diff --git a/src/fp.cpp b/src/fp.cpp
index 5be1bcb..d8425a0 100644
--- a/src/fp.cpp
+++ b/src/fp.cpp
@@ -569,38 +569,16 @@ bool Op::init(const mpz_class& _p, size_t maxBitSize, int _xi_a, Mode mode, size
 	}
 #endif
 	switch (N) {
-	case 1:  setOp<1>(*this, mode); break;
-	case 2:  setOp<2>(*this, mode); break;
-	case 3:  setOp<3>(*this, mode); break;
-	case 4:  setOp<4>(*this, mode); break; // 256 if 64-bit
-#if MCL_MAX_UNIT_SIZE >= 6
-	case 5:  setOp<5>(*this, mode); break;
-	case 6:  setOp<6>(*this, mode); break;
+	case 192/CYBOZU_OS_BIT:  setOp<192/CYBOZU_OS_BIT>(*this, mode); break;
+#if CYBOZU_OS_BIT == 32
+	case 224/CYBOZU_OS_BIT:  setOp<224/CYBOZU_OS_BIT>(*this, mode); break;
 #endif
-#if MCL_MAX_UNIT_SIZE >= 8
-	case 7:  setOp<7>(*this, mode); break;
-	case 8:  setOp<8>(*this, mode); break;
+	case 256/CYBOZU_OS_BIT:  setOp<256/CYBOZU_OS_BIT>(*this, mode); break;
+#if MCL_MAX_BIT_SIZE >= 384
+	case 384/CYBOZU_OS_BIT:  setOp<384/CYBOZU_OS_BIT>(*this, mode); break;
 #endif
-#if MCL_MAX_UNIT_SIZE >= 9
-	case 9:  setOp<9>(*this, mode); break; // 521 if 64-bit
-#endif
-#if MCL_MAX_UNIT_SIZE >= 10
-	case 10: setOp<10>(*this, mode); break;
-#endif
-#if MCL_MAX_UNIT_SIZE >= 12
-	case 11: setOp<11>(*this, mode); break;
-	case 12: setOp<12>(*this, mode); break; // 768 if 64-bit
-#endif
-#if MCL_MAX_UNIT_SIZE >= 14
-	case 13: setOp<13>(*this, mode); break;
-	case 14: setOp<14>(*this, mode); break;
-#endif
-#if MCL_MAX_UNIT_SIZE >= 16
-	case 15: setOp<15>(*this, mode); break;
-	case 16: setOp<16>(*this, mode); break; // 1024 if 64-bit
-#endif
-#if MCL_MAX_UNIT_SIZE >= 17
-	case 17: setOp<17>(*this, mode); break; // 521 if 32-bit
+#if MCL_MAX_BIT_SIZE >= 512
+	case 512/CYBOZU_OS_BIT:  setOp<512/CYBOZU_OS_BIT>(*this, mode); break;
 #endif
 	default:
 		return false;
@@ -618,10 +596,12 @@ bool Op::init(const mpz_class& _p, size_t maxBitSize, int _xi_a, Mode mode, size
 		fp_sqr = &mcl_fp_sqr_NIST_P192L;
 		fpDbl_mod = &mcl_fpDbl_mod_NIST_P192L;
 	}
+#if MCL_MAX_BIT_SIZE >= 521
 	if (primeMode == PM_NIST_P521) {
 		fpDbl_mod = &mcl_fpDbl_mod_NIST_P521L;
 	}
 #endif
+#endif
 #if defined(MCL_USE_VINT) && MCL_SIZEOF_UNIT == 8
 	if (primeMode == PM_SECP256K1) {
 		fp_mul = &mcl::vint::mcl_fp_mul_SECP256K1;
diff --git a/src/gen.cpp b/src/gen.cpp
index e0f736c..1bae367 100644
--- a/src/gen.cpp
+++ b/src/gen.cpp
@@ -669,7 +669,7 @@ struct Code : public mcl::Generator {
 		Operand z(Int, bu);
 		Operand px(IntPtr, unit);
 		Operand y(Int, unit);
-		std::string name = "mulPv" + cybozu::itoa(bit) + "x" + cybozu::itoa(unit);
+		std::string name = "mulPv" + cybozu::itoa(bit) + "x" + cybozu::itoa(unit) + suf;
 		mulPvM[bit] = Function(name, z, px, y);
 		// workaround at https://github.com/herumi/mcl/pull/82
 //		mulPvM[bit].setPrivate();
@@ -1006,6 +1006,23 @@ struct Code : public mcl::Generator {
 		gen_mulUU();
 #else
 		gen_once();
+#if 1
+		int bitTbl[] = {
+			192,
+			224,
+			256,
+			384,
+			512
+		};
+		for (size_t i = 0; i < CYBOZU_NUM_OF_ARRAY(bitTbl); i++) {
+			uint32_t bit = bitTbl[i];
+			if (unit == 64 && bit == 224) continue;
+			setBit(bit);
+			gen_mul();
+			gen_all();
+			gen_addsub();
+		}
+#else
 		uint32_t end = ((maxBitSize + unit - 1) / unit);
 		for (uint32_t n = 1; n <= end; n++) {
 			setBit(n * unit);
@@ -1013,6 +1030,7 @@ struct Code : public mcl::Generator {
 			gen_all();
 			gen_addsub();
 		}
+#endif
 		if (unit == 64 && maxBitSize == 768) {
 			for (uint32_t i = maxBitSize + unit * 2; i <= maxBitSize * 2; i += unit * 2) {
 				setBit(i);
diff --git a/src/low_func_llvm.hpp b/src/low_func_llvm.hpp
index 117288f..c305ed1 100644
--- a/src/low_func_llvm.hpp
+++ b/src/low_func_llvm.hpp
@@ -52,41 +52,29 @@ template<>const void4u DblSub<n, tag>::f = &mcl_fpDbl_sub ## n ## suf; \
 	MCL_DEF_LLVM_FUNC2(n, Ltag, L)
 #endif
 
