optimize divBy2

include/mcl/fp.hpp

@@ -341,7 +341,15 @@ public:
 	static inline void sqr(FpT& y, const FpT& x) { op_.fp_sqr(y.v_, x.v_, op_.p); }
 	static inline void divBy2(FpT& y, const FpT& x)
 	{
+#if 0
 		mul(y, x, inv2_); // QQQ : optimize later
+#else
+		bool odd = (x.v_[0] & 1) != 0;
+		op_.fp_shr1(y.v_, x.v_);
+		if (odd) {
+			op_.fp_addPre(y.v_, y.v_, op_.half);
+		}
+#endif
 	}
 	bool isZero() const { return op_.fp_isZero(v_); }
 	bool isOne() const { return fp::isEqualArray(v_, op_.oneRep, op_.N); }

include/mcl/op.hpp

@@ -92,6 +92,7 @@ struct Op {
 	bool (*fp_isZero)(const Unit*);
 	void1u fp_clear;
 	void2u fp_copy;
+	void2u fp_shr1;
 	void3u fp_neg;
 	void4u fp_add;
 	void4u fp_sub;
@@ -159,6 +160,7 @@ struct Op {
 		fp_isZero = 0;
 		fp_clear = 0;
 		fp_copy = 0;
+		fp_shr1 = 0;
 		fp_neg = 0;
 		fp_add = 0;
 		fp_sub = 0;

sample/rawbench.cpp

@@ -38,13 +38,8 @@ void benchRaw(const char *p, mcl::fp::Mode mode)
 	double fp2_sqrT, fp2_mulT;
 	CYBOZU_BENCH_T(fp_addT, op.fp_add, uz, ux, uy, op.p);
 	CYBOZU_BENCH_T(fp_subT, op.fp_sub, uz, uy, ux, op.p);
-	if (op.fp_addPre) {
+	CYBOZU_BENCH_T(fp_addPreT, op.fp_addPre, uz, ux, uy);
-		CYBOZU_BENCH_T(fp_addPreT, op.fp_addPre, uz, ux, uy);
+	CYBOZU_BENCH_T(fp_subPreT, op.fp_subPre, uz, uy, ux);
-		CYBOZU_BENCH_T(fp_subPreT, op.fp_subPre, uz, uy, ux);
-	} else {
-		fp_addPreT = 0;
-		fp_subPreT = 0;
-	}
 	CYBOZU_BENCH_T(fp_sqrT, op.fp_sqr, uz, ux, op.p);
 	CYBOZU_BENCH_T(fp_mulT, op.fp_mul, uz, ux, uy, op.p);
 	CYBOZU_BENCH_T(fp_mulUnitT, op.fp_mulUnit, uz, ux, 12345678, op.p);

src/fp.cpp

@@ -153,16 +153,17 @@ template<size_t N>
 struct SetFpDbl<N, true> {
 	static inline void exec(Op& op)
 	{
-		if (!op.isFullBit) {
+//		if (!op.isFullBit) {
 			op.fpDbl_addPre = AddPre<N * 2, Ltag>::f;
 			op.fpDbl_subPre = SubPre<N * 2, Ltag>::f;
-		}
+//		}
 	}
 };
 
 template<size_t N, class Tag, bool enableFpDbl>
 void setOpSub(Op& op)
 {
+	op.fp_shr1 = Shr1<N, Tag>::f;
 	op.fp_neg = Neg<N, Tag>::f;
 	op.fp_add = Add<N, Tag>::f;
 	op.fp_sub = Sub<N, Tag>::f;
@@ -182,10 +183,8 @@ void setOpSub(Op& op)
 	op.fpN1_mod = N1_Mod<N, Tag>::f;
 	op.fpDbl_add = DblAdd<N, Tag>::f;
 	op.fpDbl_sub = DblSub<N, Tag>::f;
-	if (!op.isFullBit) {
+	op.fp_addPre = AddPre<N, Tag>::f;
-		op.fp_addPre = AddPre<N, Tag>::f;
+	op.fp_subPre = SubPre<N, Tag>::f;
-		op.fp_subPre = SubPre<N, Tag>::f;
-	}
 	SetFpDbl<N, enableFpDbl>::exec(op);
 }
 

src/fp_generator.hpp

@@ -197,17 +197,12 @@ struct FpGenerator : Xbyak::CodeGenerator {
 		op.fp_sub = getCurr<void4u>();
 		gen_fp_sub();
 
-		if (op.isFullBit) {
+		align(16);
-			op.fp_addPre = 0;
+		op.fp_addPre = getCurr<u3u>();
-			op.fp_subPre = 0;
+		gen_addSubPre(true, pn_);
-		} else {
+		align(16);
-			align(16);
+		op.fp_subPre = getCurr<u3u>();
-			op.fp_addPre = getCurr<u3u>();
+		gen_addSubPre(false, pn_);
-			gen_addSubPre(true, pn_);
-			align(16);
-			op.fp_subPre = getCurr<u3u>();
-			gen_addSubPre(false, pn_);
-		}
 		align(16);
 		shr1_ = getCurr<void2op>();
 		gen_shr1();

src/fp_proto.hpp

@@ -43,7 +43,6 @@ template<size_t N, class Tag = Gtag>
 struct AddPre {
 	static inline Unit func(Unit *z, const Unit *x, const Unit *y)
 	{
-		if (N == 0) return 0;
 		return mpn_add_n((mp_limb_t*)z, (const mp_limb_t*)x, (const mp_limb_t*)y, N);
 	}
 	static const u3u f;
@@ -87,6 +86,19 @@ struct SubPre {
 template<size_t N, class Tag>
 const u3u SubPre<N, Tag>::f = SubPre<N, Tag>::func;
 
+// y[N] <- (x[N] >> 1)
+template<size_t N, class Tag = Gtag>
+struct Shr1 {
+	static inline void func(Unit *y, const Unit *x)
+	{
+		mpn_rshift((mp_limb_t*)y, (const mp_limb_t*)x, (int)N, 1);
+	}
+	static const void2u f;
+};
+
+template<size_t N, class Tag>
+const void2u Shr1<N, Tag>::f = Shr1<N, Tag>::func;
+
 // y[N] <- (-x[N]) % p[N]
 template<size_t N, class Tag = Gtag>
 struct Neg {