first working version

1 files changed, 285 insertions, 68 deletions

diff --git a/sha1/msauth.asm b/sha1/msauth.asm
index e4d0929..5e03d48 100644
--- a/sha1/msauth.asm
+++ b/sha1/msauth.asm
@@ -5,11 +5,13 @@
 
 
 idx		EQU	$60
-rot_cnt		EQU	$61
+lidx		EQU	$61
 k		EQU	$62
 src1		EQU	$63
 src2		EQU	$64
 dest		EQU	$65
+i		EQU	$66
+j		EQU	$67
 
 
 ; BYTES 60-67 are persisent RAM for our use
@@ -88,26 +90,197 @@ handle_st0_event:
 
 
 test:
-		jsr sha1
-		stop
 
-sha1:	
-		ldx #0
 
-sha1_copy_h_to_ae_loop:
-		lda 	H,x
-		sta	VA,x
+		lda	#0
+		sta	lidx
+
+		; data starts off as all zeros
+		; fixme correct time here
+		lda	#$a6
+		sta	DATA1+4
+		lda	#$a7
+		sta	DATA1+5
+
+		; don't need to init H from IH already there
+		jsr	sha1_block
+		inc	lidx
+		jsr	sha1_block
+
+		; Copy H->DATA, IH->H
+		ldx	#20
+test_l1:
+		lda	(H-1),x
+		sta	(DATA1-1),x
+		lda	(IH-1),x
+		sta	(H-1),x
+		decx
+		bne test_l1
+
+		lda 	#$80
+		sta	DATA1+23
+
+		lda	#$a0
+		sta	DATA2
+
+		; set KEYOR = 5c5c5c5c
+		lda	#$5c
+		ldx 	#4
+test_l2:
+		sta	(KEYOR-1),x
+		decx
+		bne	test_l2
+
+		dec	lidx
+		jsr	sha1_block
+		inc	lidx
+		jsr	sha1_block
+
+		lda 	H+16
+		and	#$F
+		tax
+		add	#(H-BASE_START)
+		sta	src1
+		lda	H+3,x
+		and	#$7F
+		sta	H+3,x
+		lda	#(T1-BASE_START)
+		sta	dest
+		lda	#(ZERO-BASE_START)
+		jsr 	add4
+		
+		
+
+end:		bra end
+
+
+
+get_w32_015_to_t2:
+		lda	#(T2-BASE_START)
+		sta	dest
+		lda	#(ZERO-BASE_START)
+		sta	src1
+
+		lda	lidx
+		cmp 	#0
+		bne 	get_w32_015_to_t2_l1
+
+			lda	j
+			cmp 	#3
+			bhs	get_w32_015_to_t2_z1
+				lsla
+				lsla
+				add	#(KEY-BASE_START)
+				sta	src1
+get_w32_015_to_t2_z1:
+			lda	#(ZERO-BASE_START)
+			jsr	add4
+			lda	#(T2-BASE_START)
+			sta	src1
+			lda	#(KEYOR-BASE_START)
+			jsr 	xor4
+		rts
+get_w32_015_to_t2_l1:
+		lda	j
+		cmp	#6
+		bhs	get_w32_015_to_t2_l2
+			lsla
+			lsla
+			add	#(DATA1-BASE_START)
+			bra	get_w32_015_to_t2_e
+get_w32_015_to_t2_l2:
+		cmp	#15
+		bne	get_w32_015_to_t2_l3
+			lda	#(DATA2-BASE_START)
+			bra	get_w32_015_to_t2_e
+get_w32_015_to_t2_l3:	
+			lda	#(ZERO-BASE_START)
+get_w32_015_to_t2_e:
+		jsr	add4
+		rts
+
+get_w32_to_t2:
+	lda	idx
+	cmp	#16
+	bhs	get_w32_to_t2_l0
+		sta	j			; <16 use our function
+		jsr	get_w32_015_to_t2
+		rts
+get_w32_to_t2_l0:
+	sub 	#15				; store number of iterations needed in i
+	sta	i
+
+	; fill W
+	lda	#0
+	sta	j
+get_w32_to_t2_l1:
+		jsr	get_w32_015_to_t2	
+		lda	j
+		lsla
+		lsla
+		add	#(WBUF-BASE_START)
+		sta 	dest
+		lda	#(T2-BASE_START)
+		sta	src1
+		lda	#(ZERO-BASE_START)
+		jsr 	add4
+
+	inc	j
+	lda	j	
+	cmp 	#16
+	bne	get_w32_to_t2_l1	
+
+
+get_w32_to_t2_l2:
+		lda	#(T2-BASE_START)
+		sta	dest
+		lda	#(WBUF13 - BASE_START)
+		sta	src1
+		lda	#(WBUF8 - BASE_START)
+		jsr	xor4		; T2 =W13 ^ W8
+		lda	#(T2-BASE_START)
+		sta	src1
+		lda	#(WBUF2 - BASE_START)
+		jsr	xor4		; T2 = T2 ^ W2
+		lda	#(WBUF0 - BASE_START)
+		jsr	xor4		; T2= T2 ^ W0
+		lda	#1
+		jsr	rot_left	; T2= ROTLEFT(T2,1)
+
+
+		; for (x=0;x<14;++x) wbuf[x]=wbuf[x+1]
+		ldx	#((WBUF0 + $FF - 3 ) - WBUF14)
+get_w32_to_t2_l3:
+			lda 	(WBUF15 + 3 - $FF),x
+			sta 	(WBUF14 + 3 - $FF),x
 		incx
-		cpx 	#20
-		bne	sha1_copy_h_to_ae_loop
+		bne	get_w32_to_t2_l3
+		lda	#(WBUF15-BASE_START)
+		sta	dest
+		lda	#(ZERO-BASE_START)
+		jsr	add4		; WBUF15=T2
+		
+	dec i
+	bne get_w32_to_t2_l2
+	rts
+
+
+sha1_block:	
+		ldx #20
+
+sha1_block_copy_h_to_ae_loop:
+		lda 	(H-1),x
+		sta	(VA-1),x
+		decx
+		bne	sha1_block_copy_h_to_ae_loop
 