-MCL_DEF_LLVM_FUNC(1)
-MCL_DEF_LLVM_FUNC(2)
-MCL_DEF_LLVM_FUNC(3)
-MCL_DEF_LLVM_FUNC(4)
-#if MCL_MAX_UNIT_SIZE >= 6
-MCL_DEF_LLVM_FUNC(5)
+#if CYBOZU_OS_BIT == 32
+
 MCL_DEF_LLVM_FUNC(6)
-#endif
-#if MCL_MAX_UNIT_SIZE >= 8
 MCL_DEF_LLVM_FUNC(7)
 MCL_DEF_LLVM_FUNC(8)
-#endif
-#if MCL_MAX_UNIT_SIZE >= 9
-MCL_DEF_LLVM_FUNC(9)
-#endif
-#if MCL_MAX_UNIT_SIZE >= 10
-MCL_DEF_LLVM_FUNC(10)
-#endif
 #if MCL_MAX_UNIT_SIZE >= 12
-MCL_DEF_LLVM_FUNC(11)
 MCL_DEF_LLVM_FUNC(12)
 #endif
-#if MCL_MAX_UNIT_SIZE >= 14
-MCL_DEF_LLVM_FUNC(13)
-MCL_DEF_LLVM_FUNC(14)
-#endif
 #if MCL_MAX_UNIT_SIZE >= 16
-MCL_DEF_LLVM_FUNC(15)
-#if MCL_SIZEOF_UNIT == 4
 MCL_DEF_LLVM_FUNC(16)
-#else
 #endif
+
+#else // 64
+
+MCL_DEF_LLVM_FUNC(3)
+MCL_DEF_LLVM_FUNC(4)
+#if MCL_MAX_UNIT_SIZE >= 6
+MCL_DEF_LLVM_FUNC(6)
 #endif
-#if MCL_MAX_UNIT_SIZE >= 17
-MCL_DEF_LLVM_FUNC(17)
+#if MCL_MAX_UNIT_SIZE >= 8
+MCL_DEF_LLVM_FUNC(8)
+#endif
+
 #endif
 
 } } // mcl::fp
diff --git a/src/proto.hpp b/src/proto.hpp
index c9f78a5..70588f7 100644
--- a/src/proto.hpp
+++ b/src/proto.hpp
@@ -38,38 +38,29 @@ void mcl_fpDbl_mod_NIST_P521 ## suf(mcl::fp::Unit* z, const mcl::fp::Unit* xy, c
 
 extern "C" {
 
-MCL_FP_DEF_FUNC(1)
-MCL_FP_DEF_FUNC(2)
-MCL_FP_DEF_FUNC(3)
-MCL_FP_DEF_FUNC(4)
-#if MCL_MAX_UNIT_SIZE >= 6
-MCL_FP_DEF_FUNC(5)
+#if CYBOZU_OS_BIT == 32
+
 MCL_FP_DEF_FUNC(6)
-#endif
-#if MCL_MAX_UNIT_SIZE >= 8
 MCL_FP_DEF_FUNC(7)
 MCL_FP_DEF_FUNC(8)
-#endif
-#if MCL_MAX_UNIT_SIZE >= 9
-MCL_FP_DEF_FUNC(9)
-#endif
-#if MCL_MAX_UNIT_SIZE >= 10
-MCL_FP_DEF_FUNC(10)
-#endif
 #if MCL_MAX_UNIT_SIZE >= 12
-MCL_FP_DEF_FUNC(11)
 MCL_FP_DEF_FUNC(12)
 #endif
-#if MCL_MAX_UNIT_SIZE >= 14
-MCL_FP_DEF_FUNC(13)
-MCL_FP_DEF_FUNC(14)
-#endif
 #if MCL_MAX_UNIT_SIZE >= 16
-MCL_FP_DEF_FUNC(15)
 MCL_FP_DEF_FUNC(16)
 #endif
-#if MCL_MAX_UNIT_SIZE >= 17
-MCL_FP_DEF_FUNC(17)
+
+#else // 64
+
+MCL_FP_DEF_FUNC(3)
+MCL_FP_DEF_FUNC(4)
+#if MCL_MAX_UNIT_SIZE >= 6
+MCL_FP_DEF_FUNC(6)
+#endif
+#if MCL_MAX_UNIT_SIZE >= 8
+MCL_FP_DEF_FUNC(8)
+#endif
+
 #endif
 
 MCL_FP_DEF_FUNC_SPECIAL(L)