OpenSSL 0.9.5a merge

9 files changed, 17 insertions, 2789 deletions

diff --git a/lib/libcrypto/bn/Makefile.ssl b/lib/libcrypto/bn/Makefile.ssl
index fa23a43fa0c..beb9c1b5231 100644
--- a/lib/libcrypto/bn/Makefile.ssl
+++ b/lib/libcrypto/bn/Makefile.ssl
@@ -5,6 +5,7 @@
 DIR=	bn
 TOP=	../..
 CC=	cc
+CPP=    $(CC) -E
 INCLUDES= -I.. -I../../include
 CFLAG=-g
 INSTALL_PREFIX=
@@ -72,10 +73,10 @@ lib:	$(LIBOBJ)
 
 # elf
 asm/bn86-elf.o: asm/bn86unix.cpp
-	$(CPP) -DELF asm/bn86unix.cpp | as -o asm/bn86-elf.o
+	$(CPP) -DELF -x c asm/bn86unix.cpp | as -o asm/bn86-elf.o
 
 asm/co86-elf.o: asm/co86unix.cpp
-	$(CPP) -DELF asm/co86unix.cpp | as -o asm/co86-elf.o
+	$(CPP) -DELF -x c asm/co86unix.cpp | as -o asm/co86-elf.o
 
 # solaris
 asm/bn86-sol.o: asm/bn86unix.cpp
@@ -102,10 +103,10 @@ asm/bn86bsdi.o: asm/bn86unix.cpp
 asm/co86bsdi.o: asm/co86unix.cpp
 	$(CPP) -DBSDI asm/co86unix.cpp | sed 's/ :/:/' | as -o asm/co86bsdi.o
 
-asm/bn86unix.cpp: asm/bn-586.pl
+asm/bn86unix.cpp: asm/bn-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) bn-586.pl cpp >bn86unix.cpp )
 
-asm/co86unix.cpp: asm/co-586.pl
+asm/co86unix.cpp: asm/co-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) co-586.pl cpp >co86unix.cpp )
 
 asm/sparcv8.o: asm/sparcv8.S
diff --git a/lib/libcrypto/bn/asm/bn-win32.asm b/lib/libcrypto/bn/asm/bn-win32.asm
index 871bd88d77f..e69de29bb2d 100644
--- a/lib/libcrypto/bn/asm/bn-win32.asm
+++ b/lib/libcrypto/bn/asm/bn-win32.asm
@@ -1,2122 +0,0 @@
-	; Don't even think of reading this code
-	; It was automatically generated by bn-586.pl
-	; Which is a perl program used to generate the x86 assember for
-	; any of elf, a.out, BSDI,Win32, or Solaris
-	; eric <eay@cryptsoft.com>
-	; 
-	TITLE	bn-586.asm
-        .386
-.model FLAT
-_TEXT	SEGMENT
-PUBLIC	_bn_mul_add_words
-
-_bn_mul_add_words PROC NEAR
-	push	ebp
-	push	ebx
-	push	esi
-	push	edi
-	; 
-	xor	esi,		esi
-	mov	edi,		DWORD PTR 20[esp]
-	mov	ecx,		DWORD PTR 28[esp]
-	mov	ebx,		DWORD PTR 24[esp]
-	and	ecx,		4294967288
-	mov	ebp,		DWORD PTR 32[esp]
-	push	ecx
-	jz	$L000maw_finish
-L001maw_loop:
-	mov	DWORD PTR [esp],ecx
-	; Round 0
-	mov	eax,		DWORD PTR [ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR [edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR [edi],eax
-	mov	esi,		edx
-	; Round 4
-	mov	eax,		DWORD PTR 4[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 4[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 4[edi],eax
-	mov	esi,		edx
-	; Round 8
-	mov	eax,		DWORD PTR 8[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 8[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 8[edi],eax
-	mov	esi,		edx
-	; Round 12
-	mov	eax,		DWORD PTR 12[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 12[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 12[edi],eax
-	mov	esi,		edx
-	; Round 16
-	mov	eax,		DWORD PTR 16[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 16[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 16[edi],eax
-	mov	esi,		edx
-	; Round 20
-	mov	eax,		DWORD PTR 20[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 20[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 20[edi],eax
-	mov	esi,		edx
-	; Round 24
-	mov	eax,		DWORD PTR 24[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 24[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 24[edi],eax
-	mov	esi,		edx
-	; Round 28
-	mov	eax,		DWORD PTR 28[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 28[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 28[edi],eax
-	mov	esi,		edx
-	; 
-	mov	ecx,		DWORD PTR [esp]
-	add	ebx,		32
-	add	edi,		32
-	sub	ecx,		8
-	jnz	L001maw_loop
-$L000maw_finish:
-	mov	ecx,		DWORD PTR 32[esp]
-	and	ecx,		7
-	jnz	$L002maw_finish2
-	jmp	$L003maw_end
-$L002maw_finish2:
-	; Tail Round 0
-	mov	eax,		DWORD PTR [ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR [edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	dec	ecx
-	mov	DWORD PTR [edi],eax
-	mov	esi,		edx
-	jz	$L003maw_end
-	; Tail Round 1
-	mov	eax,		DWORD PTR 4[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 4[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	dec	ecx
-	mov	DWORD PTR 4[edi],eax
-	mov	esi,		edx
-	jz	$L003maw_end
-	; Tail Round 2
-	mov	eax,		DWORD PTR 8[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 8[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	dec	ecx
-	mov	DWORD PTR 8[edi],eax
-	mov	esi,		edx
-	jz	$L003maw_end
-	; Tail Round 3
-	mov	eax,		DWORD PTR 12[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 12[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	dec	ecx
-	mov	DWORD PTR 12[edi],eax
-	mov	esi,		edx
-	jz	$L003maw_end
-	; Tail Round 4
-	mov	eax,		DWORD PTR 16[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 16[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	dec	ecx
-	mov	DWORD PTR 16[edi],eax
-	mov	esi,		edx
-	jz	$L003maw_end
-	; Tail Round 5
-	mov	eax,		DWORD PTR 20[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 20[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	dec	ecx
-	mov	DWORD PTR 20[edi],eax
-	mov	esi,		edx
-	jz	$L003maw_end
-	; Tail Round 6
-	mov	eax,		DWORD PTR 24[ebx]
-	mul	ebp
-	add	eax,		esi
-	mov	esi,		DWORD PTR 24[edi]
-	adc	edx,		0
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 24[edi],eax
-	mov	esi,		edx
-$L003maw_end:
-	mov	eax,		esi
-	pop	ecx
-	pop	edi
-	pop	esi
-	pop	ebx
-	pop	ebp
-	ret
-_bn_mul_add_words ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_mul_words
-
-_bn_mul_words PROC NEAR
-	push	ebp
-	push	ebx
-	push	esi
-	push	edi
-	; 
-	xor	esi,		esi
-	mov	edi,		DWORD PTR 20[esp]
-	mov	ebx,		DWORD PTR 24[esp]
-	mov	ebp,		DWORD PTR 28[esp]
-	mov	ecx,		DWORD PTR 32[esp]
-	and	ebp,		4294967288
-	jz	$L004mw_finish
-L005mw_loop:
-	; Round 0
-	mov	eax,		DWORD PTR [ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR [edi],eax
-	mov	esi,		edx
-	; Round 4
-	mov	eax,		DWORD PTR 4[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 4[edi],eax
-	mov	esi,		edx
-	; Round 8
-	mov	eax,		DWORD PTR 8[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 8[edi],eax
-	mov	esi,		edx
