add wrapper for mcl_fp_mont

include/mcl/fp.hpp

@@ -37,12 +37,6 @@ bool strToMpzArray(size_t *pBitSize, Unit *y, size_t maxBitSize, mpz_class& x, c
 
 void copyAndMask(Unit *y, const void *x, size_t xByteSize, const Op& op, bool doMask);
 
-enum Mode {
-	FP_AUTO,
-	FP_LLVM,
-	FP_XBYAK
-};
-
 } // mcl::fp
 
 template<class tag = fp::TagDefault, size_t maxBitSize = MCL_MAX_OP_BIT_SIZE>
@@ -67,7 +61,6 @@ public:
 	}
 	static inline void setModulo(const std::string& mstr, int base = 0, fp::Mode mode = fp::FP_AUTO)
 	{
-		cybozu::disable_warning_unused_variable(mode);
 		assert(maxBitSize <= MCL_MAX_OP_BIT_SIZE);
 		assert(sizeof(mp_limb_t) == sizeof(Unit));
 		// set default wrapper function
@@ -75,7 +68,13 @@ public:
 		op_.add = addW;
 		op_.sub = subW;
 		op_.mul = mulW;
-		op_.init(mstr, base, maxBitSize);
+#ifdef MCL_USE_LLVM
+//		op_.useMont = !(mode == fp::FP_LLVM);
+		if (op_.useMont) {
+			op_.mul = montW;
+		}
+#endif
+		op_.init(mstr, base, maxBitSize, mode);
 	}
 	static inline void getModulo(std::string& pstr)
 	{
@@ -344,6 +343,11 @@ public:
 	{
 		op_.negP(y, x, op_.p);
 	}
+	// wrapper function for mcl_fp_mont by LLVM
+	static inline void montW(Unit *z, const Unit *x, const Unit *y)
+	{
+		op_.mont(z, x, y, op_.p, op_.rp);
+	}
 };
 
 template<class tag, size_t maxBitSize> fp::Op FpT<tag, maxBitSize>::op_;

include/mcl/op.hpp

@@ -39,6 +39,13 @@ struct Block {
 	Unit v_[maxOpUnitSize];
 };
 
+enum Mode {
+	FP_AUTO,
+	FP_LLVM,
+	FP_LLVM_MONT,
+	FP_XBYAK
+};
+
 struct Op {
 	mpz_class mp;
 	mcl::SquareRoot sq;
@@ -66,6 +73,9 @@ struct Op {
 	// for Montgomery
 	bool useMont;
 	int2u preInv;
+	// these two members are for mcl_fp_mont
+	Unit rp;
+	void (*mont)(Unit *z, const Unit *x, const Unit *y, const Unit *p, Unit rp);
 	// require p
 	void3u negP;
 	void2uOp invOp;
@@ -79,6 +89,7 @@ struct Op {
 		, isZero(0), clear(0), copy(0)
 		, neg(0), add(0), sub(0), mul(0)
 		, useMont(false), preInv(0)
+		, rp(0), mont(0)
 		, negP(0), invOp(0), addP(0), subP(0), mulPreP(0), modP(0)
 		, fg(createFpGenerator())
 	{
@@ -95,7 +106,7 @@ struct Op {
 	{
 		mul(y, x, RR);
 	}
-	void init(const std::string& mstr, int base, size_t maxBitSize);
+	void init(const std::string& mstr, int base, size_t maxBitSize, Mode mode);
 	static FpGenerator* createFpGenerator();
 	static void destroyFpGenerator(FpGenerator *fg);
 private:

src/fp.cpp

@@ -168,7 +168,8 @@ struct OpeFunc {
 	#define SET_OP_LLVM(n) \
 		addP = mcl_fp_add ## n ##S; \
 		subP = mcl_fp_sub ## n ##S; \
-		mulPreP = mcl_fp_mulPre ## n;
+		mulPreP = mcl_fp_mulPre ## n; \
+		mont = mcl_fp_mont ## n;
 #else
 	#define SET_OP_LLVM(n)
 #endif
@@ -184,7 +185,7 @@ struct OpeFunc {
 		subP = OpeFunc<n>::subC; \
 		mulPreP = OpeFunc<n>::mulPreC; \
 		modP = OpeFunc<n>::modC; \
-		SET_OP_LLVM(n)
+		SET_OP_LLVM(n) 
 
 #ifdef USE_MONT_FP
 inline void invOpForMont(Unit *y, const Unit *x, const Op& op)
@@ -239,52 +240,51 @@ static void initForMont(Op& op, const Unit *p)
 }
 #endif
 
-void Op::init(const std::string& mstr, int base, size_t maxBitSize)
+void Op::init(const std::string& mstr, int base, size_t maxBitSize, Mode mode)
 {
 	bool isMinus = fp::strToMpzArray(&bitSize, p, maxBitSize, mp, mstr, base);
 	if (isMinus) throw cybozu::Exception("Op:init:mstr is minus") << mstr;
 	if (mp == 0) throw cybozu::Exception("Op:init:mstr is zero") << mstr;
 
-	if (bitSize <= 128) {
-		SET_OP(128)
-	} else
-#if CYBOZU_OS_BIT == 32
-	if (bitSize <= 160) {
-		SET_OP(160)
-	} else
-#endif
-	if (bitSize <= 192) {
-		SET_OP(192)
-	} else
-#if CYBOZU_OS_BIT == 32
-	if (bitSize <= 224) {
-		SET_OP(224)
-	} else
+	const size_t roundBit = (bitSize + UnitBitSize - 1) & ~(UnitBitSize - 1);
+#ifdef MCL_USE_LLVM
+	rp = getMontgomeryCoeff(p[0]);
 #endif
-	if (bitSize <= 256) {
-		SET_OP(256)
-	} else
-	if (bitSize <= 384) {
-		SET_OP(384)
-	} else
+	switch (roundBit) {
+	case 32:
+	case 64:
+	case 96:
+	case 128: SET_OP(128); break;
+	case 192: SET_OP(192); break;
+	case 256: SET_OP(256); break;
+	case 320: SET_OP(320); break;
+	case 384: SET_OP(384); break;
+	case 448: SET_OP(448); break;
+	case 512: SET_OP(512); break;
 #if CYBOZU_OS_BIT == 64
-	if (bitSize <= 576) {
-		SET_OP(576)
-	}
+	case 576: SET_OP(576); break;
 #else
-	if (bitSize <= 544) {
-		SET_OP(544)
-	}
+	case 160: SET_OP(160); break;
+	case 224: SET_OP(224); break;
+	case 288: SET_OP(288); break;
+	case 352: SET_OP(352); break;
+	case 416: SET_OP(416); break;
+	case 480: SET_OP(480); break;
+	case 544: SET_OP(544); break;
 #endif
-
+	default:
+		throw cybozu::Exception("Op::init:not:support") << mstr;
+	}
 #ifdef MCL_USE_LLVM
 	if (mp == mpz_class("0xfffffffffffffffffffffffffffffffeffffffffffffffff")) {
 		mul = &mcl_fp_mul_NIST_P192; // slower than MontFp192
 	}
 #endif
+	if (mode == FP_XBYAK) {
 #ifdef USE_MONT_FP
-	fp::initForMont(*this, p);
+		fp::initForMont(*this, p);
 #endif
+	}
 	sq.set(mp);
 }