add OP_CHECKZKPVERIFY test, optimize non-asm code through flto and in…

…line

src/bls12-381/Makefile.am

@@ -1,12 +1,8 @@
-ACLOCAL_AMFLAGS = -I build-aux/m4
 
-AM_CXXFLAGS = -std=c++11 -fpic -O2
-AM_CPPFLAGS = -std=c++11 -fpic -O2
+AM_CXXFLAGS = $(CXXFLAGS)
 
-AM_CFLAGS = --fpic -O2
+pkgconfigdir = $(libdir)/pkgconfig
+pkgconfig_DATA = libbls.pc
 
 lib_LTLIBRARIES = libbls.la
 libbls_la_SOURCES = src/groth16.cpp src/arithmetic.cpp src/fp.cpp src/g.cpp src/pairing.cpp src/scalar.cpp src/utils.cpp
-
-pkgconfigdir = $(libdir)/pkgconfig
-pkgconfig_DATA = libbls.pc

src/bls12-381/configure.ac

@@ -13,55 +13,48 @@ AC_INIT([libbls], m4_join([.], _PKG_VERSION_MAJOR, _PKG_VERSION_MINOR, _PKG_VERS
 AC_SUBST(LIB_VERSION_CURRENT, _LIB_VERSION_CURRENT)
 AC_SUBST(LIB_VERSION_REVISION, _LIB_VERSION_REVISION)
 AC_SUBST(LIB_VERSION_AGE, _LIB_VERSION_AGE)
-AC_CONFIG_AUX_DIR([build-aux])
-AC_CONFIG_MACRO_DIR([build-aux/m4])
-AC_CANONICAL_HOST
-AH_TOP([#ifndef LIBMCL_CONFIG_H])
-AH_TOP([#define LIBMCL_CONFIG_H])
-AH_BOTTOM([#endif /*LIBMCL_CONFIG_H*/])
 
 AM_INIT_AUTOMAKE([1.11.2 foreign subdir-objects])
 
-m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
+AC_ARG_ENABLE([blsasm], [AS_HELP_STRING([--enable-blsasm=yes|no], [Enable blsasm option])], [], [enable_blsasm=no])
 
 AC_PROG_CXX
-AC_PROG_CC
 AM_PROG_AR
 
 LT_INIT([win32-dll])
 
+
+CXXFLAGS="-O3 -I include"
+
 AC_CANONICAL_HOST
-case "$host_os" in
-    *mingw64*)
-        CXXFLAGS="$CXXFLAGS -D__USE_MINGW_ANSI_STDIO=1"
-        ;;
-    *cygwin*)
-        ;;
-    *darwin*)
-        AC_DEFINE([DARWIN], [1], [Define if on Darwin/Mac OS X])
-        ;;
-    *openbsd*)
-        CXXFLAGS="$CXXFLAGS -I/usr/local/include"
-        LDFLAGS="$LDFLAGS -L/usr/local/lib"
-        ;;
-    *freebsd*)
-        CXXFLAGS="$CXXFLAGS -I/usr/local/include"
-        LDFLAGS="$LDFLAGS -L/usr/local/lib"
-        ;;
-    *linux*)
-        CXXFLAGS="$CXXFLAGS -I/usr/local/include"
-        LDFLAGS="$LDFLAGS -L/usr/local/lib"
-        ;;
+AC_CANONICAL_TARGET
+case $target_cpu in 
+    x86_64*) 
+    AC_MSG_NOTICE(["The target platform is $target_cpu"])
+    if test "x$enable_blsasm" = "xyes"; then
+        AC_MSG_NOTICE(["Enable blsasm"])
+        CXXFLAGS="$CXXFLAGS -D__x86_64_asm__"
+    else
+        AC_MSG_NOTICE(["Disable blsasm $host_os"])
+        case $host_os in
+            darwin*)
+            CXXFLAGS="$CXXFLAGS"
+            ;;
+            *)
+            CXXFLAGS="$CXXFLAGS -flto"
+            ;;
+        esac
+    fi
+    ;;
+    *)
+    AC_MSG_NOTICE([The target platform is not x86_64])
+    CXXFLAGS="$CXXFLAGS -flto"
+    ;;
 esac
 