-	; Round 12
-	mov	eax,		DWORD PTR 12[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 12[edi],eax
-	mov	esi,		edx
-	; Round 16
-	mov	eax,		DWORD PTR 16[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 16[edi],eax
-	mov	esi,		edx
-	; Round 20
-	mov	eax,		DWORD PTR 20[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 20[edi],eax
-	mov	esi,		edx
-	; Round 24
-	mov	eax,		DWORD PTR 24[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 24[edi],eax
-	mov	esi,		edx
-	; Round 28
-	mov	eax,		DWORD PTR 28[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 28[edi],eax
-	mov	esi,		edx
-	; 
-	add	ebx,		32
-	add	edi,		32
-	sub	ebp,		8
-	jz	$L004mw_finish
-	jmp	L005mw_loop
-$L004mw_finish:
-	mov	ebp,		DWORD PTR 28[esp]
-	and	ebp,		7
-	jnz	$L006mw_finish2
-	jmp	$L007mw_end
-$L006mw_finish2:
-	; Tail Round 0
-	mov	eax,		DWORD PTR [ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR [edi],eax
-	mov	esi,		edx
-	dec	ebp
-	jz	$L007mw_end
-	; Tail Round 1
-	mov	eax,		DWORD PTR 4[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 4[edi],eax
-	mov	esi,		edx
-	dec	ebp
-	jz	$L007mw_end
-	; Tail Round 2
-	mov	eax,		DWORD PTR 8[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 8[edi],eax
-	mov	esi,		edx
-	dec	ebp
-	jz	$L007mw_end
-	; Tail Round 3
-	mov	eax,		DWORD PTR 12[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 12[edi],eax
-	mov	esi,		edx
-	dec	ebp
-	jz	$L007mw_end
-	; Tail Round 4
-	mov	eax,		DWORD PTR 16[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 16[edi],eax
-	mov	esi,		edx
-	dec	ebp
-	jz	$L007mw_end
-	; Tail Round 5
-	mov	eax,		DWORD PTR 20[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 20[edi],eax
-	mov	esi,		edx
-	dec	ebp
-	jz	$L007mw_end
-	; Tail Round 6
-	mov	eax,		DWORD PTR 24[ebx]
-	mul	ecx
-	add	eax,		esi
-	adc	edx,		0
-	mov	DWORD PTR 24[edi],eax
-	mov	esi,		edx
-$L007mw_end:
-	mov	eax,		esi
-	pop	edi
-	pop	esi
-	pop	ebx
-	pop	ebp
-	ret
-_bn_mul_words ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_sqr_words
-
-_bn_sqr_words PROC NEAR
-	push	ebp
-	push	ebx
-	push	esi
-	push	edi
-	; 
-	mov	esi,		DWORD PTR 20[esp]
-	mov	edi,		DWORD PTR 24[esp]
-	mov	ebx,		DWORD PTR 28[esp]
-	and	ebx,		4294967288
-	jz	$L008sw_finish
-L009sw_loop:
-	; Round 0
-	mov	eax,		DWORD PTR [edi]
-	mul	eax
-	mov	DWORD PTR [esi],eax
-	mov	DWORD PTR 4[esi],edx
-	; Round 4
-	mov	eax,		DWORD PTR 4[edi]
-	mul	eax
-	mov	DWORD PTR 8[esi],eax
-	mov	DWORD PTR 12[esi],edx
-	; Round 8
-	mov	eax,		DWORD PTR 8[edi]
-	mul	eax
-	mov	DWORD PTR 16[esi],eax
-	mov	DWORD PTR 20[esi],edx
-	; Round 12
-	mov	eax,		DWORD PTR 12[edi]
-	mul	eax
-	mov	DWORD PTR 24[esi],eax
-	mov	DWORD PTR 28[esi],edx
-	; Round 16
-	mov	eax,		DWORD PTR 16[edi]
-	mul	eax
-	mov	DWORD PTR 32[esi],eax
-	mov	DWORD PTR 36[esi],edx
-	; Round 20
-	mov	eax,		DWORD PTR 20[edi]
-	mul	eax
-	mov	DWORD PTR 40[esi],eax
-	mov	DWORD PTR 44[esi],edx
-	; Round 24
-	mov	eax,		DWORD PTR 24[edi]
-	mul	eax
-	mov	DWORD PTR 48[esi],eax
-	mov	DWORD PTR 52[esi],edx
-	; Round 28
-	mov	eax,		DWORD PTR 28[edi]
-	mul	eax
-	mov	DWORD PTR 56[esi],eax
-	mov	DWORD PTR 60[esi],edx
-	; 
-	add	edi,		32
-	add	esi,		64
-	sub	ebx,		8
-	jnz	L009sw_loop
-$L008sw_finish:
-	mov	ebx,		DWORD PTR 28[esp]
-	and	ebx,		7
-	jz	$L010sw_end
-	; Tail Round 0
-	mov	eax,		DWORD PTR [edi]
-	mul	eax
-	mov	DWORD PTR [esi],eax
-	dec	ebx
-	mov	DWORD PTR 4[esi],edx
-	jz	$L010sw_end
-	; Tail Round 1
-	mov	eax,		DWORD PTR 4[edi]
-	mul	eax
-	mov	DWORD PTR 8[esi],eax
-	dec	ebx
-	mov	DWORD PTR 12[esi],edx
-	jz	$L010sw_end
-	; Tail Round 2
-	mov	eax,		DWORD PTR 8[edi]
-	mul	eax
-	mov	DWORD PTR 16[esi],eax
-	dec	ebx
-	mov	DWORD PTR 20[esi],edx
-	jz	$L010sw_end
-	; Tail Round 3
-	mov	eax,		DWORD PTR 12[edi]
-	mul	eax
-	mov	DWORD PTR 24[esi],eax
-	dec	ebx
-	mov	DWORD PTR 28[esi],edx
-	jz	$L010sw_end
-	; Tail Round 4
-	mov	eax,		DWORD PTR 16[edi]
-	mul	eax
-	mov	DWORD PTR 32[esi],eax
-	dec	ebx
-	mov	DWORD PTR 36[esi],edx
-	jz	$L010sw_end
-	; Tail Round 5
-	mov	eax,		DWORD PTR 20[edi]
-	mul	eax
-	mov	DWORD PTR 40[esi],eax
-	dec	ebx
-	mov	DWORD PTR 44[esi],edx
-	jz	$L010sw_end
-	; Tail Round 6
-	mov	eax,		DWORD PTR 24[edi]
-	mul	eax
-	mov	DWORD PTR 48[esi],eax
-	mov	DWORD PTR 52[esi],edx
-$L010sw_end:
-	pop	edi
-	pop	esi
-	pop	ebx
-	pop	ebp
-	ret
-_bn_sqr_words ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_div_words
-
-_bn_div_words PROC NEAR
-	push	ebp
-	push	ebx
-	push	esi
-	push	edi
-	mov	edx,		DWORD PTR 20[esp]
-	mov	eax,		DWORD PTR 24[esp]
-	mov	ebx,		DWORD PTR 28[esp]
-	div	ebx
-	pop	edi
-	pop	esi
-	pop	ebx
-	pop	ebp
-	ret
-_bn_div_words ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_add_words
-
-_bn_add_words PROC NEAR
-	push	ebp
-	push	ebx
-	push	esi
-	push	edi
-	; 
-	mov	ebx,		DWORD PTR 20[esp]
-	mov	esi,		DWORD PTR 24[esp]
-	mov	edi,		DWORD PTR 28[esp]
-	mov	ebp,		DWORD PTR 32[esp]
-	xor	eax,		eax
-	and	ebp,		4294967288
-	jz	$L011aw_finish
-L012aw_loop:
-	; Round 0
-	mov	ecx,		DWORD PTR [esi]
-	mov	edx,		DWORD PTR [edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR [ebx],ecx
-	; Round 1
-	mov	ecx,		DWORD PTR 4[esi]
-	mov	edx,		DWORD PTR 4[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 4[ebx],ecx
-	; Round 2
-	mov	ecx,		DWORD PTR 8[esi]
-	mov	edx,		DWORD PTR 8[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 8[ebx],ecx
-	; Round 3
-	mov	ecx,		DWORD PTR 12[esi]
-	mov	edx,		DWORD PTR 12[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 12[ebx],ecx
-	; Round 4
-	mov	ecx,		DWORD PTR 16[esi]
-	mov	edx,		DWORD PTR 16[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 16[ebx],ecx
-	; Round 5
-	mov	ecx,		DWORD PTR 20[esi]
-	mov	edx,		DWORD PTR 20[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 20[ebx],ecx
-	; Round 6
-	mov	ecx,		DWORD PTR 24[esi]
-	mov	edx,		DWORD PTR 24[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 24[ebx],ecx
-	; Round 7
-	mov	ecx,		DWORD PTR 28[esi]
-	mov	edx,		DWORD PTR 28[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 28[ebx],ecx
-	; 
-	add	esi,		32
-	add	edi,		32
-	add	ebx,		32
-	sub	ebp,		8
-	jnz	L012aw_loop
-$L011aw_finish:
-	mov	ebp,		DWORD PTR 32[esp]
