
+		if (op.fp2_subA_ == 0) {
-		neg = fp::func_ptr_cast<void (*)(Fp2T& y, const Fp2T& x)>(op.fp2_negA_);
+			op.fp2_subA_ = subA;
-		if (neg == 0) neg = negC;
+		}
-		mul = fp::func_ptr_cast<void (*)(Fp2T& z, const Fp2T& x, const Fp2T& y)>(op.fp2_mulA_);
+		if (op.fp2_negA_ == 0) {
-		if (mul == 0) mul = mulC;
+			op.fp2_negA_ = negA;
-		sqr = fp::func_ptr_cast<void (*)(Fp2T& y, const Fp2T& x)>(op.fp2_sqrA_);
+		}
-		if (sqr == 0) sqr = sqrC;
+		if (op.fp2_mulA_ == 0) {
-		mul_xi = fp::func_ptr_cast<void (*)(Fp2T&, const Fp2T&)>(op.fp2_mul_xiA_);
+			op.fp2_mulA_ = mulA;
+		}
+		if (op.fp2_sqrA_ == 0) {
+			op.fp2_sqrA_ = sqrA;
+		}
+		if (op.fp2_mul2A_ == 0) {
+			op.fp2_mul2A_ = mul2A;
+		}
 #endif
-		op.fp2_inv = fp2_invW;
+		if (op.fp2_mul_xiA_ == 0) {
-		if (mul_xi == 0) {
 			if (op.xi_a == 1) {
-				mul_xi = fp2_mul_xi_1_1iC;
+				op.fp2_mul_xiA_ = fp2_mul_xi_1_1iA;
 			} else {
-				mul_xi = fp2_mul_xiC;
+				op.fp2_mul_xiA_ = fp2_mul_xiA;
 			}
 		}
 		FpDblT<Fp>::init();
@@ -446,8 +497,15 @@ public:
 			Fp2T::mul(g2[i], t, g[i]);
 			g3[i] = g[i] * g2[i];
 		}
+		*pb = true;
 	}
 #ifndef CYBOZU_DONT_USE_EXCEPTION
+	static void init()
+	{
+		bool b;
+		init(&b);
+		if (!b) throw cybozu::Exception("Fp2::init");
+	}
 	template<class InputStream>
 	void load(InputStream& is, int ioMode = IoSerialize)
 	{
@@ -484,75 +542,65 @@ public:
 	}
 #endif
 private:
+	static Fp2T& cast(Unit *x) { return *reinterpret_cast<Fp2T*>(x); }
+	static const Fp2T& cast(const Unit *x) { return *reinterpret_cast<const Fp2T*>(x); }
 	/*
 		default Fp2T operator
 		Fp2T = Fp[i]/(i^2 + 1)
 	*/
-	static void addC(Fp2T& z, const Fp2T& x, const Fp2T& y)
+	static void addA(Unit *pz, const Unit *px, const Unit *py)
 	{
+		Fp2T& z = cast(pz);
+		const Fp2T& x = cast(px);
+		const Fp2T& y = cast(py);
 		Fp::add(z.a, x.a, y.a);
 		Fp::add(z.b, x.b, y.b);
 	}
-	static void subC(Fp2T& z, const Fp2T& x, const Fp2T& y)
+	static void subA(Unit *pz, const Unit *px, const Unit *py)
 	{
+		Fp2T& z = cast(pz);
+		const Fp2T& x = cast(px);
+		const Fp2T& y = cast(py);
 		Fp::sub(z.a, x.a, y.a);
 		Fp::sub(z.b, x.b, y.b);
 	}
-	static void negC(Fp2T& y, const Fp2T& x)
+	static void negA(Unit *py, const Unit *px)
 	{
+		Fp2T& y = cast(py);
+		const Fp2T& x = cast(px);
 		Fp::neg(y.a, x.a);
 		Fp::neg(y.b, x.b);
 	}
-#if 0
+	static void mulA(Unit *pz, const Unit *px, const Unit *py)
-	/*
-		x = a + bi, y = c + di, i^2 = -1
-		z = xy = (a + bi)(c + di) = (ac - bd) + (ad + bc)i
-		ad+bc = (a + b)(c + d) - ac - bd
-		# of mod = 3
-	*/
-	static void fp2_mulW(Unit *z, const Unit *x, const Unit *y)
-	{
-		const Fp *px = reinterpret_cast<const Fp*>(x);
-		const Fp *py = reinterpret_cast<const Fp*>(y);
-		const Fp& a = px[0];
-		const Fp& b = px[1];
-		const Fp& c = py[0];
-		const Fp& d = py[1];
-		Fp *pz = reinterpret_cast<Fp*>(z);
-		Fp t1, t2, ac, bd;
-		Fp::add(t1, a, b);
-		Fp::add(t2, c, d);
-		t1 *= t2; // (a + b)(c + d)
-		Fp::mul(ac, a, c);
-		Fp::mul(bd, b, d);
-		Fp::sub(pz[0], ac, bd); // ac - bd
-		Fp::sub(pz[1], t1, ac);
-		pz[1] -= bd;
-	}
-	static void fp2_mulNFW(Fp2T& z, const Fp2T& x, const Fp2T& y)
-	{
-		const fp::Op& op = Fp::op_;
-		op.fp2_mulNF((Unit*)&z, (const Unit*)&x, (const Unit*)&y, op.p);
-	}
-#endif
-	static void mulC(Fp2T& z, const Fp2T& x, const Fp2T& y)
 	{
+		Fp2T& z = cast(pz);
+		const Fp2T& x = cast(px);
+		const Fp2T& y = cast(py);
 		Fp2Dbl d;
 		Fp2Dbl::mulPre(d, x, y);
 		FpDbl::mod(z.a, d.a);
 		FpDbl::mod(z.b, d.b);
 	}
+	static void mul2A(Unit *py, const Unit *px)
+	{
+		Fp2T& y = cast(py);
+		const Fp2T& x = cast(px);
+		Fp::mul2(y.a, x.a);
+		Fp::mul2(y.b, x.b);
+	}
 	/*
 		x = a + bi, i^2 = -1
 		y = x^2 = (a + bi)^2 = (a + b)(a - b) + 2abi
 	*/
-	static void sqrC(Fp2T& y, const Fp2T& x)
+	static void sqrA(Unit *py, const Unit *px)
 	{
+		Fp2T& y = cast(py);
+		const Fp2T& x = cast(px);
 		const Fp& a = x.a;
 		const Fp& b = x.b;
 #if 1 // faster than using FpDbl
 		Fp t1, t2, t3;
-		Fp::add(t1, b, b); // 2b
+		Fp::mul2(t1, b);
 		t1 *= a; // 2ab
 		Fp::add(t2, a, b); // a + b
 		Fp::sub(t3, a, b); // a - b
@@ -576,8 +624,10 @@ private:
 		y = (a + bi)xi = (a + bi)(xi_a + i)
 		=(a * x_ia - b) + (a + b xi_a)i
 	*/
-	static void fp2_mul_xiC(Fp2T& y, const Fp2T& x)
+	static void fp2_mul_xiA(Unit *py, const Unit *px)
 	{
+		Fp2T& y = cast(py);
+		const Fp2T& x = cast(px);
 		const Fp& a = x.a;
 		const Fp& b = x.b;
 		Fp t;
@@ -591,8 +641,10 @@ private:
 		xi = 1 + i ; xi_a = 1
 		y = (a + bi)xi = (a - b) + (a + b)i
 	*/
-	static void fp2_mul_xi_1_1iC(Fp2T& y, const Fp2T& x)
+	static void fp2_mul_xi_1_1iA(Unit *py, const Unit *px)
 	{
+		Fp2T& y = cast(py);
+		const Fp2T& x = cast(px);
 		const Fp& a = x.a;
 		const Fp& b = x.b;
 		Fp t;
@@ -600,38 +652,11 @@ private:
 		Fp::sub(y.a, a, b);
 		y.b = t;
 	}
-	/*
-		x = a + bi
-		1 / x = (a - bi) / (a^2 + b^2)
-	*/
-	static void fp2_invW(Unit *y, const Unit *x)
-	{
-		const Fp *px = reinterpret_cast<const Fp*>(x);
-		Fp *py = reinterpret_cast<Fp*>(y);
-		const Fp& a = px[0];
-		const Fp& b = px[1];
-		Fp aa, bb;
-		Fp::sqr(aa, a);
-		Fp::sqr(bb, b);
-		aa += bb;
-		Fp::inv(aa, aa); // aa = 1 / (a^2 + b^2)
-		Fp::mul(py[0], a, aa);
-		Fp::mul(py[1], b, aa);
-		Fp::neg(py[1], py[1]);
-	}
 };
 
-#ifdef MCL_XBYAK_DIRECT_CALL
-template<class Fp_> void (*Fp2T<Fp_>::add)(Fp2T& z, const Fp2T& x, const Fp2T& y);
-template<class Fp_> void (*Fp2T<Fp_>::sub)(Fp2T& z, const Fp2T& x, const Fp2T& y);
-template<class Fp_> void (*Fp2T<Fp_>::neg)(Fp2T& y, const Fp2T& x);
-template<class Fp_> void (*Fp2T<Fp_>::mul)(Fp2T& z, const Fp2T& x, const Fp2T& y);
-template<class Fp_> void (*Fp2T<Fp_>::sqr)(Fp2T& y, const Fp2T& x);
-#endif
-template<class Fp_> void (*Fp2T<Fp_>::mul_xi)(Fp2T& y, const Fp2T& x);
-
 template<class Fp>
 struct Fp2DblT {
+	typedef Fp2DblT<Fp> Fp2Dbl;
 	typedef FpDblT<Fp> FpDbl;
 	typedef Fp2T<Fp> Fp2;
 	typedef fp::Unit Unit;
@@ -656,30 +681,32 @@ struct Fp2DblT {
 		FpDbl::subPre(z.a, x.a, y.a);
 		FpDbl::subPre(z.b, x.b, y.b);
 	}
+	/*
+		imaginary part of Fp2Dbl::mul uses only add,
+		so it does not require mod.
+	*/
+	static void subSpecial(Fp2DblT& y, const Fp2DblT& x)
+	{
+		FpDbl::sub(y.a, y.a, x.a);
+		FpDbl::subPre(y.b, y.b, x.b);
+	}
 	static void neg(Fp2DblT& y, const Fp2DblT& x)
 	{
 		FpDbl::neg(y.a, x.a);
 		FpDbl::neg(y.b, x.b);
 	}
+	static void mulPre(Fp2DblT& z, const Fp2& x, const Fp2& y)
+	{
+		Fp::getOp().fp2Dbl_mulPreA_(z.a.v_, x.getUnit(), y.getUnit());
+	}
+	static void sqrPre(Fp2DblT& y, const Fp2& x)
+	{
+		Fp::getOp().fp2Dbl_sqrPreA_(y.a.v_, x.getUnit());
+	}
 	static void mul_xi(Fp2DblT& y, const Fp2DblT& x)
 	{
-		const uint32_t xi_a = Fp2::get_xi_a();
+		Fp::getOp().fp2Dbl_mul_xiA_(y.a.v_, x.a.getUnit());
-		if (xi_a == 1) {
-			FpDbl t;
-			FpDbl::add(t, x.a, x.b);
-			FpDbl::sub(y.a, x.a, x.b);
-			y.b = t;
-		} else {
-			FpDbl t;
-			FpDbl::mulUnit(t, x.a, xi_a);
-			FpDbl::sub(t, t, x.b);
-			FpDbl::mulUnit(y.b, x.b, xi_a);
-			FpDbl::add(y.b, y.b, x.a);
-			y.a = t;
-		}
 	}
-	static void (*mulPre)(Fp2DblT&, const Fp2&, const Fp2&);
-	static void (*sqrPre)(Fp2DblT&, const Fp2&);
 	static void mod(Fp2& y, const Fp2DblT& x)
 	{
 		FpDbl::mod(y.a, x.a);
@@ -695,33 +722,38 @@ struct Fp2DblT {
 	void operator-=(const Fp2DblT& x) { sub(*this, *this, x); }
 	static void init()
  	{
-		const mcl::fp::Op& op = Fp::getOp();
+		assert(!Fp::getOp().isFullBit);
-		if (op.fp2Dbl_mulPreA_) {
+		mcl::fp::Op& op = Fp::getOpNonConst();
-			mulPre = fp::func_ptr_cast<void (*)(Fp2DblT&, const Fp2&, const Fp2&)>(op.fp2Dbl_mulPreA_);
+		if (op.fp2Dbl_mulPreA_ == 0) {
-		} else {
+			op.fp2Dbl_mulPreA_ = mulPreA;
-			if (op.isFullBit) {
-				mulPre = fp2Dbl_mulPreW<true>;
-			} else {
-				mulPre = fp2Dbl_mulPreW<false>;
-			}
 		}
-		if (op.fp2Dbl_sqrPreA_) {
+		if (op.fp2Dbl_sqrPreA_ == 0) {
-			sqrPre = fp::func_ptr_cast<void (*)(Fp2DblT&, const Fp2&)>(op.fp2Dbl_sqrPreA_);
+			op.fp2Dbl_sqrPreA_ = sqrPreA;
-		} else {
+		}
-			if (op.isFullBit) {
+		if (op.fp2Dbl_mul_xiA_ == 0) {
-				sqrPre = fp2Dbl_sqrPreW<true>;
+			const uint32_t xi_a = Fp2::get_xi_a();
+			if (xi_a == 1) {
+				op.fp2Dbl_mul_xiA_ = mul_xi_1A;
 			} else {
-				sqrPre = fp2Dbl_sqrPreW<false>;
+				op.fp2Dbl_mul_xiA_ = mul_xi_genericA;
 			}
 		}
 	}
+private:
+	static Fp2 cast(Unit *x) { return *reinterpret_cast<Fp2*>(x); }
+	static const Fp2 cast(const Unit *x) { return *reinterpret_cast<const Fp2*>(x); }
+	static Fp2Dbl& castD(Unit *x) { return *reinterpret_cast<Fp2Dbl*>(x); }
+	static const Fp2Dbl& castD(const Unit *x) { return *reinterpret_cast<const Fp2Dbl*>(x); }
 	/*
 		Fp2Dbl::mulPre by FpDblT
 		@note mod of NIST_P192 is fast
 	*/
-	template<bool isFullBit>
+	static void mulPreA(Unit *pz, const Unit *px, const Unit *py)
-	static void fp2Dbl_mulPreW(Fp2DblT& z, const Fp2& x, const Fp2& y)
 	{
+		Fp2Dbl& z = castD(pz);
+		const Fp2& x = cast(px);
+		const Fp2& y = cast(py);
+		assert(!Fp::getOp().isFullBit);
 		const Fp& a = x.a;
 		const Fp& b = x.b;
 		const Fp& c = y.a;
@@ -730,45 +762,50 @@ struct Fp2DblT {
 		FpDbl& d1 = z.b;
 		FpDbl d2;
 		Fp s, t;
-		if (isFullBit) {
+		Fp::addPre(s, a, b);
-			Fp::add(s, a, b);
+		Fp::addPre(t, c, d);
-			Fp::add(t, c, d);
-		} else {
-			Fp::addPre(s, a, b);
-			Fp::addPre(t, c, d);
-		}
 		FpDbl::mulPre(d1, s, t); // (a + b)(c + d)
 		FpDbl::mulPre(d0, a, c);
 		FpDbl::mulPre(d2, b, d);
-		if (isFullBit) {
+		FpDbl::subPre(d1, d1, d0);
-			FpDbl::sub(d1, d1, d0); // (a + b)(c + d) - ac
+		FpDbl::subPre(d1, d1, d2);
-			FpDbl::sub(d1, d1, d2); // (a + b)(c + d) - ac - bd
-		} else {
-			FpDbl::subPre(d1, d1, d0);
-			FpDbl::subPre(d1, d1, d2);
-		}
 		FpDbl::sub(d0, d0, d2); // ac - bd
 	}
-	template<bool isFullBit>
+	static void sqrPreA(Unit *py, const Unit *px)
-	static void fp2Dbl_sqrPreW(Fp2DblT& y, const Fp2& x)
 	{
+		assert(!Fp::getOp().isFullBit);
+		Fp2Dbl& y = castD(py);
+		const Fp2& x = cast(px);
 		Fp t1, t2;
-		if (isFullBit) {
+		Fp::addPre(t1, x.b, x.b); // 2b
-			Fp::add(t1, x.b, x.b); // 2b
+		Fp::addPre(t2, x.a, x.b); // a + b
-			Fp::add(t2, x.a, x.b); // a + b
-		} else {
-			Fp::addPre(t1, x.b, x.b); // 2b
-			Fp::addPre(t2, x.a, x.b); // a + b
-		}
 		FpDbl::mulPre(y.b, t1, x.a); // 2ab
 		Fp::sub(t1, x.a, x.b); // a - b
 		FpDbl::mulPre(y.a, t1, t2); // (a + b)(a - b)
 	}
+	static void mul_xi_1A(Unit *py, const Unit *px)
+	{
+		Fp2Dbl& y = castD(py);
+		const Fp2Dbl& x = castD(px);
+		FpDbl t;
+		FpDbl::add(t, x.a, x.b);
+		FpDbl::sub(y.a, x.a, x.b);
+		y.b = t;
+	}
+	static void mul_xi_genericA(Unit *py, const Unit *px)
+	{
+		const uint32_t xi_a = Fp2::get_xi_a();
+		Fp2Dbl& y = castD(py);
+		const Fp2Dbl& x = castD(px);
+		FpDbl t;
+		FpDbl::mulUnit(t, x.a, xi_a);
+		FpDbl::sub(t, t, x.b);
+		FpDbl::mulUnit(y.b, x.b, xi_a);
+		FpDbl::add(y.b, y.b, x.a);
+		y.a = t;
+	}
 };
 
-template<class Fp> void (*Fp2DblT<Fp>::mulPre)(Fp2DblT&, const Fp2T<Fp>&, const Fp2T<Fp>&);
-template<class Fp> void (*Fp2DblT<Fp>::sqrPre)(Fp2DblT&, const Fp2T<Fp>&);
-
 template<class Fp> Fp2T<Fp> Fp2T<Fp>::g[Fp2T<Fp>::gN];
 template<class Fp> Fp2T<Fp> Fp2T<Fp>::g2[Fp2T<Fp>::gN];
 template<class Fp> Fp2T<Fp> Fp2T<Fp>::g3[Fp2T<Fp>::gN];
@@ -889,34 +926,24 @@ struct Fp6T : public fp::Serializable<Fp6T<_Fp>,
 		Fp2::neg(y.b, x.b);
 		Fp2::neg(y.c, x.c);
 	}
-	/*
+	static void mul2(Fp6T& y, const Fp6T& x)
-		x = a + bv + cv^2, v^3 = xi
+	{
-		x^2 = (a^2 + 2bc xi) + (c^2 xi + 2ab)v + (b^2 + 2ac)v^2
+		Fp2::mul2(y.a, x.a);
-
+		Fp2::mul2(y.b, x.b);
-		b^2 + 2ac = (a + b + c)^2 - a^2 - 2bc - c^2 - 2ab
+		Fp2::mul2(y.c, x.c);
-	*/
+	}
 	static void sqr(Fp6T& y, const Fp6T& x)
 	{
-		Fp2 t1, t2, t3;
+		Fp6Dbl XX;
-		Fp2::mul(t1, x.a, x.b);
+		Fp6Dbl::sqrPre(XX, x);
-		t1 += t1; // 2ab
+		Fp6Dbl::mod(y, XX);
-		Fp2::mul(t2, x.b, x.c);
+	}
-		t2 += t2; // 2bc
+	static inline void mul(Fp6T& z, const Fp6T& x, const Fp6T& y)
-		Fp2::sqr(t3, x.c); // c^2
+	{
-		Fp2::add(y.c, x.a, x.c); // a + c, destroy y.c
+		Fp6Dbl XY;
-		y.c += x.b; // a + b + c
+		Fp6Dbl::mulPre(XY, x, y);
-		Fp2::sqr(y.b, y.c); // (a + b + c)^2, destroy y.b
+		Fp6Dbl::mod(z, XY);
-		y.b -= t2; // (a + b + c)^2 - 2bc
+	}
-		Fp2::mul_xi(t2, t2); // 2bc xi
-		Fp2::sqr(y.a, x.a); // a^2, destroy y.a
-		y.b -= y.a; // (a + b + c)^2 - 2bc - a^2
-		y.a += t2; // a^2 + 2bc xi
-		Fp2::sub(y.c, y.b, t3); // (a + b + c)^2 - 2bc - a^2 - c^2
-		Fp2::mul_xi(y.b, t3); // c^2 xi
-		y.b += t1; // c^2 xi + 2ab
-		y.c -= t1; // b^2 + 2ac
-	}
-	static inline void mul(Fp6T& z, const Fp6T& x, const Fp6T& y);
 	/*
 		x = a + bv + cv^2, v^3 = xi
 		y = 1/x = p/q where
@@ -929,27 +956,34 @@ struct Fp6T : public fp::Serializable<Fp6T<_Fp>,
 		const Fp2& a = x.a;
 		const Fp2& b = x.b;
 		const Fp2& c = x.c;
-		Fp2 aa, bb, cc, ab, bc, ac;
+		Fp2Dbl aa, bb, cc, ab, bc, ac;
-		Fp2::sqr(aa, a);
+		Fp2Dbl::sqrPre(aa, a);
-		Fp2::sqr(bb, b);
+		Fp2Dbl::sqrPre(bb, b);
-		Fp2::sqr(cc, c);
+		Fp2Dbl::sqrPre(cc, c);
-		Fp2::mul(ab, a, b);
+		Fp2Dbl::mulPre(ab, a, b);
-		Fp2::mul(bc, b, c);
+		Fp2Dbl::mulPre(bc, b, c);
-		Fp2::mul(ac, c, a);
+		Fp2Dbl::mulPre(ac, c, a);
 
