a little optimize of power

include/mcl/ec.hpp

@@ -368,12 +368,16 @@ public:
 	}
 	static inline void mulArray(EcT& z, const EcT& x, const fp::Unit *y, size_t yn, bool isNegative)
 	{
-		EcT out;
-		fp::powerGeneric(out, x, y, yn, EcT::add, EcT::dbl);
+		EcT tmp;
+		const EcT *px = &x;
+		if (&z == &x) {
+			tmp = x;
+			px = &tmp;
+		}
+		z.clear();
+		fp::powerGeneric(z, *px, y, yn, EcT::add, EcT::dbl);
 		if (isNegative) {
-			neg(z, out);
-		} else {
-			z = out;
+			neg(z, z);
 		}
 	}
 	template<class tag, size_t maxBitSize, template<class _tag, size_t _maxBitSize>class FpT>

include/mcl/fp.hpp

@@ -270,12 +270,16 @@ public:
 	}
 	static inline void powerArray(FpT& z, const FpT& x, const Unit *y, size_t yn, bool isNegative)
 	{
-		FpT out(1);
-		fp::powerGeneric(out, x, y, yn, FpT::mul, FpT::square);
+		FpT tmp;
+		const FpT *px = &x;
+		if (&z == &x) {
+			tmp = x;
+			px = &tmp;
+		}
+		z = 1;
+		fp::powerGeneric(z, *px, y, yn, FpT::mul, FpT::square);
 		if (isNegative) {
-			FpT::inv(z, out);
-		} else {
-			z = out;
+			FpT::inv(z, z);
 		}
 	}
 	template<class tag2, size_t maxBitSize2>

include/mcl/util.hpp

@@ -142,6 +142,34 @@ void getRandVal(T *out, RG& rg, const T *in, size_t bitSize)
 */
 template<class G, class T>
 void powerGeneric(G& out, const G& x, const T *y, size_t n, void mul(G&, const G&, const G&) , void square(G&, const G&)){
+#if 1
+	assert(&out != &x);
+	G t(x);
+	while (n > 0) {
+		if (y[n - 1]) break;
+		n--;
+	}
+	if (n == 0) return;
+	out = x;
+	int m = cybozu::bsr<T>(y[n - 1]);
+	if (m == 0) {
+		if (n == 1) return;
+		n--;
+		m = (int)sizeof(T) * 8;
+	}
+	for (int i = (int)n - 1; i >= 0; i--) {
+		T v = y[i];
+		if (i < n - 1) {
+			m = (int)sizeof(T) * 8;
+		}
+		for (int j = m - 1; j >= 0; j--) {
+			square(out, out);
+			if (v & (T(1) << j)) {
+				mul(out, out, t);
+			}
+		}
+	}
+#else
 	G t(x);
 	while (n > 0) {
 		if (y[n - 1]) break;
@@ -161,6 +189,7 @@ void powerGeneric(G& out, const G& x, const T *y, size_t n, void mul(G&, const G
 			square(t, t);
 		}
 	}
+#endif
 }
 
 } } // mcl::fp

sample/Makefile

@@ -3,7 +3,7 @@ include ../common.mk
 TARGET=$(TEST_FILE)
 LIBS=
 
-SRC=$(wildcard *.cpp)
+SRC=bench.cpp ecdh.cpp random.cpp vote.cpp
 
 all: $(TARGET)