vms.mar

優化過的openssl加解密引擎,適合嵌入式開發環境使用,非常容易移植!

	.title	vax_bn_mul_add_word  unsigned multiply & add, 32*32+32+32=>64
;
; w.j.m. 15-jan-1999
;
; it's magic ...
;
; ULONG bn_mul_add_words(ULONG r[],ULONG a[],int n,ULONG w) {
;	ULONG c = 0;
;	int i;
;	for(i = 0; i < n; i++) <c,r[i]> := r[i] + c + a[i] * w ;
;	return c;
; }

r=4 ;(AP)
a=8 ;(AP)
n=12 ;(AP)	n	by value (input)
w=16 ;(AP)	w	by value (input)


	.psect	code,nowrt

.entry	bn_mul_add_words,^m<r2,r3,r4,r5,r6>

	moval	@r(ap),r2
	moval	@a(ap),r3
	movl	n(ap),r4	; assumed >0 by C code
	movl	w(ap),r5
	clrl	r6		; c

0$:
	emul	r5,(r3),(r2),r0		; w, a[], r[] considered signed

	; fixup for "negative" r[]
	tstl	(r2)
	bgeq	10$
	incl	r1
10$:

	; add in c
	addl2	r6,r0
	adwc	#0,r1

	; combined fixup for "negative" w, a[]
	tstl	r5
	bgeq	20$
	addl2	(r3),r1
20$:
	tstl	(r3)
	bgeq	30$
	addl2	r5,r1
30$:

	movl	r0,(r2)+		; store lo result in r[] & advance
	addl	#4,r3			; advance a[]
	movl	r1,r6			; store hi result => c

	sobgtr	r4,0$

	movl	r6,r0			; return c
	ret

	.title	vax_bn_mul_word  unsigned multiply & add, 32*32+32=>64
;
; w.j.m. 15-jan-1999
;
; it's magic ...
;
; ULONG bn_mul_words(ULONG r[],ULONG a[],int n,ULONG w) {
;	ULONG c = 0;
;	int i;
;	for(i = 0; i < num; i++) <c,r[i]> := a[i] * w + c ;
;	return(c);
; }

r=4 ;(AP)
a=8 ;(AP)
n=12 ;(AP)	n	by value (input)
w=16 ;(AP)	w	by value (input)


	.psect	code,nowrt

.entry	bn_mul_words,^m<r2,r3,r4,r5,r6>

	moval	@r(ap),r2	; r2 -> r[]
	moval	@a(ap),r3	; r3 -> a[]
	movl	n(ap),r4	; r4 = loop count (assumed >0 by C code)
	movl	w(ap),r5	; r5 = w
	clrl	r6		; r6 = c

0$:
	; <r1,r0> := w * a[] + c
	emul	r5,(r3),r6,r0		; w, a[], c considered signed

	; fixup for "negative" c
	tstl	r6			; c
	bgeq	10$
	incl	r1
10$:

	; combined fixup for "negative" w, a[]
	tstl	r5			; w
	bgeq	20$
	addl2	(r3),r1			; a[]
20$:
	tstl	(r3)			; a[]
	bgeq	30$
	addl2	r5,r1			; w
30$:

	movl	r0,(r2)+		; store lo result in r[] & advance
	addl	#4,r3			; advance a[]
	movl	r1,r6			; store hi result => c

	sobgtr	r4,0$

	movl	r6,r0			; return c
	ret

	.title	vax_bn_sqr_words  unsigned square, 32*32=>64
;
; w.j.m. 15-jan-1999
;
; it's magic ...
;
; void bn_sqr_words(ULONG r[],ULONG a[],int n) {
;	int i;
;	for(i = 0; i < n; i++) <r[2*i+1],r[2*i]> := a[i] * a[i] ;
; }

r=4 ;(AP)
a=8 ;(AP)
n=12 ;(AP)	n	by value (input)