-	and	ebp,		7
-	jz	$L013aw_end
-	; Tail Round 0
-	mov	ecx,		DWORD PTR [esi]
-	mov	edx,		DWORD PTR [edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR [ebx],ecx
-	jz	$L013aw_end
-	; Tail Round 1
-	mov	ecx,		DWORD PTR 4[esi]
-	mov	edx,		DWORD PTR 4[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 4[ebx],ecx
-	jz	$L013aw_end
-	; Tail Round 2
-	mov	ecx,		DWORD PTR 8[esi]
-	mov	edx,		DWORD PTR 8[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 8[ebx],ecx
-	jz	$L013aw_end
-	; Tail Round 3
-	mov	ecx,		DWORD PTR 12[esi]
-	mov	edx,		DWORD PTR 12[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 12[ebx],ecx
-	jz	$L013aw_end
-	; Tail Round 4
-	mov	ecx,		DWORD PTR 16[esi]
-	mov	edx,		DWORD PTR 16[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 16[ebx],ecx
-	jz	$L013aw_end
-	; Tail Round 5
-	mov	ecx,		DWORD PTR 20[esi]
-	mov	edx,		DWORD PTR 20[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 20[ebx],ecx
-	jz	$L013aw_end
-	; Tail Round 6
-	mov	ecx,		DWORD PTR 24[esi]
-	mov	edx,		DWORD PTR 24[edi]
-	add	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	add	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 24[ebx],ecx
-$L013aw_end:
-	pop	edi
-	pop	esi
-	pop	ebx
-	pop	ebp
-	ret
-_bn_add_words ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_sub_words
-
-_bn_sub_words PROC NEAR
-	push	ebp
-	push	ebx
-	push	esi
-	push	edi
-	; 
-	mov	ebx,		DWORD PTR 20[esp]
-	mov	esi,		DWORD PTR 24[esp]
-	mov	edi,		DWORD PTR 28[esp]
-	mov	ebp,		DWORD PTR 32[esp]
-	xor	eax,		eax
-	and	ebp,		4294967288
-	jz	$L014aw_finish
-L015aw_loop:
-	; Round 0
-	mov	ecx,		DWORD PTR [esi]
-	mov	edx,		DWORD PTR [edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR [ebx],ecx
-	; Round 1
-	mov	ecx,		DWORD PTR 4[esi]
-	mov	edx,		DWORD PTR 4[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 4[ebx],ecx
-	; Round 2
-	mov	ecx,		DWORD PTR 8[esi]
-	mov	edx,		DWORD PTR 8[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 8[ebx],ecx
-	; Round 3
-	mov	ecx,		DWORD PTR 12[esi]
-	mov	edx,		DWORD PTR 12[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 12[ebx],ecx
-	; Round 4
-	mov	ecx,		DWORD PTR 16[esi]
-	mov	edx,		DWORD PTR 16[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 16[ebx],ecx
-	; Round 5
-	mov	ecx,		DWORD PTR 20[esi]
-	mov	edx,		DWORD PTR 20[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 20[ebx],ecx
-	; Round 6
-	mov	ecx,		DWORD PTR 24[esi]
-	mov	edx,		DWORD PTR 24[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 24[ebx],ecx
-	; Round 7
-	mov	ecx,		DWORD PTR 28[esi]
-	mov	edx,		DWORD PTR 28[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 28[ebx],ecx
-	; 
-	add	esi,		32
-	add	edi,		32
-	add	ebx,		32
-	sub	ebp,		8
-	jnz	L015aw_loop
-$L014aw_finish:
-	mov	ebp,		DWORD PTR 32[esp]
-	and	ebp,		7
-	jz	$L016aw_end
-	; Tail Round 0
-	mov	ecx,		DWORD PTR [esi]
-	mov	edx,		DWORD PTR [edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR [ebx],ecx
-	jz	$L016aw_end
-	; Tail Round 1
-	mov	ecx,		DWORD PTR 4[esi]
-	mov	edx,		DWORD PTR 4[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 4[ebx],ecx
-	jz	$L016aw_end
-	; Tail Round 2
-	mov	ecx,		DWORD PTR 8[esi]
-	mov	edx,		DWORD PTR 8[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 8[ebx],ecx
-	jz	$L016aw_end
-	; Tail Round 3
-	mov	ecx,		DWORD PTR 12[esi]
-	mov	edx,		DWORD PTR 12[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 12[ebx],ecx
-	jz	$L016aw_end
-	; Tail Round 4
-	mov	ecx,		DWORD PTR 16[esi]
-	mov	edx,		DWORD PTR 16[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 16[ebx],ecx
-	jz	$L016aw_end
-	; Tail Round 5
-	mov	ecx,		DWORD PTR 20[esi]
-	mov	edx,		DWORD PTR 20[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	dec	ebp
-	mov	DWORD PTR 20[ebx],ecx
-	jz	$L016aw_end
-	; Tail Round 6
-	mov	ecx,		DWORD PTR 24[esi]
-	mov	edx,		DWORD PTR 24[edi]
-	sub	ecx,		eax
-	mov	eax,		0
-	adc	eax,		eax
-	sub	ecx,		edx
-	adc	eax,		0
-	mov	DWORD PTR 24[ebx],ecx
-$L016aw_end:
-	pop	edi
-	pop	esi
-	pop	ebx
-	pop	ebp
-	ret
-_bn_sub_words ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_mul_comba8
-
-_bn_mul_comba8 PROC NEAR
-	push	esi
-	mov	esi,		DWORD PTR 12[esp]
-	push	edi
-	mov	edi,		DWORD PTR 20[esp]
-	push	ebp
-	push	ebx
-	xor	ebx,		ebx
-	mov	eax,		DWORD PTR [esi]
-	xor	ecx,		ecx
-	mov	edx,		DWORD PTR [edi]
-	; ################## Calculate word 0
-	xor	ebp,		ebp
-	; mul a[0]*b[0]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebp,		0
-	mov	DWORD PTR [eax],ebx
-	mov	eax,		DWORD PTR 4[esi]
-	; saved r[0]
-	; ################## Calculate word 1
-	xor	ebx,		ebx
-	; mul a[1]*b[0]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebx,		0
-	; mul a[0]*b[1]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebx,		0
-	mov	DWORD PTR 4[eax],ecx
-	mov	eax,		DWORD PTR 8[esi]
-	; saved r[1]
-	; ################## Calculate word 2
-	xor	ecx,		ecx
-	; mul a[2]*b[0]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ecx,		0
-	; mul a[1]*b[1]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ecx,		0
-	; mul a[0]*b[2]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ecx,		0
-	mov	DWORD PTR 8[eax],ebp
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[2]
-	; ################## Calculate word 3
-	xor	ebp,		ebp
-	; mul a[3]*b[0]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebp,		0
-	; mul a[2]*b[1]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebp,		0
-	; mul a[1]*b[2]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebp,		0
-	; mul a[0]*b[3]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebp,		0
-	mov	DWORD PTR 12[eax],ebx
-	mov	eax,		DWORD PTR 16[esi]
-	; saved r[3]
-	; ################## Calculate word 4
-	xor	ebx,		ebx
-	; mul a[4]*b[0]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebx,		0
-	; mul a[3]*b[1]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebx,		0
-	; mul a[2]*b[2]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebx,		0
-	; mul a[1]*b[3]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ebx,		0
-	; mul a[0]*b[4]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebx,		0
-	mov	DWORD PTR 16[eax],ecx