 		Fp6T p;
-		Fp2::mul_xi(p.a, bc);
+		Fp2Dbl T;
-		Fp2::sub(p.a, aa, p.a); // a^2 - bc xi
+		Fp2Dbl::mul_xi(T, bc);
-		Fp2::mul_xi(p.b, cc);
+		Fp2Dbl::sub(T, aa, T); // a^2 - bc xi
-		p.b -= ab; // c^2 xi - ab
+		Fp2Dbl::mod(p.a, T);
-		Fp2::sub(p.c, bb, ac); // b^2 - ac
+		Fp2Dbl::mul_xi(T, cc);
-		Fp2 q, t;
+		Fp2Dbl::sub(T, T, ab); // c^2 xi - ab
-		Fp2::mul(q, p.b, c);
+		Fp2Dbl::mod(p.b, T);
-		Fp2::mul(t, p.c, b);
+		Fp2Dbl::sub(T, bb, ac); // b^2 - ac
-		q += t;
+		Fp2Dbl::mod(p.c, T);
-		Fp2::mul_xi(q, q);
+
-		Fp2::mul(t, p.a, a);
+		Fp2Dbl T2;
-		q += t;
+		Fp2Dbl::mulPre(T, p.b, c);
+		Fp2Dbl::mulPre(T2, p.c, b);
+		Fp2Dbl::add(T, T, T2);
+		Fp2Dbl::mul_xi(T, T);
+		Fp2Dbl::mulPre(T2, p.a, a);
+		Fp2Dbl::addPre(T, T, T2);
+		Fp2 q;
+		Fp2Dbl::mod(q, T);
 		Fp2::inv(q, q);
 
 		Fp2::mul(y.a, p.a, q);
@@ -962,6 +996,7 @@ template<class Fp>
 struct Fp6DblT {
 	typedef Fp2T<Fp> Fp2;
 	typedef Fp6T<Fp> Fp6;
+	typedef FpDblT<Fp> FpDbl;
 	typedef Fp2DblT<Fp> Fp2Dbl;
 	typedef Fp6DblT<Fp> Fp6Dbl;
 	typedef fp::Unit Unit;
@@ -984,51 +1019,75 @@ struct Fp6DblT {
 		bf + ce = (b + c)(e + f) - be - cf
 		ae + bd = (a + b)(e + d) - ad - be
 		af + cd = (a + c)(d + f) - ad - cf
+		assum p < W/4 where W = 1 << (sizeof(Unit) * 8 * N)
+		then (b + c)(e + f) < 4p^2 < pW
 	*/
 	static void mulPre(Fp6DblT& z, const Fp6& x, const Fp6& y)
 	{
-//clk.begin();
 		const Fp2& a = x.a;
 		const Fp2& b = x.b;
 		const Fp2& c = x.c;
 		const Fp2& d = y.a;
 		const Fp2& e = y.b;
 		const Fp2& f = y.c;
-		Fp2Dbl& za = z.a;
+		Fp2Dbl& ZA = z.a;
-		Fp2Dbl& zb = z.b;
+		Fp2Dbl& ZB = z.b;
-		Fp2Dbl& zc = z.c;
+		Fp2Dbl& ZC = z.c;
-		Fp2Dbl BE;
+		Fp2 t1, t2;
-		Fp2Dbl::mulPre(za, a, d);
+		Fp2Dbl BE, CF, AD;
+		Fp2::addPre(t1, b, c);
+		Fp2::addPre(t2, e, f);
+		Fp2Dbl::mulPre(ZA, t1, t2);
+		Fp2::addPre(t1, a, b);
+		Fp2::addPre(t2, e, d);
+		Fp2Dbl::mulPre(ZB, t1, t2);
+		Fp2::addPre(t1, a, c);
+		Fp2::addPre(t2, d, f);
+		Fp2Dbl::mulPre(ZC, t1, t2);
 		Fp2Dbl::mulPre(BE, b, e);
-		Fp2Dbl::mulPre(zb, c, f);
+		Fp2Dbl::mulPre(CF, c, f);
-
+		Fp2Dbl::mulPre(AD, a, d);
-		Fp2 t1, t2, t3, t4;
+		Fp2Dbl::subSpecial(ZA, BE);
-		Fp2::add(t1, b, c);
+		Fp2Dbl::subSpecial(ZA, CF);
-		Fp2::add(t2, e, f);
+		Fp2Dbl::subSpecial(ZB, AD);
-		Fp2Dbl T1;
+		Fp2Dbl::subSpecial(ZB, BE);
-		Fp2Dbl::mulPre(T1, t1, t2);
+		Fp2Dbl::subSpecial(ZC, AD);
-		Fp2Dbl::sub(T1, T1, BE);
+		Fp2Dbl::subSpecial(ZC, CF);
-		Fp2Dbl::sub(T1, T1, zb);
+		Fp2Dbl::mul_xi(ZA, ZA);
-		Fp2Dbl::mul_xi(T1, T1);
+		Fp2Dbl::add(ZA, ZA, AD);
-
+		Fp2Dbl::mul_xi(CF, CF);
-		Fp2::add(t2, a, b);
+		Fp2Dbl::add(ZB, ZB, CF);
-		Fp2::add(t3, e, d);
+		Fp2Dbl::add(ZC, ZC, BE);
-		Fp2Dbl T2;
+	}
-		Fp2Dbl::mulPre(T2, t2, t3);
+	/*
-		Fp2Dbl::sub(T2, T2, za);
+		x = a + bv + cv^2, v^3 = xi
-		Fp2Dbl::sub(T2, T2, BE);
+		x^2 = (a^2 + 2bc xi) + (c^2 xi + 2ab)v + (b^2 + 2ac)v^2
-
-		Fp2::add(t3, a, c);
-		Fp2::add(t4, d, f);
-		Fp2Dbl::mulPre(zc, t3, t4);
-		Fp2Dbl::sub(zc, zc, za);
-		Fp2Dbl::sub(zc, zc, zb);
 
-		Fp2Dbl::add(za, za, T1);
+		b^2 + 2ac = (a + b + c)^2 - a^2 - 2bc - c^2 - 2ab
-		Fp2Dbl::mul_xi(zb, zb);
+	*/
-		Fp2Dbl::add(zb, zb, T2);
+	static void sqrPre(Fp6DblT& y, const Fp6& x)
-		Fp2Dbl::add(zc, zc, BE);
+	{
-//clk.end();
+		const Fp2& a = x.a;
+		const Fp2& b = x.b;
+		const Fp2& c = x.c;
+		Fp2 t;
+		Fp2Dbl BC2, AB2, AA, CC, T;
+		Fp2::mul2(t, b);
+		Fp2Dbl::mulPre(BC2, t, c); // 2bc
+		Fp2Dbl::mulPre(AB2, t, a); // 2ab
+		Fp2Dbl::sqrPre(AA, a);
+		Fp2Dbl::sqrPre(CC, c);
+		Fp2::add(t, a, b);
+		Fp2::add(t, t, c);
+		Fp2Dbl::sqrPre(T, t); // (a + b + c)^2
+		Fp2Dbl::sub(T, T, AA);
+		Fp2Dbl::sub(T, T, BC2);
+		Fp2Dbl::sub(T, T, CC);
+		Fp2Dbl::sub(y.c, T, AB2);
+		Fp2Dbl::mul_xi(BC2, BC2);
+		Fp2Dbl::add(y.a, AA, BC2);
+		Fp2Dbl::mul_xi(CC, CC);
+		Fp2Dbl::add(y.b, CC, AB2);
 	}
 	static void mod(Fp6& y, const Fp6Dbl& x)
 	{
@@ -1038,14 +1097,6 @@ struct Fp6DblT {
 	}
 };
 
-template<class Fp>
-inline void Fp6T<Fp>::mul(Fp6T<Fp>& z, const Fp6T<Fp>& x, const Fp6T<Fp>& y)
-{
-	Fp6DblT<Fp> Z;
-	Fp6DblT<Fp>::mulPre(Z, x, y);
-	Fp6DblT<Fp>::mod(z, Z);
-}
-
 /*
 	Fp12T = Fp6[w] / (w^2 - v)
 	x = a + b w
@@ -1148,7 +1199,6 @@ struct Fp12T : public fp::Serializable<Fp12T<Fp>,
 		Fp6 t1, t2;
 		Fp6::add(t1, a, b);
 		Fp6::add(t2, c, d);
-#if 1
 		Fp6Dbl T, AC, BD;
 		Fp6Dbl::mulPre(AC, a, c);
 		Fp6Dbl::mulPre(BD, b, d);
@@ -1158,15 +1208,6 @@ struct Fp12T : public fp::Serializable<Fp12T<Fp>,
 		Fp6Dbl::sub(T, T, AC);
 		Fp6Dbl::sub(T, T, BD);
 		Fp6Dbl::mod(z.b, T);
-#else
-		Fp6 ac, bd;
-		t1 *= t2; // (a + b)(c + d)
-		Fp6::mul(ac, a, c);
-		Fp6::mul(bd, b, d);
-		mulVadd(z.a, bd, ac);
-		t1 -= ac;
-		Fp6::sub(z.b, t1, bd);
-#endif
 	}
 	/*
 		x = a + bw, w^2 = v
@@ -1182,7 +1223,7 @@ struct Fp12T : public fp::Serializable<Fp12T<Fp>,
 		mulVadd(t1, b, a); // bv + a
 		t0 *= t1; // (a + b)(bv + a)
 		Fp6::mul(t1, a, b); // ab
-		Fp6::add(y.b, t1, t1); // 2ab
+		Fp6::mul2(y.b, t1); // 2ab
 		mulVadd(y.a, t1, t1); // abv + ab
 		Fp6::sub(y.a, t0, y.a);
 	}
@@ -1194,16 +1235,18 @@ struct Fp12T : public fp::Serializable<Fp12T<Fp>,
 	{
 		const Fp6& a = x.a;
 		const Fp6& b = x.b;
-		Fp6 t0, t1;
+		Fp6Dbl AA, BB;
-		Fp6::sqr(t0, a);
+		Fp6Dbl::sqrPre(AA, a);
-		Fp6::sqr(t1, b);
+		Fp6Dbl::sqrPre(BB, b);
-		Fp2::mul_xi(t1.c, t1.c);
+		Fp2Dbl::mul_xi(BB.c, BB.c);
-		t0.a -= t1.c;
+		Fp2Dbl::sub(AA.a, AA.a, BB.c);
-		t0.b -= t1.a;
+		Fp2Dbl::sub(AA.b, AA.b, BB.a);
-		t0.c -= t1.b; // t0 = a^2 - b^2v
+		Fp2Dbl::sub(AA.c, AA.c, BB.b); // a^2 - b^2 v
-		Fp6::inv(t0, t0);
+		Fp6 t;
-		Fp6::mul(y.a, x.a, t0);
+		Fp6Dbl::mod(t, AA);
-		Fp6::mul(y.b, x.b, t0);
+		Fp6::inv(t, t);
+		Fp6::mul(y.a, x.a, t);
+		Fp6::mul(y.b, x.b, t);
 		Fp6::neg(y.b, y.b);
 	}
 	/*
@@ -1336,6 +1379,12 @@ struct GroupMtoA : public T {
 	{
 		T::mul(castT(z), castT(x), castT(y));
 	}
+	static void sub(GroupMtoA& z, const GroupMtoA& x, const GroupMtoA& y)
+	{
+		T r;
+		T::unitaryInv(r, castT(y));
+		T::mul(castT(z), castT(x), r);
+	}
 	static void dbl(GroupMtoA& y, const GroupMtoA& x)
 	{
 		T::sqr(castT(y), castT(x));
@@ -1363,6 +1412,10 @@ struct GroupMtoA : public T {
 	{
 		add(*this, *this, rhs);
 	}
+	void operator-=(const GroupMtoA& rhs)
+	{
+		sub(*this, *this, rhs);
+	}
 	void normalize() {}
 private:
 	bool isOne() const;

include/mcl/gmp_util.hpp

@@ -10,10 +10,14 @@
 #include <stdlib.h>
 #include <assert.h>
 #include <stdint.h>
+#include <cybozu/bit_operation.hpp>
 #ifndef CYBOZU_DONT_USE_EXCEPTION
 #include <cybozu/exception.hpp>
 #endif
 #include <mcl/randgen.hpp>
+#include <mcl/config.hpp>
+#include <mcl/conversion.hpp>
+
 #ifdef _MSC_VER
 	#pragma warning(push)
 	#pragma warning(disable : 4616)
@@ -26,9 +30,6 @@
 #if defined(__EMSCRIPTEN__) || defined(__wasm__)
 	#define MCL_USE_VINT
 #endif
-#ifndef MCL_MAX_BIT_SIZE
-	#define MCL_MAX_BIT_SIZE 521
-#endif
 #ifdef MCL_USE_VINT
 #include <mcl/vint.hpp>
 typedef mcl::Vint mpz_class;
@@ -40,27 +41,8 @@ typedef mcl::Vint mpz_class;
 #endif
 #endif
 
-#ifndef MCL_SIZEOF_UNIT
-	#if defined(CYBOZU_OS_BIT) && (CYBOZU_OS_BIT == 32)
-		#define MCL_SIZEOF_UNIT 4
-	#else
-		#define MCL_SIZEOF_UNIT 8
-	#endif
-#endif
-
 namespace mcl {
 
-namespace fp {
-
-#if MCL_SIZEOF_UNIT == 8
-typedef uint64_t Unit;
-#else
-typedef uint32_t Unit;
-#endif
-#define MCL_UNIT_BIT_SIZE (MCL_SIZEOF_UNIT * 8)
-
-} // mcl::fp
-
 namespace gmp {
 
 typedef mpz_class ImplType;
@@ -81,24 +63,20 @@ void setArray(bool *pb, mpz_class& z, const T *buf, size_t n)
 	buf[0, size) = x
 	buf[size, maxSize) with zero
 */
-template<class T, class U>
-bool getArray_(T *buf, size_t maxSize, const U *x, int xn)//const mpz_srcptr x)
-{
-	const size_t bufByteSize = sizeof(T) * maxSize;
-	if (xn < 0) return false;
-	size_t xByteSize = sizeof(*x) * xn;
-	if (xByteSize > bufByteSize) return false;
-	memcpy(buf, x, xByteSize);
-	memset((char*)buf + xByteSize, 0, bufByteSize - xByteSize);
-	return true;
-}
 template<class T>
 void getArray(bool *pb, T *buf, size_t maxSize, const mpz_class& x)
 {
 #ifdef MCL_USE_VINT
-	*pb = getArray_(buf, maxSize, x.getUnit(), x.getUnitSize());
+	const fp::Unit *src = x.getUnit();
+	const size_t n = x.getUnitSize();
+	*pb = fp::convertArrayAsLE(buf, maxSize, src, n);
 #else
-	*pb = getArray_(buf, maxSize, x.get_mpz_t()->_mp_d, x.get_mpz_t()->_mp_size);
+	int n = x.get_mpz_t()->_mp_size;
+	if (n < 0) {
+		*pb = false;
+		return;
+	}
+	*pb = fp::convertArrayAsLE(buf, maxSize, x.get_mpz_t()->_mp_d, n);
 #endif
 }
 inline void set(mpz_class& z, uint64_t x)
@@ -434,6 +412,36 @@ inline size_t getUnitSize(const mpz_class& x)
 	return std::abs(x.get_mpz_t()->_mp_size);
 #endif
 }
+
+/*
+	get the number of lower zeros
+*/
+template<class T>
+size_t getLowerZeroBitNum(const T *x, size_t n)
+{
+	size_t ret = 0;
+	for (size_t i = 0; i < n; i++) {
+		T v = x[i];
+		if (v == 0) {
+			ret += sizeof(T) * 8;
+		} else {
+			ret += cybozu::bsf<T>(v);
+			break;
+		}
+	}
+	return ret;
+}
+
+/*
+	get the number of lower zero
+	@note x != 0
+*/
+inline size_t getLowerZeroBitNum(const mpz_class& x)
+{
+	assert(!isZero(x));
+	return getLowerZeroBitNum(getUnit(x), getUnitSize(x));
+}
+
 inline mpz_class abs(const mpz_class& x)
 {
 #ifdef MCL_USE_VINT
@@ -576,6 +584,53 @@ bool getNAF(Vec& v, const mpz_class& x)
 	}
 }
 
+/*
+	v = naf[i]
+	v = 0 or (|v| <= 2^(w-1) - 1 and odd)
+*/
+template<class Vec>
+void getNAFwidth(bool *pb, Vec& naf, mpz_class x, size_t w)
+{
+	assert(w > 0);
+	*pb = true;
+	naf.clear();
+	bool negative = false;
+	if (x < 0) {
+		negative = true;
+		x = -x;
+	}
+	size_t zeroNum = 0;
+	const int signedMaxW = 1 << (w - 1);
+	const int maxW = signedMaxW * 2;
+	const int maskW = maxW - 1;
+	while (!isZero(x)) {
+		size_t z = gmp::getLowerZeroBitNum(x);
+		if (z) {
+			x >>= z;
+			zeroNum += z;
+		}
+		for (size_t i = 0; i < zeroNum; i++) {
+			naf.push(pb, 0);
+			if (!*pb) return;
+		}
+		assert(!isZero(x));
+		int v = getUnit(x)[0] & maskW;
+		x >>= w;
+		if (v & signedMaxW) {
+			x++;
+			v -= maxW;
+		}
+		naf.push(pb, typename Vec::value_type(v));
+		if (!*pb) return;
+		zeroNum = w - 1;
+	}
+	if (negative) {
+		for (size_t i = 0; i < naf.size(); i++) {
+			naf[i] = -naf[i];
+		}
+	}
+}
+
 #ifndef CYBOZU_DONT_USE_EXCEPTION
 inline void setStr(mpz_class& z, const std::string& str, int base = 0)
 {
@@ -864,6 +919,86 @@ public:
 #endif
 };
 