	.psect	code,nowrt

.entry	bn_sqr_words,^m<r2,r3,r4,r5>

	moval	@r(ap),r2	; r2 -> r[]
	moval	@a(ap),r3	; r3 -> a[]
	movl	n(ap),r4	; r4 = n (assumed >0 by C code)

0$:
	movl	(r3)+,r5		; r5 = a[] & advance

	; <r1,r0> := a[] * a[]
	emul	r5,r5,#0,r0		; a[] considered signed

	; fixup for "negative" a[]
	tstl	r5			; a[]
	bgeq	30$
	addl2	r5,r1			; a[]
	addl2	r5,r1			; a[]
30$:

	movl	r0,(r2)+		; store lo result in r[] & advance
	movl	r1,(r2)+		; store hi result in r[] & advance

	sobgtr	r4,0$

	movl	#1,r0			; return SS$_NORMAL
	ret

	.title	(generated)

	.psect	code,nowrt

.entry	BN_DIV_WORDS,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10>
	subl2	#4,sp

	clrl	r9
	movl	#2,r8

	tstl	12(ap)
	bneq	noname.2
	mnegl	#1,r10
	brw	noname.3
	tstl	r0
	nop	
noname.2:

	pushl	12(ap)
	calls	#1,BN_NUM_BITS_WORD
	movl	r0,r7

	cmpl	r7,#32
	beql	noname.4
	ashl	r7,#1,r2
	cmpl	4(ap),r2
	blequ	noname.4

	pushl	r7
	calls	#1,BN_DIV_WORDS_ABORT
noname.4:

	subl3	r7,#32,r7

	movl	12(ap),r2
	cmpl	4(ap),r2
	blssu	noname.5
	subl2	r2,4(ap)
noname.5:

	tstl	r7
	beql	noname.6

	ashl	r7,r2,12(ap)

	ashl	r7,4(ap),r4
	subl3	r7,#32,r3
	subl3	r3,#32,r2
	extzv	r3,r2,8(ap),r2
	bisl3	r4,r2,4(ap)

	ashl	r7,8(ap),8(ap)
noname.6:

	bicl3	#65535,12(ap),r2
	extzv	#16,#16,r2,r5

	bicl3	#-65536,12(ap),r6

noname.7:

	moval	4(ap),r2
	movzwl	2(r2),r0
	cmpl	r0,r5
	bneq	noname.8

	movzwl	#65535,r4
	brb	noname.9
noname.8:

	clrl	r1
	movl	(r2),r0
	movl	r5,r2
	bgeq	vcg.1
	cmpl	r2,r0
	bgtru	vcg.2
	incl	r1
	brb	vcg.2
	nop	
vcg.1:
	ediv	r2,r0,r1,r0
vcg.2:
	movl	r1,r4
noname.9:

noname.10:

	mull3	r5,r4,r0
	subl3	r0,4(ap),r3

	bicl3	#65535,r3,r0
	bneq	noname.13
	mull3	r6,r4,r2
	ashl	#16,r3,r1
	bicl3	#65535,8(ap),r0
	extzv	#16,#16,r0,r0
	addl2	r0,r1
	cmpl	r2,r1
	bgtru	noname.12
noname.11:

	brb	noname.13
	nop	
noname.12:

	decl	r4
	brb	noname.10
noname.13:

	mull3	r5,r4,r1

	mull3	r6,r4,r0

	extzv	#16,#16,r0,r3

	ashl	#16,r0,r2
	bicl3	#65535,r2,r0

	addl2	r3,r1

	moval	8(ap),r3
	cmpl	(r3),r0
	bgequ	noname.15
	incl	r1
noname.15:

	subl2	r0,(r3)

	cmpl	4(ap),r1
	bgequ	noname.16

	addl2	12(ap),4(ap)

	decl	r4
noname.16:

	subl2	r1,4(ap)

	decl	r8
	beql	noname.18
noname.17:

