add karatuba(not enabled)

src/gen.cpp

@@ -42,7 +42,7 @@ struct Code : public mcl::Generator {
 
 		x = zext(x, unit2);
 		y = zext(y, unit2);
-		z= mul(x, y);
+		z = mul(x, y);
 		ret(z);
 		endFunc();
 	}
@@ -549,27 +549,99 @@ struct Code : public mcl::Generator {
 	}
 	void generic_fpDbl_mul(const Operand& pz, const Operand& px, const Operand& py)
 	{
-		const int bu = bit + unit;
-		Operand y = load(py);
-		Operand xy = call(mulPvM[bit], px, y);
-		store(trunc(xy, unit), pz);
-		Operand t = lshr(xy, unit);
-		Operand z, pzi;
-		for (uint32_t i = 1; i < N; i++) {
-			Operand pyi = getelementptr(py, makeImm(32, i));
-			y = load(pyi);
-			xy = call(mulPvM[bit], px, y);
-			t = add(t, xy);
-			z = trunc(t, unit);
-			pzi = getelementptr(pz, makeImm(32, i));
-			if (i < N - 1) {
-				store(z, pzi);
-				t = lshr(t, unit);
+		if (N == 1) {
+			Operand x = load(px);
+			Operand y = load(py);
+			x = zext(x, unit * 2);
+			y = zext(y, unit * 2);
+			Operand z = mul(x, y);
+			store(z, bitcast(pz, Operand(IntPtr, unit * 2)));
+			ret(Void);
+		} else if (N >= 32 && (N % 2) == 0) {
+			/*
+				W = 1 << half
+				(aW + b)(cW + d) = acW^2 + (ad + bc)W + bd
+				ad + bc = (a + b)(c + d) - ac - bd
+			*/
+			const int half = bit / 2;
+			Operand pxW = getelementptr(px, makeImm(32, N / 2));
+			Operand pyW = getelementptr(py, makeImm(32, N / 2));
+			Operand pzWW = getelementptr(pz, makeImm(32, N));
+			call(mcl_fpDbl_mulPreM[half], pz, px, py); // bd
+			call(mcl_fpDbl_mulPreM[half], pzWW, pxW, pyW); // ac
+
+			Operand pa = bitcast(pxW, Operand(IntPtr, half));
+			Operand pb = bitcast(px, Operand(IntPtr, half));
+			Operand pc = bitcast(pyW, Operand(IntPtr, half));
+			Operand pd = bitcast(py, Operand(IntPtr, half));
+			Operand a = zext(load(pa), half + unit);
+			Operand b = zext(load(pb), half + unit);
+			Operand c = zext(load(pc), half + unit);
+			Operand d = zext(load(pd), half + unit);
+			Operand t1 = add(a, b);
+			Operand t2 = add(c, d);
+			Operand buf = _alloca(unit, N);
+			Operand t1L = trunc(t1, half);
+			Operand t2L = trunc(t2, half);
+			Operand c1 = trunc(lshr(t1, half), 1);
+			Operand c2 = trunc(lshr(t2, half), 1);
+			Operand c0 = _and(c1, c2);
+			c1 = select(c1, t2L, makeImm(half, 0));
+			c2 = select(c2, t1L, makeImm(half, 0));
+			Operand buf1 = _alloca(half, 1);
+			Operand buf2 = _alloca(half, 1);
+			store(t1L, buf1);
+			store(t2L, buf2);
+			buf1 = bitcast(buf1, Operand(IntPtr, unit));
+			buf2 = bitcast(buf2, Operand(IntPtr, unit));
+			call(mcl_fpDbl_mulPreM[half], buf, buf1, buf2);
+			buf = bitcast(buf, Operand(IntPtr, bit));
+			Operand t = load(buf);
+			t = zext(t, bit + unit);
+			c0 = zext(c0, bit + unit);
+			c0 = shl(c0, bit);
+			t = _or(t, c0);
+			c1 = zext(c1, bit + unit);
+			c2 = zext(c2, bit + unit);
+			c1 = shl(c1, half);
+			c2 = shl(c2, half);
+			t = add(t, c1);
+			t = add(t, c2);
+			Operand pzL = bitcast(pz, Operand(IntPtr, bit));
+			Operand pzH = getelementptr(pzL, makeImm(32, 1));
+			t = sub(t, zext(load(pzL), bit + unit));
+			t = sub(t, zext(load(pzH), bit + unit));
+			pzL = getelementptr(pz, makeImm(32, N / 2));
+			pzL = bitcast(pzL, Operand(IntPtr, bit + half));
+			if (bit + half > t.bit) {
+				t = zext(t, bit + half);
 			}
+			t = add(t, load(pzL));
+			store(t, pzL);
+			ret(Void);
+		} else {
+			const int bu = bit + unit;
+			Operand y = load(py);
+			Operand xy = call(mulPvM[bit], px, y);
+			store(trunc(xy, unit), pz);
+			Operand t = lshr(xy, unit);
+			Operand z, pzi;
+			for (uint32_t i = 1; i < N; i++) {
+				Operand pyi = getelementptr(py, makeImm(32, i));
+				y = load(pyi);
+				xy = call(mulPvM[bit], px, y);
+				t = add(t, xy);
+				z = trunc(t, unit);
+				pzi = getelementptr(pz, makeImm(32, i));
+				if (i < N - 1) {
+					store(z, pzi);
+					t = lshr(t, unit);
+				}
+			}
+			pzi = bitcast(pzi, Operand(IntPtr, bu));
+			store(t, pzi);
+			ret(Void);
 		}
-		pzi = bitcast(pzi, Operand(IntPtr, bu));
-		store(t, pzi);
-		ret(Void);
 	}
 	void gen_mcl_fpDbl_mulPre()
 	{