bn-win32.asm

老外写的加密库cryptlib(版本3.1)

	mov	edi,		[20+esp]
	push	ebp
	push	ebx
	xor	ebx,		ebx
	mov	eax,		[esi]
	xor	ecx,		ecx
	mov	edx,		[edi]
	; ################## Calculate word 0
	xor	ebp,		ebp
	; mul a[0]*b[0]
	mul	edx
	add	ebx,		eax
	mov	eax,		[20+esp]
	adc	ecx,		edx
	mov	edx,		[edi]
	adc	ebp,		0
	mov	[eax],		ebx
	mov	eax,		[4+esi]
	; saved r[0]
	; ################## Calculate word 1
	xor	ebx,		ebx
	; mul a[1]*b[0]
	mul	edx
	add	ecx,		eax
	mov	eax,		[esi]
	adc	ebp,		edx
	mov	edx,		[4+edi]
	adc	ebx,		0
	; mul a[0]*b[1]
	mul	edx
	add	ecx,		eax
	mov	eax,		[20+esp]
	adc	ebp,		edx
	mov	edx,		[edi]
	adc	ebx,		0
	mov	[4+eax],	ecx
	mov	eax,		[8+esi]
	; saved r[1]
	; ################## Calculate word 2
	xor	ecx,		ecx
	; mul a[2]*b[0]
	mul	edx
	add	ebp,		eax
	mov	eax,		[4+esi]
	adc	ebx,		edx
	mov	edx,		[4+edi]
	adc	ecx,		0
	; mul a[1]*b[1]
	mul	edx
	add	ebp,		eax
	mov	eax,		[esi]
	adc	ebx,		edx
	mov	edx,		[8+edi]
	adc	ecx,		0
	; mul a[0]*b[2]
	mul	edx
	add	ebp,		eax
	mov	eax,		[20+esp]
	adc	ebx,		edx
	mov	edx,		[edi]
	adc	ecx,		0
	mov	[8+eax],	ebp
	mov	eax,		[12+esi]
	; saved r[2]
	; ################## Calculate word 3
	xor	ebp,		ebp
	; mul a[3]*b[0]
	mul	edx
	add	ebx,		eax
	mov	eax,		[8+esi]
	adc	ecx,		edx
	mov	edx,		[4+edi]
	adc	ebp,		0
	; mul a[2]*b[1]
	mul	edx
	add	ebx,		eax
	mov	eax,		[4+esi]
	adc	ecx,		edx
	mov	edx,		[8+edi]
	adc	ebp,		0
	; mul a[1]*b[2]
	mul	edx
	add	ebx,		eax
	mov	eax,		[esi]
	adc	ecx,		edx
	mov	edx,		[12+edi]
	adc	ebp,		0
	; mul a[0]*b[3]
	mul	edx
	add	ebx,		eax
	mov	eax,		[20+esp]
	adc	ecx,		edx
	mov	edx,		[4+edi]
	adc	ebp,		0
	mov	[12+eax],	ebx
	mov	eax,		[12+esi]
	; saved r[3]
	; ################## Calculate word 4
	xor	ebx,		ebx
	; mul a[3]*b[1]
	mul	edx
	add	ecx,		eax
	mov	eax,		[8+esi]
	adc	ebp,		edx
	mov	edx,		[8+edi]
	adc	ebx,		0
	; mul a[2]*b[2]
	mul	edx
	add	ecx,		eax
	mov	eax,		[4+esi]
	adc	ebp,		edx
	mov	edx,		[12+edi]
	adc	ebx,		0
	; mul a[1]*b[3]
	mul	edx
	add	ecx,		eax
	mov	eax,		[20+esp]
	adc	ebp,		edx
	mov	edx,		[8+edi]
	adc	ebx,		0
	mov	[16+eax],	ecx
	mov	eax,		[12+esi]
	; saved r[4]
	; ################## Calculate word 5
	xor	ecx,		ecx
	; mul a[3]*b[2]
	mul	edx
	add	ebp,		eax
	mov	eax,		[8+esi]
	adc	ebx,		edx
	mov	edx,		[12+edi]
	adc	ecx,		0
	; mul a[2]*b[3]
	mul	edx
	add	ebp,		eax
	mov	eax,		[20+esp]
	adc	ebx,		edx
	mov	edx,		[12+edi]
	adc	ecx,		0
	mov	[20+eax],	ebp
	mov	eax,		[12+esi]
	; saved r[5]
	; ################## Calculate word 6
	xor	ebp,		ebp
	; mul a[3]*b[3]
	mul	edx
	add	ebx,		eax
	mov	eax,		[20+esp]
	adc	ecx,		edx
	adc	ebp,		0
	mov	[24+eax],	ebx
	; saved r[6]
	; save r[7]
	mov	[28+eax],	ecx
	pop	ebx
	pop	ebp
	pop	edi
	pop	esi
	ret
global	_bn_sqr_comba8
_bn_sqr_comba8:
	push	esi
	push	edi
	push	ebp
	push	ebx
	mov	edi,		[20+esp]
	mov	esi,		[24+esp]
	xor	ebx,		ebx
	xor	ecx,		ecx
	mov	eax,		[esi]
	; ############### Calculate word 0
	xor	ebp,		ebp
	; sqr a[0]*a[0]
	mul	eax
	add	ebx,		eax
	adc	ecx,		edx
	mov	edx,		[esi]
	adc	ebp,		0
	mov	[edi],		ebx
	mov	eax,		[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,		[8+esi]
	adc	ebx,		0
	mov	[4+edi],	ecx
	mov	edx,		[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,		[4+esi]
	adc	ecx,		0
	; sqr a[1]*a[1]
	mul	eax
	add	ebp,		eax
	adc	ebx,		edx
	mov	edx,		[esi]
	adc	ecx,		0
	mov	[8+edi],	ebp
	mov	eax,		[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,		[8+esi]
	adc	ebp,		0
	mov	edx,		[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,		[16+esi]
	adc	ebp,		0
	mov	[12+edi],	ebx
	mov	edx,		[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,		[12+esi]
	adc	ebx,		0
	mov	edx,		[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,		[8+esi]
	adc	ebx,		0
	; sqr a[2]*a[2]
	mul	eax
	add	ecx,		eax
	adc	ebp,		edx
	mov	edx,		[esi]
	adc	ebx,		0
	mov	[16+edi],	ecx
	mov	eax,		[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,		[16+esi]