+/*
+	x mod p for a small value x < (pMulTblN * p).
+*/
+struct SmallModp {
+	typedef mcl::fp::Unit Unit;
+	static const size_t unitBitSize = sizeof(Unit) * 8;
+	static const size_t maxTblSize = (MCL_MAX_BIT_SIZE + unitBitSize - 1) / unitBitSize + 1;
+	static const size_t maxMulN = 9;
+	static const size_t pMulTblN = maxMulN + 1;
+	uint32_t N_;
+	uint32_t shiftL_;
+	uint32_t shiftR_;
+	uint32_t maxIdx_;
+	// pMulTbl_[i] = (p * i) >> (pBitSize_ - 1)
+	Unit pMulTbl_[pMulTblN][maxTblSize];
+	// idxTbl_[x] = (x << (pBitSize_ - 1)) / p
+	uint8_t idxTbl_[pMulTblN * 2];
+	// return x >> (pBitSize_ - 1)
+	SmallModp()
+		: N_(0)
+		, shiftL_(0)
+		, shiftR_(0)
+		, maxIdx_(0)
+		, pMulTbl_()
+		, idxTbl_()
+	{
+	}
+	// return argmax { i : x > i * p }
+	uint32_t approxMul(const Unit *x) const
+	{
+		uint32_t top = getTop(x);
+		assert(top <= maxIdx_);
+		return idxTbl_[top];
+	}
+	const Unit *getPmul(size_t v) const
+	{
+		assert(v < pMulTblN);
+		return pMulTbl_[v];
+	}
+	uint32_t getTop(const Unit *x) const
+	{
+		if (shiftR_ == 0) return x[N_ - 1];
+		return (x[N_ - 1] >> shiftR_) | (x[N_] << shiftL_);
+	}
+	uint32_t cvtInt(const mpz_class& x) const
+	{
+		assert(mcl::gmp::getUnitSize(x) <= 1);
+		if (x == 0) {
+			return 0;
+		} else {
+			return uint32_t(mcl::gmp::getUnit(x)[0]);
+		}
+	}
+	void init(const mpz_class& p)
+	{
+		size_t pBitSize = mcl::gmp::getBitSize(p);
+		N_ = uint32_t((pBitSize + unitBitSize - 1) / unitBitSize);
+		shiftR_ = (pBitSize - 1) % unitBitSize;
+		shiftL_ = unitBitSize - shiftR_;
+		mpz_class t = 0;
+		for (size_t i = 0; i < pMulTblN; i++) {
+			bool b;
+			mcl::gmp::getArray(&b, pMulTbl_[i], maxTblSize, t);
+			assert(b);
+			(void)b;
+			if (i == pMulTblN - 1) {
+				maxIdx_ = getTop(pMulTbl_[i]);
+				assert(maxIdx_ < CYBOZU_NUM_OF_ARRAY(idxTbl_));
+				break;
+			}
+			t += p;
+		}
+
+		for (uint32_t i = 0; i <= maxIdx_; i++) {
+			idxTbl_[i] = cvtInt((mpz_class(int(i)) << (pBitSize - 1)) / p);
+		}
+	}
+};
+
+
 /*
 	Barrett Reduction
 	for non GMP version

include/mcl/impl/bn_c_impl.hpp

@@ -77,12 +77,18 @@ int mclBn_init(int curve, int compiledTimeVar)
 		initG1only(&b, *para);
 		return b ? 0 : -1;
 	}
-	const mcl::CurveParam& cp = mcl::getCurveParam(curve);
+	const mcl::CurveParam* cp = mcl::getCurveParam(curve);
+	if (cp == 0) return -1;
 	bool b;
-	initPairing(&b, cp);
+	initPairing(&b, *cp);
 	return b ? 0 : -1;
 }
 
+int mclBn_getCurveType()
+{
+	return mcl::bn::BN::param.cp.curveType;
+}
+
 int mclBn_getOpUnitSize()
 {
 	return (int)Fp::getUnitSize() * sizeof(mcl::fp::Unit) / sizeof(uint64_t);
@@ -113,9 +119,21 @@ mclSize mclBn_getFieldOrder(char *buf, mclSize maxBufSize)
 	return Fp::getModulo(buf, maxBufSize);
 }
 
-void mclBn_setETHserialization(int ETHserialization)
+void mclBn_setETHserialization(int enable)
+{
+	if (mclBn_getCurveType() != MCL_BLS12_381) return;
+	Fp::setETHserialization(enable == 1);
+	Fr::setETHserialization(enable == 1);
+}
+
+int mclBn_getETHserialization()
+{
+	return Fp::getETHserialization() ? 1 : 0;
+}
+
+int mclBn_setMapToMode(int mode)
 {
-	Fp::setETHserialization(ETHserialization == 1);
+	return setMapToMode(mode) ? 0 : -1;
 }
 
 ////////////////////////////////////////////////
@@ -141,13 +159,24 @@ int mclBnFr_setStr(mclBnFr *x, const char *buf, mclSize bufSize, int ioMode)
 }
 int mclBnFr_setLittleEndian(mclBnFr *x, const void *buf, mclSize bufSize)
 {
-	cast(x)->setArrayMask((const char *)buf, bufSize);
+	cast(x)->setArrayMask((const uint8_t *)buf, bufSize);
 	return 0;
 }
+int mclBnFr_setBigEndianMod(mclBnFr *x, const void *buf, mclSize bufSize)
+{
+	bool b;
+	cast(x)->setBigEndianMod(&b, (const uint8_t*)buf, bufSize);
+	return b ? 0 : -1;
+}
+
+mclSize mclBnFr_getLittleEndian(void *buf, mclSize maxBufSize, const mclBnFr *x)
+{
+	return cast(x)->getLittleEndian((uint8_t*)buf, maxBufSize);
+}
 int mclBnFr_setLittleEndianMod(mclBnFr *x, const void *buf, mclSize bufSize)
 {
 	bool b;
-	cast(x)->setArray(&b, (const char *)buf, bufSize, mcl::fp::Mod);
+	cast(x)->setArrayMod(&b, (const uint8_t *)buf, bufSize);
 	return b ? 0 : -1;
 }
 mclSize mclBnFr_deserialize(mclBnFr *x, const void *buf, mclSize bufSize)
@@ -171,6 +200,14 @@ int mclBnFr_isOne(const mclBnFr *x)
 {
 	return cast(x)->isOne();
 }
+int mclBnFr_isOdd(const mclBnFr *x)
+{
+	return cast(x)->isOdd();
+}
+int mclBnFr_isNegative(const mclBnFr *x)
+{
+	return cast(x)->isNegative();
+}
 
 #ifndef MCL_DONT_USE_CSRPNG
 int mclBnFr_setByCSPRNG(mclBnFr *x)
@@ -179,6 +216,12 @@ int mclBnFr_setByCSPRNG(mclBnFr *x)
 	cast(x)->setByCSPRNG(&b);
 	return b ? 0 : -1;
 }
+int mclBnFp_setByCSPRNG(mclBnFp *x)
+{
+	bool b;
+	cast(x)->setByCSPRNG(&b);
+	return b ? 0 : -1;
+}
 void mclBn_setRandFunc(void *self, unsigned int (*readFunc)(void *self, void *buf, unsigned int bufSize))
 {
 	mcl::fp::RandGen::setRandFunc(self, readFunc);
@@ -230,6 +273,77 @@ void mclBnFr_div(mclBnFr *z, const mclBnFr *x, const mclBnFr *y)
 	Fr::div(*cast(z),*cast(x), *cast(y));
 }
 
+void mclBnFp_neg(mclBnFp *y, const mclBnFp *x)
+{
+	Fp::neg(*cast(y), *cast(x));
+}
+void mclBnFp_inv(mclBnFp *y, const mclBnFp *x)
+{
+	Fp::inv(*cast(y), *cast(x));
+}
+void mclBnFp_sqr(mclBnFp *y, const mclBnFp *x)
+{
+	Fp::sqr(*cast(y), *cast(x));
+}
+void mclBnFp_add(mclBnFp *z, const mclBnFp *x, const mclBnFp *y)
+{
+	Fp::add(*cast(z),*cast(x), *cast(y));
+}
+void mclBnFp_sub(mclBnFp *z, const mclBnFp *x, const mclBnFp *y)
+{
+	Fp::sub(*cast(z),*cast(x), *cast(y));
+}
+void mclBnFp_mul(mclBnFp *z, const mclBnFp *x, const mclBnFp *y)
+{
+	Fp::mul(*cast(z),*cast(x), *cast(y));
+}
+void mclBnFp_div(mclBnFp *z, const mclBnFp *x, const mclBnFp *y)
+{
+	Fp::div(*cast(z),*cast(x), *cast(y));
+}
+
+void mclBnFp2_neg(mclBnFp2 *y, const mclBnFp2 *x)
+{
+	Fp2::neg(*cast(y), *cast(x));
+}
+void mclBnFp2_inv(mclBnFp2 *y, const mclBnFp2 *x)
+{
+	Fp2::inv(*cast(y), *cast(x));
+}
+void mclBnFp2_sqr(mclBnFp2 *y, const mclBnFp2 *x)
+{
+	Fp2::sqr(*cast(y), *cast(x));
+}
+void mclBnFp2_add(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y)
+{
+	Fp2::add(*cast(z),*cast(x), *cast(y));
+}
+void mclBnFp2_sub(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y)
+{
+	Fp2::sub(*cast(z),*cast(x), *cast(y));
+}
+void mclBnFp2_mul(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y)
+{
+	Fp2::mul(*cast(z),*cast(x), *cast(y));
+}
+void mclBnFp2_div(mclBnFp2 *z, const mclBnFp2 *x, const mclBnFp2 *y)
+{
+	Fp2::div(*cast(z),*cast(x), *cast(y));
+}
+
+int mclBnFr_squareRoot(mclBnFr *y, const mclBnFr *x)
+{
+	return Fr::squareRoot(*cast(y), *cast(x)) ? 0 : -1;
+}
+int mclBnFp_squareRoot(mclBnFp *y, const mclBnFp *x)
+{
+	return Fp::squareRoot(*cast(y), *cast(x)) ? 0 : -1;
+}
+int mclBnFp2_squareRoot(mclBnFp2 *y, const mclBnFp2 *x)
+{
+	return Fp2::squareRoot(*cast(y), *cast(x)) ? 0 : -1;
+}
+
 ////////////////////////////////////////////////
 // set zero
 void mclBnG1_clear(mclBnG1 *x)
@@ -443,6 +557,10 @@ void mclBnGT_neg(mclBnGT *y, const mclBnGT *x)
 	Fp12::neg(*cast(y), *cast(x));
 }
 void mclBnGT_inv(mclBnGT *y, const mclBnGT *x)
+{
+	Fp12::unitaryInv(*cast(y), *cast(x));
+}
+void mclBnGT_invGeneric(mclBnGT *y, const mclBnGT *x)
 {
 	Fp12::inv(*cast(y), *cast(x));
 }
@@ -476,6 +594,19 @@ void mclBnGT_powGeneric(mclBnGT *z, const mclBnGT *x, const mclBnFr *y)
 	Fp12::powGeneric(*cast(z), *cast(x), *cast(y));
 }
 
+void mclBnG1_mulVec(mclBnG1 *z, const mclBnG1 *x, const mclBnFr *y, mclSize n)
+{
+	G1::mulVec(*cast(z), cast(x), cast(y), n);
+}
+void mclBnG2_mulVec(mclBnG2 *z, const mclBnG2 *x, const mclBnFr *y, mclSize n)
+{
+	G2::mulVec(*cast(z), cast(x), cast(y), n);
+}
+void mclBnGT_powVec(mclBnGT *z, const mclBnGT *x, const mclBnFr *y, mclSize n)
+{
+	GT::powVec(*cast(z), cast(x), cast(y), n);
+}
+
 void mclBn_pairing(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y)
 {
 	pairing(*cast(z), *cast(x), *cast(y));
@@ -488,6 +619,10 @@ void mclBn_millerLoop(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y)
 {
 	millerLoop(*cast(z), *cast(x), *cast(y));
 }
+void mclBn_millerLoopVec(mclBnGT *z, const mclBnG1 *x, const mclBnG2 *y, mclSize n)
+{
+	millerLoopVec(*cast(z), cast(x), cast(y), n);
+}
 int mclBn_getUint64NumToPrecompute(void)
 {
 	return int(BN::param.precomputedQcoeffSize * sizeof(Fp6) / sizeof(uint64_t));
@@ -560,6 +695,15 @@ void mclBn_verifyOrderG2(int doVerify)
 	verifyOrderG2(doVerify != 0);
 }
 
+void mclBnFp_setInt(mclBnFp *y, mclInt x)
+{
+	*cast(y) = x;
+}
+void mclBnFp_setInt32(mclBnFp *y, int x)
+{
+	*cast(y) = x;
+}
+
 mclSize mclBnFp_getStr(char *buf, mclSize maxBufSize, const mclBnFp *x, int ioMode)
 {
 	return cast(x)->getStr(buf, maxBufSize, ioMode);
@@ -585,20 +729,52 @@ void mclBnFp_clear(mclBnFp *x)
 
 int mclBnFp_setLittleEndian(mclBnFp *x, const void *buf, mclSize bufSize)
 {
-	cast(x)->setArrayMask((const char *)buf, bufSize);
+	cast(x)->setArrayMask((const uint8_t *)buf, bufSize);
 	return 0;
 }
 
 int mclBnFp_setLittleEndianMod(mclBnFp *x, const void *buf, mclSize bufSize)
 {
 	bool b;
-	cast(x)->setArray(&b, (const char *)buf, bufSize, mcl::fp::Mod);
+	cast(x)->setLittleEndianMod(&b, (const uint8_t*)buf, bufSize);
 	return b ? 0 : -1;
 }
+
+int mclBnFp_setBigEndianMod(mclBnFp *x, const void *buf, mclSize bufSize)
+{
+	bool b;
+	cast(x)->setBigEndianMod(&b, (const uint8_t*)buf, bufSize);
+	return b ? 0 : -1;
+}
+
+mclSize mclBnFp_getLittleEndian(void *buf, mclSize maxBufSize, const mclBnFp *x)
+{
+	return cast(x)->getLittleEndian((uint8_t*)buf, maxBufSize);
+}
+int mclBnFp_isValid(const mclBnFp *x)
+{
+	return cast(x)->isValid();
+}
 int mclBnFp_isEqual(const mclBnFp *x, const mclBnFp *y)
 {
 	return *cast(x) == *cast(y);
 }
+int mclBnFp_isZero(const mclBnFp *x)
+{
+	return cast(x)->isZero();
+}
+int mclBnFp_isOne(const mclBnFp *x)
+{
+	return cast(x)->isOne();
+}
+int mclBnFp_isOdd(const mclBnFp *x)
+{
+	return cast(x)->isOdd();
+}
+int mclBnFp_isNegative(const mclBnFp *x)
+{
+	return cast(x)->isNegative();
+}
 
 int mclBnFp_setHashOf(mclBnFp *x, const void *buf, mclSize bufSize)
 {
@@ -632,6 +808,14 @@ int mclBnFp2_isEqual(const mclBnFp2 *x, const mclBnFp2 *y)
 {
 	return *cast(x) == *cast(y);
 }
+int mclBnFp2_isZero(const mclBnFp2 *x)
+{
+	return cast(x)->isZero();
+}
+int mclBnFp2_isOne(const mclBnFp2 *x)
+{
+	return cast(x)->isOne();
+}
 