	ashl	#16,r4,r9

	ashl	#16,4(ap),r2
	movzwl	2(r3),r0
	bisl2	r0,r2
	bicl3	#0,r2,4(ap)

	bicl3	#-65536,(r3),r0
	ashl	#16,r0,(r3)
	brw	noname.7
	nop	
noname.18:

	bisl2	r4,r9

	movl	r9,r10

noname.3:
	movl	r10,r0
	ret	
	tstl	r0


	.psect	code,nowrt

.entry	BN_ADD_WORDS,^m<r2,r3,r4,r5,r6,r7>

	tstl	16(ap)
	bgtr	noname.21
	clrl	r7
	brw	noname.22
noname.21:

	clrl	r4

	tstl	r0
noname.23:

	movl	8(ap),r6
	addl3	r4,(r6),r2

	bicl2	#0,r2

	clrl	r0
	cmpl	r2,r4
	bgequ	vcg.3
	incl	r0
vcg.3:
	movl	r0,r4

	movl	12(ap),r5
	addl3	(r5),r2,r1
	bicl2	#0,r1

	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.4
	incl	r0
vcg.4:
	addl2	r0,r4

	movl	4(ap),r3
	movl	r1,(r3)

	decl	16(ap)
	bgtr	gen.1
	brw	noname.25
gen.1:
noname.24:

	addl3	r4,4(r6),r2

	bicl2	#0,r2

	clrl	r0
	cmpl	r2,r4
	bgequ	vcg.5
	incl	r0
vcg.5:
	movl	r0,r4

	addl3	4(r5),r2,r1
	bicl2	#0,r1

	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.6
	incl	r0
vcg.6:
	addl2	r0,r4

	movl	r1,4(r3)

	decl	16(ap)
	bleq	noname.25
noname.26:

	addl3	r4,8(r6),r2

	bicl2	#0,r2

	clrl	r0
	cmpl	r2,r4
	bgequ	vcg.7
	incl	r0
vcg.7:
	movl	r0,r4

	addl3	8(r5),r2,r1
	bicl2	#0,r1

	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.8
	incl	r0
vcg.8:
	addl2	r0,r4

	movl	r1,8(r3)

	decl	16(ap)
	bleq	noname.25
noname.27:

	addl3	r4,12(r6),r2

	bicl2	#0,r2

	clrl	r0
	cmpl	r2,r4
	bgequ	vcg.9
	incl	r0
vcg.9:
	movl	r0,r4

	addl3	12(r5),r2,r1
	bicl2	#0,r1

	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.10
	incl	r0
vcg.10:
	addl2	r0,r4

	movl	r1,12(r3)

	decl	16(ap)
	bleq	noname.25
noname.28:

	addl3	#16,r6,8(ap)

	addl3	#16,r5,12(ap)

	addl3	#16,r3,4(ap)
	brw	noname.23
	tstl	r0
noname.25:

	movl	r4,r7

noname.22:
	movl	r7,r0
	ret	
	nop	



;r=4 ;(AP)
;a=8 ;(AP)
;b=12 ;(AP)
;n=16 ;(AP)	n	by value (input)

	.psect	code,nowrt

.entry	BN_SUB_WORDS,^m<r2,r3,r4,r5,r6,r7>

	clrl	r6

	tstl	16(ap)
	bgtr	noname.31
	clrl	r7
	brw	noname.32
	tstl	r0
noname.31:

noname.33:

	movl	8(ap),r5
	movl	(r5),r1
	movl	12(ap),r4
	movl	(r4),r2

	movl	4(ap),r3
	subl3	r2,r1,r0
	subl2	r6,r0
	bicl3	#0,r0,(r3)

	cmpl	r1,r2
	beql	noname.34
	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.11
	incl	r0
vcg.11:
	movl	r0,r6
noname.34:

	decl	16(ap)
	bgtr	gen.2
	brw	noname.36
gen.2:
noname.35:

