From dcf160282882fcd2a1e10569a5d684c5845f7a6b Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Sat, 5 Mar 2016 19:41:31 +0900
Subject: [PATCH] add op

---
 include/mcl/fp_generator.hpp | 12 ++++++------
 src/fp.cpp                   |  2 +-
 test/fp_generator_test.cpp   |  3 ++-
 3 files changed, 9 insertions(+), 8 deletions(-)

diff --git a/include/mcl/fp_generator.hpp b/include/mcl/fp_generator.hpp
index a499ca2..04457b3 100644
--- a/include/mcl/fp_generator.hpp
+++ b/include/mcl/fp_generator.hpp
@@ -195,7 +195,7 @@ struct FpGenerator : Xbyak::CodeGenerator {
 		@param p [in] pointer to prime
 		@param pn [in] length of prime
 	*/
-	void init(const uint64_t *p, int pn)
+	void init(Op& /*op*/, const uint64_t *p, int pn)
 	{
 		if (pn < 2) throw cybozu::Exception("mcl:FpGenerator:small pn") << pn;
 		p_ = p;
@@ -244,7 +244,7 @@ struct FpGenerator : Xbyak::CodeGenerator {
 	{
 		StackFrame sf(this, 3);
 		if (isAdd) {
-			gen_raw_add(sf.p[0], sf.p[1], sf.p[2], rax);
+			gen_raw_add(sf.p[0], sf.p[1], sf.p[2], rax, pn_);
 		} else {
 			gen_raw_sub(sf.p[0], sf.p[1], sf.p[2], rax);
 		}
@@ -254,12 +254,12 @@ struct FpGenerator : Xbyak::CodeGenerator {
 	/*
 		pz[] = px[] + py[]
 	*/
-	void gen_raw_add(const RegExp& pz, const RegExp& px, const RegExp& py, const Reg64& t)
+	void gen_raw_add(const RegExp& pz, const RegExp& px, const RegExp& py, const Reg64& t, int n)
 	{
 		mov(t, ptr [px]);
 		add(t, ptr [py]);
 		mov(ptr [pz], t);
-		for (int i = 1; i < pn_; i++) {
+		for (int i = 1; i < n; i++) {
 			mov(t, ptr [px + i * 8]);
 			adc(t, ptr [py + i * 8]);
 			mov(ptr [pz + i * 8], t);
@@ -474,7 +474,7 @@ struct FpGenerator : Xbyak::CodeGenerator {
 		const Xbyak::CodeGenerator::LabelType jmpMode = pn_ < 5 ? T_AUTO : T_NEAR;
 
 		inLocalLabel();
-		gen_raw_add(pz, px, py, rax);
+		gen_raw_add(pz, px, py, rax, pn_);
 		mov(px, (size_t)p_); // destroy px
 		if (isFullBit_) {
 			jc(".over", jmpMode);
@@ -518,7 +518,7 @@ struct FpGenerator : Xbyak::CodeGenerator {
 		gen_raw_sub(pz, px, py, rax);
 		jnc(".exit", jmpMode);
 		mov(px, (size_t)p_);
-		gen_raw_add(pz, pz, px, rax);
+		gen_raw_add(pz, pz, px, rax, pn_);
 	L(".exit");
 		outLocalLabel();
 	}
diff --git a/src/fp.cpp b/src/fp.cpp
index 2c7450e..49e8e36 100644
--- a/src/fp.cpp
+++ b/src/fp.cpp
@@ -363,7 +363,7 @@ static void initForMont(Op& op, const Unit *p, Mode mode)
 #ifdef MCL_USE_XBYAK
 	FpGenerator *fg = op.fg;
 	if (fg == 0) return;
-	fg->init(p, (int)N);
+	fg->init(op, p, (int)N);
 
 	op.fp_neg = Xbyak::CastTo<void2u>(fg->neg_);
 	op.fp_add = Xbyak::CastTo<void3u>(fg->add_);
diff --git a/test/fp_generator_test.cpp b/test/fp_generator_test.cpp
index 3209fa3..5930727 100644
--- a/test/fp_generator_test.cpp
+++ b/test/fp_generator_test.cpp
@@ -189,7 +189,8 @@ void test(const char *pStr)
 	const int pn = mcl::fp::getNonZeroArraySize(p, MAX_N);
 	printf("pn=%d\n", pn);
 	mcl::fp::FpGenerator fg;
-	fg.init(p, pn);
+	mcl::fp::Op op;
+	fg.init(op, p, pn);
 	testAddSub(fg, pn);
 	testNeg(fg, pn);
 	testMulI(fg, pn);