 int mclBnFp2_mapToG2(mclBnG2 *y, const mclBnFp2 *x)
 {

include/mcl/mapto_wb19.hpp

@@ -0,0 +1,572 @@
+#pragma once
+/**
+	@file
+	@brief map to G2 on BLS12-381 (must be included from mcl/bn.hpp)
+	@author MITSUNARI Shigeo(@herumi)
+	@license modified new BSD license
+	http://opensource.org/licenses/BSD-3-Clause
+	ref. https://eprint.iacr.org/2019/403 , https://github.com/algorand/bls_sigs_ref
+*/
+namespace mcl {
+
+namespace local {
+
+// y^2 = x^3 + 4(1 + i)
+template<class F>
+struct PointT {
+	typedef F Fp;
+	F x, y, z;
+	static F a_;
+	static F b_;
+	static int specialA_;
+	bool isZero() const
+	{
+		return z.isZero();
+	}
+	void clear()
+	{
+		x.clear();
+		y.clear();
+		z.clear();
+	}
+	bool isEqual(const PointT<F>& rhs) const
+	{
+		return ec::isEqualJacobi(*this, rhs);
+	}
+};
+
+template<class F> F PointT<F>::a_;
+template<class F> F PointT<F>::b_;
+template<class F> int PointT<F>::specialA_;
+
+} // mcl::local
+
+template<class Fp, class G1, class Fp2, class G2>
+struct MapTo_WB19 {
+	typedef local::PointT<Fp> E1;
+	typedef local::PointT<Fp2> E2;
+	mpz_class sqrtConst; // (p^2 - 9) / 16
+	Fp2 g2A;
+	Fp2 g2B;
+	Fp2 root4[4];
+	Fp2 etas[4];
+	Fp2 xnum[4];
+	Fp2 xden[3];
+	Fp2 ynum[4];
+	Fp2 yden[4];
+	Fp g1A, g1B, g1c1, g1c2;
+	Fp g1xnum[12];
+	Fp g1xden[11];
+	Fp g1ynum[16];
+	Fp g1yden[16];
+	mpz_class g1cofactor;
+	int g1Z;
+	void init()
+	{
+		bool b;
+		g2A.a = 0;
+		g2A.b = 240;
+		g2B.a = 1012;
+		g2B.b = 1012;
+		E1::a_.clear();
+		E1::b_ = 4;
+		E1::specialA_ = ec::Zero;
+		E2::a_.clear();
+		E2::b_.a = 4;
+		E2::b_.b = 4;
+		E2::specialA_ = ec::Zero;
+		sqrtConst = Fp::getOp().mp;
+		sqrtConst *= sqrtConst;
+		sqrtConst -= 9;
+		sqrtConst /= 16;
+		const char *rv1Str = "0x6af0e0437ff400b6831e36d6bd17ffe48395dabc2d3435e77f76e17009241c5ee67992f72ec05f4c81084fbede3cc09";
+		root4[0].a = 1;
+		root4[0].b.clear();
+		root4[1].a.clear();
+		root4[1].b = 1;
+		root4[2].a.setStr(&b, rv1Str);
+		assert(b); (void)b;
+		root4[2].b = root4[2].a;
+		root4[3].a = root4[2].a;
+		Fp::neg(root4[3].b, root4[3].a);
+		const char *ev1Str = "0x699be3b8c6870965e5bf892ad5d2cc7b0e85a117402dfd83b7f4a947e02d978498255a2aaec0ac627b5afbdf1bf1c90";
+		const char *ev2Str = "0x8157cd83046453f5dd0972b6e3949e4288020b5b8a9cc99ca07e27089a2ce2436d965026adad3ef7baba37f2183e9b5";
+		const char *ev3Str = "0xab1c2ffdd6c253ca155231eb3e71ba044fd562f6f72bc5bad5ec46a0b7a3b0247cf08ce6c6317f40edbc653a72dee17";
+		const char *ev4Str = "0xaa404866706722864480885d68ad0ccac1967c7544b447873cc37e0181271e006df72162a3d3e0287bf597fbf7f8fc1";
+		Fp& ev1 = etas[0].a;
+		Fp& ev2 = etas[0].b;
+		Fp& ev3 = etas[2].a;
+		Fp& ev4 = etas[2].b;
+		ev1.setStr(&b, ev1Str);
+		assert(b); (void)b;
+		ev2.setStr(&b, ev2Str);
+		assert(b); (void)b;
+		Fp::neg(etas[1].a, ev2);
+		etas[1].b = ev1;
+		ev3.setStr(&b, ev3Str);
+		assert(b); (void)b;
+		ev4.setStr(&b, ev4Str);
+		assert(b); (void)b;
+		Fp::neg(etas[3].a, ev4);
+		etas[3].b = ev3;
+		init_iso3();
+		{
+			const char *A = "0x144698a3b8e9433d693a02c96d4982b0ea985383ee66a8d8e8981aefd881ac98936f8da0e0f97f5cf428082d584c1d";
+			const char *B = "0x12e2908d11688030018b12e8753eee3b2016c1f0f24f4070a0b9c14fcef35ef55a23215a316ceaa5d1cc48e98e172be0";
+			const char *c1 = "0x680447a8e5ff9a692c6e9ed90d2eb35d91dd2e13ce144afd9cc34a83dac3d8907aaffffac54ffffee7fbfffffffeaaa";
+			const char *c2 = "0x3d689d1e0e762cef9f2bec6130316806b4c80eda6fc10ce77ae83eab1ea8b8b8a407c9c6db195e06f2dbeabc2baeff5";
+			g1A.setStr(&b, A);
+			assert(b); (void)b;
+			g1B.setStr(&b, B);
+			assert(b); (void)b;
+			g1c1.setStr(&b, c1);
+			assert(b); (void)b;
+			g1c2.setStr(&b, c2);
+			assert(b); (void)b;
+			g1Z = 11;
+			gmp::setStr(&b, g1cofactor, "d201000000010001", 16);
+			assert(b); (void)b;
+		}
+		init_iso11();
+	}
+	void initArray(Fp *dst, const char **s, size_t n) const
+	{
+		bool b;
+		for (size_t i = 0; i < n; i++) {
+			dst[i].setStr(&b, s[i]);
+			assert(b);
+			(void)b;
+		}
+	}
+	void init_iso3()
+	{
+		const char *tbl[] = {
+			"0x5c759507e8e333ebb5b7a9a47d7ed8532c52d39fd3a042a88b58423c50ae15d5c2638e343d9c71c6238aaaaaaaa97d6",
+			"0x11560bf17baa99bc32126fced787c88f984f87adf7ae0c7f9a208c6b4f20a4181472aaa9cb8d555526a9ffffffffc71a",
+			"0x11560bf17baa99bc32126fced787c88f984f87adf7ae0c7f9a208c6b4f20a4181472aaa9cb8d555526a9ffffffffc71e",
+			"0x8ab05f8bdd54cde190937e76bc3e447cc27c3d6fbd7063fcd104635a790520c0a395554e5c6aaaa9354ffffffffe38d",
+			"0x171d6541fa38ccfaed6dea691f5fb614cb14b4e7f4e810aa22d6108f142b85757098e38d0f671c7188e2aaaaaaaa5ed1",
+			"0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffaa63",
+			"0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffaa9f",
+			"0x1530477c7ab4113b59a4c18b076d11930f7da5d4a07f649bf54439d87d27e500fc8c25ebf8c92f6812cfc71c71c6d706",
+			"0x5c759507e8e333ebb5b7a9a47d7ed8532c52d39fd3a042a88b58423c50ae15d5c2638e343d9c71c6238aaaaaaaa97be",
+			"0x11560bf17baa99bc32126fced787c88f984f87adf7ae0c7f9a208c6b4f20a4181472aaa9cb8d555526a9ffffffffc71c",
+			"0x8ab05f8bdd54cde190937e76bc3e447cc27c3d6fbd7063fcd104635a790520c0a395554e5c6aaaa9354ffffffffe38f",
+			"0x124c9ad43b6cf79bfbf7043de3811ad0761b0f37a1e26286b0e977c69aa274524e79097a56dc4bd9e1b371c71c718b10",
+			"0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffa8fb",
+			"0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffa9d3",
+			"0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffaa99",
+		};
+		bool b;
+		xnum[0].a.setStr(&b, tbl[0]); assert(b); (void)b;
+		xnum[0].b = xnum[0].a;
+		xnum[1].a.clear();
+		xnum[1].b.setStr(&b, tbl[1]); assert(b); (void)b;
+		xnum[2].a.setStr(&b, tbl[2]); assert(b); (void)b;
+		xnum[2].b.setStr(&b, tbl[3]); assert(b); (void)b;
+		xnum[3].a.setStr(&b, tbl[4]); assert(b); (void)b;
+		xnum[3].b.clear();
+		xden[0].a.clear();
+		xden[0].b.setStr(&b, tbl[5]); assert(b); (void)b;
+		xden[1].a = 0xc;
+		xden[1].b.setStr(&b, tbl[6]); assert(b); (void)b;
+		xden[2].a = 1;
+		xden[2].b = 0;
+		ynum[0].a.setStr(&b, tbl[7]); assert(b); (void)b;
+		ynum[0].b = ynum[0].a;
+		ynum[1].a.clear();
+		ynum[1].b.setStr(&b, tbl[8]); assert(b); (void)b;
+		ynum[2].a.setStr(&b, tbl[9]); assert(b); (void)b;
+		ynum[2].b.setStr(&b, tbl[10]); assert(b); (void)b;
+		ynum[3].a.setStr(&b, tbl[11]); assert(b); (void)b;
+		ynum[3].b.clear();
+		yden[0].a.setStr(&b, tbl[12]); assert(b); (void)b;
+		yden[0].b = yden[0].a;
+		yden[1].a.clear();
+		yden[1].b.setStr(&b, tbl[13]); assert(b); (void)b;
+		yden[2].a = 0x12;
+		yden[2].b.setStr(&b, tbl[14]); assert(b); (void)b;
+		yden[3].a = 1;
+		yden[3].b.clear();
+	}
+	void init_iso11()
+	{
+		const char *xnumStr[] = {
+			"0x11a05f2b1e833340b809101dd99815856b303e88a2d7005ff2627b56cdb4e2c85610c2d5f2e62d6eaeac1662734649b7",
+			"0x17294ed3e943ab2f0588bab22147a81c7c17e75b2f6a8417f565e33c70d1e86b4838f2a6f318c356e834eef1b3cb83bb",
+			"0xd54005db97678ec1d1048c5d10a9a1bce032473295983e56878e501ec68e25c958c3e3d2a09729fe0179f9dac9edcb0",
+			"0x1778e7166fcc6db74e0609d307e55412d7f5e4656a8dbf25f1b33289f1b330835336e25ce3107193c5b388641d9b6861",
+			"0xe99726a3199f4436642b4b3e4118e5499db995a1257fb3f086eeb65982fac18985a286f301e77c451154ce9ac8895d9",
+			"0x1630c3250d7313ff01d1201bf7a74ab5db3cb17dd952799b9ed3ab9097e68f90a0870d2dcae73d19cd13c1c66f652983",
+			"0xd6ed6553fe44d296a3726c38ae652bfb11586264f0f8ce19008e218f9c86b2a8da25128c1052ecaddd7f225a139ed84",
+			"0x17b81e7701abdbe2e8743884d1117e53356de5ab275b4db1a682c62ef0f2753339b7c8f8c8f475af9ccb5618e3f0c88e",
+			"0x80d3cf1f9a78fc47b90b33563be990dc43b756ce79f5574a2c596c928c5d1de4fa295f296b74e956d71986a8497e317",
+			"0x169b1f8e1bcfa7c42e0c37515d138f22dd2ecb803a0c5c99676314baf4bb1b7fa3190b2edc0327797f241067be390c9e",
+			"0x10321da079ce07e272d8ec09d2565b0dfa7dccdde6787f96d50af36003b14866f69b771f8c285decca67df3f1605fb7b",
+			"0x6e08c248e260e70bd1e962381edee3d31d79d7e22c837bc23c0bf1bc24c6b68c24b1b80b64d391fa9c8ba2e8ba2d229",
+		};
+		const char *xdenStr[] = {
+			"0x8ca8d548cff19ae18b2e62f4bd3fa6f01d5ef4ba35b48ba9c9588617fc8ac62b558d681be343df8993cf9fa40d21b1c",
+			"0x12561a5deb559c4348b4711298e536367041e8ca0cf0800c0126c2588c48bf5713daa8846cb026e9e5c8276ec82b3bff",
+			"0xb2962fe57a3225e8137e629bff2991f6f89416f5a718cd1fca64e00b11aceacd6a3d0967c94fedcfcc239ba5cb83e19",
+			"0x3425581a58ae2fec83aafef7c40eb545b08243f16b1655154cca8abc28d6fd04976d5243eecf5c4130de8938dc62cd8",
+			"0x13a8e162022914a80a6f1d5f43e7a07dffdfc759a12062bb8d6b44e833b306da9bd29ba81f35781d539d395b3532a21e",
+			"0xe7355f8e4e667b955390f7f0506c6e9395735e9ce9cad4d0a43bcef24b8982f7400d24bc4228f11c02df9a29f6304a5",
+			"0x772caacf16936190f3e0c63e0596721570f5799af53a1894e2e073062aede9cea73b3538f0de06cec2574496ee84a3a",
+			"0x14a7ac2a9d64a8b230b3f5b074cf01996e7f63c21bca68a81996e1cdf9822c580fa5b9489d11e2d311f7d99bbdcc5a5e",
+			"0xa10ecf6ada54f825e920b3dafc7a3cce07f8d1d7161366b74100da67f39883503826692abba43704776ec3a79a1d641",
+			"0x95fc13ab9e92ad4476d6e3eb3a56680f682b4ee96f7d03776df533978f31c1593174e4b4b7865002d6384d168ecdd0a",
+			"0x1",
+		};
+		const char *ynumStr[] = {
+			"0x90d97c81ba24ee0259d1f094980dcfa11ad138e48a869522b52af6c956543d3cd0c7aee9b3ba3c2be9845719707bb33",
+			"0x134996a104ee5811d51036d776fb46831223e96c254f383d0f906343eb67ad34d6c56711962fa8bfe097e75a2e41c696",
+			"0xcc786baa966e66f4a384c86a3b49942552e2d658a31ce2c344be4b91400da7d26d521628b00523b8dfe240c72de1f6",
+			"0x1f86376e8981c217898751ad8746757d42aa7b90eeb791c09e4a3ec03251cf9de405aba9ec61deca6355c77b0e5f4cb",
+			"0x8cc03fdefe0ff135caf4fe2a21529c4195536fbe3ce50b879833fd221351adc2ee7f8dc099040a841b6daecf2e8fedb",
+			"0x16603fca40634b6a2211e11db8f0a6a074a7d0d4afadb7bd76505c3d3ad5544e203f6326c95a807299b23ab13633a5f0",
+			"0x4ab0b9bcfac1bbcb2c977d027796b3ce75bb8ca2be184cb5231413c4d634f3747a87ac2460f415ec961f8855fe9d6f2",
+			"0x987c8d5333ab86fde9926bd2ca6c674170a05bfe3bdd81ffd038da6c26c842642f64550fedfe935a15e4ca31870fb29",
+			"0x9fc4018bd96684be88c9e221e4da1bb8f3abd16679dc26c1e8b6e6a1f20cabe69d65201c78607a360370e577bdba587",
+			"0xe1bba7a1186bdb5223abde7ada14a23c42a0ca7915af6fe06985e7ed1e4d43b9b3f7055dd4eba6f2bafaaebca731c30",
+			"0x19713e47937cd1be0dfd0b8f1d43fb93cd2fcbcb6caf493fd1183e416389e61031bf3a5cce3fbafce813711ad011c132",
+			"0x18b46a908f36f6deb918c143fed2edcc523559b8aaf0c2462e6bfe7f911f643249d9cdf41b44d606ce07c8a4d0074d8e",
+			"0xb182cac101b9399d155096004f53f447aa7b12a3426b08ec02710e807b4633f06c851c1919211f20d4c04f00b971ef8",
+			"0x245a394ad1eca9b72fc00ae7be315dc757b3b080d4c158013e6632d3c40659cc6cf90ad1c232a6442d9d3f5db980133",
+			"0x5c129645e44cf1102a159f748c4a3fc5e673d81d7e86568d9ab0f5d396a7ce46ba1049b6579afb7866b1e715475224b",
+			"0x15e6be4e990f03ce4ea50b3b42df2eb5cb181d8f84965a3957add4fa95af01b2b665027efec01c7704b456be69c8b604",
+		};
+		const char *ydenStr[] = {
+			"0x16112c4c3a9c98b252181140fad0eae9601a6de578980be6eec3232b5be72e7a07f3688ef60c206d01479253b03663c1",
+			"0x1962d75c2381201e1a0cbd6c43c348b885c84ff731c4d59ca4a10356f453e01f78a4260763529e3532f6102c2e49a03d",
+			"0x58df3306640da276faaae7d6e8eb15778c4855551ae7f310c35a5dd279cd2eca6757cd636f96f891e2538b53dbf67f2",
+			"0x16b7d288798e5395f20d23bf89edb4d1d115c5dbddbcd30e123da489e726af41727364f2c28297ada8d26d98445f5416",
+			"0xbe0e079545f43e4b00cc912f8228ddcc6d19c9f0f69bbb0542eda0fc9dec916a20b15dc0fd2ededda39142311a5001d",
+			"0x8d9e5297186db2d9fb266eaac783182b70152c65550d881c5ecd87b6f0f5a6449f38db9dfa9cce202c6477faaf9b7ac",
+			"0x166007c08a99db2fc3ba8734ace9824b5eecfdfa8d0cf8ef5dd365bc400a0051d5fa9c01a58b1fb93d1a1399126a775c",
+			"0x16a3ef08be3ea7ea03bcddfabba6ff6ee5a4375efa1f4fd7feb34fd206357132b920f5b00801dee460ee415a15812ed9",
+			"0x1866c8ed336c61231a1be54fd1d74cc4f9fb0ce4c6af5920abc5750c4bf39b4852cfe2f7bb9248836b233d9d55535d4a",
+			"0x167a55cda70a6e1cea820597d94a84903216f763e13d87bb5308592e7ea7d4fbc7385ea3d529b35e346ef48bb8913f55",
+			"0x4d2f259eea405bd48f010a01ad2911d9c6dd039bb61a6290e591b36e636a5c871a5c29f4f83060400f8b49cba8f6aa8",
+			"0xaccbb67481d033ff5852c1e48c50c477f94ff8aefce42d28c0f9a88cea7913516f968986f7ebbea9684b529e2561092",
+			"0xad6b9514c767fe3c3613144b45f1496543346d98adf02267d5ceef9a00d9b8693000763e3b90ac11e99b138573345cc",
+			"0x2660400eb2e4f3b628bdd0d53cd76f2bf565b94e72927c1cb748df27942480e420517bd8714cc80d1fadc1326ed06f7",
+			"0xe0fa1d816ddc03e6b24255e0d7819c171c40f65e273b853324efcd6356caa205ca2f570f13497804415473a1d634b8f",
+			"0x1",
+		};
+		initArray(g1xnum, xnumStr, CYBOZU_NUM_OF_ARRAY(xnumStr));
+		initArray(g1xden, xdenStr, CYBOZU_NUM_OF_ARRAY(xdenStr));
+		initArray(g1ynum, ynumStr, CYBOZU_NUM_OF_ARRAY(ynumStr));
+		initArray(g1yden, ydenStr, CYBOZU_NUM_OF_ARRAY(ydenStr));
+	}
+	template<class F, size_t N>
+	void evalPoly(F& y, const F& x, const F *zpows, const F (&cof)[N]) const
+	{
+		y = cof[N - 1]; // always zpows[0] = 1
+		for (size_t i = 1; i < N; i++) {
+			y *= x;
+			F t;
+			F::mul(t, zpows[i - 1], cof[N - 1 - i]);
+			y += t;
+		}
+	}
+	// refer (xnum, xden, ynum, yden)
+	void iso3(G2& Q, const E2& P) const
+	{
+		Fp2 zpows[3];
+		Fp2::sqr(zpows[0], P.z);
+		Fp2::sqr(zpows[1], zpows[0]);
+		Fp2::mul(zpows[2], zpows[1], zpows[0]);
+		Fp2 mapvals[4];
+		evalPoly(mapvals[0], P.x, zpows, xnum);
+		evalPoly(mapvals[1], P.x, zpows, xden);
+		evalPoly(mapvals[2], P.x, zpows, ynum);
+		evalPoly(mapvals[3], P.x, zpows, yden);
+		mapvals[1] *= zpows[0];
+		mapvals[2] *= P.y;
+		mapvals[3] *= zpows[0];
+		mapvals[3] *= P.z;
+		Fp2::mul(Q.z, mapvals[1], mapvals[3]);
+		Fp2::mul(Q.x, mapvals[0], mapvals[3]);
+		Q.x *= Q.z;
+		Fp2 t;
+		Fp2::sqr(t, Q.z);
+		Fp2::mul(Q.y, mapvals[2], mapvals[1]);
+		Q.y *= t;
+	}
+	template<class X, class C, size_t N>
+	X evalPoly2(const X& x, const C (&c)[N]) const
+	{
+		X ret = c[N - 1];
+		for (size_t i = 1; i < N; i++) {
+			ret *= x;
+			ret += c[N - 1 - i];
+		}
+		return ret;
+	}
+	// refer (g1xnum, g1xden, g1ynum, g1yden)
+	void iso11(G1& Q, E1& P) const
+	{
+		ec::normalizeJacobi(P);
+		Fp xn, xd, yn, yd;
+		xn = evalPoly2(P.x, g1xnum);
+		xd = evalPoly2(P.x, g1xden);
+		yn = evalPoly2(P.x, g1ynum);
+		yd = evalPoly2(P.x, g1yden);
+		/*
+			[xn/xd:y * yn/yd:1] = [xn xd yd^2:y yn xd^3 yd^2:xd yd]
+			=[xn yd z:y yn xd z^2:z] where z = xd yd
+		*/
+		Fp::mul(Q.z, xd, yd);
+		Fp::mul(Q.x, xn, yd);
+		Q.x *= Q.z;
+		Fp::mul(Q.y, P.y, yn);
+		Q.y *= xd;
+		Fp::sqr(xd, Q.z);
+		Q.y *= xd;
+	}
+	/*
+		xi = -2-i
+		(a+bi)*(-2-i) = (b-2a)-(a+2b)i
+	*/
+	void mul_xi(Fp2& y, const Fp2& x) const
+	{
+		Fp t;
+		Fp::sub(t, x.b, x.a);
+		t -= x.a;
+		Fp::add(y.b, x.b, x.b);
+		y.b += x.a;
+		Fp::neg(y.b, y.b);
+		y.a = t;
+	}
+	bool isNegSign(const Fp& x) const
+	{
+		return x.isOdd();
+	}
+	bool isNegSign(const Fp2& x) const
+	{
+		bool sign0 = isNegSign(x.a);
+		bool zero0 = x.a.isZero();
+		bool sign1 = isNegSign(x.b);
+		return sign0 || (zero0 & sign1);
+	}
+	// https://tools.ietf.org/html/draft-irtf-cfrg-hash-to-curve-07#appendix-D.3.5
+	void sswuG1(Fp& xn, Fp& xd, Fp& y, const Fp& u) const
+	{
+		const Fp& A = g1A;
+		const Fp& B = g1B;
+		const Fp& c1 = g1c1;
+		const Fp& c2 = g1c2;
+		const int Z = g1Z;
+		Fp u2, u2Z, t, t2, t3;
+
+		Fp::sqr(u2, u);
+		Fp::mulUnit(u2Z, u2, Z);
+		Fp::sqr(t, u2Z);
+		Fp::add(xd, t, u2Z);
+		if (xd.isZero()) {
+			Fp::mulUnit(xd, A, Z);
+			xn = B;
+		} else {
+			Fp::add(xn, xd, Fp::one());
+			xn *= B;
+			xd *= A;
+			Fp::neg(xd, xd);
+		}
+		Fp::sqr(t, xd);
+		Fp::mul(t2, t, xd);
+		t *= A;
+		Fp::sqr(t3, xn);
+		t3 += t;
+		t3 *= xn;
+		Fp::mul(t, t2, B);
+		t3 += t;
+		Fp::sqr(y, t2);
+		Fp::mul(t, t3, t2);
+		y *= t;
+		Fp::pow(y, y, c1);
+		y *= t;
+		Fp::sqr(t, y);
+		t *= t2;
+		if (t != t3) {
+			xn *= u2Z;
+			y *= c2;
+			y *= u2;
+			y *= u;
+		}
+		if (isNegSign(u) != isNegSign(y)) {
+			Fp::neg(y, y);
+		}
+	}
+	void sswuG1(E1& pt, const Fp& u) const
+	{
+		Fp xn, y;
+		Fp& xd = pt.z;
+		sswuG1(xn, xd, y, u);
+		Fp::mul(pt.x, xn, xd);
+		Fp::sqr(pt.y, xd);
+		pt.y *= xd;
+		pt.y *= y;
+	}
+	// https://github.com/algorand/bls_sigs_ref
+	void sswuG2(E2& P, const Fp2& t) const
+	{
+		Fp2 t2, t2xi;
+		Fp2::sqr(t2, t);
+		Fp2 den, den2;
+		mul_xi(t2xi, t2);
+		den = t2xi;
+		Fp2::sqr(den2, den);
+		// (t^2 * xi)^2 + (t^2 * xi)
+		den += den2;
+		Fp2 x0_num, x0_den;
+		Fp2::add(x0_num, den, 1);
+		x0_num *= g2B;
+		if (den.isZero()) {
+			mul_xi(x0_den, g2A);
+		} else {
+			Fp2::mul(x0_den, -g2A, den);
+		}
+		Fp2 x0_den2, x0_den3, gx0_den, gx0_num;
+		Fp2::sqr(x0_den2, x0_den);
+		Fp2::mul(x0_den3, x0_den2, x0_den);
+		gx0_den = x0_den3;
+
+		Fp2::mul(gx0_num, g2B, gx0_den);
+		Fp2 tmp, tmp1, tmp2;
+		Fp2::mul(tmp, g2A, x0_num);
+		tmp *= x0_den2;
+		gx0_num += tmp;
+		Fp2::sqr(tmp, x0_num);
+		tmp *= x0_num;
+		gx0_num += tmp;
+
+		Fp2::sqr(tmp1, gx0_den); // x^2
+		Fp2::sqr(tmp2, tmp1); // x^4
+		tmp1 *= tmp2;
+		tmp1 *= gx0_den; // x^7
+		Fp2::mul(tmp2, gx0_num, tmp1);
+		tmp1 *= tmp2;
+		tmp1 *= gx0_den;
+		Fp2 candi;
+		Fp2::pow(candi, tmp1, sqrtConst);
+		candi *= tmp2;
+		bool isNegT = isNegSign(t);
+		for (size_t i = 0; i < CYBOZU_NUM_OF_ARRAY(root4); i++) {
+			Fp2::mul(P.y, candi, root4[i]);
+			Fp2::sqr(tmp, P.y);
+			tmp *= gx0_den;
+			if (tmp == gx0_num) {
+				if (isNegSign(P.y) != isNegT) {
+					Fp2::neg(P.y, P.y);
+				}
+				Fp2::mul(P.x, x0_num, x0_den);
+				P.y *= x0_den3;
+				P.z = x0_den;
+				return;
+			}
+		}
+		Fp2 x1_num, x1_den, gx1_num, gx1_den;
+		Fp2::mul(x1_num, t2xi, x0_num);
+		x1_den = x0_den;
+		Fp2::mul(gx1_num, den2, t2xi);
+		gx1_num *= gx0_num;
+		gx1_den = gx0_den;
+		candi *= t2;
+		candi *= t;
+		for (size_t i = 0; i < CYBOZU_NUM_OF_ARRAY(etas); i++) {
+			Fp2::mul(P.y, candi, etas[i]);
+			Fp2::sqr(tmp, P.y);
+			tmp *= gx1_den;
+			if (tmp == gx1_num) {
+				if (isNegSign(P.y) != isNegT) {
+					Fp2::neg(P.y, P.y);
+				}
+				Fp2::mul(P.x, x1_num, x1_den);
+				Fp2::sqr(tmp, x1_den);
+				P.y *= tmp;
+				P.y *= x1_den;
+				P.z = x1_den;
+				return;
+			}
+		}
+		assert(0);
+	}
+	template<class T>
+	void put(const T& P) const
+	{
+		const int base = 10;
+		printf("x=%s\n", P.x.getStr(base).c_str());
+		printf("y=%s\n", P.y.getStr(base).c_str());
+		printf("z=%s\n", P.z.getStr(base).c_str());
+	}
+	void Fp2ToG2(G2& P, const Fp2& t, const Fp2 *t2 = 0) const
+	{
+		E2 Pp;
+		sswuG2(Pp, t);
+		if (t2) {
+			E2 P2;
+			sswuG2(P2, *t2);
+			ec::addJacobi(Pp, Pp, P2);
+		}
+		iso3(P, Pp);
+		mcl::local::mulByCofactorBLS12fast(P, P);
+	}
+	void hashToFp2(Fp2 out[2], const void *msg, size_t msgSize, const void *dst, size_t dstSize) const
+	{
+		uint8_t md[256];
+		mcl::fp::expand_message_xmd(md, sizeof(md), msg, msgSize, dst, dstSize);
+		Fp *x = out[0].getFp0();
+		for (size_t i = 0; i < 4; i++) {
+			bool b;
+			x[i].setBigEndianMod(&b, &md[64 * i], 64);
+			assert(b); (void)b;
+		}
+	}
+	void msgToG2(G2& out, const void *msg, size_t msgSize, const void *dst, size_t dstSize) const
+	{
+		Fp2 t[2];
+		hashToFp2(t, msg, msgSize, dst, dstSize);
+		Fp2ToG2(out, t[0], &t[1]);
+	}
+	void msgToG2(G2& out, const void *msg, size_t msgSize) const
+	{
+		const char *dst = "BLS_SIG_BLS12381G2_XMD:SHA-256_SSWU_RO_POP_";
+		const size_t dstSize = strlen(dst);
+		msgToG2(out, msg, msgSize, dst, dstSize);
+	}
+	void FpToG1(G1& out, const Fp& u0, const Fp *u1 = 0) const
+	{
+		E1 P1;
+		sswuG1(P1, u0);
+		if (u1) {
+			E1 P2;
+			sswuG1(P2, *u1);
+			ec::addJacobi(P1, P1, P2);
+		}
+		iso11(out, P1);
+		G1::mulGeneric(out, out, g1cofactor);
+	}
+	void msgToG1(G1& out, const void *msg, size_t msgSize, const char *dst, size_t dstSize) const
+	{
+		uint8_t md[128];
+		mcl::fp::expand_message_xmd(md, sizeof(md), msg, msgSize, dst, dstSize);
+		Fp u[2];
+		for (size_t i = 0; i < 2; i++) {
+			bool b;
+			u[i].setBigEndianMod(&b, &md[64 * i], 64);
+			assert(b); (void)b;
+		}
+		FpToG1(out, u[0], &u[1]);
+	}
+
+	void msgToG1(G1& out, const void *msg, size_t msgSize) const
+	{
+		const char *dst = "BLS_SIG_BLS12381G1_XMD:SHA-256_SSWU_RO_POP_";
+		const size_t dstSize = strlen(dst);
+		msgToG1(out, msg, msgSize, dst, dstSize);
+	}
+};
+
+} // mcl
+

