short.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 short.txt mul.txt
+ASM_SRC_TXT=long.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

@@ -348,6 +348,7 @@ struct Code : public mcl::Generator {
 	}
 	void gen_mcl_fpDbl_add()
 	{
+		// QQQ : generate unnecessary memory copy for large bit
 		const int bu = bit + unit;
 		const int b2 = bit * 2;
 		const int b2u = b2 + unit;
@@ -381,6 +382,39 @@ struct Code : public mcl::Generator {
 		ret(Void);
 		endFunc();
 	}
+	void gen_mcl_fpDbl_sub()
+	{
+		// QQQ : rol is used?
+		const int b2 = bit * 2;
+		const int b2u = b2 + unit;
+		resetGlobalIdx();
+		Operand pz(IntPtr, bit);
+		Operand px(IntPtr, b2);
+		Operand py(IntPtr, b2);
+		Operand pp(IntPtr, bit);
+		std::string name = "mcl_fpDbl_sub" + cybozu::itoa(bit);
+		Function f(name, Void, pz, px, py, pp);
+		beginFunc(f);
+		Operand x = load(px);
+		Operand y = load(py);
+		x = zext(x, b2u);
+		y = zext(y, b2u);
+		Operand vc = sub(x, y); // x - y = [H:L]
+		Operand L = trunc(vc, bit);
+		store(L, pz);
+
+		Operand H = lshr(vc, bit);
+		H = trunc(H, bit);
+		Operand c = lshr(vc, b2u - 1);
+		c = trunc(c, 1);
+		Operand p = load(pp);
+		c = select(c, p, makeImm(bit, 0));
+		Operand t = add(H, c);
+		pz = getelementptr(pz, makeImm(32, 1));
+		store(t, pz);
+		ret(Void);
+		endFunc();
+	}
 	void gen_all()
 	{
 		gen_mcl_fp_addsubNC(true);
@@ -391,6 +425,7 @@ struct Code : public mcl::Generator {
 		gen_mcl_fp_add();
 		gen_mcl_fp_sub();
 		gen_mcl_fpDbl_add();
+		gen_mcl_fpDbl_sub();
 	}
 	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, 'short.txt', unitL, bitLL)
 	gen(fo, 'long.txt', unitL, bitLL)
 	gen(fo, 'mul.txt', unitL, bitLL[1:])
 	fo.close()

src/short.txt

@@ -1,23 +0,0 @@
-@define bu = bit + unit
-@define b2 = bit * 2
-@define b2u = b2 + unit
-
-define void @mcl_fpDbl_sub$(bit)(i$(bit)* %pz, i$(b2)* %px, i$(b2)* %py, i$(bit)* %pp) {
-	%x = load i$(b2)* %px
-	%y = load i$(b2)* %py
-	%x1 = zext i$(b2) %x to i$(b2u)
-	%y1 = zext i$(b2) %y to i$(b2u)
-	%vc = sub i$(b2u) %x1, %y1 ; x - y = [H:L]
-	%L = trunc i$(b2u) %vc to i$(bit) ; L
-	store i$(bit) %L, i$(bit)* %pz
-	%vc1 = lshr i$(b2u) %vc, $(bit)
-	%H = trunc i$(b2u) %vc1 to i$(bit) ; H
-	%c = lshr i$(b2u) %vc, $(b2u-1)
-	%c1 = trunc i$(b2u) %c to i1
-	%p = load i$(bit)* %pp
-	%a = select i1 %c1, i$(bit) %p, i$(bit) 0
-	%v1 = add i$(bit) %H, %a
-	%pz1 = getelementptr i$(bit)* %pz, i32 1
-	store i$(bit) %v1, i$(bit)* %pz1
-	ret void
-}