-	mov	eax,		DWORD PTR 20[esi]
-	; saved r[4]
-	; ################## Calculate word 5
-	xor	ecx,		ecx
-	; mul a[5]*b[0]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ecx,		0
-	; mul a[4]*b[1]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ecx,		0
-	; mul a[3]*b[2]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ecx,		0
-	; mul a[2]*b[3]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ecx,		0
-	; mul a[1]*b[4]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ecx,		0
-	; mul a[0]*b[5]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ecx,		0
-	mov	DWORD PTR 20[eax],ebp
-	mov	eax,		DWORD PTR 24[esi]
-	; saved r[5]
-	; ################## Calculate word 6
-	xor	ebp,		ebp
-	; mul a[6]*b[0]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebp,		0
-	; mul a[5]*b[1]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebp,		0
-	; mul a[4]*b[2]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebp,		0
-	; mul a[3]*b[3]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ebp,		0
-	; mul a[2]*b[4]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ebp,		0
-	; mul a[1]*b[5]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ebp,		0
-	; mul a[0]*b[6]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebp,		0
-	mov	DWORD PTR 24[eax],ebx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[6]
-	; ################## Calculate word 7
-	xor	ebx,		ebx
-	; mul a[7]*b[0]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebx,		0
-	; mul a[6]*b[1]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebx,		0
-	; mul a[5]*b[2]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebx,		0
-	; mul a[4]*b[3]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ebx,		0
-	; mul a[3]*b[4]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ebx,		0
-	; mul a[2]*b[5]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ebx,		0
-	; mul a[1]*b[6]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ebx,		0
-	; mul a[0]*b[7]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebx,		0
-	mov	DWORD PTR 28[eax],ecx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[7]
-	; ################## Calculate word 8
-	xor	ecx,		ecx
-	; mul a[7]*b[1]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ecx,		0
-	; mul a[6]*b[2]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ecx,		0
-	; mul a[5]*b[3]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ecx,		0
-	; mul a[4]*b[4]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ecx,		0
-	; mul a[3]*b[5]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ecx,		0
-	; mul a[2]*b[6]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ecx,		0
-	; mul a[1]*b[7]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ecx,		0
-	mov	DWORD PTR 32[eax],ebp
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[8]
-	; ################## Calculate word 9
-	xor	ebp,		ebp
-	; mul a[7]*b[2]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebp,		0
-	; mul a[6]*b[3]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ebp,		0
-	; mul a[5]*b[4]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ebp,		0
-	; mul a[4]*b[5]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ebp,		0
-	; mul a[3]*b[6]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ebp,		0
-	; mul a[2]*b[7]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebp,		0
-	mov	DWORD PTR 36[eax],ebx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[9]
-	; ################## Calculate word 10
-	xor	ebx,		ebx
-	; mul a[7]*b[3]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ebx,		0
-	; mul a[6]*b[4]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ebx,		0
-	; mul a[5]*b[5]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ebx,		0
-	; mul a[4]*b[6]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ebx,		0
-	; mul a[3]*b[7]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 16[edi]
-	adc	ebx,		0
-	mov	DWORD PTR 40[eax],ecx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[10]
-	; ################## Calculate word 11
-	xor	ecx,		ecx
-	; mul a[7]*b[4]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ecx,		0
-	; mul a[6]*b[5]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ecx,		0
-	; mul a[5]*b[6]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ecx,		0
-	; mul a[4]*b[7]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 20[edi]
-	adc	ecx,		0
-	mov	DWORD PTR 44[eax],ebp
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[11]
-	; ################## Calculate word 12
-	xor	ebp,		ebp
-	; mul a[7]*b[5]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ebp,		0
-	; mul a[6]*b[6]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ebp,		0
-	; mul a[5]*b[7]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 24[edi]
-	adc	ebp,		0
-	mov	DWORD PTR 48[eax],ebx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[12]
-	; ################## Calculate word 13
-	xor	ebx,		ebx
-	; mul a[7]*b[6]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ebx,		0
-	; mul a[6]*b[7]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 28[edi]
-	adc	ebx,		0
-	mov	DWORD PTR 52[eax],ecx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[13]
-	; ################## Calculate word 14
-	xor	ecx,		ecx
-	; mul a[7]*b[7]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	adc	ecx,		0
-	mov	DWORD PTR 56[eax],ebp
-	; saved r[14]
-	; save r[15]
-	mov	DWORD PTR 60[eax],ebx
-	pop	ebx
-	pop	ebp
-	pop	edi
-	pop	esi
-	ret
-_bn_mul_comba8 ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_mul_comba4
-
-_bn_mul_comba4 PROC NEAR
-	push	esi
-	mov	esi,		DWORD PTR 12[esp]
-	push	edi
-	mov	edi,		DWORD PTR 20[esp]
-	push	ebp
-	push	ebx
-	xor	ebx,		ebx
-	mov	eax,		DWORD PTR [esi]
-	xor	ecx,		ecx
-	mov	edx,		DWORD PTR [edi]
-	; ################## Calculate word 0
-	xor	ebp,		ebp
-	; mul a[0]*b[0]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebp,		0
-	mov	DWORD PTR [eax],ebx
-	mov	eax,		DWORD PTR 4[esi]
-	; saved r[0]
-	; ################## Calculate word 1
-	xor	ebx,		ebx
-	; mul a[1]*b[0]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebx,		0
-	; mul a[0]*b[1]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ebx,		0