include/mcl/op.hpp

@@ -14,8 +14,11 @@
 	#define MCL_DONT_USE_XBYAK
 	#define MCL_DONT_USE_OPENSSL
 #endif
-#if !defined(MCL_DONT_USE_XBYAK) && (defined(_WIN64) || defined(__x86_64__)) && (MCL_SIZEOF_UNIT == 8)
+#if !defined(MCL_DONT_USE_XBYAK) && (defined(_WIN64) || defined(__x86_64__)) && (MCL_SIZEOF_UNIT == 8) && !defined(MCL_STATIC_CODE)
 	#define MCL_USE_XBYAK
+#endif
+#if defined(MCL_USE_XBYAK) || defined(MCL_STATIC_CODE)
+	#define MCL_X64_ASM
 	#define MCL_XBYAK_DIRECT_CALL
 #endif
 
@@ -23,7 +26,7 @@
 
 namespace mcl {
 
-static const int version = 0x094; /* 0xABC = A.BC */
+static const int version = 0x151; /* 0xABC = A.BC */
 
 /*
 	specifies available string format mode for X::setIoMode()
@@ -101,16 +104,29 @@ enum IoMode {
 	IoSerialize = 512, // use MBS for 1-bit y
 	IoFixedSizeByteSeq = IoSerialize, // obsolete
 	IoEcProj = 1024, // projective or jacobi coordinate
-	IoSerializeHexStr = 2048 // printable hex string
+	IoSerializeHexStr = 2048, // printable hex string
+	IoEcAffineSerialize = 4096 // serialize [x:y]
 };
 
 namespace fp {
 
+inline bool isIoSerializeMode(int ioMode)
+{
+	return ioMode & (IoArray | IoArrayRaw | IoSerialize | IoEcAffineSerialize | IoSerializeHexStr);
+}
+
 const size_t UnitBitSize = sizeof(Unit) * 8;
 
 const size_t maxUnitSize = (MCL_MAX_BIT_SIZE + UnitBitSize - 1) / UnitBitSize;
 #define MCL_MAX_UNIT_SIZE ((MCL_MAX_BIT_SIZE + MCL_UNIT_BIT_SIZE - 1) / MCL_UNIT_BIT_SIZE)
 
+const size_t maxMulVecN = 32; // inner loop of mulVec
+
+#ifndef MCL_MAX_MUL_VEC_NGLV
+	#define MCL_MAX_MUL_VEC_NGLV 16
+#endif
+const size_t maxMulVecNGLV = MCL_MAX_MUL_VEC_NGLV; // inner loop of mulVec with GLV
+
 struct FpGenerator;
 struct Op;
 
@@ -157,13 +173,6 @@ enum PrimeMode {
 	PM_NIST_P521
 };
 
-enum MaskMode {
-	NoMask = 0, // throw if greater or equal
-	SmallMask = 1, // 1-bit smaller mask if greater or equal
-	MaskAndMod = 2, // mask and substract if greater or equal
-	Mod = 3 // mod p
-};
-
 struct Op {
 	/*
 		don't change the layout of rp and p
@@ -175,6 +184,7 @@ struct Op {
 	uint32_t pmod4;
 	mcl::SquareRoot sq;
 	mcl::Modp modp;
+	mcl::SmallModp smallModp;
 	Unit half[maxUnitSize]; // (p + 1) / 2
 	Unit oneRep[maxUnitSize]; // 1(=inv R if Montgomery)
 	/*
@@ -189,6 +199,8 @@ struct Op {
 	Unit R3[maxUnitSize];
 #ifdef MCL_USE_XBYAK
 	FpGenerator *fg;
+#endif
+#ifdef MCL_X64_ASM
 	mcl::Array<Unit> invTbl;
 #endif
 	void3u fp_addA_;
@@ -196,18 +208,20 @@ struct Op {
 	void2u fp_negA_;
 	void3u fp_mulA_;
 	void2u fp_sqrA_;
+	void2u fp_mul2A_;
+	void2u fp_mul9A_;
 	void3u fp2_addA_;
 	void3u fp2_subA_;
 	void2u fp2_negA_;
 	void3u fp2_mulA_;
 	void2u fp2_sqrA_;
+	void2u fp2_mul2A_;
 	void3u fpDbl_addA_;
 	void3u fpDbl_subA_;
-	void3u fpDbl_mulPreA_;
-	void2u fpDbl_sqrPreA_;
 	void2u fpDbl_modA_;
 	void3u fp2Dbl_mulPreA_;
 	void2u fp2Dbl_sqrPreA_;
+	void2u fp2Dbl_mul_xiA_;
 	size_t maxN;
 	size_t N;
 	size_t bitSize;
@@ -220,6 +234,7 @@ struct Op {
 	void4u fp_sub;
 	void4u fp_mul;
 	void3u fp_sqr;
+	void3u fp_mul2;
 	void2uOp fp_invOp;
 	void2uIu fp_mulUnit; // fpN1_mod + fp_mulUnitPre
 
@@ -243,7 +258,6 @@ struct Op {
 	*/
 	int xi_a; // xi = xi_a + u
 	void4u fp2_mulNF;
-	void2u fp2_inv;
 	void2u fp2_mul_xiA_;
 	uint32_t (*hash)(void *out, uint32_t maxOutSize, const void *msg, uint32_t msgSize);
 
@@ -275,7 +289,7 @@ struct Op {
 		memset(one, 0, sizeof(one));
 		memset(R2, 0, sizeof(R2));
 		memset(R3, 0, sizeof(R3));
-#ifdef MCL_USE_XBYAK
+#ifdef MCL_X64_ASM
 		invTbl.clear();
 #endif
 		fp_addA_ = 0;
@@ -283,18 +297,20 @@ struct Op {
 		fp_negA_ = 0;
 		fp_mulA_ = 0;
 		fp_sqrA_ = 0;
+		fp_mul2A_ = 0;
+		fp_mul9A_ = 0;
 		fp2_addA_ = 0;
 		fp2_subA_ = 0;
 		fp2_negA_ = 0;
 		fp2_mulA_ = 0;
 		fp2_sqrA_ = 0;
+		fp2_mul2A_ = 0;
 		fpDbl_addA_ = 0;
 		fpDbl_subA_ = 0;
-		fpDbl_mulPreA_ = 0;
-		fpDbl_sqrPreA_ = 0;
 		fpDbl_modA_ = 0;
 		fp2Dbl_mulPreA_ = 0;
 		fp2Dbl_sqrPreA_ = 0;
+		fp2Dbl_mul_xiA_ = 0;
 		maxN = 0;
 		N = 0;
 		bitSize = 0;
@@ -307,6 +323,7 @@ struct Op {
 		fp_sub = 0;
 		fp_mul = 0;
 		fp_sqr = 0;
+		fp_mul2 = 0;
 		fp_invOp = 0;
 		fp_mulUnit = 0;
 
@@ -327,7 +344,6 @@ struct Op {
 
 		xi_a = 0;
 		fp2_mulNF = 0;
-		fp2_inv = 0;
 		fp2_mul_xiA_ = 0;
 		hash = 0;
 
@@ -363,13 +379,14 @@ private:
 
 inline const char* getIoSeparator(int ioMode)
 {
-	return (ioMode & (IoArray | IoArrayRaw | IoSerialize | IoSerializeHexStr)) ? "" : " ";
+	return (ioMode & (IoArray | IoArrayRaw | IoSerialize | IoSerializeHexStr | IoEcAffineSerialize)) ? "" : " ";
 }
 
-inline void dump(const char *s, size_t n)
+inline void dump(const void *buf, size_t n)
 {
+	const uint8_t *s = (const uint8_t *)buf;
 	for (size_t i = 0; i < n; i++) {
-		printf("%02x ", (uint8_t)s[i]);
+		printf("%02x ", s[i]);
 	}
 	printf("\n");
 }

include/mcl/operator.hpp

@@ -84,12 +84,39 @@ struct Operator : public E {
 	{
 		powArray(z, x, gmp::getUnit(y), gmp::getUnitSize(y), y < 0, true);
 	}
-	static void setPowArrayGLV(void f(T& z, const T& x, const Unit *y, size_t yn, bool isNegative, bool constTime))
+	static void setPowArrayGLV(void f(T& z, const T& x, const Unit *y, size_t yn, bool isNegative, bool constTime), size_t g(T& z, const T *xVec, const mpz_class *yVec, size_t n) = 0)
 	{
 		powArrayGLV = f;
+		powVecNGLV = g;
+	}
+	static const size_t powVecMaxN = 16;
+	template<class tag, size_t maxBitSize, template<class _tag, size_t _maxBitSize>class FpT>
+	static void powVec(T& z, const T* xVec, const FpT<tag, maxBitSize> *yVec, size_t n)
+	{
+		assert(powVecNGLV);
+		T r;
+		r.setOne();
+		const size_t N = mcl::fp::maxMulVecNGLV;
+		mpz_class myVec[N];
+		while (n > 0) {
+			T t;
+			size_t tn = fp::min_(n, N);
+			for (size_t i = 0; i < tn; i++) {
+				bool b;
+				yVec[i].getMpz(&b, myVec[i]);
+				assert(b); (void)b;
+			}
+			size_t done = powVecNGLV(t, xVec, myVec, tn);
+			r *= t;
+			xVec += done;
+			yVec += done;
+			n -= done;
+		}
+		z = r;
 	}
 private:
 	static void (*powArrayGLV)(T& z, const T& x, const Unit *y, size_t yn, bool isNegative, bool constTime);
+	static size_t (*powVecNGLV)(T& z, const T* xVec, const mpz_class *yVec, size_t n);
 	static void powArray(T& z, const T& x, const Unit *y, size_t yn, bool isNegative, bool constTime)
 	{
 		if (powArrayGLV && (constTime || yn > 1)) {
@@ -117,6 +144,9 @@ private:
 template<class T, class E>
 void (*Operator<T, E>::powArrayGLV)(T& z, const T& x, const Unit *y, size_t yn, bool isNegative, bool constTime);
 
+template<class T, class E>
+size_t (*Operator<T, E>::powVecNGLV)(T& z, const T* xVec, const mpz_class *yVec, size_t n);
+
 /*
 	T must have save and load
 */
@@ -154,6 +184,20 @@ struct Serializable : public E {
 		getStr(str, ioMode);
 		return str;
 	}
+	std::string serializeToHexStr() const
+	{
+		std::string str(sizeof(T) * 2, 0);
+		size_t n = serialize(&str[0], str.size(), IoSerializeHexStr);
+		str.resize(n);
+		return str;
+	}
+#ifndef CYBOZU_DONT_USE_EXCEPTION
+	void deserializeHexStr(const std::string& str)
+	{
+		size_t n = deserialize(str.c_str(), str.size(), IoSerializeHexStr);
+		if (n == 0) throw cybozu::Exception("bad str") << str;
+	}
+#endif
 #endif
 	// return written bytes
 	size_t serialize(void *buf, size_t maxBufSize, int ioMode = IoSerialize) const

include/mcl/she.h

@@ -74,6 +74,18 @@ typedef struct {
 typedef struct {
 	mclBnFr d[7];
 } sheZkpBinEq;
+
+typedef struct {
+	mclBnFr d[2];
+} sheZkpDec;
+
+typedef struct {
+	mclBnGT d[4];
+} sheAuxiliaryForZkpDecGT;
+
+typedef struct {
+	mclBnFr d[4];
+} sheZkpDecGT;
 /*
 	initialize this library
 	call this once before using the other functions
@@ -96,6 +108,8 @@ MCLSHE_DLL_API mclSize sheCipherTextGTSerialize(void *buf, mclSize maxBufSize, c
 MCLSHE_DLL_API mclSize sheZkpBinSerialize(void *buf, mclSize maxBufSize, const sheZkpBin *zkp);
 MCLSHE_DLL_API mclSize sheZkpEqSerialize(void *buf, mclSize maxBufSize, const sheZkpEq *zkp);
 MCLSHE_DLL_API mclSize sheZkpBinEqSerialize(void *buf, mclSize maxBufSize, const sheZkpBinEq *zkp);
+MCLSHE_DLL_API mclSize sheZkpDecSerialize(void *buf, mclSize maxBufSize, const sheZkpDec *zkp);
+MCLSHE_DLL_API mclSize sheZkpDecGTSerialize(void *buf, mclSize maxBufSize, const sheZkpDecGT *zkp);
 
 // return read byte size if sucess else 0
 MCLSHE_DLL_API mclSize sheSecretKeyDeserialize(sheSecretKey* sec, const void *buf, mclSize bufSize);
@@ -106,6 +120,8 @@ MCLSHE_DLL_API mclSize sheCipherTextGTDeserialize(sheCipherTextGT* c, const void
 MCLSHE_DLL_API mclSize sheZkpBinDeserialize(sheZkpBin* zkp, const void *buf, mclSize bufSize);
 MCLSHE_DLL_API mclSize sheZkpEqDeserialize(sheZkpEq* zkp, const void *buf, mclSize bufSize);
 MCLSHE_DLL_API mclSize sheZkpBinEqDeserialize(sheZkpBinEq* zkp, const void *buf, mclSize bufSize);
+MCLSHE_DLL_API mclSize sheZkpDecDeserialize(sheZkpDec* zkp, const void *buf, mclSize bufSize);
+MCLSHE_DLL_API mclSize sheZkpDecGTDeserialize(sheZkpDecGT* zkp, const void *buf, mclSize bufSize);
 
 /*
 	set secretKey if system has /dev/urandom or CryptGenRandom
@@ -115,6 +131,8 @@ MCLSHE_DLL_API int sheSecretKeySetByCSPRNG(sheSecretKey *sec);
 
 MCLSHE_DLL_API void sheGetPublicKey(shePublicKey *pub, const sheSecretKey *sec);
 
+MCLSHE_DLL_API void sheGetAuxiliaryForZkpDecGT(sheAuxiliaryForZkpDecGT *aux, const shePublicKey *pub);
+
 /*
 	make table to decode DLP
 	return 0 if success
@@ -192,6 +210,13 @@ MCLSHE_DLL_API int shePrecomputedPublicKeyEncWithZkpBinEq(sheCipherTextG1 *c1, s
 MCLSHE_DLL_API int sheEncWithZkpEq(sheCipherTextG1 *c1, sheCipherTextG2 *c2, sheZkpEq *zkp, const shePublicKey *pub, mclInt m);
 MCLSHE_DLL_API int shePrecomputedPublicKeyEncWithZkpEq(sheCipherTextG1 *c1, sheCipherTextG2 *c2, sheZkpEq *zkp, const shePrecomputedPublicKey *ppub, mclInt m);
 
