add gen_mcl_fp_mont

src/gen.cpp

@@ -25,6 +25,7 @@ struct Code : public mcl::Generator {
 	FunctionMap mcl_fp_mul_UnitPreM;
 	FunctionMap mcl_fpDbl_mulPreM;
 	FunctionMap mcl_fpDbl_sqrPreM;
+	FunctionMap mcl_fp_montM;
 	FunctionMap mcl_fp_montRedM;
 	Code() : unit(0), unit2(0), bit(0), N(0) { }
 
@@ -593,6 +594,49 @@ struct Code : public mcl::Generator {
 		generic_fpDbl_mul(py, px, px);
 		endFunc();
 	}
+	void gen_mcl_fp_mont()
+	{
+		const int bu = bit + unit;
+		const int bu2 = bit + unit * 2;
+		resetGlobalIdx();
+		Operand pz(IntPtr, bit);
+		Operand px(IntPtr, unit);
+		Operand py(IntPtr, unit);
+		Operand pp(IntPtr, unit);
+		Operand r(Int, unit);
+		std::string name = "mcl_fp_mont" + cybozu::itoa(bit);
+		mcl_fp_montM[bit] = Function(name, Void, pz, px, py, pp, r);
+		beginFunc(mcl_fp_montM[bit]);
+		Operand p = load(bitcast(pp, Operand(IntPtr, bit)));
+		Operand z, s, a;
+		for (uint32_t i = 0; i < N; i++) {
+			Operand y = load(getelementptr(py, makeImm(unit, i)));
+			Operand xy = call(mulPvM[bit], px, y);
+			Operand at;
+			if (i == 0) {
+				a = zext(xy, bu2);
+				at = trunc(xy, unit);
+			} else {
+				xy = zext(xy, bu2);
+				a = add(s, xy);
+				at = trunc(a, unit);
+			}
+			Operand q = mul(at, r);
+			Operand pq = call(mulPvM[bit], pp, q);
+			pq = zext(pq, bu2);
+			Operand t = add(a, pq);
+			s = lshr(t, unit);
+		}
+		s = trunc(s, bu);
+		p = zext(p, bu);
+		Operand vc = sub(s, p);
+		Operand c = trunc(lshr(vc, bit), 1);
+		z = select(c, s, vc);
+		z = trunc(z, bit);
+		store(z, pz);
+		ret(Void);
+		endFunc();
+	}
 	void gen_mcl_fp_montRed()
 	{
 		const int bu = bit + unit;
@@ -651,6 +695,7 @@ struct Code : public mcl::Generator {
 		gen_mcl_fp_mul_UnitPre();
 		gen_mcl_fpDbl_mulPre();
 		gen_mcl_fpDbl_sqrPre();
+		gen_mcl_fp_mont();
 		gen_mcl_fp_montRed();
 	}
 	void setBit(uint32_t bit)

src/llvm_gen.hpp

@@ -149,6 +149,7 @@ struct Generator {
 	Eval call(const Function& f, const Operand& op1, const Operand& op2);
 	Eval call(const Function& f, const Operand& op1, const Operand& op2, const Operand& op3);
 	Eval call(const Function& f, const Operand& op1, const Operand& op2, const Operand& op3, const Operand& op4);
+	Eval call(const Function& f, const Operand& op1, const Operand& op2, const Operand& op3, const Operand& op4, const Operand& op5);
 
 	Operand makeImm(uint32_t bit, int64_t imm);
 };
@@ -290,6 +291,14 @@ struct Generator::Function {
 			opv.push_back(op3);
 			opv.push_back(op4);
 	}
+	Function(const std::string& name, const Operand& ret, const Operand& op1, const Operand& op2, const Operand& op3, const Operand& op4, const Operand& op5)
+		: name(name), ret(ret), isPrivate(false) {
+			opv.push_back(op1);
+			opv.push_back(op2);
+			opv.push_back(op3);
+			opv.push_back(op4);
+			opv.push_back(op5);
+	}
 	void setPrivate()
 	{
 		isPrivate = true;
@@ -544,6 +553,12 @@ inline Generator::Eval Generator::call(const Generator::Function& f, const Gener
 	return impl::callSub(f, tbl, CYBOZU_NUM_OF_ARRAY(tbl));
 }
 
+inline Generator::Eval Generator::call(const Generator::Function& f, const Generator::Operand& op1, const Generator::Operand& op2, const Generator::Operand& op3, const Generator::Operand& op4, const Generator::Operand& opt5)
+{
+	const Operand *tbl[] = { &op1, &op2, &op3, &op4, &opt5 };
+	return impl::callSub(f, tbl, CYBOZU_NUM_OF_ARRAY(tbl));
+}
+
 #define MCL_GEN_FUNCTION(name, ...) Function name(#name, __VA_ARGS__)
 
 } // mcl

src/mul.txt

@@ -2,37 +2,3 @@
 
 @define bu = bit + unit
 @define bu2 = bit + unit * 2
-define void @mcl_fp_mont$(bit)(i$(bit)* %pz, i$(unit)* %px, i$(unit)* %py, i$(unit)* %pp, i$(unit) %r) {
-	%ppt = bitcast i$(unit)* %pp to i$(bit)*
-	%p = load i$(bit)* %ppt
-
-@for i, 0, N
-	%py$(i) = getelementptr i$(unit)* %py, i$(unit) $(i)
-	%y$(i) = load i$(unit)* %py$(i)
-	%xy$(i) = call i$(bu) @mulPv$(bit)x$(unit)(i$(unit)* %px, i$(unit) %y$(i))
-@if i == 0
-	%a0 = zext i$(bu) %xy0 to i$(bu2)
-
-	%at$(i) = trunc i$(bu) %xy$(i) to i$(unit)
-@else
-	%xye$(i) = zext i$(bu) %xy$(i) to i$(bu2)
-	%a$(i) = add i$(bu2) %s$(i-1), %xye$(i)
-	%at$(i) = trunc i$(bu2) %a$(i) to i$(unit)
-@endif
-	%q$(i) = mul i$(unit) %at$(i), %r
-	%pq$(i) = call i$(bu) @mulPv$(bit)x$(unit)(i$(unit)* %pp, i$(unit) %q$(i))
-	%pqe$(i) = zext i$(bu) %pq$(i) to i$(bu2)
-	%t$(i) = add i$(bu2) %a$(i), %pqe$(i)
-	%s$(i) = lshr i$(bu2) %t$(i), $(unit)
-@endfor
-	%v = trunc i$(bu2) %s$(N-1) to i$(bu)
-	%pe = zext i$(bit) %p to i$(bu)
-	%vc = sub i$(bu) %v, %pe
-	%c = lshr i$(bu) %vc, $(bit)
-	%c1 = trunc i$(bu) %c to i1
-	%z = select i1 %c1, i$(bu) %v, i$(bu) %vc
-	%zt = trunc i$(bu) %z to i$(bit)
-	store i$(bit) %zt, i$(bit)* %pz
-	ret void
-}
-