-	mov	DWORD PTR 4[eax],ecx
-	mov	eax,		DWORD PTR 8[esi]
-	; saved r[1]
-	; ################## Calculate word 2
-	xor	ecx,		ecx
-	; mul a[2]*b[0]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ecx,		0
-	; mul a[1]*b[1]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ecx,		0
-	; mul a[0]*b[2]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR [edi]
-	adc	ecx,		0
-	mov	DWORD PTR 8[eax],ebp
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[2]
-	; ################## Calculate word 3
-	xor	ebp,		ebp
-	; mul a[3]*b[0]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebp,		0
-	; mul a[2]*b[1]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebp,		0
-	; mul a[1]*b[2]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR [esi]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebp,		0
-	; mul a[0]*b[3]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 4[edi]
-	adc	ebp,		0
-	mov	DWORD PTR 12[eax],ebx
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[3]
-	; ################## Calculate word 4
-	xor	ebx,		ebx
-	; mul a[3]*b[1]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebx,		0
-	; mul a[2]*b[2]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ebx,		0
-	; mul a[1]*b[3]
-	mul	edx
-	add	ecx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 8[edi]
-	adc	ebx,		0
-	mov	DWORD PTR 16[eax],ecx
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[4]
-	; ################## Calculate word 5
-	xor	ecx,		ecx
-	; mul a[3]*b[2]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ecx,		0
-	; mul a[2]*b[3]
-	mul	edx
-	add	ebp,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 12[edi]
-	adc	ecx,		0
-	mov	DWORD PTR 20[eax],ebp
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[5]
-	; ################## Calculate word 6
-	xor	ebp,		ebp
-	; mul a[3]*b[3]
-	mul	edx
-	add	ebx,		eax
-	mov	eax,		DWORD PTR 20[esp]
-	adc	ecx,		edx
-	adc	ebp,		0
-	mov	DWORD PTR 24[eax],ebx
-	; saved r[6]
-	; save r[7]
-	mov	DWORD PTR 28[eax],ecx
-	pop	ebx
-	pop	ebp
-	pop	edi
-	pop	esi
-	ret
-_bn_mul_comba4 ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_sqr_comba8
-
-_bn_sqr_comba8 PROC NEAR
-	push	esi
-	push	edi
-	push	ebp
-	push	ebx
-	mov	edi,		DWORD PTR 20[esp]
-	mov	esi,		DWORD PTR 24[esp]
-	xor	ebx,		ebx
-	xor	ecx,		ecx
-	mov	eax,		DWORD PTR [esi]
-	; ############### Calculate word 0
-	xor	ebp,		ebp
-	; sqr a[0]*a[0]
-	mul	eax
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [esi]
-	adc	ebp,		0
-	mov	DWORD PTR [edi],ebx
-	mov	eax,		DWORD PTR 4[esi]
-	; saved r[0]
-	; ############### Calculate word 1
-	xor	ebx,		ebx
-	; sqr a[1]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		0
-	mov	DWORD PTR 4[edi],ecx
-	mov	edx,		DWORD PTR [esi]
-	; saved r[1]
-	; ############### Calculate word 2
-	xor	ecx,		ecx
-	; sqr a[2]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ecx,		0
-	; sqr a[1]*a[1]
-	mul	eax
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR [esi]
-	adc	ecx,		0
-	mov	DWORD PTR 8[edi],ebp
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[2]
-	; ############### Calculate word 3
-	xor	ebp,		ebp
-	; sqr a[3]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebp,		0
-	mov	edx,		DWORD PTR 4[esi]
-	; sqr a[2]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebp,		0
-	mov	DWORD PTR 12[edi],ebx
-	mov	edx,		DWORD PTR [esi]
-	; saved r[3]
-	; ############### Calculate word 4
-	xor	ebx,		ebx
-	; sqr a[4]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebx,		0
-	mov	edx,		DWORD PTR 4[esi]
-	; sqr a[3]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		0
-	; sqr a[2]*a[2]
-	mul	eax
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR [esi]
-	adc	ebx,		0
-	mov	DWORD PTR 16[edi],ecx
-	mov	eax,		DWORD PTR 20[esi]
-	; saved r[4]
-	; ############### Calculate word 5
-	xor	ecx,		ecx
-	; sqr a[5]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ecx,		0
-	mov	edx,		DWORD PTR 4[esi]
-	; sqr a[4]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ecx,		0
-	mov	edx,		DWORD PTR 8[esi]
-	; sqr a[3]*a[2]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ecx,		0
-	mov	DWORD PTR 20[edi],ebp
-	mov	edx,		DWORD PTR [esi]
-	; saved r[5]
-	; ############### Calculate word 6
-	xor	ebp,		ebp
-	; sqr a[6]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebp,		0
-	mov	edx,		DWORD PTR 4[esi]
-	; sqr a[5]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebp,		0
-	mov	edx,		DWORD PTR 8[esi]
-	; sqr a[4]*a[2]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebp,		0
-	; sqr a[3]*a[3]
-	mul	eax
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [esi]
-	adc	ebp,		0
-	mov	DWORD PTR 24[edi],ebx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[6]
-	; ############### Calculate word 7
-	xor	ebx,		ebx
-	; sqr a[7]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebx,		0
-	mov	edx,		DWORD PTR 4[esi]
-	; sqr a[6]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebx,		0
-	mov	edx,		DWORD PTR 8[esi]
-	; sqr a[5]*a[2]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ebx,		0
-	mov	edx,		DWORD PTR 12[esi]
-	; sqr a[4]*a[3]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 28[esi]
-	adc	ebx,		0
-	mov	DWORD PTR 28[edi],ecx
-	mov	edx,		DWORD PTR 4[esi]
-	; saved r[7]
-	; ############### Calculate word 8
-	xor	ecx,		ecx
-	; sqr a[7]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ecx,		0
-	mov	edx,		DWORD PTR 8[esi]
-	; sqr a[6]*a[2]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ecx,		0
-	mov	edx,		DWORD PTR 12[esi]
-	; sqr a[5]*a[3]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 16[esi]
-	adc	ecx,		0
-	; sqr a[4]*a[4]
-	mul	eax
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR 8[esi]
-	adc	ecx,		0
-	mov	DWORD PTR 32[edi],ebp
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[8]
-	; ############### Calculate word 9
-	xor	ebp,		ebp
-	; sqr a[7]*a[2]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebp,		0
-	mov	edx,		DWORD PTR 12[esi]
-	; sqr a[6]*a[3]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebp,		0
-	mov	edx,		DWORD PTR 16[esi]