+/*
+	Zkp s.t. Dec(c) = m
+	return 0 if success
+*/
+MCLSHE_DLL_API int sheDecWithZkpDecG1(mclInt *m, sheZkpDec *zkp, const sheSecretKey *sec, const sheCipherTextG1 *c, const shePublicKey *pub);
+MCLSHE_DLL_API int sheDecWithZkpDecGT(mclInt *m, sheZkpDecGT *zkp, const sheSecretKey *sec, const sheCipherTextGT *c, const sheAuxiliaryForZkpDecGT *aux);
+
 /*
 	decode c and set m
 	return 0 if success
@@ -211,6 +236,8 @@ MCLSHE_DLL_API int shePrecomputedPublicKeyVerifyZkpBinG1(const shePrecomputedPub
 MCLSHE_DLL_API int shePrecomputedPublicKeyVerifyZkpBinG2(const shePrecomputedPublicKey *ppub, const sheCipherTextG2 *c, const sheZkpBin *zkp);
 MCLSHE_DLL_API int shePrecomputedPublicKeyVerifyZkpEq(const shePrecomputedPublicKey *ppub, const sheCipherTextG1 *c1, const sheCipherTextG2 *c2, const sheZkpEq *zkp);
 MCLSHE_DLL_API int shePrecomputedPublicKeyVerifyZkpBinEq(const shePrecomputedPublicKey *ppub, const sheCipherTextG1 *c1, const sheCipherTextG2 *c2, const sheZkpBinEq *zkp);
+MCLSHE_DLL_API int sheVerifyZkpDecG1(const shePublicKey *pub, const sheCipherTextG1 *c1, mclInt m, const sheZkpDec *zkp);
+MCLSHE_DLL_API int sheVerifyZkpDecGT(const sheAuxiliaryForZkpDecGT *aux, const sheCipherTextGT *ct, mclInt m, const sheZkpDecGT *zkp);
 /*
 	decode c via GT and set m
 	return 0 if success

include/mcl/she.hpp

@@ -26,6 +26,8 @@
 #include <mcl/window_method.hpp>
 #include <cybozu/endian.hpp>
 #include <cybozu/serializer.hpp>
+#include <cybozu/sha2.hpp>
+#include <mcl/ecparam.hpp>
 
 namespace mcl { namespace she {
 
@@ -324,6 +326,26 @@ int log(const G& P, const G& xP)
 	throw cybozu::Exception("she:log:not found");
 }
 
+struct Hash {
+	cybozu::Sha256 h_;
+	template<class T>
+	Hash& operator<<(const T& t)
+	{
+		char buf[sizeof(T)];
+		cybozu::MemoryOutputStream os(buf, sizeof(buf));
+		t.save(os);
+		h_.update(buf, os.getPos());
+		return *this;
+	}
+	template<class F>
+	void get(F& x)
+	{
+		uint8_t md[32];
+		h_.digest(md, sizeof(md), 0, 0);
+		x.setArrayMask(md, sizeof(md));
+	}
+};
+
 } // mcl::she::local
 
 template<size_t dummyInpl = 0>
@@ -366,11 +388,17 @@ private:
 	public:
 		const G& getS() const { return S_; }
 		const G& getT() const { return T_; }
+		G& getNonConstRefS() { return S_; }
+		G& getNonConstRefT() { return T_; }
 		void clear()
 		{
 			S_.clear();
 			T_.clear();
 		}
+		bool isValid() const
+		{
+			return S_.isValid() && T_.isValid();
+		}
 		static void add(CipherTextAT& z, const CipherTextAT& x, const CipherTextAT& y)
 		{
 			/*
@@ -537,6 +565,8 @@ private:
 	struct ZkpBinTag;
 	struct ZkpEqTag; // d_[] = { c, sp, ss, sm }
 	struct ZkpBinEqTag; // d_[] = { d0, d1, sp0, sp1, ss, sp, sm }
+	struct ZkpDecTag; // d_[] = { c, h }
+	struct ZkpDecGTTag; // d_[] = { d1, d2, d3, h }
 public:
 	/*
 		Zkp for m = 0 or 1
@@ -550,9 +580,66 @@ public:
 		Zkp for (m = 0 or 1) and decG1(c1) == decG2(c2)
 	*/
 	typedef ZkpT<ZkpBinEqTag, 7> ZkpBinEq;
+	/*
+		Zkp for Dec(c) = m for c in G1
+	*/
+	typedef ZkpT<ZkpDecTag, 2> ZkpDec;
+	/*
+		Zkp for Dec(c) = m for c in GT
+	*/
+	typedef ZkpT<ZkpDecGTTag, 4> ZkpDecGT;
 
 	typedef CipherTextAT<G1> CipherTextG1;
 	typedef CipherTextAT<G2> CipherTextG2;
+	/*
+		auxiliary for ZkpDecGT
+		@note GT is multiplicative group though treating GT as additive group in comment
+	*/
+	struct AuxiliaryForZkpDecGT {
+		GT R_[4]; // [R = e(R, Q), xR, yR, xyR]
+
+		// dst = v[1] a[0] + v[0] a[1] - v[2] a[2]
+		void f(GT& dst, const GT *v, const Fr *a) const
+		{
+			GT t;
+			GT::pow(dst, v[0], a[1]);
+			GT::pow(t, v[1], a[0]);
+			dst *= t;
+			GT::pow(t, v[2], a[2]);
+			GT::unitaryInv(t, t);
+			dst *= t;
+		}
+		bool verify(const CipherTextGT& c, int64_t m, const ZkpDecGT& zkp) const
+		{
+			const Fr *d = &zkp.d_[0];
+			const Fr &h = zkp.d_[3];
+
+			GT A[4];
+			GT t;
+			GT::pow(t, R_[0], m); // m R
+			GT::unitaryInv(t, t);
+			GT::mul(A[0], c.g_[0], t);
+			A[1] = c.g_[1];
+			A[2] = c.g_[2];
+			A[3] = c.g_[3];
+			GT B[3], X;
+			for (int i = 0; i < 3; i++) {
+				GT::pow(B[i], R_[0], d[i]);
+				GT::pow(t, R_[i+1], h);
+				GT::unitaryInv(t, t);
+				B[i] *= t;
+			}
+			f(X, A + 1, zkp.d_);
+			GT::pow(t, A[0], h);
+			GT::unitaryInv(t, t);
+			X *= t;
+			local::Hash hash;
+			hash << R_[1] << R_[2] << R_[3] << A[0] << A[1] << A[2] << A[3] << B[0] << B[1] << B[2] << X;
+			Fr h2;
+			hash.get(h2);
+			return h == h2;
+		}
+	};
 
 	static void init(const mcl::CurveParam& cp = mcl::BN254, size_t hashSize = 1024, size_t tryNum = local::defaultTryNum)
 	{
@@ -576,13 +663,7 @@ public:
 	*/
 	static void initG1only(const mcl::EcParam& para, size_t hashSize = 1024, size_t tryNum = local::defaultTryNum)
 	{
-		Fp::init(para.p);
+		mcl::initCurve<G1, Fr>(para.curveType, &P_);
-		Fr::init(para.n);
-		G1::init(para.a, para.b);
-		const Fp x0(para.gx);
-		const Fp y0(para.gy);
-		P_.set(x0, y0);
-
 		setRangeForG1DLP(hashSize);
 		useDecG1ViaGT_ = false;
 		useDecG2ViaGT_ = false;
@@ -774,6 +855,88 @@ public:
 				return isZero(c.a_);
 			}
 		}
+		int64_t decWithZkpDec(bool *pok, ZkpDec& zkp, const CipherTextG1& c, const PublicKey& pub) const
+		{
+			/*
+				c = (S, T)
+				S = mP + rxP
+				T = rP
+				R = S - xT = mP
+			*/
+			G1 R;
+			G1::mul(R, c.T_, x_);
+			G1::sub(R, c.S_, R);
+			int64_t m = PhashTbl_.log(R, pok);
+			if (!*pok) return 0;
+			const G1& P1 = P_;
+			const G1& P2 = c.T_; // rP
+			const G1& A1 = pub.xP_;
+			G1 A2;
+			G1::sub(A2, c.S_, R); // rxP
+			Fr b;
+			b.setRand();
+			G1 B1, B2;
+			G1::mul(B1, P1, b);
+			G1::mul(B2, P2, b);
+			Fr& d = zkp.d_[0];
+			Fr& h = zkp.d_[1];
+			local::Hash hash;
+			hash << P2 << A1 << A2 << B1 << B2;
+			hash.get(h);
+			Fr::mul(d, h, x_);
+			d += b;
+			return m;
+		}
+		// @note GT is multiplicative group though treating GT as additive group in comment
+		int64_t decWithZkpDec(bool *pok, ZkpDecGT& zkp, const CipherTextGT& c, const AuxiliaryForZkpDecGT& aux) const
+		{
+			int64_t m = dec(c, pok);
+			if (!*pok) return 0;
+			// A = c - Enc(m; 0, 0, 0) = c - (m R, 0, 0, 0)
+			GT A[4];
+			GT t;
+			GT::pow(t, aux.R_[0], m); // m R
+			GT::unitaryInv(t, t);
+			GT::mul(A[0], c.g_[0], t);
+			A[1] = c.g_[1];
+			A[2] = c.g_[2];
+			A[3] = c.g_[3];
+			// dec(A) = 0
+
+			Fr b[3];
+			GT B[3], X;
+			for (int i = 0; i < 3; i++) {
+				b[i].setByCSPRNG();
+				GT::pow(B[i], aux.R_[0], b[i]);
+			}
+			aux.f(X, A + 1, b);
+			local::Hash hash;
+			hash << aux.R_[1] << aux.R_[2] << aux.R_[3] << A[0] << A[1] << A[2] << A[3] << B[0] << B[1] << B[2] << X;
+			Fr *d = &zkp.d_[0];
+			Fr &h = zkp.d_[3];
+			hash.get(h);
+			Fr::mul(d[0], h, x_); // h x
+			Fr::mul(d[1], h, y_); // h y
+			Fr::mul(d[2], d[1], x_); // h xy
+			for (int i = 0; i < 3; i++) {
+				d[i] += b[i];
+			}
+			return m;
+		}
+		int64_t decWithZkpDec(ZkpDec& zkp, const CipherTextG1& c, const PublicKey& pub) const
+		{
+			bool b;
+			int64_t ret = decWithZkpDec(&b, zkp, c, pub);
+			if (!b) throw cybozu::Exception("she:SecretKey:decWithZkpDec");
+			return ret;
+		}
+		int64_t decWithZkpDec(ZkpDecGT& zkp, const CipherTextGT& c, const AuxiliaryForZkpDecGT& aux) const
+		{
+			bool b;
+			int64_t ret = decWithZkpDec(&b, zkp, c, aux);
+			if (!b) throw cybozu::Exception("she:SecretKey:decWithZkpDec");
+			return ret;
+		}
 		template<class InputStream>
 		void load(bool *pb, InputStream& is, int ioMode = IoSerialize)
 		{
@@ -886,16 +1049,10 @@ private:
 		r.setRand();
 		Pmul.mul(static_cast<I&>(R[0][m]), r); // R[0][m] = r P
 		xPmul.mul(R[1][m], r); // R[1][m] = r xP
-		char buf[sizeof(G) * 2];
-		cybozu::MemoryOutputStream os(buf, sizeof(buf));
-		S.save(os);
-		T.save(os);
-		R[0][0].save(os);
-		R[0][1].save(os);
-		R[1][0].save(os);
-		R[1][1].save(os);
 		Fr c;
-		c.setHashOf(buf, os.getPos());
+		local::Hash hash;
+		hash << S << T << R[0][0] << R[0][1] << R[1][0] << R[1][1];
+		hash.get(c);
 		d[m] = c - d[1-m];
 		s[m] = r + d[m] * encRand;
 	}
@@ -925,16 +1082,10 @@ private:
 		G::sub(T2, S, P);
 		G::mul(T2, T2, d[1]);
 		G::sub(R[1][1], T1, T2);
-		char buf[sizeof(G) * 2];
-		cybozu::MemoryOutputStream os(buf, sizeof(buf));
-		S.save(os);
-		T.save(os);
-		R[0][0].save(os);
-		R[0][1].save(os);
-		R[1][0].save(os);
-		R[1][1].save(os);
 		Fr c;
-		c.setHashOf(buf, os.getPos());
+		local::Hash hash;
+		hash << S << T << R[0][0] << R[0][1] << R[1][0] << R[1][1];
+		hash.get(c);
 		return c == d[0] + d[1];
 	}
 	/*
@@ -956,21 +1107,13 @@ private:
 		G2 R3, R4;
 		ElGamalEnc(R1, R2, rm, Pmul, xPmul, &rp);
 		ElGamalEnc(R3, R4, rm, Qmul, yQmul, &rs);
-		char buf[sizeof(G1) * 4 + sizeof(G2) * 4];
-		cybozu::MemoryOutputStream os(buf, sizeof(buf));
-		S1.save(os);
-		T1.save(os);
-		S2.save(os);
-		T2.save(os);
-		R1.save(os);
-		R2.save(os);
-		R3.save(os);
-		R4.save(os);
 		Fr& c = zkp.d_[0];
 		Fr& sp = zkp.d_[1];
 		Fr& ss = zkp.d_[2];
 		Fr& sm = zkp.d_[3];
-		c.setHashOf(buf, os.getPos());
+		local::Hash hash;
+		hash << S1 << T1 << S2 << T2 << R1 << R2 << R3 << R4;
+		hash.get(c);
 		Fr::mul(sp, c, p);
 		sp += rp;
 		Fr::mul(ss, c, s);
@@ -997,18 +1140,10 @@ private:
 		R3 -= X2;
 		G2::mul(X2, T2, c);
 		R4 -= X2;
-		char buf[sizeof(G1) * 4 + sizeof(G2) * 4];
-		cybozu::MemoryOutputStream os(buf, sizeof(buf));
-		S1.save(os);
-		T1.save(os);
-		S2.save(os);
-		T2.save(os);
-		R1.save(os);
-		R2.save(os);
-		R3.save(os);
-		R4.save(os);
 		Fr c2;
-		c2.setHashOf(buf, os.getPos());
+		local::Hash hash;
+		hash << S1 << T1 << S2 << T2 << R1 << R2 << R3 << R4;
+		hash.get(c2);
 		return c == c2;
 	}
 	/*
@@ -1052,20 +1187,10 @@ private:
 		G2 R5, R6;
 		ElGamalEnc(R4, R3, rm, Pmul, xPmul, &rp);
 		ElGamalEnc(R6, R5, rm, Qmul, yQmul, &rs);
-		char buf[sizeof(Fp) * 12];
-		cybozu::MemoryOutputStream os(buf, sizeof(buf));
-		S1.save(os);
-		T1.save(os);
-		R1[0].save(os);
-		R1[1].save(os);
-		R2[0].save(os);
-		R2[1].save(os);
-		R3.save(os);
-		R4.save(os);
-		R5.save(os);
-		R6.save(os);
 		Fr c;
-		c.setHashOf(buf, os.getPos());
+		local::Hash hash;
+		hash << S1 << T1 << R1[0] << R1[1] << R2[0] << R2[1] << R3 << R4 << R5 << R6;
+		hash.get(c);
 		Fr::sub(d[m], c, d[1-m]);
 		Fr::mul(spm[m], d[m], p);
 		spm[m] += rpm;
@@ -1112,20 +1237,10 @@ private:
 		R5 -= X2;
 		G2::mul(X2, S2, c);
 		R6 -= X2;
-		char buf[sizeof(Fp) * 12];
-		cybozu::MemoryOutputStream os(buf, sizeof(buf));
-		S1.save(os);
-		T1.save(os);
-		R1[0].save(os);
-		R1[1].save(os);
-		R2[0].save(os);
-		R2[1].save(os);
-		R3.save(os);
-		R4.save(os);
-		R5.save(os);
-		R6.save(os);
 		Fr c2;
-		c2.setHashOf(buf, os.getPos());
+		local::Hash hash;
+		hash << S1 << T1 << R1[0] << R1[1] << R2[0] << R2[1] << R3 << R4 << R5 << R6;
+		hash.get(c2);
 		return c == c2;
 	}
 	/*
@@ -1266,6 +1381,13 @@ public:
 			ElGamalEnc(c.S_, c.T_, m, QhashTbl_.getWM(), yQmul);
 		}
 public:
+		void getAuxiliaryForZkpDecGT(AuxiliaryForZkpDecGT& aux) const
+		{
+			aux.R_[0] = ePQ_;
+			pairing(aux.R_[1], xP_, Q_);
+			pairing(aux.R_[2], P_, yQ_);
+			pairing(aux.R_[3], xP_, yQ_);
+		}
 		void encWithZkpBin(CipherTextG1& c, ZkpBin& zkp, int m) const
 		{
 			Fr encRand;
@@ -1287,6 +1409,33 @@ public:
 			const MulG<G1> xPmul(xP_);
 			return verifyZkpBin(c.S_, c.T_, P_, zkp, PhashTbl_.getWM(), xPmul);
 		}
+		bool verify(const CipherTextG1& c, int64_t m, const ZkpDec& zkp) const
+		{
+			/*
+				Enc(m;r) - Enc(m;0) = (S, T) - (mP, 0) = (S - mP, T)
+			*/
+			const Fr& d = zkp.d_[0];
+			const Fr& h = zkp.d_[1];
+			const G1& P1 = P_;
+			const G1& P2 = c.T_; // rP
+			const G1& A1 = xP_;
+			G1 A2;
+			G1::mul(A2, P_, m);
+//			PhashTbl_.getWM().mul(A2, m);
+			G1::sub(A2, c.S_, A2); // S - mP = xrP
+			G1 B1, B2, T;
+			G1::mul(B1, P1, d);
+			G1::mul(B2, P2, d);
+			G1::mul(T, A1, h);
+			B1 -= T;
+			G1::mul(T, A2, h);
+			B2 -= T;
+			Fr h2;
+			local::Hash hash;
+			hash << P2 << A1 << A2 << B1 << B2;
+			hash.get(h2);
+			return h == h2;
+		}
 		bool verify(const CipherTextG2& c, const ZkpBin& zkp) const
 		{
 			const MulG<G2> yQmul(yQ_);
@@ -1599,6 +1748,7 @@ public:
 		friend class PublicKey;
 		friend class PrecomputedPublicKey;
 		friend class CipherTextA;
+		friend struct AuxiliaryForZkpDecGT;
 		template<class T>
 		friend struct PublicKeyMethod;
 	public:
@@ -1809,7 +1959,14 @@ public:
 		template<class InputStream>
 		void load(bool *pb, InputStream& is, int ioMode = IoSerialize)
 		{
-			cybozu::writeChar(pb, isMultiplied_ ? '0' : '1', is); if (!*pb) return;
+			char c;
+			if (!cybozu::readChar(&c, is)) return;
+			if (c == '0' || c == '1') {
+				isMultiplied_ = c == '0';
+			} else {
+				*pb = false;
+				return;
+			}
 			if (isMultiplied()) {
 				m_.load(pb, is, ioMode);
 			} else {
@@ -1819,14 +1976,7 @@ public:
 		template<class OutputStream>
 		void save(bool *pb, OutputStream& os, int ioMode = IoSerialize) const
 		{
-			char c;
+			cybozu::writeChar(pb, os, isMultiplied_ ? '0' : '1'); if (!*pb) return;
-			if (!cybozu::readChar(&c, os)) return;
-			if (c == '0' || c == '1') {
-				isMultiplied_ = c == '0';
-			} else {
-				*pb = false;
-				return;
-			}
 			if (isMultiplied()) {
 				m_.save(pb, os, ioMode);
 			} else {
@@ -1895,6 +2045,9 @@ typedef SHE::CipherText CipherText;
 typedef SHE::ZkpBin ZkpBin;
 typedef SHE::ZkpEq ZkpEq;
 typedef SHE::ZkpBinEq ZkpBinEq;
+typedef SHE::ZkpDec ZkpDec;
+typedef SHE::AuxiliaryForZkpDecGT AuxiliaryForZkpDecGT;
+typedef SHE::ZkpDecGT ZkpDecGT;
 
 inline void init(const mcl::CurveParam& cp = mcl::BN254, size_t hashSize = 1024, size_t tryNum = local::defaultTryNum)
 {

include/mcl/util.hpp

@@ -17,8 +17,21 @@
 namespace mcl { namespace fp {
 
 // some environments do not have utility
-template<class T>
+inline uint32_t abs_(int32_t x)
-T abs_(T x) { return x < 0 ? -x : x; }
+{
+	if (x >= 0) return uint32_t(x);
+	// avoid undefined behavior
+	if (x == -2147483647 - 1) return 2147483648u;
+	return uint32_t(-x);
+}
+
+inline uint64_t abs_(int64_t x)
+{
+	if (x >= 0) return uint64_t(x);
+	// avoid undefined behavior
+	if (x == -9223372036854775807ll - 1) return 9223372036854775808ull;
+	return uint64_t(-x);
+}
 
 template<class T>
 T min_(T x, T y) { return x < y ? x : y; }
@@ -157,7 +170,6 @@ void maskArray(T *x, size_t n, size_t bitSize)
 template<class T>
 size_t getNonZeroArraySize(const T *x, size_t n)
 {
-	assert(n > 0);
 	while (n > 0) {
 		if (x[n - 1]) return n;
 		n--;
@@ -165,6 +177,62 @@ size_t getNonZeroArraySize(const T *x, size_t n)
 	return 1;
 }
 
