use textbook mul for N = 8

src/gen.cpp

@@ -715,11 +715,12 @@ struct Code : public mcl::Generator {
 			Operand z = mul(x, y);
 			storeN(z, pz);
 			ret(Void);
-		} else if (N >= 8 && (N % 2) == 0) {
+		} else if (N > 8 && (N % 2) == 0) {
 			/*
 				W = 1 << half
 				(aW + b)(cW + d) = acW^2 + (ad + bc)W + bd
 				ad + bc = (a + b)(c + d) - ac - bd
+				@note Karatsuba is slower for N = 8
 			*/
 			const int H = N / 2;
 			const int half = bit / 2;

test/llvm_test.cpp

@@ -70,7 +70,7 @@ void bench(Unit *x, Unit *y, const Unit *p)
 {
 	printf("N=%zd\n", N);
 	Unit xx[N * 2], yy[N * 2];
-	const int C = 1000;
+	const int C = 10000;
 	CYBOZU_BENCH_C("mulPre", C, mulPre<N>, xx, x, y);
 	CYBOZU_BENCH_C("sqrPre", C, sqrPre<N>, yy, x);
 	CYBOZU_BENCH_C("mod   ", C, mod<N>, yy, xx, p);