-	; sqr a[5]*a[4]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 28[esi]
-	adc	ebp,		0
-	mov	DWORD PTR 36[edi],ebx
-	mov	edx,		DWORD PTR 12[esi]
-	; saved r[9]
-	; ############### Calculate word 10
-	xor	ebx,		ebx
-	; sqr a[7]*a[3]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebx,		0
-	mov	edx,		DWORD PTR 16[esi]
-	; sqr a[6]*a[4]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 20[esi]
-	adc	ebx,		0
-	; sqr a[5]*a[5]
-	mul	eax
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 16[esi]
-	adc	ebx,		0
-	mov	DWORD PTR 40[edi],ecx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[10]
-	; ############### Calculate word 11
-	xor	ecx,		ecx
-	; sqr a[7]*a[4]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ecx,		0
-	mov	edx,		DWORD PTR 20[esi]
-	; sqr a[6]*a[5]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 28[esi]
-	adc	ecx,		0
-	mov	DWORD PTR 44[edi],ebp
-	mov	edx,		DWORD PTR 20[esi]
-	; saved r[11]
-	; ############### Calculate word 12
-	xor	ebp,		ebp
-	; sqr a[7]*a[5]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 24[esi]
-	adc	ebp,		0
-	; sqr a[6]*a[6]
-	mul	eax
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR 24[esi]
-	adc	ebp,		0
-	mov	DWORD PTR 48[edi],ebx
-	mov	eax,		DWORD PTR 28[esi]
-	; saved r[12]
-	; ############### Calculate word 13
-	xor	ebx,		ebx
-	; sqr a[7]*a[6]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 28[esi]
-	adc	ebx,		0
-	mov	DWORD PTR 52[edi],ecx
-	; saved r[13]
-	; ############### Calculate word 14
-	xor	ecx,		ecx
-	; sqr a[7]*a[7]
-	mul	eax
-	add	ebp,		eax
-	adc	ebx,		edx
-	adc	ecx,		0
-	mov	DWORD PTR 56[edi],ebp
-	; saved r[14]
-	mov	DWORD PTR 60[edi],ebx
-	pop	ebx
-	pop	ebp
-	pop	edi
-	pop	esi
-	ret
-_bn_sqr_comba8 ENDP
-_TEXT	ENDS
-_TEXT	SEGMENT
-PUBLIC	_bn_sqr_comba4
-
-_bn_sqr_comba4 PROC NEAR
-	push	esi
-	push	edi
-	push	ebp
-	push	ebx
-	mov	edi,		DWORD PTR 20[esp]
-	mov	esi,		DWORD PTR 24[esp]
-	xor	ebx,		ebx
-	xor	ecx,		ecx
-	mov	eax,		DWORD PTR [esi]
-	; ############### Calculate word 0
-	xor	ebp,		ebp
-	; sqr a[0]*a[0]
-	mul	eax
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	edx,		DWORD PTR [esi]
-	adc	ebp,		0
-	mov	DWORD PTR [edi],ebx
-	mov	eax,		DWORD PTR 4[esi]
-	; saved r[0]
-	; ############### Calculate word 1
-	xor	ebx,		ebx
-	; sqr a[1]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		0
-	mov	DWORD PTR 4[edi],ecx
-	mov	edx,		DWORD PTR [esi]
-	; saved r[1]
-	; ############### Calculate word 2
-	xor	ecx,		ecx
-	; sqr a[2]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 4[esi]
-	adc	ecx,		0
-	; sqr a[1]*a[1]
-	mul	eax
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	edx,		DWORD PTR [esi]
-	adc	ecx,		0
-	mov	DWORD PTR 8[edi],ebp
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[2]
-	; ############### Calculate word 3
-	xor	ebp,		ebp
-	; sqr a[3]*a[0]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebp,		0
-	mov	edx,		DWORD PTR 4[esi]
-	; sqr a[2]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebp,		0
-	add	ebx,		eax
-	adc	ecx,		edx
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ebp,		0
-	mov	DWORD PTR 12[edi],ebx
-	mov	edx,		DWORD PTR 4[esi]
-	; saved r[3]
-	; ############### Calculate word 4
-	xor	ebx,		ebx
-	; sqr a[3]*a[1]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ebx,		0
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	eax,		DWORD PTR 8[esi]
-	adc	ebx,		0
-	; sqr a[2]*a[2]
-	mul	eax
-	add	ecx,		eax
-	adc	ebp,		edx
-	mov	edx,		DWORD PTR 8[esi]
-	adc	ebx,		0
-	mov	DWORD PTR 16[edi],ecx
-	mov	eax,		DWORD PTR 12[esi]
-	; saved r[4]
-	; ############### Calculate word 5
-	xor	ecx,		ecx
-	; sqr a[3]*a[2]
-	mul	edx
-	add	eax,		eax
-	adc	edx,		edx
-	adc	ecx,		0
-	add	ebp,		eax
-	adc	ebx,		edx
-	mov	eax,		DWORD PTR 12[esi]
-	adc	ecx,		0
-	mov	DWORD PTR 20[edi],ebp
-	; saved r[5]
-	; ############### Calculate word 6
-	xor	ebp,		ebp
-	; sqr a[3]*a[3]
-	mul	eax
-	add	ebx,		eax
-	adc	ecx,		edx
-	adc	ebp,		0
-	mov	DWORD PTR 24[edi],ebx
-	; saved r[6]
-	mov	DWORD PTR 28[edi],ecx
-	pop	ebx
-	pop	ebp
-	pop	edi
-	pop	esi
-	ret
-_bn_sqr_comba4 ENDP
-_TEXT	ENDS
-END
diff --git a/lib/libcrypto/bn/asm/x86w16.asm b/lib/libcrypto/bn/asm/x86w16.asm
index 80a9ed6eef1..e69de29bb2d 100644
--- a/lib/libcrypto/bn/asm/x86w16.asm
+++ b/lib/libcrypto/bn/asm/x86w16.asm
@@ -1,297 +0,0 @@
-;	Static Name Aliases
-;
-	TITLE   bn_mulw.c
-	.8087
-F_TEXT	SEGMENT  WORD PUBLIC 'CODE'
-F_TEXT	ENDS
-_DATA	SEGMENT  WORD PUBLIC 'DATA'
-_DATA	ENDS
-_CONST	SEGMENT  WORD PUBLIC 'CONST'
-_CONST	ENDS
-_BSS	SEGMENT  WORD PUBLIC 'BSS'
-_BSS	ENDS
-DGROUP	GROUP	_CONST, _BSS, _DATA
-	ASSUME DS: DGROUP, SS: DGROUP
-F_TEXT      SEGMENT
-	ASSUME	CS: F_TEXT
-	PUBLIC	_bn_mul_add_words
-_bn_mul_add_words	PROC FAR
-; Line 58
-	push	bp
-	push	bx
-	push	si
-	push	di
-	push	ds
-	push	es
-	mov	bp,sp
-;	w = 26
-;	num = 24
-;	ap = 20
-;	rp = 16
-	xor	si,si			;c=0;
-	mov	di,WORD PTR [bp+16]	; load r
-	mov	ds,WORD PTR [bp+18]	; load r
-	mov	bx,WORD PTR [bp+20]	; load a
-	mov	es,WORD PTR [bp+22]	; load a
-	mov	cx,WORD PTR [bp+26]	; load w
-	mov	bp,WORD PTR [bp+24]	; load num
-
-	shr	bp,1	; div count by 4 and do groups of 4
-	shr	bp,1
-	je	$L555
-
-$L546:
-	mov	ax,cx
-	mul	WORD PTR es:[bx]	; w* *a
-	add	ax,WORD PTR ds:[di]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di],ax
-	mov	si,dx
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+2]	; w* *a
-	add	ax,WORD PTR ds:[di+2]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+2],ax
-	mov	si,dx
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+4]	; w* *a
-	add	ax,WORD PTR ds:[di+4]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+4],ax
-	mov	si,dx
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+6]	; w* *a
-	add	ax,WORD PTR ds:[di+6]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+6],ax
-	mov	si,dx
-	;
-	add	bx,8
-	add	di,8
-	;
-	dec	bp
-	je	$L555
-	jmp	$L546
-;
-;
-$L555:
-	mov	bp,sp
-	mov	bp,WORD PTR [bp+24]	; load num
-	and	bp,3
-	dec	bp
-	js	$L547
-
-	mov	ax,cx
-	mul	WORD PTR es:[bx]	; w* *a
-	add	ax,WORD PTR ds:[di]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di],ax
-	mov	si,dx
-	dec	bp