	adc	ecx,		0
	mov	edx,		[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,		[12+esi]
	adc	ecx,		0
	mov	edx,		[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,		[24+esi]
	adc	ecx,		0
	mov	[20+edi],	ebp
	mov	edx,		[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,		[20+esi]
	adc	ebp,		0
	mov	edx,		[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,		[16+esi]
	adc	ebp,		0
	mov	edx,		[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,		[12+esi]
	adc	ebp,		0
	; sqr a[3]*a[3]
	mul	eax
	add	ebx,		eax
	adc	ecx,		edx
	mov	edx,		[esi]
	adc	ebp,		0
	mov	[24+edi],	ebx
	mov	eax,		[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,		[24+esi]
	adc	ebx,		0
	mov	edx,		[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,		[20+esi]
	adc	ebx,		0
	mov	edx,		[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,		[16+esi]
	adc	ebx,		0
	mov	edx,		[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,		[28+esi]
	adc	ebx,		0
	mov	[28+edi],	ecx
	mov	edx,		[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,		[24+esi]
	adc	ecx,		0
	mov	edx,		[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,		[20+esi]
	adc	ecx,		0
	mov	edx,		[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,		[16+esi]
	adc	ecx,		0
	; sqr a[4]*a[4]
	mul	eax
	add	ebp,		eax
	adc	ebx,		edx
	mov	edx,		[8+esi]
	adc	ecx,		0
	mov	[32+edi],	ebp
	mov	eax,		[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,		[24+esi]
	adc	ebp,		0
	mov	edx,		[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,		[20+esi]
	adc	ebp,		0
	mov	edx,		[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,		[28+esi]
	adc	ebp,		0
	mov	[36+edi],	ebx
	mov	edx,		[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,		[24+esi]
	adc	ebx,		0
	mov	edx,		[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,		[20+esi]
	adc	ebx,		0
	; sqr a[5]*a[5]
	mul	eax
	add	ecx,		eax
	adc	ebp,		edx
	mov	edx,		[16+esi]
	adc	ebx,		0
	mov	[40+edi],	ecx
	mov	eax,		[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,		[24+esi]
	adc	ecx,		0
	mov	edx,		[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,		[28+esi]
	adc	ecx,		0
	mov	[44+edi],	ebp
	mov	edx,		[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,		[24+esi]
	adc	ebp,		0
	; sqr a[6]*a[6]
	mul	eax
	add	ebx,		eax
	adc	ecx,		edx
	mov	edx,		[24+esi]
	adc	ebp,		0
	mov	[48+edi],	ebx
	mov	eax,		[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,		[28+esi]
	adc	ebx,		0
	mov	[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	[56+edi],	ebp
	; saved r[14]
	mov	[60+edi],	ebx
	pop	ebx
	pop	ebp
	pop	edi
	pop	esi
	ret
global	_bn_sqr_comba4
_bn_sqr_comba4:
	push	esi
	push	edi
	push	ebp
	push	ebx
	mov	edi,		[20+esp]
	mov	esi,		[24+esp]
	xor	ebx,		ebx
	xor	ecx,		ecx
	mov	eax,		[esi]
	; ############### Calculate word 0
	xor	ebp,		ebp
	; sqr a[0]*a[0]
	mul	eax
	add	ebx,		eax
	adc	ecx,		edx
	mov	edx,		[esi]
	adc	ebp,		0
	mov	[edi],		ebx
	mov	eax,		[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,		[8+esi]
	adc	ebx,		0
	mov	[4+edi],	ecx
	mov	edx,		[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,		[4+esi]
	adc	ecx,		0
	; sqr a[1]*a[1]
	mul	eax
	add	ebp,		eax
	adc	ebx,		edx
	mov	edx,		[esi]
	adc	ecx,		0
	mov	[8+edi],	ebp
	mov	eax,		[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,		[8+esi]
	adc	ebp,		0
	mov	edx,		[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,		[12+esi]
	adc	ebp,		0
	mov	[12+edi],	ebx
	mov	edx,		[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,		[8+esi]
	adc	ebx,		0
	; sqr a[2]*a[2]
	mul	eax
	add	ecx,		eax
	adc	ebp,		edx
	mov	edx,		[8+esi]
	adc	ebx,		0
	mov	[16+edi],	ecx
	mov	eax,		[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,		[12+esi]
	adc	ecx,		0
	mov	[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	[24+edi],	ebx
	; saved r[6]
	mov	[28+edi],	ecx
	pop	ebx
	pop	ebp
	pop	edi
	pop	esi
	ret