a little faster fp2_mul for llvm

5 files changed, 43 insertions, 1 deletions

diff --git a/include/mcl/fp_tower.hpp b/include/mcl/fp_tower.hpp
index a0edde5..28f1772 100644
--- a/include/mcl/fp_tower.hpp
+++ b/include/mcl/fp_tower.hpp
@@ -214,7 +214,11 @@ public:
 		if (op.isFastMod) {
 			op.fp2_mul = fp2_mulW;
 		} else if (!op.isFullBit) {
-			op.fp2_mul = fp2_mulUseDblUseNCW;
+			if (sizeof(Fp) * 8 == op.N * fp::UnitBitSize && op.fp2_mulNF) {
+				op.fp2_mul = fp2_mulNFW;
+			} else {
+				op.fp2_mul = fp2_mulUseDblUseNCW;
+			}
 		} else {
 			op.fp2_mul = fp2_mulUseDblW;
 		}
@@ -306,6 +310,11 @@ private:
 		FpDbl::sub(d1, d1, d2); // ac - bd
 		FpDbl::mod(pz[0], d1); // set z0
 	}
+	static void fp2_mulNFW(Unit *z, const Unit *x, const Unit *y)
+	{
+		const fp::Op& op = Fp::op_;
+		op.fp2_mulNF(z, x, y, op.p);
+	}
 	static void fp2_mulUseDblUseNCW(Unit *z, const Unit *x, const Unit *y)
 	{
 		const Fp *px = reinterpret_cast<const Fp*>(x);
diff --git a/include/mcl/op.hpp b/include/mcl/op.hpp
index 0e5cba9..159047a 100644
--- a/include/mcl/op.hpp
+++ b/include/mcl/op.hpp
@@ -123,6 +123,7 @@ struct Op {
 	void3u fp2_add;
 	void3u fp2_sub;
 	void3u fp2_mul;
+	void4u fp2_mulNF;
 	void2u fp2_neg;
 	void2u fp2_inv;
 	void2u fp2_sqr;
@@ -188,6 +189,7 @@ struct Op {
 		fp2_add = 0;
 		fp2_sub = 0;
 		fp2_mul = 0;
+		fp2_mulNF = 0;
 		fp2_neg = 0;
 		fp2_inv = 0;
 		fp2_sqr = 0;
diff --git a/src/fp.cpp b/src/fp.cpp
index 1817a38..107b99d 100644
--- a/src/fp.cpp
+++ b/src/fp.cpp
@@ -211,6 +211,7 @@ void setOp2(Op& op)
 	op.fpDbl_sub = DblSub<N, Tag>::f;
 	op.fp_addPre = AddPre<N, Tag>::f;
 	op.fp_subPre = SubPre<N, Tag>::f;
+	op.fp2_mulNF = Fp2MulNF<N, Tag>::f;
 	SetFpDbl<N, enableFpDbl>::exec(op);
 }
 
diff --git a/src/fp_generator.hpp b/src/fp_generator.hpp
index 9036f86..bc318bf 100644
--- a/src/fp_generator.hpp
+++ b/src/fp_generator.hpp
@@ -229,6 +229,7 @@ struct FpGenerator : Xbyak::CodeGenerator {
 		}
 		// setup fp_tower
 		if (op.N > 4) return;
+		op.fp2_mulNF = 0;
 		align(16);
 		op.fpDbl_add = getCurr<void4u>();
 		gen_fpDbl_add();
diff --git a/src/low_func.hpp b/src/low_func.hpp
index 8684131..cc75464 100644
--- a/src/low_func.hpp
+++ b/src/low_func.hpp
@@ -638,6 +638,35 @@ struct Sqr {
 template<size_t N, class Tag>
 const void3u Sqr<N, Tag>::f = Sqr<N, Tag>::func;
 
+template<size_t N, class Tag = Gtag>
+struct Fp2MulNF {
+	static inline void func(Unit *z, const Unit *x, const Unit *y, const Unit *p)
+	{
+		const Unit *const a = x;
+		const Unit *const b = x + N;
+		const Unit *const c = y;
+		const Unit *const d = y + N;
+		Unit d0[N * 2];
+		Unit d1[N * 2];
+		Unit d2[N * 2];
+		Unit s[N];
+		Unit t[N];
+		AddPre<N, Tag>::f(s, a, b);
+		AddPre<N, Tag>::f(t, c, d);
+		MulPre<N, Tag>::f(d0, s, t);
+		MulPre<N, Tag>::f(d1, a, c);
+		MulPre<N, Tag>::f(d2, b, d);
+		SubPre<N * 2, Tag>::f(d0, d0, d1);
+		SubPre<N * 2, Tag>::f(d0, d0, d2);
+		MontRed<N, Tag>::f(z + N, d0, p);
+		DblSub<N, Tag>::f(d1, d1, d2, p);
+		MontRed<N, Tag>::f(z, d1, p);
+	}
+	static const void4u f;
+};
+template<size_t N, class Tag>
+const void4u Fp2MulNF<N, Tag>::f = Fp2MulNF<N, Tag>::func;
+
 } } // mcl::fp
 
 #ifdef _WIN32