-	js	$L547			; Note that we are now testing for -1
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+2]	; w* *a
-	add	ax,WORD PTR ds:[di+2]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+2],ax
-	mov	si,dx
-	dec	bp
-	js	$L547
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+4]	; w* *a
-	add	ax,WORD PTR ds:[di+4]	; + *r
-	adc	dx,0
-	adc	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+4],ax
-	mov	si,dx
-$L547:
-	mov	ax,si
-	pop	es
-	pop	ds
-	pop	di
-	pop	si
-	pop	bx
-	pop	bp
-	ret	
-	nop	
-
-_bn_mul_add_words	ENDP
-	PUBLIC	_bn_mul_words
-_bn_mul_words	PROC FAR
-; Line 76
-	push	bp
-	push	bx
-	push	si
-	push	di
-	push	ds
-	push	es
-	xor	si,si
-	mov	bp,sp
-	mov	di,WORD PTR [bp+16]	; r
-	mov	ds,WORD PTR [bp+18]
-	mov	bx,WORD PTR [bp+20]	; a
-	mov	es,WORD PTR [bp+22]
-	mov	cx,WORD PTR [bp+26]	; w
-	mov	bp,WORD PTR [bp+24]	; num 
-$FC743:
-	mov	ax,cx
-	mul	WORD PTR es:[bx]
-	add	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di],ax
-	mov	si,dx
-	dec	bp
-	je	$L764
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+2]
-	add	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+2],ax
-	mov	si,dx
-	dec	bp
-	je	$L764
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+4]
-	add	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+4],ax
-	mov	si,dx
-	dec	bp
-	je	$L764
-	;
-	mov	ax,cx
-	mul	WORD PTR es:[bx+6]
-	add	ax,si
-	adc	dx,0
-	mov	WORD PTR ds:[di+6],ax
-	mov	si,dx
-	dec	bp
-	je	$L764
-	;
-	add	bx,8
-	add	di,8
-	jmp	$FC743
-	nop
-$L764:
-	mov	ax,si
-	pop	es
-	pop	ds
-	pop	di
-	pop	si
-	pop	bx
-	pop	bp
-	ret	
-	nop	
-_bn_mul_words	ENDP
-	PUBLIC	_bn_sqr_words
-_bn_sqr_words	PROC FAR
-; Line 92
-	push	bp
-	push	bx
-	push	si
-	push	di
-	push	ds
-	push	es
-	mov	bp,sp
-	mov	si,WORD PTR [bp+16]
-	mov	ds,WORD PTR [bp+18]
-	mov	di,WORD PTR [bp+20]
-	mov	es,WORD PTR [bp+22]
-	mov	bx,WORD PTR [bp+24]
-
-	mov	bp,bx	; save a memory lookup later
-	shr	bx,1	; div count by 4 and do groups of 4
-	shr	bx,1
-	je	$L666
-
-$L765:
-	mov	ax,WORD PTR es:[di]
-	mul	ax
-	mov	WORD PTR ds:[si],ax
-	mov	WORD PTR ds:[si+2],dx
-	;
-	mov	ax,WORD PTR es:[di+2]
-	mul	ax
-	mov	WORD PTR ds:[si+4],ax
-	mov	WORD PTR ds:[si+6],dx
-	;
-	mov	ax,WORD PTR es:[di+4]
-	mul	ax
-	mov	WORD PTR ds:[si+8],ax
-	mov	WORD PTR ds:[si+10],dx
-	;
-	mov	ax,WORD PTR es:[di+6]
-	mul	ax
-	mov	WORD PTR ds:[si+12],ax
-	mov	WORD PTR ds:[si+14],dx
-	;
-	add	di,8
-	add	si,16
-	dec	bx
-	je	$L666
-	jmp	$L765
-$L666:
-	and	bp,3
-	dec	bp	; The copied value of bx (num)
-	js	$L645
-	;
-	mov	ax,WORD PTR es:[di]
-	mul	ax
-	mov	WORD PTR ds:[si],ax
-	mov	WORD PTR ds:[si+2],dx
-	dec	bp
-	js	$L645
-	;
-	mov	ax,WORD PTR es:[di+2]
-	mul	ax
-	mov	WORD PTR ds:[si+4],ax
-	mov	WORD PTR ds:[si+6],dx
-	dec	bp
-	js	$L645
-	;
-	mov	ax,WORD PTR es:[di+4]
-	mul	ax
-	mov	WORD PTR ds:[si+8],ax
-	mov	WORD PTR ds:[si+10],dx
-$L645:
-	pop	es
-	pop	ds
-	pop	di
-	pop	si
-	pop	bx
-	pop	bp
-	ret	
-
-_bn_sqr_words	ENDP
-	PUBLIC	_bn_div64
-_bn_div64	PROC FAR
-	push	bp
-	mov	bp,sp
-	mov	dx, WORD PTR [bp+6]
-	mov	ax, WORD PTR [bp+8]
-	div	WORD PTR [bp+10]
-	pop	bp
-	ret	
-_bn_div64	ENDP
-F_TEXT	ENDS
-END
diff --git a/lib/libcrypto/bn/asm/x86w32.asm b/lib/libcrypto/bn/asm/x86w32.asm
index 957d71e3b18..e69de29bb2d 100644
--- a/lib/libcrypto/bn/asm/x86w32.asm
+++ b/lib/libcrypto/bn/asm/x86w32.asm
@@ -1,360 +0,0 @@
-;	Static Name Aliases
-;
-	TITLE   bn_mulw.c
-	.386
-F_TEXT	SEGMENT  WORD USE16 PUBLIC 'CODE'
-F_TEXT	ENDS
-_DATA	SEGMENT  WORD USE16 PUBLIC 'DATA'
-_DATA	ENDS
-_CONST	SEGMENT  WORD USE16 PUBLIC 'CONST'
-_CONST	ENDS
-_BSS	SEGMENT  WORD USE16 PUBLIC 'BSS'
-_BSS	ENDS
-DGROUP	GROUP	_CONST, _BSS, _DATA
-	ASSUME DS: DGROUP, SS: DGROUP
-F_TEXT      SEGMENT
-	ASSUME	CS: F_TEXT
-	PUBLIC	_bn_mul_add_words
-_bn_mul_add_words	PROC FAR
-; Line 58
-	push	bp
-	push	bx
-	push	esi
-	push	di
-	push	ds
-	push	es
-	mov	bp,sp
-;	w = 28
-;	num = 26
-;	ap = 22
-;	rp = 18
-	xor	esi,esi			;c=0;
-	mov	di,WORD PTR [bp+18]	; load r
-	mov	ds,WORD PTR [bp+20]	; load r
-	mov	bx,WORD PTR [bp+22]	; load a
-	mov	es,WORD PTR [bp+24]	; load a
-	mov	ecx,DWORD PTR [bp+28]	; load w
-	mov	bp,WORD PTR [bp+26]	; load num
-	shr	bp,1	; div count by 4 and do groups of 4
-	shr	bp,1
-	je	$L555
-
-$L546:
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx]	; w* *a
-	add	eax,DWORD PTR ds:[di]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di],eax
-	mov	esi,edx
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+4]	; w* *a
-	add	eax,DWORD PTR ds:[di+4]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+4],eax
-	mov	esi,edx
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+8]	; w* *a
-	add	eax,DWORD PTR ds:[di+8]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+8],eax
-	mov	esi,edx
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+12]	; w* *a
-	add	eax,DWORD PTR ds:[di+12]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+12],eax
-	mov	esi,edx
-	;
-	add	bx,16
-	add	di,16
-	;
-	dec	bp
-	je	$L555
-	jmp	$L546
-;
-;
-$L555:
-	mov	bp,sp
-	mov	bp,WORD PTR [bp+26]	; load num
-	and	bp,3
-	dec	bp
-	js	$L547m
-
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx]	; w* *a
-	add	eax,DWORD PTR ds:[di]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di],eax
-	mov	esi,edx
-	dec	bp
-	js	$L547m			; Note that we are now testing for -1
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+4]	; w* *a
-	add	eax,DWORD PTR ds:[di+4]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+4],eax
-	mov	esi,edx
-	dec	bp
-	js	$L547m
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+8]	; w* *a
-	add	eax,DWORD PTR ds:[di+8]	; + *r
-	adc	edx,0
-	adc	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+8],eax
-	mov	esi,edx
-$L547m:
-	mov	eax,esi
-	mov	edx,esi
-	shr	edx,16
-	pop	es
-	pop	ds
-	pop	di
-	pop	esi
-	pop	bx
-	pop	bp
-	ret	
-	nop	
-_bn_mul_add_words	ENDP
-
-	PUBLIC	_bn_mul_words
-_bn_mul_words	PROC FAR
-; Line 76
-	push	bp
-	push	bx
-	push	esi
-	push	di
-	push	ds
-	push	es
-	xor	esi,esi
-	mov	bp,sp
-	mov	di,WORD PTR [bp+18]	; r
-	mov	ds,WORD PTR [bp+20]
-	mov	bx,WORD PTR [bp+22]	; a
-	mov	es,WORD PTR [bp+24]
-	mov	ecx,DWORD PTR [bp+28]	; w
-	mov	bp,WORD PTR [bp+26]	; num 
