From fa17bab40dad2cccc17c6c9418f4f6d2cf1fd060 Mon Sep 17 00:00:00 2001
From: Diederik Huys <diederik.b.huys@gmail.com>
Date: Wed, 27 Mar 2013 14:13:52 +0100
Subject: [PATCH] Port to more generally used YASM assembler

---
 Makefile  |  2 +-
 field.cpp |  4 ++--
 lin64.asm | 30 ++++++++++++++----------------
 lin64.h   |  4 ++--
 4 files changed, 19 insertions(+), 21 deletions(-)

diff --git a/Makefile b/Makefile
index 90f5402..bbe02e6 100644
--- a/Makefile
+++ b/Makefile
@@ -28,7 +28,7 @@ LIBS:=-lgmp obj/lin64.o
 SECP256K1_FILES := $(SECP256K1_FILES) num_gmp.h num_gmp.cpp obj/lin64.o
 
 obj/lin64.o: lin64.asm
-	/tmp/jwasm -Fo obj/lin64.o -elf64 lin64.asm
+	yasm -f elf64 -o obj/lin64.o lin64.asm
 endif
 endif
 endif
diff --git a/field.cpp b/field.cpp
index ad3aef1..de51953 100644
--- a/field.cpp
+++ b/field.cpp
@@ -168,7 +168,7 @@ void FieldElem::SetMult(const FieldElem &a, const FieldElem &b) {
 #endif
 
 #ifdef INLINE_ASM
-    _ExSetMult((uint64_t *) a.n,(uint64_t *) b.n, (uint64_t *) n);
+    ExSetMult((uint64_t *) a.n,(uint64_t *) b.n, (uint64_t *) n);
 #else
     unsigned __int128 c = (__int128)a.n[0] * b.n[0];
     uint64_t t0 = c & 0xFFFFFFFFFFFFFULL; c = c >> 52; // c max 0FFFFFFFFFFFFFE0
@@ -232,7 +232,7 @@ void FieldElem::SetSquare(const FieldElem &a) {
 #endif
 
 #ifdef INLINE_ASM
-    _ExSetSquare((uint64_t *)a.n,(uint64_t *)n);
+    ExSetSquare((uint64_t *)a.n,(uint64_t *)n);
 #else
     __int128 c = (__int128)a.n[0] * a.n[0];
     uint64_t t0 = c & 0xFFFFFFFFFFFFFULL; c = c >> 52; // c max 0FFFFFFFFFFFFFE0
diff --git a/lin64.asm b/lin64.asm
index 7d320dc..a59ce4d 100644
--- a/lin64.asm
+++ b/lin64.asm
@@ -4,19 +4,13 @@
 	;; 	ExSetMult
 	;; 	ExSetSquare
 	;;
-	;; Needed tools: JWASM (http://www.japheth.de/JWasm.html)
-	;;
-	;; !!! WARNING !!!   !!! WARNING !!!   !!! WARNING !!!
-	;;
-	;; Please note that recompiling this binary (jwasm) under a 64-bit OS
-	;; may yield unexpected results and create a corrupted ELF64 header.
+	;; Needed tools: YASM (http://www.japheth.de/JWasm.html)
 	;;
 	;; 
 
-	.x64			
-QTEST	EQU 1
-	.code
-	
+	BITS 64
+
+	;;  Procedure ExSetMult
 	;;  Register Layout:
 	;;  INPUT: 	rdi	= a.n
 	;; 	   	rsi  	= b.n
@@ -31,7 +25,10 @@ QTEST	EQU 1
 	;; 		rcx	 = b.n[3] / t7
 	;; 		rbp	 = Constant 0FFFFFFFFFFFFFh / t8
 	;; 		rsi	 = b.n / b.n[4] / t9
-ExSetMult	PROC C PUBLIC USES rbx rbp r12 r13 r14 r15
+
+	GLOBAL ExSetMult
+	ALIGN 32
+ExSetMult:
 	push rdx
 	mov r14,[rsi+8*0]	; preload b.n[0]. This will be the case until
 				; b.n[0] is no longer needed, then we reassign
@@ -217,7 +214,7 @@ ExSetMult	PROC C PUBLIC USES rbx rbp r12 r13 r14 r15
 	mov rsi,r8		; load c into t9 and destroy b.n[4]
 
 	;; *******************************************************
-common_exit_norm::
+common_exit_norm:
 	mov rdi,01000003D10h	; load constant
 
 	mov rax,r15		; get t5
@@ -290,9 +287,9 @@ common_exit_norm::
 	add r8,r11
 	mov [rbx+1*8],r8	; -> this.n[1]
 	ret
-ExSetMult	ENDP
-
 
+	
+	;;  PROC ExSetSquare
 	;;  Register Layout:
 	;;  INPUT: 	rdi	 = a.n
 	;; 	   	rsi  	 = this.a
@@ -305,7 +302,9 @@ ExSetMult	ENDP
 	;; 		rcx	 = a.n[3] / t7
 	;; 		rbp	 = 0FFFFFFFFFFFFFh / t8
 	;; 		rsi	 = a.n[4] / a.n[4] /t9
-ExSetSquare	PROC C PUBLIC USES rbx rbp r12 r13 r14 r15
+	GLOBAL ExSetSquare
+	ALIGN 32
+ExSetSquare:
 	push rsi
 	mov rbp,0FFFFFFFFFFFFFh
 	
@@ -440,7 +439,6 @@ ExSetSquare	PROC C PUBLIC USES rbx rbp r12 r13 r14 r15
 
 	;; *******************************************************
 	jmp common_exit_norm
-ExSetSquare	ENDP
 	end
 
 	
diff --git a/lin64.h b/lin64.h
index 6dc1367..b5c53ac 100644
--- a/lin64.h
+++ b/lin64.h
@@ -2,8 +2,8 @@
 #define _SECP256K1_LIN64
 
 #ifdef INLINE_ASM
-extern "C" void _ExSetMult(uint64_t *, uint64_t *, uint64_t *);
-extern "C" void _ExSetSquare(uint64_t *, uint64_t *);
+extern "C" void __attribute__ ((sysv_abi)) ExSetMult(uint64_t *, uint64_t *, uint64_t *);
+extern "C" void __attribute__ ((sysv_abi)) ExSetSquare(uint64_t *, uint64_t *);
 #endif
 
 #endif