 
-CXXFLAGS_WARN="-Wall -Wextra -Wformat=2 -Wcast-qual -Wcast-align -Wwrite-strings -Wfloat-equal -Wpointer-arith -Wundef"
-CXXFLAGS="$CXXFLAGS $CXXFLAGS_WARN -I include -I src"
-
-AC_SUBST(CXXFLAGS)
-AC_SUBST(LDFLAGS)
 AC_SUBST(CXX)
-AC_SUBST(CC)
+AC_SUBST(CXXFLAGS)
 
 AC_CONFIG_FILES([Makefile libbls.pc])
 
@@ -76,4 +69,5 @@ echo "  CPPFLAGS      = $CPPFLAGS"
 echo "  CXX           = $CXX"
 echo "  CXXFLAGS      = $CXXFLAGS"
 echo "  LDFLAGS       = $LDFLAGS"
+echo "  ENABLE_BLSASM = $enable_blsasm"
 echo

src/bls12-381/include/bls12-381/arithmetic.hpp

@@ -30,7 +30,7 @@ void _lsubtract(fp* z, const fp* x, const fp* y);
 // The "smaller than 4p" here means the montgomery form itself as number is less than 4p.
 // Therefore, at most ONE _ladd/_lsubstract/_ldouble is allowed before passing the result to _multiply,
 // unless the algorithm makes sure the number is small.
-#if defined(__x86_64__)
+#if defined(__x86_64_asm__)
 extern void (*_multiply)(fp*, const fp*, const fp*);
 #else
 void _multiply(fp*, const fp*, const fp*);
@@ -41,7 +41,7 @@ void _multiply(fp*, const fp*, const fp*);
 // The carryOut output is guaranteed to be 0 or 1.
 //
 // This function's execution time does not depend on the inputs.
-std::tuple<uint64_t, uint64_t> Add64(
+inline std::tuple<uint64_t, uint64_t> Add64(
     const uint64_t& x,
     const uint64_t& y,
     const uint64_t& carry
@@ -63,12 +63,12 @@ std::tuple<uint64_t, uint64_t> Sub64(
 // half returned in lo.
 //
 // This function's execution time does not depend on the inputs.
-std::tuple<uint64_t, uint64_t> Mul64(
+inline std::tuple<uint64_t, uint64_t> Mul64(
     const uint64_t& x,
     const uint64_t& y
 );
 
-std::tuple<uint64_t, uint64_t, uint64_t> madd(
+inline std::tuple<uint64_t, uint64_t, uint64_t> madd(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& t,
@@ -77,57 +77,57 @@ std::tuple<uint64_t, uint64_t, uint64_t> madd(
 );
 
 // madd0 hi = a*b + c (discards lo bits)
-uint64_t madd0(
+inline uint64_t madd0(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& c
 );
 
 // madd1 hi, lo = a*b + c
-std::tuple<uint64_t, uint64_t> madd1(
+inline std::tuple<uint64_t, uint64_t> madd1(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& c
 );
 
 // madd2 hi, lo = a*b + c + d
-std::tuple<uint64_t, uint64_t> madd2(
+inline std::tuple<uint64_t, uint64_t> madd2(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& c,
     const uint64_t& d
 );
 
 // madd2s superhi, hi, lo = 2*a*b + c + d + e
-std::tuple<uint64_t, uint64_t, uint64_t> madd2s(
+inline std::tuple<uint64_t, uint64_t, uint64_t> madd2s(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& c,
     const uint64_t& d,
     const uint64_t& e
 );
 
-std::tuple<uint64_t, uint64_t, uint64_t> madd1s(
+inline std::tuple<uint64_t, uint64_t, uint64_t> madd1s(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& d,
     const uint64_t& e
 );
 