-
-$FC743:
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx]
-	add	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di],eax
-	mov	esi,edx
-	dec	bp
-	je	$L764
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+4]
-	add	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+4],eax
-	mov	esi,edx
-	dec	bp
-	je	$L764
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+8]
-	add	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+8],eax
-	mov	esi,edx
-	dec	bp
-	je	$L764
-	;
-	mov	eax,ecx
-	mul	DWORD PTR es:[bx+12]
-	add	eax,esi
-	adc	edx,0
-	mov	DWORD PTR ds:[di+12],eax
-	mov	esi,edx
-	dec	bp
-	je	$L764
-	;
-	add	bx,16
-	add	di,16
-	jmp	$FC743
-	nop
-$L764:
-	mov	eax,esi
-	mov	edx,esi
-	shr	edx,16
-	pop	es
-	pop	ds
-	pop	di
-	pop	esi
-	pop	bx
-	pop	bp
-	ret	
-	nop	
-_bn_mul_words	ENDP
-	PUBLIC	_bn_sqr_words
-_bn_sqr_words	PROC FAR
-; Line 92
-	push	bp
-	push	bx
-	push	si
-	push	di
-	push	ds
-	push	es
-	mov	bp,sp
-	mov	si,WORD PTR [bp+16]
-	mov	ds,WORD PTR [bp+18]
-	mov	di,WORD PTR [bp+20]
-	mov	es,WORD PTR [bp+22]
-	mov	bx,WORD PTR [bp+24]
-
-	mov	bp,bx	; save a memory lookup later
-	shr	bx,1	; div count by 4 and do groups of 4
-	shr	bx,1
-	je	$L666
-
-$L765:
-	mov	eax,DWORD PTR es:[di]
-	mul	eax
-	mov	DWORD PTR ds:[si],eax
-	mov	DWORD PTR ds:[si+4],edx
-	;
-	mov	eax,DWORD PTR es:[di+4]
-	mul	eax
-	mov	DWORD PTR ds:[si+8],eax
-	mov	DWORD PTR ds:[si+12],edx
-	;
-	mov	eax,DWORD PTR es:[di+8]
-	mul	eax
-	mov	DWORD PTR ds:[si+16],eax
-	mov	DWORD PTR ds:[si+20],edx
-	;
-	mov	eax,DWORD PTR es:[di+12]
-	mul	eax
-	mov	DWORD PTR ds:[si+24],eax
-	mov	DWORD PTR ds:[si+28],edx
-	;
-	add	di,16
-	add	si,32
-	dec	bx
-	je	$L666
-	jmp	$L765
-$L666:
-	and	bp,3
-	dec	bp	; The copied value of bx (num)
-	js	$L645
-	;
-	mov	eax,DWORD PTR es:[di]
-	mul	eax
-	mov	DWORD PTR ds:[si],eax
-	mov	DWORD PTR ds:[si+4],edx
-	dec	bp
-	js	$L645
-	;
-	mov	eax,DWORD PTR es:[di+4]
-	mul	eax
-	mov	DWORD PTR ds:[si+8],eax
-	mov	DWORD PTR ds:[si+12],edx
-	dec	bp
-	js	$L645
-	;
-	mov	eax,DWORD PTR es:[di+8]
-	mul	eax
-	mov	DWORD PTR ds:[si+16],eax
-	mov	DWORD PTR ds:[si+20],edx
-$L645:
-	pop	es
-	pop	ds
-	pop	di
-	pop	si
-	pop	bx
-	pop	bp
-	ret	
-_bn_sqr_words	ENDP
-
-	PUBLIC	_bn_div64
-_bn_div64	PROC FAR
-	push	bp
-	mov	bp,sp
-	mov	edx, DWORD PTR [bp+6]
-	mov	eax, DWORD PTR [bp+10]
-	div	DWORD PTR [bp+14]
-	mov	edx,eax
-	shr	edx,16
-	pop	bp
-	ret	
-_bn_div64	ENDP
-
-	PUBLIC	_bn_add_words
-_bn_add_words	PROC FAR
-; Line 58
-	push	bp
-	push	bx
-	push	esi
-	push	di
-	push	ds
-	push	es
-	mov	bp,sp
-;	w = 28
-;	num = 26
-;	ap = 22
-;	rp = 18
-	xor	esi,esi			;c=0;
-	mov	bx,WORD PTR [bp+18]	; load low r
-	mov	si,WORD PTR [bp+22]	; load a
-	mov	es,WORD PTR [bp+24]	; load a
-	mov	di,WORD PTR [bp+26]	; load b
-	mov	ds,WORD PTR [bp+28]	; load b
-
-	mov	dx,WORD PTR [bp+30]	; load num
-	xor	ecx,ecx
-	dec	dx
-	js	$L547a
-
-$L5477:
-	mov	eax,DWORD PTR es:[si]	; *a
-	add	eax,ecx
-	mov	ecx,0
-	adc	ecx,0
-	add	si,4			; a++
-	add	eax,DWORD PTR ds:[di]	; + *b
-	adc	ecx,0
-	mov	ds,WORD PTR [bp+20]
-	add	di,4
-	mov	DWORD PTR ds:[bx],eax
-	mov	ds,WORD PTR [bp+28]
-	add	bx,4
-	dec	dx
-	js	$L547a			; Note that we are now testing for -1
-	jmp	$L5477
-	;
-$L547a:
-	mov	eax,ecx
-	mov	edx,ecx
-	shr	edx,16
-	pop	es
-	pop	ds
-	pop	di
-	pop	esi
-	pop	bx
-	pop	bp
-	ret	
-	nop	
-_bn_add_words	ENDP
-F_TEXT	ENDS
-END
diff --git a/lib/libcrypto/bn/bn.h b/lib/libcrypto/bn/bn.h
index d8822610dfa..009b0eb6856 100644
--- a/lib/libcrypto/bn/bn.h
+++ b/lib/libcrypto/bn/bn.h
@@ -89,7 +89,7 @@ extern "C" {
  * For machines with only one compiler (or shared libraries), this should
  * be on.  Again this in only really a problem on machines
  * using "long long's", are 32bit, and are not using my assembler code. */
-#if defined(MSDOS) || defined(WINDOWS) || defined(linux)
+#if defined(MSDOS) || defined(WINDOWS) || defined(WIN32) || defined(linux)
 #define BN_DIV2W
 #endif
 
diff --git a/lib/libcrypto/bn/bn_err.c b/lib/libcrypto/bn/bn_err.c
index f3b9497dca0..988270bcf4f 100644
--- a/lib/libcrypto/bn/bn_err.c
+++ b/lib/libcrypto/bn/bn_err.c
@@ -54,7 +54,8 @@
  */
 
 /* NOTE: this file was auto generated by the mkerr.pl script: any changes
- * made to it will be overwritten when the script next updates this file.
+ * made to it will be overwritten when the script next updates this file,
+ * only reason strings will be preserved.
  */
 
 #include <stdio.h>
diff --git a/lib/libcrypto/bn/bn_mont.c b/lib/libcrypto/bn/bn_mont.c
index 7bb0b91223c..598fecbf0c8 100644
--- a/lib/libcrypto/bn/bn_mont.c
+++ b/lib/libcrypto/bn/bn_mont.c
@@ -73,6 +73,7 @@ int BN_mod_mul_montgomery(BIGNUM *r, BIGNUM *a, BIGNUM *b,
 			  BN_MONT_CTX *mont, BN_CTX *ctx)
 	{
 	BIGNUM *tmp,*tmp2;
+	int ret=0;
 
 	BN_CTX_start(ctx);
 	tmp = BN_CTX_get(ctx);
@@ -101,10 +102,10 @@ int BN_mod_mul_montgomery(BIGNUM *r, BIGNUM *a, BIGNUM *b,
 		}
 	/* reduce from aRR to aR */
 	if (!BN_from_montgomery(r,tmp,mont,ctx)) goto err;
-	BN_CTX_end(ctx);
-	return(1);
+	ret=1;
 err:
-	return(0);
+	BN_CTX_end(ctx);
+	return(ret);
 	}
 
 int BN_from_montgomery(BIGNUM *ret, BIGNUM *a, BN_MONT_CTX *mont,
diff --git a/lib/libcrypto/bn/bn_mul.c b/lib/libcrypto/bn/bn_mul.c
index eb007e19e9a..3e8baaad9a0 100644
--- a/lib/libcrypto/bn/bn_mul.c
+++ b/lib/libcrypto/bn/bn_mul.c
@@ -61,6 +61,9 @@
 #include "bn_lcl.h"
 
 #ifdef BN_RECURSION
+/* Karatsuba recursive multiplication algorithm
+ * (cf. Knuth, The Art of Computer Programming, Vol. 2) */
+
 /* r is 2*n2 words in size,
  * a and b are both n2 words in size.
  * n2 must be a power of 2.
diff --git a/lib/libcrypto/bn/bntest.c b/lib/libcrypto/bn/bntest.c
index 41c22f5954d..0a97af69c58 100644
--- a/lib/libcrypto/bn/bntest.c
+++ b/lib/libcrypto/bn/bntest.c
@@ -234,7 +234,8 @@ int main(int argc, char *argv[])
 /**/
 	exit(0);
 err:
-	BIO_puts(out,"1\n"); /* make sure bc fails if we are piping to it */
+	BIO_puts(out,"1\n"); /* make sure the Perl script fed by bc notices
+	                      * the failure, see test_bn in test/Makefile.ssl*/
 	BIO_flush(out);
 	ERR_load_crypto_strings();
 	ERR_print_errors_fp(stderr);