add test for NIST_P192

include/mcl/fp_tower.hpp

@@ -247,6 +247,7 @@ private:
 	/*
 		x = a + bi, i^2 = -1
 		y = x^2 = (a + bi)^2 = (a^2 - b^2) + 2abi
+		  = (a + b)(a - b) + 2abi
 	*/
 	static inline void fp2_sqrW(Unit *y, const Unit *x)
 	{
@@ -254,12 +255,22 @@ private:
 		Fp *py = reinterpret_cast<Fp*>(y);
 		const Fp& a = px[0];
 		const Fp& b = px[1];
+#if 1
 		Fp aa, bb, t;
 		Fp::sqr(aa, a);
 		Fp::sqr(bb, b);
 		Fp::mul(t, a, b);
 		Fp::sub(py[0], aa, bb); // a^2 - b^2
 		Fp::add(py[1], t, t); // 2ab
+#else
+		Fp t1, t2;
+		Fp::addNC(t1, b, b); // 2b
+		t1 *= a; // 2ab
+		Fp::sub(t2, a, b); // a - b
+		Fp::addNC(py[0], a, b); // a + b
+		py[0] *= t2; // (a + b)(a - b)
+		py[1] = t1; // 2ab
+#endif
 	}
 	/*
 		xi = xi_c + i

src/fp.cpp

@@ -411,7 +411,7 @@ void Op::init(const std::string& mstr, int base, size_t maxBitSize, Mode mode)
 		throw cybozu::Exception("Op::init:not:support") << mstr;
 	}
 #ifdef MCL_USE_LLVM
-	if (mode == FP_AUTO && mp == mpz_class("0xfffffffffffffffffffffffffffffffeffffffffffffffff")) {
+	if ((mode == FP_AUTO || mode == FP_LLVM) && mp == mpz_class("0xfffffffffffffffffffffffffffffffeffffffffffffffff")) {
 		fp_mul = &mcl_fp_mul_NIST_P192;
 		isMont = false;
 	}

test/fp_tower_test.cpp

@@ -138,15 +138,13 @@ void testFpDbl()
 			"0", "1", "123456", "123456789012345668909", pstr.c_str(),
 		};
 		for (size_t i = 0; i < CYBOZU_NUM_OF_ARRAY(tbl); i++) {
-			mpz_class mx(tbl[i]), my;
+			mpz_class mx(tbl[i]);
 			FpDbl x;
 			x.setMpz(mx);
-			x.getMpz(my);
-			CYBOZU_TEST_EQUAL(mx, my);
 			for (size_t j = 0; j < CYBOZU_NUM_OF_ARRAY(tbl); j++) {
 				FpDbl y, z;
 				mpz_class mz, mo;
-				my = tbl[j];
+				mpz_class my(tbl[j]);
 				y.setMpz(my);
 				FpDbl::add(z, x, y);
 				mcl::Gmp::addMod(mo, mx, my, mp);
@@ -284,6 +282,7 @@ void testAll()
 		"0x30000000000000000000000000000000000000000000002b",
 		"0x70000000000000000000000000000000000000000000001f",
 		"0x800000000000000000000000000000000000000000000005",
+		"0xfffffffffffffffffffffffffffffffeffffffffffffffff",
 		"0xfffffffffffffffffffffffffffffffffffffffeffffee37",
 		"0xfffffffffffffffffffffffe26f2fc170f69466a74defd8d",
 		"0xffffffffffffffffffffffffffffffffffffffffffffff13", // max prime