-std::tuple<uint64_t, uint64_t, uint64_t> madd2sb(
+inline std::tuple<uint64_t, uint64_t, uint64_t> madd2sb(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& c,
     const uint64_t& e
 );
 
-std::tuple<uint64_t, uint64_t, uint64_t> madd1sb(
+inline std::tuple<uint64_t, uint64_t, uint64_t> madd1sb(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& e
 );
 
-std::tuple<uint64_t, uint64_t> madd3(
+inline std::tuple<uint64_t, uint64_t> madd3(
     const uint64_t& a,
     const uint64_t& b,
     const uint64_t& c,

src/bls12-381/libbls.pc.in

@@ -3,7 +3,7 @@ exec_prefix=@exec_prefix@
 libdir=@libdir@
 includedir=@includedir@
 
-Name: libmcl
+Name: libbls
 Description: Pairing Library
 URL: https://github.com/QEDProtocol/bls12-381
 Version: @PACKAGE_VERSION@

src/bls12-381/src/arithmetic.cpp

@@ -1,5 +1,5 @@
 #include <bls12-381/bls12-381.hpp>
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 #include <cpuid.h>
 #endif
 
@@ -8,7 +8,7 @@ using namespace std;
 namespace bls12_381
 {
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _add(fp* z, const fp* x, const fp* y)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -96,7 +96,7 @@ void _add(fp* z, const fp* x, const fp* y)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _ladd(fp* z, const fp* x, const fp* y)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -136,7 +136,7 @@ void _ladd(fp* z, const fp* x, const fp* y)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _double(fp* z, const fp* x)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -223,7 +223,7 @@ void _double(fp* z, const fp* x)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _ldouble(fp* z, const fp* x)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -263,7 +263,7 @@ void _ldouble(fp* z, const fp* x)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _subtract(fp* z, const fp* x, const fp* y)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -342,7 +342,7 @@ void _subtract(fp* z, const fp* x, const fp* y)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _lsubtract(fp* z, const fp* x, const fp* y)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -382,7 +382,7 @@ void _lsubtract(fp* z, const fp* x, const fp* y)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void __negate(fp* z, const fp* x)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -436,7 +436,7 @@ void _negate(fp* z, const fp* x)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void __multiply(fp* z, const fp* x, const fp* y)
 {
     // x86_64 calling convention (https://en.wikipedia.org/wiki/X86_calling_conventions#System_V_AMD64_ABI):
@@ -1624,7 +1624,7 @@ void _multiply(fp* z, const fp* x, const fp* y)
 }
 #endif
 
-#ifdef __x86_64__
+#ifdef __x86_64_asm__
 void _square(fp* z, const fp* x)
 {
     #ifdef __clang__

src/test/script_tests.cpp

@@ -1085,6 +1085,27 @@ BOOST_AUTO_TEST_CASE(script_CHECKMULTISIG12)
     BOOST_CHECK_MESSAGE(err == SCRIPT_ERR_EVAL_FALSE, ScriptErrorString(err));
 }
 
+BOOST_AUTO_TEST_CASE(script_CHECKZKPVERIFY)
+{
+    const int flags = SCRIPT_VERIFY_P2SH;
+
+    CMutableTransaction txCredit;
+    DecodeHexTx(txCredit, "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");
+    CMutableTransaction txSpend;
+    DecodeHexTx(txSpend, "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");
+
+    ScriptError err;
+    bool success = VerifyScript(txSpend.vin[0].scriptSig,
+        txCredit.vout[0].scriptPubKey,
+        &txSpend.vin[0].scriptWitness,
+        flags,
+        MutableTransactionSignatureChecker(&txSpend, 0, txCredit.vout[0].nValue),
+        &err);
+
+    BOOST_CHECK_MESSAGE(err == SCRIPT_ERR_OK, ScriptErrorString(err));
+    BOOST_CHECK(success);
+}
+
 BOOST_AUTO_TEST_CASE(script_CHECKMULTISIG23)
 {
     ScriptError err;