diff --git a/src/gen.cpp b/src/gen.cpp
index e19a992..83728a6 100644
--- a/src/gen.cpp
+++ b/src/gen.cpp
@@ -24,6 +24,7 @@ struct Code : public mcl::Generator {
 	FunctionMap mulPvM;
 	FunctionMap mcl_fp_mul_UnitPreM;
 	FunctionMap mcl_fpDbl_mulPreM;
+	FunctionMap mcl_fpDbl_sqrPreM;
 	Code() : unit(0), unit2(0), bit(0), N(0) { }
 
 	void gen_mulUU()
@@ -544,16 +545,9 @@ struct Code : public mcl::Generator {
 		ret(Void);
 		endFunc();
 	}
-	void gen_mcl_fpDbl_mulPre()
+	void generic_fpDbl_mul(Operand& pz, Operand& px, Operand& py)
 	{
 		const int bu = bit + unit;
-		resetGlobalIdx();
-		Operand pz(IntPtr, unit);
-		Operand px(IntPtr, unit);
-		Operand py(IntPtr, unit);
-		std::string name = "mcl_fpDbl_mulPre" + cybozu::itoa(bit);
-		mcl_fpDbl_mulPreM[bit] = Function(name, Void, pz, px, py);
-		beginFunc(mcl_fpDbl_mulPreM[bit]);
 		Operand y = load(py);
 		Operand xy = call(mulPvM[bit], px, y);
 		store(trunc(xy, unit), pz);
@@ -574,6 +568,28 @@ struct Code : public mcl::Generator {
 		pz = bitcast(pz, Operand(IntPtr, bu));
 		store(t, pz);
 		ret(Void);
+	}
+	void gen_mcl_fpDbl_mulPre()
+	{
+		resetGlobalIdx();
+		Operand pz(IntPtr, unit);
+		Operand px(IntPtr, unit);
+		Operand py(IntPtr, unit);
+		std::string name = "mcl_fpDbl_mulPre" + cybozu::itoa(bit);
+		mcl_fpDbl_mulPreM[bit] = Function(name, Void, pz, px, py);
+		beginFunc(mcl_fpDbl_mulPreM[bit]);
+		generic_fpDbl_mul(pz, px, py);
+		endFunc();
+	}
+	void gen_mcl_fpDbl_sqrPre()
+	{
+		resetGlobalIdx();
+		Operand py(IntPtr, unit);
+		Operand px(IntPtr, unit);
+		std::string name = "mcl_fpDbl_sqrPre" + cybozu::itoa(bit);
+		mcl_fpDbl_sqrPreM[bit] = Function(name, Void, py, px);
+		beginFunc(mcl_fpDbl_sqrPreM[bit]);
+		generic_fpDbl_mul(py, px, px);
 		endFunc();
 	}
 	void gen_all()
@@ -598,6 +614,7 @@ struct Code : public mcl::Generator {
 		gen_mulPv();
 		gen_mcl_fp_mul_UnitPre();
 		gen_mcl_fpDbl_mulPre();
+		gen_mcl_fpDbl_sqrPre();
 	}
 	void setBit(uint32_t bit)
 	{
diff --git a/src/mul.txt b/src/mul.txt
index 37a3f93..98151c7 100644
--- a/src/mul.txt
+++ b/src/mul.txt
@@ -1,30 +1,3 @@
-@define bu = bit + unit
-
-define void @mcl_fpDbl_sqrPre$(bit)(i$(unit)* %py, i$(unit)* %px) {
-  %x0 = load i$(unit)* %px
-  %xx0 = call i$(bu) @mulPv$(bit)x$(unit)(i$(unit) *%px, i$(unit) %x0)
-  %s0 = trunc i$(bu) %xx0 to i$(unit)
-  store i$(unit) %s0, i$(unit)* %py
-  %t0 = lshr i$(bu) %xx0, $(unit)
-
-@for i, 1, N
-  %px$(i) = getelementptr i$(unit)* %px, i32 $(i)
-  %x$(i) = load i$(unit)* %px$(i)
-  %xx$(i) = call i$(bu) @mulPv$(bit)x$(unit)(i$(unit)* %px, i$(unit) %x$(i))
-  %a$(i) = add i$(bu) %t$(i-1), %xx$(i)
-  %s$(i) = trunc i$(bu) %a$(i) to i$(unit)
-  %py$(i) = getelementptr i$(unit)* %py, i32 $(i)
-  @if i < N - 1
-    store i$(unit) %s$(i), i$(unit)* %py$(i)
-    %t$(i) = lshr i$(bu) %a$(i), $(unit)
-  @endif
-@endfor
-
-  %py$(N-1)e = bitcast i$(unit)* %py$(N-1) to i$(bu)*
-  store i$(bu) %a$(N-1), i$(bu)* %py$(N-1)e
-
-  ret void
-}
 
 @define bu = bit + unit
 @define bu2 = bit + unit * 2