+template<class T>
+class BitIterator {
+	const T *x_;
+	size_t bitPos_;
+	size_t bitSize_;
+	static const size_t TbitSize = sizeof(T) * 8;
+public:
+	BitIterator(const T *x, size_t n)
+		: x_(x)
+		, bitPos_(0)
+	{
+		assert(n > 0);
+		n = getNonZeroArraySize(x, n);
+		if (n == 1 && x[0] == 0) {
+			bitSize_ = 1;
+		} else {
+			assert(x_[n - 1]);
+			bitSize_ = (n - 1) * sizeof(T) * 8 + 1 + cybozu::bsr<T>(x_[n - 1]);
+		}
+	}
+	bool hasNext() const { return bitPos_ < bitSize_; }
+	T getNext(size_t w)
+	{
+		assert(0 < w && w <= TbitSize);
+		assert(hasNext());
+		const size_t q = bitPos_ / TbitSize;
+		const size_t r = bitPos_ % TbitSize;
+		const size_t remain = bitSize_ - bitPos_;
+		if (w > remain) w = remain;
+		T v = x_[q] >> r;
+		if (r + w > TbitSize) {
+			v |= x_[q + 1] << (TbitSize - r);
+		}
+		bitPos_ += w;
+		return v & mask(w);
+	}
+	// whethere next bit is 1 or 0 (bitPos is not moved)
+	bool peekBit() const
+	{
+		assert(hasNext());
+		const size_t q = bitPos_ / TbitSize;
+		const size_t r = bitPos_ % TbitSize;
+		return (x_[q] >> r) & 1;
+	}
+	void skipBit()
+	{
+		assert(hasNext());
+		bitPos_++;
+	}
+	T mask(size_t w) const
+	{
+		assert(w <= TbitSize);
+		return (w == TbitSize ? T(0) : (T(1) << w)) - 1;
+	}
+};
+
 /*
 	@param out [inout] : set element of G ; out = x^y[]
 	@param x [in]
@@ -231,7 +299,7 @@ void powGeneric(G& out, const G& x, const T *y, size_t n, const Mul& mul, const
 		out = x;
 	}
 	for (int i = (int)n - 1; i >= 0; i--) {
-		T v = y[i];
+		v = y[i];
 		for (int j = m - 2; j >= 0; j -= 2) {
 			sqr(out, out);
 			sqr(out, out);
@@ -272,6 +340,8 @@ bool mulSmallUnit(T& z, const T& x, U y)
 	case 8: T::add(z, x, x); T::add(z, z, z); T::add(z, z, z); break;
 	case 9: { T t; T::add(t, x, x); T::add(t, t, t); T::add(t, t, t); T::add(z, t, x); break; }
 	case 10: { T t; T::add(t, x, x); T::add(t, t, t); T::add(t, t, x); T::add(z, t, t); break; }
+	case 11: { T t; T::add(t, x, x); T::add(t, t, x); T::add(t, t, t); T::add(t, t, t); T::sub(z, t, x); break; }
+	case 12: { T t; T::add(t, x, x); T::add(t, t, t); T::add(z, t, t); T::add(z, z, t); break; }
 	default:
 		return false;
 	}

include/mcl/vint.hpp

@@ -11,6 +11,7 @@
 #ifndef CYBOZU_DONT_USE_STRING
 #include <iostream>
 #endif
+#include <mcl/config.hpp>
 #include <mcl/array.hpp>
 #include <mcl/util.hpp>
 #include <mcl/randgen.hpp>
@@ -24,23 +25,11 @@
 	#error "define MCL_MAX_BIT_SZIE"
 #endif
 
-#ifndef MCL_SIZEOF_UNIT
-	#if defined(CYBOZU_OS_BIT) && (CYBOZU_OS_BIT == 32)
-		#define MCL_SIZEOF_UNIT 4
-	#else
-		#define MCL_SIZEOF_UNIT 8
-	#endif
-#endif
-
 namespace mcl {
 
 namespace vint {
 
-#if MCL_SIZEOF_UNIT == 8
+typedef fp::Unit Unit;
-typedef uint64_t Unit;
-#else
-typedef uint32_t Unit;
-#endif
 
 template<size_t x>
 struct RoundUp {
@@ -90,35 +79,19 @@ inline uint32_t mulUnit(uint32_t *pH, uint32_t x, uint32_t y)
 inline uint64_t mulUnit(uint64_t *pH, uint64_t x, uint64_t y)
 {
 #ifdef MCL_VINT_64BIT_PORTABLE
-	uint32_t a = uint32_t(x >> 32);
+	const uint64_t mask = 0xffffffff;
-	uint32_t b = uint32_t(x);
+	uint64_t v = (x & mask) * (y & mask);
-	uint32_t c = uint32_t(y >> 32);
+	uint64_t L = uint32_t(v);
-	uint32_t d = uint32_t(y);
+	uint64_t H = v >> 32;
-
+	uint64_t ad = (x & mask) * uint32_t(y >> 32);
-	uint64_t ad = uint64_t(d) * a;
+	uint64_t bc = uint32_t(x >> 32) * (y & mask);
-	uint64_t bd = uint64_t(d) * b;
+	H += uint32_t(ad);
-	uint64_t L = uint32_t(bd);
+	H += uint32_t(bc);
-	ad += bd >> 32; // [ad:L]
-
-	uint64_t ac = uint64_t(c) * a;
-	uint64_t bc = uint64_t(c) * b;
-	uint64_t H = uint32_t(bc);
-	ac += bc >> 32; // [ac:H]
-	/*
-		  adL
-		 acH
-	*/
-	uint64_t t = (ac << 32) | H;
-	ac >>= 32;
-	H = t + ad;
-	if (H < t) {
-		ac++;
-	}
-	/*
-		ac:H:L
-	*/
 	L |= H << 32;
-	H = (ac << 32) | uint32_t(H >> 32);
+	H >>= 32;
+	H += ad >> 32;
+	H += bc >> 32;
+	H += (x >> 32) * (y >> 32);
 	*pH = H;
 	return L;
 #elif defined(_WIN64) && !defined(__INTEL_COMPILER)
@@ -142,6 +115,7 @@ void divNM(T *q, size_t qn, T *r, const T *x, size_t xn, const T *y, size_t yn);
 */
 inline uint32_t divUnit(uint32_t *pr, uint32_t H, uint32_t L, uint32_t y)
 {
+	assert(y != 0);
 	uint64_t t = make64(H, L);
 	uint32_t q = uint32_t(t / y);
 	*pr = uint32_t(t % y);
@@ -150,7 +124,8 @@ inline uint32_t divUnit(uint32_t *pr, uint32_t H, uint32_t L, uint32_t y)
 #if MCL_SIZEOF_UNIT == 8
 inline uint64_t divUnit(uint64_t *pr, uint64_t H, uint64_t L, uint64_t y)
 {
-#if defined(MCL_VINT_64BIT_PORTABLE)
+	assert(y != 0);
+#if defined(MCL_VINT_64BIT_PORTABLE) || (defined(_MSC_VER) && _MSC_VER < 1920)
 	uint32_t px[4] = { uint32_t(L), uint32_t(L >> 32), uint32_t(H), uint32_t(H >> 32) };
 	uint32_t py[2] = { uint32_t(y), uint32_t(y >> 32) };
 	size_t xn = 4;
@@ -162,7 +137,7 @@ inline uint64_t divUnit(uint64_t *pr, uint64_t H, uint64_t L, uint64_t y)
 	*pr = make64(r[1], r[0]);
 	return make64(q[1], q[0]);
 #elif defined(_MSC_VER)
-	#error "divUnit for uint64_t is not supported"
+	return _udiv128(H, L, y, pr);
 #else
 	typedef __attribute__((mode(TI))) unsigned int uint128;
 	uint128 t = (uint128(H) << 64) | L;
@@ -213,14 +188,11 @@ T addN(T *z, const T *x, const T *y, size_t n)
 	T c = 0;
 	for (size_t i = 0; i < n; i++) {
 		T xc = x[i] + c;
-		if (xc < c) {
+		c = xc < c;
-			// x[i] = Unit(-1) and c = 1
+		T yi = y[i];
-			z[i] = y[i];
+		xc += yi;
-		} else {
+		c += xc < yi;
-			xc += y[i];
+		z[i] = xc;
-			c = y[i] > xc ? 1 : 0;
-			z[i] = xc;
-		}
 	}
 	return c;
 }
@@ -301,14 +273,12 @@ T subN(T *z, const T *x, const T *y, size_t n)
 	assert(n > 0);
 	T c = 0;
 	for (size_t i = 0; i < n; i++) {
-		T yc = y[i] + c;
+		T yi = y[i];
-		if (yc < c) {
+		yi += c;
-			// y[i] = T(-1) and c = 1
+		c = yi < c;
-			z[i] = x[i];
+		T xi = x[i];
-		} else {
+		c += xi < yi;
-			c = x[i] < yc ? 1 : 0;
+		z[i] = xi - yi;
-			z[i] = x[i] - yc;
-		}
 	}
 	return c;
 }
@@ -552,15 +522,6 @@ size_t getRealSize(const T *x, size_t xn)
 	return 1;
 }
 
-template<class T>
-size_t getBitSize(const T *x, size_t n)
-{
-	if (n == 1 && x[0] == 0) return 1;
-	T v = x[n - 1];
-	assert(v);
-	return (n - 1) * sizeof(T) * 8 + 1 + cybozu::bsr<Unit>(v);
-}
-
 /*
 	q[qn] = x[xn] / y[yn] ; qn == xn - yn + 1 if xn >= yn if q
 	r[rn] = x[xn] % y[yn] ; rn = yn before getRealSize
@@ -1204,7 +1165,7 @@ public:
 	}
 	/*
 		set positive value
-		@note assume little endian system
+		@note x is treated as a little endian
 	*/
 	template<class S>
 	void setArray(bool *pb, const S *x, size_t size)
@@ -1218,15 +1179,9 @@ public:
 		size_t unitSize = (sizeof(S) * size + sizeof(Unit) - 1) / sizeof(Unit);
 		buf_.alloc(pb, unitSize);
 		if (!*pb) return;
-		char *dst = (char *)&buf_[0];
+		bool b = fp::convertArrayAsLE(&buf_[0], unitSize, x, size);
-		const char *src = (const char *)x;
+		assert(b);
-		size_t i = 0;
+		(void)b;
-		for (; i < sizeof(S) * size; i++) {
-			dst[i] = src[i];
-		}
-		for (; i < sizeof(Unit) * unitSize; i++) {
-			dst[i] = 0;
-		}
 		trim(unitSize);
 	}
 	/*
@@ -1510,9 +1465,10 @@ public:
 	*/
 	static void divMod(VintT *q, VintT& r, const VintT& x, const VintT& y)
 	{
-		bool qsign = x.isNeg_ ^ y.isNeg_;
+		bool xNeg = x.isNeg_;
+		bool qsign = xNeg ^ y.isNeg_;
 		udiv(q, r, x.buf_, x.size(), y.buf_, y.size());
-		r.isNeg_ = x.isNeg_;
+		r.isNeg_ = xNeg;
 		if (q) q->isNeg_ = qsign;
 	}
 	static void div(VintT& q, const VintT& x, const VintT& y)
@@ -1563,10 +1519,12 @@ public:
 	*/
 	static void quotRem(VintT *q, VintT& r, const VintT& x, const VintT& y)
 	{
+		assert(q != &r);
 		VintT yy = y;
-		bool qsign = x.isNeg_ ^ y.isNeg_;
+		bool yNeg = y.isNeg_;
+		bool qsign = x.isNeg_ ^ yNeg;
 		udiv(q, r, x.buf_, x.size(), y.buf_, y.size());
-		r.isNeg_ = y.isNeg_;
+		r.isNeg_ = yNeg;
 		if (q) q->isNeg_ = qsign;
 		if (!r.isZero() && qsign) {
 			if (q) {
@@ -1755,6 +1713,46 @@ public:
 	static void invMod(VintT& y, const VintT& x, const VintT& m)
 	{
 		assert(!x.isZero() && !m.isZero());
+#if 0
+		VintT u = x;
+		VintT v = m;
+		VintT x1 = 1, x2 = 0;
+		VintT t;
+		while (u != 1 && v != 1) {
+			while (u.isEven()) {
+				u >>= 1;
+				if (x1.isOdd()) {
+					x1 += m;
+				}
+				x1 >>= 1;
+			}
+			while (v.isEven()) {
+				v >>= 1;
+				if (x2.isOdd()) {
+					x2 += m;
+				}
+				x2 >>= 1;
+			}
+			if (u >= v) {
+				u -= v;
+				x1 -= x2;
+				if (x1 < 0) {
+					x1 += m;
+				}
+			} else {
+				v -= u;
+				x2 -= x1;
+				if (x2 < 0) {
+					x2 += m;
+				}
+			}
+		}
+		if (u == 1) {
+			y = x1;
+		} else {
+			y = x2;
+		}
+#else
 		if (x == 1) {
 			y = 1;
 			return;
@@ -1787,6 +1785,7 @@ public:
 			}
 			b -= a * q;
 		}
+#endif
 	}
 	/*
 		Miller-Rabin

include/mcl/window_method.hpp

@@ -23,35 +23,33 @@ struct ArrayIterator {
 		, bitSize(bitSize)
 		, w(w)
 		, pos(0)
-		, mask((w == TbitSize ? 0 : (T(1) << w)) - 1)
+		, mask(makeMask(w))
 	{
 		assert(w <= TbitSize);
 	}
+	T makeMask(size_t w) const
+	{
+		return (w == TbitSize) ? ~T(0) : (T(1) << w) - 1;
+	}
 	bool hasNext() const { return bitSize > 0; }
 	T getNext()
 	{
-		if (w == TbitSize) {
+		if (bitSize < w) {
-			bitSize -= w;
+			w = bitSize;
-			return *x++;
+			mask = makeMask(w);
 		}
-		if (pos + w < TbitSize) {
+		if (pos + w <= TbitSize) {
-			T v = (*x >> pos) & mask;
+			T v = x[0] >> pos;
-			pos += w;
+			if (pos + w < TbitSize) {
-			if (bitSize < w) {
+				pos += w;
-				bitSize = 0;
+				v &= mask;
 			} else {
-				bitSize -= w;
+				pos = 0;
+				x++;
 			}
+			bitSize -= w;
 			return v;
 		}
-		if (pos + bitSize <= TbitSize) {
-			assert(bitSize <= w);
-			T v = *x >> pos;
-			assert((v >> bitSize) == 0);
-			bitSize = 0;
-			return v & mask;
-		}
-		assert(pos > 0);
 		T v = (x[0] >> pos) | (x[1] << (TbitSize - pos));
 		v &= mask;
 		pos = (pos + w) - TbitSize;

misc/Makefile

@@ -0,0 +1,6 @@
+all: low_test
+
+CFLAGS=-I ../include/ -m32 -Ofast -Wall -Wextra -DNDEBUG
+
+low_test: low_test.cpp ../src/low_func_wasm.hpp
+	$(CXX) -o low_test low_test.cpp $(CFLAGS)

misc/low_test.cpp

@@ -0,0 +1,234 @@
+#include <stdio.h>
+#include <stdint.h>
+
+void dump(const char *msg, const uint32_t *x, size_t n)
+{
+	printf("%s", msg);
+	for (size_t i = 0; i < n; i++) {
+		printf("%08x", x[n - 1 - i]);
+	}
+	printf("\n");
+}
+#include "../src/low_func_wasm.hpp"
+
+#define MCL_USE_VINT
+#define MCL_VINT_FIXED_BUFFER
+#define MCL_SIZEOF_UNIT 4
+#define MCL_MAX_BIT_SIZE 768
+#include <mcl/vint.hpp>
+#include <cybozu/test.hpp>
+#include <cybozu/xorshift.hpp>
+#include <cybozu/benchmark.hpp>
+#include <mcl/util.hpp>
+
+const int C = 10000;
+
+template<class RG>
+void setRand(uint32_t *x, size_t n, RG& rg)
+{
+	for (size_t i = 0; i < n; i++) {
+		x[i] = rg.get32();
+	}
+}
+
+/*
+g++ -Ofast -DNDEBUG -Wall -Wextra -m32 -I ./include/ misc/low_test.cpp
+Core i7-8700
+         mulT  karatsuba
+N =  6, 182clk   225clk
+N =  8, 300clk   350clk
+N = 12, 594clk   730clk
+*/
+template<size_t N>
+void mulTest()
+{
+	printf("N=%zd (%zdbit)\n", N, N * 32);
+	cybozu::XorShift rg;
+	uint32_t x[N];
+	uint32_t y[N];
+	uint32_t z[N * 2];
+	for (size_t i = 0; i < 1000; i++) {
+		setRand(x, N, rg);
+		setRand(y, N, rg);
+		// remove MSB
+		x[N - 1] &= 0x7fffffff;
+		y[N - 1] &= 0x7fffffff;
+		mcl::Vint vx, vy;
+		vx.setArray(x, N);
+		vy.setArray(y, N);
+		vx *= vy;
+		mcl::mulT<N>(z, x, y);
+		CYBOZU_TEST_EQUAL_ARRAY(z, vx.getUnit(), N * 2);
+		memset(z, 0, sizeof(z));
+		mcl::karatsubaT<N>(z, x, y);
+		CYBOZU_TEST_EQUAL_ARRAY(z, vx.getUnit(), N * 2);
+	}
+	CYBOZU_BENCH_C("mulT", C, mcl::mulT<N>, z, x, y);
+	CYBOZU_BENCH_C("kara", C, mcl::karatsubaT<N>, z, x, y);
+}
+
+CYBOZU_TEST_AUTO(mulT)
+{
+	mulTest<8>();
+	mulTest<12>();
+}
+
+template<size_t N>
+void sqrTest()
+{
+	printf("N=%zd (%zdbit)\n", N, N * 32);
+	cybozu::XorShift rg;
+	uint32_t x[N];
+	uint32_t y[N * 2];
+	for (size_t i = 0; i < 1000; i++) {
+		setRand(x, N, rg);
+		// remove MSB
+		x[N - 1] &= 0x7fffffff;
+		mcl::Vint vx;
+		vx.setArray(x, N);
+		vx *= vx;
+		mcl::sqrT<N>(y, x);
+		CYBOZU_TEST_EQUAL_ARRAY(y, vx.getUnit(), N * 2);
+	}
+	CYBOZU_BENCH_C("sqrT", C, mcl::sqrT<N>, y, x);
+}
+
+CYBOZU_TEST_AUTO(sqrT)
+{
+	sqrTest<8>();
+	sqrTest<12>();
+}
+
+struct Montgomery {
+	mcl::Vint p_;
+	mcl::Vint R_; // (1 << (pn_ * 64)) % p
+	mcl::Vint RR_; // (R * R) % p
+	uint32_t rp_; // rp * p = -1 mod M = 1 << 64
+	size_t pn_;
+	Montgomery() {}
+	explicit Montgomery(const mcl::Vint& p)
+	{
+		p_ = p;
+		rp_ = mcl::fp::getMontgomeryCoeff(p.getUnit()[0]);
+		pn_ = p.getUnitSize();
+		R_ = 1;
+		R_ = (R_ << (pn_ * 64)) % p_;
+		RR_ = (R_ * R_) % p_;
+	}
+
+	void toMont(mcl::Vint& x) const { mul(x, x, RR_); }
+	void fromMont(mcl::Vint& x) const { mul(x, x, 1); }
+
+	void mul(mcl::Vint& z, const mcl::Vint& x, const mcl::Vint& y) const
+	{
+		const size_t ySize = y.getUnitSize();
+		mcl::Vint c = x * y.getUnit()[0];
+		uint32_t q = c.getUnit()[0] * rp_;
+		c += p_ * q;
+		c >>= sizeof(uint32_t) * 8;
+		for (size_t i = 1; i < pn_; i++) {
+			if (i < ySize) {
+				c += x * y.getUnit()[i];
+			}
+			uint32_t q = c.getUnit()[0] * rp_;
+			c += p_ * q;
+			c >>= sizeof(uint32_t) * 8;
+		}
+		if (c >= p_) {
+			c -= p_;
+		}
+		z = c;
+	}
+	void mod(mcl::Vint& z, const mcl::Vint& xy) const
+	{
+		z = xy;
+		for (size_t i = 0; i < pn_; i++) {
+			uint32_t q = z.getUnit()[0] * rp_;
+			mcl::Vint t = q;
+			z += p_ * t;
+			z >>= 32;
+		}
+		if (z >= p_) {
+			z -= p_;
+		}
+	}
+};
+
+template<size_t N>
+void mulMontTest(const char *pStr)
+{
+	mcl::Vint vp;
+	vp.setStr(pStr);
+	Montgomery mont(vp);
+
+	cybozu::XorShift rg;
+	uint32_t x[N];
+	uint32_t y[N];
+	uint32_t z[N];
+	uint32_t _p[N + 1];
+	uint32_t *const p = _p + 1;
+	vp.getArray(p, N);
+	p[-1] = mont.rp_;
+
+	for (size_t i = 0; i < 1000; i++) {
+		setRand(x, N, rg);
+		setRand(y, N, rg);
+		// remove MSB
+		x[N - 1] &= 0x7fffffff;
+		y[N - 1] &= 0x7fffffff;
+		mcl::Vint vx, vy, vz;
+		vx.setArray(x, N);
+		vy.setArray(y, N);
+		mont.mul(vz, vx, vy);
+		mcl::mulMontT<N>(z, x, y, p);
+		CYBOZU_TEST_EQUAL_ARRAY(z, vz.getUnit(), N);
+
+		mont.mul(vz, vx, vx);
+		mcl::sqrMontT<N>(z, x, p);
+		CYBOZU_TEST_EQUAL_ARRAY(z, vz.getUnit(), N);
+	}
+	CYBOZU_BENCH_C("mulMontT", C, mcl::mulMontT<N>, x, x, y, p);
+	CYBOZU_BENCH_C("sqrMontT", C, mcl::sqrMontT<N>, x, x, p);
+}
+
+template<size_t N>
+void modTest(const char *pStr)
+{
+	mcl::Vint vp;
+	vp.setStr(pStr);
+	Montgomery mont(vp);
+
+	cybozu::XorShift rg;
+	uint32_t xy[N * 2];
+	uint32_t z[N];
+	uint32_t _p[N + 1];
+	uint32_t *const p = _p + 1;
+	vp.getArray(p, N);
+	p[-1] = mont.rp_;
+
+	for (size_t i = 0; i < 1000; i++) {
+		setRand(xy, N * 2, rg);
+		// remove MSB
+		xy[N * 2 - 1] &= 0x7fffffff;
+		mcl::Vint vxy, vz;
+		vxy.setArray(xy, N * 2);
+		mont.mod(vz, vxy);
+		mcl::modT<N>(z, xy, p);
+		CYBOZU_TEST_EQUAL_ARRAY(z, vz.getUnit(), N);
+	}
+	CYBOZU_BENCH_C("modT", C, mcl::modT<N>, z, xy, p);
+}
+
+CYBOZU_TEST_AUTO(mont)
+{
+	const char *pBN254 = "0x2523648240000001ba344d80000000086121000000000013a700000000000013";
+	puts("BN254");
+	mulMontTest<8>(pBN254);
+	modTest<8>(pBN254);
+
+	const char *pBLS12_381 = "0x1a0111ea397fe69a4b1ba7b6434bacd764774b84f38512bf6730d2a0f6b0f6241eabfffeb153ffffb9feffffffffaaab";
+	puts("BLS12");
+	mulMontTest<12>(pBLS12_381);
+	modTest<12>(pBLS12_381);
+}
+

