add mcl_fp_addNC for x86-64 by nasm

4 files changed, 104 insertions, 5 deletions

diff --git a/Makefile b/Makefile
index d9539bf..b1abdd8 100644
--- a/Makefile
+++ b/Makefile
@@ -41,6 +41,11 @@ ifeq ($(HAS_BMI2),1)
   LLVM_FLAGS+=-mattr=bmi2
 endif
 
+ifneq ($(ASM),)
+  LOW_ASM_OBJ=$(LOW_ASM_SRC:.asm=.o)
+  LIB_OBJ+=$(LOW_ASM_OBJ)
+endif
+
 $(MCL_LIB): $(LIB_OBJ)
 	-$(MKDIR) $(@D)
 	$(AR) $@ $(LIB_OBJ)
@@ -57,6 +62,7 @@ $(LLVM_SRC): $(GEN_EXE) $(FUNC_LIST)
 
 $(FUNC_LIST): $(LOW_ASM_SRC)
 	$(shell awk '/global/ { print $$2}' $(LOW_ASM_SRC) > $(FUNC_LIST) || touch $(FUNC_LIST))
+	$(shell awk '/proc/ { print $$2}' $(LOW_ASM_SRC) >> $(FUNC_LIST))
 
 $(GEN_EXE): src/gen.cpp src/llvm_gen.hpp
 	$(CXX) -o $@ $< $(CFLAGS) -O0
@@ -64,6 +70,9 @@ $(GEN_EXE): src/gen.cpp src/llvm_gen.hpp
 asm: $(LLVM_SRC)
 	$(LLVM_OPT) -O3 -o - $(LLVM_SRC) | $(LLVM_LLC) -O3 $(LLVM_FLAGS) -x86-asm-syntax=intel
 
+$(LOW_ASM_OBJ): $(LOW_ASM_SRC)
+	$(ASM) $<
+
 ##################################################################
 
 VPATH=test sample src
diff --git a/common.mk b/common.mk
index ee1bb2d..d8ecaf5 100644
--- a/common.mk
+++ b/common.mk
@@ -9,6 +9,7 @@ ifeq ($(ARCH),x86_64)
   BIT=64
   BIT_OPT=-m64
   LOW_ASM_SRC=src/asm/low_x86-64.asm
+  ASM=nasm -felf64
 endif
 ifeq ($(ARCH),x86)
   CPU=x86
diff --git a/include/mcl/fp_tower.hpp b/include/mcl/fp_tower.hpp
index a86894b..22618d0 100644
--- a/include/mcl/fp_tower.hpp
+++ b/include/mcl/fp_tower.hpp
@@ -63,6 +63,7 @@ class Fp2T : public fp::Operator<Fp2T<Fp> > {
 	typedef FpDblT<Fp> FpDbl;
 	static Fp xi_a_;
 public:
+	typedef typename Fp::BaseFp BaseFp;
 	Fp a, b;
 	Fp2T() { }
 	Fp2T(int64_t a) : a(a), b(0) { }
@@ -93,15 +94,15 @@ public:
 	*/
 	friend std::ostream& operator<<(std::ostream& os, const Fp2T& self)
 	{
-		return os << self.a << mcl::getIoSeparator() << self.b;
+		return os << self.a << Fp::getIoSeparator() << self.b;
 	}
 	friend std::istream& operator>>(std::istream& is, Fp2T& self)
 	{
 		return is >> self.a >> self.b;
 	}
-	std::string getStr(int base = 10, bool withPrefix = false)
+	std::string getStr(int ioMode)
 	{
-		return a.getStr(base, withPrefix) + ' ' + b.getStr(base, withPrefix);
+		return a.getStr(ioMode) + fp::getIoSeparator(ioMode) + b.getStr(ioMode);
 	}
 	bool isZero() const { return a.isZero() && b.isZero(); }
 	bool isOne() const { return a.isOne() && b.isZero(); }
@@ -416,7 +417,7 @@ struct Fp6T : public fp::Operator<Fp6T<Fp> > {
 	bool operator!=(const Fp6T& rhs) const { return !operator==(rhs); }
 	friend std::ostream& operator<<(std::ostream& os, const Fp6T& x)
 	{
-		const char *sep = mcl::getIoSeparator();
+		const char *sep = Fp::getIoSeparator();
 		return os << x.a << sep << x.b << sep << x.c;
 	}
 	friend std::istream& operator>>(std::istream& is, Fp6T& x)
@@ -682,7 +683,7 @@ struct Fp12T : public fp::Operator<Fp12T<Fp> > {
 	}
 	friend std::ostream& operator<<(std::ostream& os, const Fp12T& self)
 	{
-		return os << self.a << mcl::getIoSeparator() << self.b;
+		return os << self.a << Fp::getIoSeparator() << self.b;
 	}
 	friend std::istream& operator>>(std::istream& is, Fp12T& self)
 	{
diff --git a/src/asm/low_x86-64.asm b/src/asm/low_x86-64.asm
index e69de29..72faa69 100644
--- a/src/asm/low_x86-64.asm
+++ b/src/asm/low_x86-64.asm
@@ -0,0 +1,88 @@
+
+; Linux     rdi     rsi    rdx    rcx
+; Win       rcx     rdx    r8     r9
+
+%ifdef _WIN64
+	%define p1org rcx
+	%define p2org rdx
+	%define p3org r8
+	%define p4org r9
+%else
+	%define p1org rdi
+	%define p2org rsi
+	%define p3org rdx
+	%define p4org rcx
+%endif
+
+%imacro proc 1
+global %1
+%1:
+%endmacro
+
+segment .text
+
+%imacro addNC 1
+	mov rax, [p2org]
+	add rax, [p3org]
+	mov [p1org], rax
+%assign i 1
+%rep %1
+	mov rax, [p2org + i * 8]
+	adc rax, [p3org + i * 8]
+	mov [p1org + i * 8], rax
+%assign i (i+1)
+%endrep
+	setc al
+	movzx eax, al
+	ret
+%endmacro
+
+proc mcl_fp_addNC64
+        addNC 0
+proc mcl_fp_addNC128
+        addNC 1
+proc mcl_fp_addNC192
+        addNC 2
+proc mcl_fp_addNC256
+        addNC 3
+proc mcl_fp_addNC320
+        addNC 4
+proc mcl_fp_addNC384
+        addNC 5
+proc mcl_fp_addNC448
+        addNC 6
+proc mcl_fp_addNC512
+        addNC 7
+proc mcl_fp_addNC576
+        addNC 8
+proc mcl_fp_addNC640
+        addNC 9
+proc mcl_fp_addNC704
+        addNC 10
+proc mcl_fp_addNC768
+        addNC 11
+proc mcl_fp_addNC832
+        addNC 12
+proc mcl_fp_addNC896
+        addNC 13
+proc mcl_fp_addNC960
+        addNC 14
+proc mcl_fp_addNC1024
+        addNC 15
+proc mcl_fp_addNC1088
+        addNC 16
+proc mcl_fp_addNC1152
+        addNC 17
+proc mcl_fp_addNC1216
+        addNC 18
+proc mcl_fp_addNC1280
+        addNC 19
+proc mcl_fp_addNC1344
+        addNC 20
+proc mcl_fp_addNC1408
+        addNC 21
+proc mcl_fp_addNC1472
+        addNC 22
+proc mcl_fp_addNC1536
+        addNC 23
+