	movl	4(r5),r2
	movl	4(r4),r1

	subl3	r1,r2,r0
	subl2	r6,r0
	bicl3	#0,r0,4(r3)

	cmpl	r2,r1
	beql	noname.37
	clrl	r0
	cmpl	r2,r1
	bgequ	vcg.12
	incl	r0
vcg.12:
	movl	r0,r6
noname.37:

	decl	16(ap)
	bleq	noname.36
noname.38:

	movl	8(r5),r1
	movl	8(r4),r2

	subl3	r2,r1,r0
	subl2	r6,r0
	bicl3	#0,r0,8(r3)

	cmpl	r1,r2
	beql	noname.39
	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.13
	incl	r0
vcg.13:
	movl	r0,r6
noname.39:

	decl	16(ap)
	bleq	noname.36
noname.40:

	movl	12(r5),r1
	movl	12(r4),r2

	subl3	r2,r1,r0
	subl2	r6,r0
	bicl3	#0,r0,12(r3)

	cmpl	r1,r2
	beql	noname.41
	clrl	r0
	cmpl	r1,r2
	bgequ	vcg.14
	incl	r0
vcg.14:
	movl	r0,r6
noname.41:

	decl	16(ap)
	bleq	noname.36
noname.42:

	addl3	#16,r5,8(ap)

	addl3	#16,r4,12(ap)

	addl3	#16,r3,4(ap)
	brw	noname.33
	tstl	r0
noname.36:

	movl	r6,r7

noname.32:
	movl	r7,r0
	ret	
	nop	



;r=4 ;(AP)
;a=8 ;(AP)
;b=12 ;(AP)
;n=16 ;(AP)	n	by value (input)

	.psect	code,nowrt

.entry	BN_MUL_COMBA8,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10,r11>
	movab	-924(sp),sp
	clrq	r8

	clrl	r10

	movl	8(ap),r6
	movzwl	2(r6),r3
	movl	12(ap),r7
	bicl3	#-65536,(r7),r2
	movzwl	2(r7),r0
	bicl2	#-65536,r0
	bicl3	#-65536,(r6),-12(fp)
	bicl3	#-65536,r3,-16(fp)
	mull3	r0,-12(fp),-4(fp)
	mull2	r2,-12(fp)
	mull3	r2,-16(fp),-8(fp)
	mull2	r0,-16(fp)
	addl3	-4(fp),-8(fp),r0
	bicl3	#0,r0,-4(fp)
	cmpl	-4(fp),-8(fp)
	bgequ	noname.45
	addl2	#65536,-16(fp)
noname.45:
	movzwl	-2(fp),r0
	bicl2	#-65536,r0
	addl2	r0,-16(fp)
	bicl3	#-65536,-4(fp),r0
	ashl	#16,r0,-8(fp)
	addl3	-8(fp),-12(fp),r0
	bicl3	#0,r0,-12(fp)
	cmpl	-12(fp),-8(fp)
	bgequ	noname.46
	incl	-16(fp)
noname.46:
	movl	-12(fp),r1
	movl	-16(fp),r2
	addl2	r1,r9
	bicl2	#0,r9
	cmpl	r9,r1
	bgequ	noname.47
	incl	r2
noname.47:
	addl2	r2,r8
	bicl2	#0,r8
	cmpl	r8,r2
	bgequ	noname.48
	incl	r10
noname.48:

	movl	4(ap),r11
	movl	r9,(r11)