misc/she/memo.txt

@@ -0,0 +1,119 @@
+P ; generator
+x ; secret key
+xP ; public key
+Enc(m;r) = (mP + rxP, rP)
+
+c = (S, T)
+dec(c) := S - xT
+dec(Enc(m;r)) = (mP + rxP) - x(rP) = mP
+DLP(mP) := m
+Dec(c) := DLP(dec(c))
+
+ZKP of dec(c) = m
+
+z := Enc(m;0) = (mP, 0)
+
+c - z = (rxP, rP) ; r is unknown
+
+ZKP of dec(c - z) = 0
+(P1, P2) := (P, rP)
+(A1, A2) := (xP, xrP)
+
+Prover shows that x(P1, P2) = (A1, A2) without revealing x.
+b ; rand
+B = (b P1, b P2)
+h = Hash(P2, A1, A2, B1, B2)
+d = b + h a
+pi = (d, h)
+
+Verifier
+Bi := d Pi - h Ai
+verify h = Hash(P2, A1, A2, B1, B2)
+-----------------------------------------------------------------------------
+CipherTextGT
+P ; generator of GT, GT=<P>
+x1, x2 ; secrect key
+(P0, P1, P2, P3) := (P, x1 P, x2 P, x1 x2 P) ; public information
+
+CipherText c = (A0, A1, A2, A3)
+dec(c) = 0 <=> A0 = x2 A1 + x1 A2 - x1 x2 A3 ; (*)
+
+F(a1, a2, a3) := a2 A1 + a1 A2 - a3 A3
+
+dec(c) = 0 <=> A0 = F(x1, x2, x1 x2)
+
+Sigma-protocol for dec(c) = 0, i.e., show (*)
+
+Prover:
+  b1, b2, b3 ; rand
+  Bi := bi P (i = 1, 2, 3)
+  X := F(b1, b2, b3)
+  send (B1, B2, B3, X) to Verfier
+
+Verifier:
+  takes h randomly and send to Prover
+
+Prover:
+  d1 := b1 + h x1
+  d2 := b2 + h x2
+  d3 := b3 + h x1 x2
+  send (d1, d2, d3) to Verifier
+
+Verifier:
+  verify
+    di P = Bi + h Pi (i = 1, 2, 3)
+    X = F(d1, d2, d3) - h A0
+    and accept it
+
+Fiat-Shamir transform:
+
+Prover:
+b1, b2, b3 ; random value
+Bi := bi P (i = 1, 2, 3)
+X := F(b1, b2, b3)
+h := Hash(P0, ..., P3, A0, ..., A3, B1, B2, B3, X)
+d1 := b1 + h x1
+d2 := b2 + h x2
+d3 := b3 + h x1 x2
+pi := (d1, d2, d3, h)
+
+Verifier:
+(pi, {Pi}, {Ai}) given
+Bi' := di P - h Pi for i = 1, 2, 3
+X' := F(d1, d2, d3) - h A0
+verify Hash({Pi}, {Ai}, {Bi'}, X') = h
+
+Completeness
+
+B1' = d1 P - h P1 = (b1 + h x1) P - h x1 P = b1 P = B1
+B2' = d2 P - h P2 = (b2 + h x2) P - h x2 P = b2 P = B2
+B3' = d3 P - h P3 = (b3 + h x1 x2) P - h x1 x2 P = B3
+X' = F(b1 + h x1, b2 + h x2, b3 + h x1 x2) - h A0
+ = F(b1, b2, b3) + h F(x1, x2, x1 x2) - h A0
+ = F(b1, b2, b3) + h (F(x1, x2, x1 x2) - A0)
+ = F(b1, b2, b3) = X
+OK
+
+Soundness
+{Ai}, pi=(d1, d2, d3, h) ; given
+compute Bi', X' as above
+Suppose Hash({Pi}, {Ai}, {Bi'}, X') = h
+
+define
+b1 := d1 - h x1
+b2 := d2 - h x2
+b3 := d3 - h x1 x2
+where x1, x2 are unknown
+d1, d2, d3 are free parameters, so b1, b2, b3 are also free.
+
+B1' = d1 P - h P1 = b1 P
+B2' = b2 P
+B3' = b3 P
+
+Y := F(x1, x2, x1 x2) - A0; unknown, but it is fixed
+X' = F(d1, d2, d3) - h A0 = F(b1 + h x1, b2 + h x2, b3 + h x1 x2) - h A0
+ = F(b1, b2, b3) + h(F(x1, x2, x1 x2) - A0)
+ = F(b1, b2, b3) + h Y
+
+Hash({Pi}, {Ai}, b1 P, b2 P, b3 P, F(b1, b2, b3) + h Y) = h
+To found {b1, b2, b3, h} to hold this equation, Y must be 0.

misc/she/she-api-ja.md

@@ -250,7 +250,7 @@ PrecomputedPublicKeyはPublicKeyの高速版
 * `CT she.sub(CT x, CT y)`(JS)
     * 暗号文xから暗号文yを引いてzにセットする(またはその値を返す)
 * `void CT::neg(CT& y, const CT& x)`(C++)
-* `void she.neg(CT x)`(JS)
+* `CT she.neg(CT x)`(JS)
     * 暗号文xの符号反転をyにセットする(またはその値を返す)
 * `void CT::mul(CT& z, const CT& x, int y)`(C++)
 * `CT she.mulInt(CT x, int y)`(JS)
@@ -275,7 +275,9 @@ PrecomputedPublicKeyはPublicKeyの高速版
 * ZkpBinEq 暗号文encG1(m1), encG2(m2)についてm1 = m2 = 0または1であることを検証できる
 
 ### API
-PK = PublicKey or PrecomputedPublicKey
+- SK = SecretKey
+- PK = PublicKey or PrecomputedPublicKey
+- AUX = AuxiliaryForZkpDecGT
 
 * `void PK::encWithZkpBin(CipherTextG1& c, Zkp& zkp, int m) const`(C++)
 * `void PK::encWithZkpBin(CipherTextG2& c, Zkp& zkp, int m) const`(C++)
@@ -290,6 +292,14 @@ PK = PublicKey or PrecomputedPublicKey
 * `[CipherTextG1, CipherTextG2, ZkpEqBin] PK::encWithZkpBinEq(m)`(JS)
     * m(=0 or 1)を暗号化して暗号文c1, c2とゼロ知識証明zkpをセットする(または[c1, c2, zkp]を返す)
     * mが0でも1でもなければ例外
+* `SK::decWithZkp(DecZkpDec& zkp, const CipherTextG1& c, const PublicKey& pub) const`(C++)
+* `[m, ZkpDecG1] SK::decWithZkpDec(c, pub)`(JS)
+  * CipherTextG1暗号文`c`を復号して`m`と`zkp`を返す. `zkp`は`dec(c) = m`の証明
+  * `pub`は計算コストを減らすために利用する
+* `SK::decWithZkpDec(ZkpDecGT& zkp, const CipherTextGT& c, const AuxiliaryForZkpDecGT& aux) const`(C++)
+* `[m, ZkpDecGT] SK::decWithZkpDecGT(c, aux)`(JS)
+  * CipherTextGT暗号文`c`を復号して`m`と`zkp`を返す. `zkp`は`dec(c) = m`の証明
+  * `aux = pub.getAuxiliaryForZkpDecGT()`. auxは計算コストを減らすために利用する
 
 ## グローバル関数
 

misc/she/she-api.md

@@ -1,6 +1,6 @@
 # she ; Two-level homomorphic encryption library for browser/Node.js by WebAssembly
 
-# Abstruct
+# Abstract
 she is a somewhat(two-level) homomorphic encryption library,
 which is based on pairings.
 This library supports polynomially many homomorphic additions and
@@ -31,7 +31,7 @@ Sum_i Enc(x_i) Enc(y_i) = Enc(Sum_i x_i y_i).
 * decrypt a ciphertext with a secret key
 
 ## Homomorphic operations
-* homomorphic addtion/substraction over ciphertexts of the same ciphertext class
+* homomorphic addition/subtraction over ciphertexts of the same ciphertext class
 * homomprphic multiplication over ciphertext of CipherTextG1 and CipherTextG2
     * The class of the result is CipherTextGT.
 
@@ -39,7 +39,7 @@ Sum_i Enc(x_i) Enc(y_i) = Enc(Sum_i x_i y_i).
 * This library requires to solve a small DLP to decrypt a ciphertext.
 * The decryption timing is O(m/s), where s is the size of table to solve DLP, and m is the size fo a plaintext.
 * call `setRangeForDLP(s)` to set the table size.
-    * The maximun `m/s` is set by `setTryNum(tryNum)`.
+    * The maximum `m/s` is set by `setTryNum(tryNum)`.
 
 ## Zero-knowledge proof class
 * A zero-knowledge proof is simultaneously created when encrypting a plaintext `m`.
@@ -66,7 +66,7 @@ and read `she.js`.
 
 ## A sample for JS
 
-```
+```js
 // initialize a library
 she.init().then(() => {
   const sec = new she.SecretKey()
@@ -103,7 +103,7 @@ she.init().then(() => {
 
 # A sample for C++
 How to build the library, see [mcl](https://github.com/herumi/mcl/#installation-requirements).
-```
+```c++
 #include <mcl/she.hpp>
 int main()
     try
@@ -255,7 +255,7 @@ PK means PublicKey or PrecomputedPublicKey
 * `CT she.sub(CT x, CT y)`(JS)
     * subtract `x` and `y` and set the value to `z`(or return the value)
 * `void CT::neg(CT& y, const CT& x)`(C++)
-* `void she.neg(CT x)`(JS)
+* `CT she.neg(CT x)`(JS)
     * negate `x` and set the value to `y`(or return the value)
 * `void CT::mul(CT& z, const CT& x, int y)`(C++)
 * `CT she.mulInt(CT x, int y)`(JS)
@@ -281,7 +281,9 @@ PK means PublicKey or PrecomputedPublicKey
 * ZkpBinEq ; verify whether `m1 = m2 = 0` or `1` for ciphertexts `encG1(m1)` and `encG2(m2)`
 
 ### API
-PK = PublicKey or PrecomputedPublicKey
+- SK = SecretKey
+- PK = PublicKey or PrecomputedPublicKey
+- AUX = AuxiliaryForZkpDecGT
 
 * `void PK::encWithZkpBin(CipherTextG1& c, Zkp& zkp, int m) const`(C++)
 * `void PK::encWithZkpBin(CipherTextG2& c, Zkp& zkp, int m) const`(C++)
@@ -296,6 +298,14 @@ PK = PublicKey or PrecomputedPublicKey
 * `[CipherTextG1, CipherTextG2, ZkpEqBin] PK::encWithZkpBinEq(m)`(JS)
     * encrypt `m`(=0 or 1) and set ciphertexts `c1`, `c2` and zero-knowledge proof `zkp`(or returns [c1, c2, zkp])
     * throw exception if m != 0 and m != 1
+* `SK::decWithZkp(DecZkpDec& zkp, const CipherTextG1& c, const PublicKey& pub) const`(C++)
+* `[m, ZkpDecG1] SK::decWithZkpDec(c, pub)`(JS)
+  * decrypt CipherTextG1 `c` and get `m` and zkp, which proves that `dec(c) = m`.
+  * `pub` is used for reducing some computation.
+* `SK::decWithZkpDec(ZkpDecGT& zkp, const CipherTextGT& c, const AuxiliaryForZkpDecGT& aux) const`(C++)
+* `[m, ZkpDecGT] SK::decWithZkpDecGT(c, aux)`(JS)
+  * decrypt CipherTextGT `c` and get `m` and zkp, which proves that `dec(c) = m`.
+  * `aux = pub.getAuxiliaryForZkpDecGT()`, which is used for reducing some computation.
 
 ## Global functions
 

misc/snark-p.py

@@ -0,0 +1,18 @@
+p=21888242871839275222246405745257275088696311157297823662689037894645226208583
+
+print("over 253 bit")
+for i in range (10):
+	print(i, (p * i) >> 253)
+
+def maxarg(x):
+	return x // p
+
+print("maxarg")
+for i in range(16):
+	print(i, maxarg(i << 253))
+
+
+x=0x2c130429c1d4802eb8703197d038ebd5109f96aee333bd027963094f5bb33ad
+
+y = x * 9
+print(hex(y))

mklib.bat

@@ -8,6 +8,10 @@ if "%1"=="dll" (
 rem nasm -f win64 -D_WIN64 src\asm\low_x86-64.asm
 rem lib /OUT:lib\mcl.lib /nodefaultlib fp.obj src\asm\low_x86-64.obj
 
+if "%1"=="dll" (
+  set CFLAGS=%CFLAGS% /DMCL_NO_AUTOLINK /DMCLBN_NO_AUTOLINK
+)
+
 echo cl /c %CFLAGS% src\fp.cpp /Foobj\fp.obj
      cl /c %CFLAGS% src\fp.cpp /Foobj\fp.obj
 echo lib /nologo /OUT:lib\mcl.lib /nodefaultlib obj\fp.obj
@@ -15,22 +19,17 @@ echo lib /nologo /OUT:lib\mcl.lib /nodefaultlib obj\fp.obj
 
 if "%1"=="dll" (
   echo cl /c %CFLAGS% src\bn_c256.cpp /Foobj\bn_c256.obj
-     cl /c %CFLAGS% src\bn_c256.cpp /Foobj\bn_c256.obj /DMCLBN_NO_AUTOLINK
+     cl /c %CFLAGS% src\bn_c256.cpp /Foobj\bn_c256.obj
   echo link /nologo /DLL /OUT:bin\mclbn256.dll obj\bn_c256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclbn256.lib
      link /nologo /DLL /OUT:bin\mclbn256.dll obj\bn_c256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclbn256.lib
 
-  echo cl /c %CFLAGS% src\bn_c384.cpp /Foobj\bn_c384.obj
+  echo cl /c %CFLAGS% src\bn_c384_256.cpp /Foobj\bn_c384_256.obj
-     cl /c %CFLAGS% src\bn_c384.cpp /Foobj\bn_c384.obj /DMCLBN_NO_AUTOLINK
+     cl /c %CFLAGS% src\bn_c384_256.cpp /Foobj\bn_c384_256.obj
-  echo link /nologo /DLL /OUT:bin\mclbn384.dll obj\bn_c384.obj obj\fp.obj %LDFLAGS% /implib:lib\mclbn384.lib
+  echo link /nologo /DLL /OUT:bin\mclbn384_256.dll obj\bn_c384_256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclbn384_256.lib
-     link /nologo /DLL /OUT:bin\mclbn384.dll obj\bn_c384.obj obj\fp.obj %LDFLAGS% /implib:lib\mclbn384.lib
+     link /nologo /DLL /OUT:bin\mclbn384_256.dll obj\bn_c384_256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclbn384_256.lib
-
-  echo cl /c %CFLAGS% src\she_c256.cpp /Foobj\she_c256.obj /DMCLBN_NO_AUTOLINK
-     cl /c %CFLAGS% src\she_c256.cpp /Foobj\she_c256.obj /DMCLBN_NO_AUTOLINK
-  echo link /nologo /DLL /OUT:bin\mclshe256.dll obj\she_c256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclshe_c256.lib
-     link /nologo /DLL /OUT:bin\mclshe256.dll obj\she_c256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclshe_c256.lib
 
-  echo cl /c %CFLAGS% src\she_c384_256.cpp /Foobj\she_c384_256.obj /DMCLBN_NO_AUTOLINK
+  echo cl /c %CFLAGS% src\she_c384_256.cpp /Foobj\she_c384_256.obj /DMCL_NO_AUTOLINK
-     cl /c %CFLAGS% src\she_c384_256.cpp /Foobj\she_c384_256.obj /DMCLBN_NO_AUTOLINK
+     cl /c %CFLAGS% src\she_c384_256.cpp /Foobj\she_c384_256.obj /DMCL_NO_AUTOLINK
   echo link /nologo /DLL /OUT:bin\mclshe384_256.dll obj\she_c384_256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclshe_c384_256.lib
      link /nologo /DLL /OUT:bin\mclshe384_256.dll obj\she_c384_256.obj obj\fp.obj %LDFLAGS% /implib:lib\mclshe_c384_256.lib
 ) else (
@@ -38,7 +37,7 @@ if "%1"=="dll" (
      cl /c %CFLAGS% src\bn_c256.cpp /Foobj\bn_c256.obj
      lib /nologo /OUT:lib\mclbn256.lib /nodefaultlib obj\bn_c256.obj lib\mcl.lib
 
-  echo cl /c %CFLAGS% src\bn_c384.cpp /Foobj\bn_c384.obj
+  echo cl /c %CFLAGS% src\bn_c384_256.cpp /Foobj\bn_c384_256.obj
-     cl /c %CFLAGS% src\bn_c384.cpp /Foobj\bn_c384.obj
+     cl /c %CFLAGS% src\bn_c384_256.cpp /Foobj\bn_c384_256.obj
-     lib /nologo /OUT:lib\mclbn384.lib /nodefaultlib obj\bn_c384.obj lib\mcl.lib
+     lib /nologo /OUT:lib\mclbn384_256.lib /nodefaultlib obj\bn_c384_256.obj lib\mcl.lib
 )