long.txt is removed

Makefile

@@ -14,7 +14,7 @@ SAMPLE_SRC=bench.cpp ecdh.cpp random.cpp rawbench.cpp vote.cpp
 MCL_LIB=$(LIB_DIR)/libmcl.a
 all: $(MCL_LIB)
 
-ASM_SRC_TXT=long.txt mul.txt
+ASM_SRC_TXT=mul.txt
 LIB_OBJ=$(OBJ_DIR)/$(CPU).o $(OBJ_DIR)/fp.o
 LLVM_VER=-3.6
 LLVM_LLC=llc$(LLVM_VER)

src/gen.cpp

@@ -379,6 +379,37 @@ struct Code : public mcl::Generator {
 		ret(Void);
 		endFunc();
 	}
+	void gen_mcl_fp_subL()
+	{
+		resetGlobalIdx();
+		Operand pz(IntPtr, bit);
+		Operand px(IntPtr, bit);
+		Operand py(IntPtr, bit);
+		Operand pp(IntPtr, bit);
+		std::string name = "mcl_fp_sub" + cybozu::itoa(bit) + "L";
+		mcl_fp_subM[bit] = Function(name, Void, pz, px, py, pp);
+		beginFunc(mcl_fp_subM[bit]);
+		Operand x = load(px);
+		Operand y = load(py);
+		x = zext(x, bit + unit);
+		y = zext(y, bit + unit);
+		Operand vc = sub(x, y);
+		Operand v = trunc(vc, bit);
+		Operand c = lshr(vc, bit + unit - 1);
+		c = trunc(c, 1);
+		store(v, pz);
+	Label carry("carry");
+	Label nocarry("nocarry");
+		br(c, carry, nocarry);
+	putLabel(nocarry);
+		ret(Void);
+	putLabel(carry);
+		Operand p = load(pp);
+		Operand t = add(v, p);
+		store(t, pz);
+		ret(Void);
+		endFunc();
+	}
 	void gen_mcl_fpDbl_add()
 	{
 		// QQQ : generate unnecessary memory copy for large bit
@@ -463,6 +494,7 @@ struct Code : public mcl::Generator {
 	void gen_long()
 	{
 		gen_mcl_fp_addL();
+		gen_mcl_fp_subL();
 	}
 	void setBit(uint32_t bit)
 	{

src/gen.py

@@ -181,7 +181,6 @@ def main():
 	fo.write(a)
 
 	bitLL = range(unitL, 576 + 1, unitL)
-	gen(fo, 'long.txt', unitL, bitLL)
 	gen(fo, 'mul.txt', unitL, bitLL[1:])
 	fo.close()
 

src/long.txt

@@ -1,21 +0,0 @@
-
-define void @mcl_fp_sub$(bit)L(i$(bit)* %pz, i$(bit)* %px, i$(bit)* %py, i$(bit)* %pp) {
-	%x = load i$(bit)* %px
-	%y = load i$(bit)* %py
-	%x1 = zext i$(bit) %x to i$(bit+unit)
-	%y1 = zext i$(bit) %y to i$(bit+unit)
-	%vc = sub i$(bit+unit) %x1, %y1
-	%v = trunc i$(bit+unit) %vc to i$(bit)
-	%c = lshr i$(bit+unit) %vc, $(bit+unit-1)
-	%c1 = trunc i$(bit+unit) %c to i1
-	store i$(bit) %v, i$(bit)* %pz
-	br i1 %c1, label %carry, label %nocarry
-nocarry:
-	ret void
-carry:
-	%p = load i$(bit)* %pp
-	%t = add i$(bit) %v, %p ; x - y + p
-	store i$(bit) %t, i$(bit)* %pz
-	ret void
-}
-