From 5dc683fea30d2217659516fdfb4f73b0684f9d19 Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Tue, 23 Feb 2021 09:35:31 +0900
Subject: [PATCH] enable Fp2Dbl::sqrPre

---
 src/fp_generator.hpp   | 13 +++----------
 src/fp_static_code.hpp |  2 +-
 2 files changed, 4 insertions(+), 11 deletions(-)

diff --git a/src/fp_generator.hpp b/src/fp_generator.hpp
index cdba525..181e103 100644
--- a/src/fp_generator.hpp
+++ b/src/fp_generator.hpp
@@ -3506,9 +3506,7 @@ private:
 	void2u gen_fp2Dbl_sqrPre()
 	{
 		if (isFullBit_) return 0;
-//		if (pn_ != 4 && !(pn_ == 6 && useMulx_ && useAdx_)) return 0;
-		// almost same for pn_ == 6
-		if (pn_ != 4) return 0;
+		if (pn_ > 6) return 0;
 		void2u func = getCurr<void2u>();
 		const RegExp y = rsp + 0 * 8;
 		const RegExp x = rsp + 1 * 8;
@@ -3520,7 +3518,7 @@ private:
 		mov(ptr [x], gp1);
 		Pack t = sf.t;
 		if (pn_ == 6) {
-			t.append(rax);
+			t.append(gp2);
 			t.append(rdx);
 		}
 		const Pack a = t.sub(0, pn_);
@@ -3544,12 +3542,7 @@ private:
 		mov(gp2, ptr [x]);
 		call(mulPreL);
 		mov(gp0, ptr [x]);
-		if (pn_ == 4) {
-			gen_raw_fp_sub(t1, gp0, gp0 + FpByte_, sf.t, false);
-		} else {
-			assert(pn_ == 6);
-			gen_raw_fp_sub6(t1, gp0, gp0, FpByte_, a, false);
-		}
+		gen_raw_fp_sub_2(t1, gp0, gp0 + FpByte_, t, false);
 		mov(gp0, ptr [y]);
 		lea(gp1, ptr [t1]);
 		lea(gp2, ptr [t2]);
diff --git a/src/fp_static_code.hpp b/src/fp_static_code.hpp
index e254562..15002b1 100644
--- a/src/fp_static_code.hpp
+++ b/src/fp_static_code.hpp
@@ -83,7 +83,7 @@ void setStaticCode(mcl::fp::Op& op)
 		op.fp2_mul2A_ = mclx_Fp2_mul2;
 		op.fp2_mul_xiA_ = mclx_Fp2_mul_xi;
 		op.fp2Dbl_mulPreA_ = mclx_Fp2Dbl_mulPre;
-		op.fp2Dbl_sqrPreA_ = 0;//mclx_Fp2Dbl_sqrPre;
+		op.fp2Dbl_sqrPreA_ = mclx_Fp2Dbl_sqrPre;
 		op.fp2Dbl_mul_xiA_ = mclx_Fp2Dbl_mul_xi;
 		op.fp_preInv = mclx_Fp_preInv;
 	} else {