	clrl	r9

	movzwl	2(r6),r2
	bicl3	#-65536,4(r7),r3
	movzwl	6(r7),r0
	bicl2	#-65536,r0
	bicl3	#-65536,(r6),-28(fp)
	bicl3	#-65536,r2,-32(fp)
	mull3	r0,-28(fp),-20(fp)
	mull2	r3,-28(fp)
	mull3	r3,-32(fp),-24(fp)
	mull2	r0,-32(fp)
	addl3	-20(fp),-24(fp),r0
	bicl3	#0,r0,-20(fp)
	cmpl	-20(fp),-24(fp)
	bgequ	noname.49
	addl2	#65536,-32(fp)
noname.49:
	movzwl	-18(fp),r0
	bicl2	#-65536,r0
	addl2	r0,-32(fp)
	bicl3	#-65536,-20(fp),r0
	ashl	#16,r0,-24(fp)
	addl3	-24(fp),-28(fp),r0
	bicl3	#0,r0,-28(fp)
	cmpl	-28(fp),-24(fp)
	bgequ	noname.50
	incl	-32(fp)
noname.50:
	movl	-28(fp),r1
	movl	-32(fp),r2
	addl2	r1,r8
	bicl2	#0,r8
	cmpl	r8,r1
	bgequ	noname.51
	incl	r2
noname.51:
	addl2	r2,r10
	bicl2	#0,r10
	cmpl	r10,r2
	bgequ	noname.52
	incl	r9
noname.52:

	movzwl	6(r6),r2
	bicl3	#-65536,(r7),r3
	movzwl	2(r7),r0
	bicl2	#-65536,r0
	bicl3	#-65536,4(r6),-44(fp)
	bicl3	#-65536,r2,-48(fp)
	mull3	r0,-44(fp),-36(fp)
	mull2	r3,-44(fp)
	mull3	r3,-48(fp),-40(fp)
	mull2	r0,-48(fp)
	addl3	-36(fp),-40(fp),r0
	bicl3	#0,r0,-36(fp)
	cmpl	-36(fp),-40(fp)
	bgequ	noname.53
	addl2	#65536,-48(fp)
noname.53:
	movzwl	-34(fp),r0
	bicl2	#-65536,r0
	addl2	r0,-48(fp)
	bicl3	#-65536,-36(fp),r0
	ashl	#16,r0,-40(fp)
	addl3	-40(fp),-44(fp),r0
	bicl3	#0,r0,-44(fp)
	cmpl	-44(fp),-40(fp)
	bgequ	noname.54
	incl	-48(fp)
noname.54:
	movl	-44(fp),r1
	movl	-48(fp),r2
	addl2	r1,r8
	bicl2	#0,r8
	cmpl	r8,r1
	bgequ	noname.55
	incl	r2
noname.55:
	addl2	r2,r10
	bicl2	#0,r10
	cmpl	r10,r2
	bgequ	noname.56
	incl	r9
noname.56:

	movl	r8,4(r11)

	clrl	r8

	movzwl	10(r6),r2
	bicl3	#-65536,(r7),r3
	movzwl	2(r7),r0
	bicl2	#-65536,r0
	bicl3	#-65536,8(r6),-60(fp)
	bicl3	#-65536,r2,-64(fp)
	mull3	r0,-60(fp),-52(fp)
	mull2	r3,-60(fp)
	mull3	r3,-64(fp),-56(fp)
	mull2	r0,-64(fp)
	addl3	-52(fp),-56(fp),r0
	bicl3	#0,r0,-52(fp)
	cmpl	-52(fp),-56(fp)
	bgequ	noname.57
	addl2	#65536,-64(fp)
noname.57:
	movzwl	-50(fp),r0
	bicl2	#-65536,r0
	addl2	r0,-64(fp)
	bicl3	#-65536,-52(fp),r0
	ashl	#16,r0,-56(fp)
	addl3	-56(fp),-60(fp),r0
	bicl3	#0,r0,-60(fp)
	cmpl	-60(fp),-56(fp)
	bgequ	noname.58
	incl	-64(fp)
noname.58:
	movl	-60(fp),r1
	movl	-64(fp),r2
	addl2	r1,r10
	bicl2	#0,r10
	cmpl	r10,r1
	bgequ	noname.59
	incl	r2
noname.59:
	addl2	r2,r9
	bicl2	#0,r9
	cmpl	r9,r2
	bgequ	noname.60
	incl	r8
noname.60:

	movzwl	6(r6),r2