add Fp2::neg for bls12

src/fp_generator.hpp

@@ -3632,19 +3632,14 @@ private:
 	{
 		align(16);
 		void2u func = getCurr<void2u>();
-		if (pn_ == 4 && !isFullBit_) {
-			gen_fp2_neg4();
+		if (pn_ <= 6) {
+			StackFrame sf(this, 2, UseRDX | pn_);
+			gen_raw_neg(sf.p[0], sf.p[1], sf.t);
+			gen_raw_neg(sf.p[0] + FpByte_, sf.p[1] + FpByte_, sf.t);
 			return func;
 		}
 		return 0;
 	}
-	void gen_fp2_neg4()
-	{
-		assert(!isFullBit_);
-		StackFrame sf(this, 2, UseRDX | pn_);
-		gen_raw_neg(sf.p[0], sf.p[1], sf.t);
-		gen_raw_neg(sf.p[0] + FpByte_, sf.p[1] + FpByte_, sf.t);
-	}
 	void3u gen_fp2_mul()
 	{
 		align(16);

test/bls12_test.cpp

@@ -700,8 +700,7 @@ if(0){
 	puts("----------");
 //	exit(1);
 }
-	CYBOZU_BENCH_C("Fp2::add", 10000000, Fp2::add, x2, x2, y2);
-	CYBOZU_BENCH_C("Fp2::sub", 10000000, Fp2::sub, x2, x2, y2);
+	CYBOZU_BENCH_C("Fp2::neg", 10000000, Fp2::neg, x2, x2);
 //	CYBOZU_BENCH_C("mulPre", 100000000, FpDbl::mulPre, dx, xv[0], yv[0]);
 //	CYBOZU_BENCH_C("sqrPre", 100000000, FpDbl::sqrPre, dx, xv[0]);
 //	CYBOZU_BENCH_C("mod   ", 100000000, FpDbl::mod, xv[0], dx);