 		lda	#0
 
-sha1_loop:	
+sha1_block_loop:	
 			; a contains idx
 			sta	idx
 			cmp	#20
-			bhs	sha1_k2
+			bhs	sha1_block_k2
 				lda	#(K1-BASE_START)
 				sta	k
 
@@ -124,17 +297,17 @@ sha1_loop:
 				lda	#(VD-BASE_START)
 				jsr	and4	;T2=T2&D
 
-				bra	sha1_t2_or_bandc ; T1=T2 | (B&C)
-sha1_k2:
+				bra	sha1_block_t2_or_bandc ; T1=T2 | (B&C)
+sha1_block_k2:
 			cmp	#40
-			bhs	sha1_k3
+			bhs	sha1_block_k3
 				lda	#(K2-BASE_START)
 				sta	k
-				bra	sha1_k24 ; T1 = B ^ C ^ D
+				bra	sha1_block_k24 ; T1 = B ^ C ^ D
 
-sha1_k3:
+sha1_block_k3:
 			cmp	#60
-			bhs	sha1_k4
+			bhs	sha1_block_k4
 				; T1 = (B & C) | (B & D) | (C & D)
 				lda	#(K3-BASE_START)
 				sta	k
@@ -156,7 +329,7 @@ sha1_k3:
 				lda	#(T1-BASE_START)
 				jsr	or4	; T2=T1|T2
 
-sha1_t2_or_bandc:
+sha1_block_t2_or_bandc:
 
 				lda	#(T1-BASE_START)
 				sta	dest
@@ -170,12 +343,12 @@ sha1_t2_or_bandc:
 				lda	#(T2-BASE_START)
 				jsr	or4	; T1= T1|T2
 
-				bra	sha1_shuffle
+				bra	sha1_block_shuffle
 
-sha1_k4:
+sha1_block_k4:
 				lda	#(K4-BASE_START)
 				sta	k
-sha1_k24:
+sha1_block_k24:
 				; T1 = B ^ C ^ D
 				lda	#(T1-BASE_START)
 				sta	dest
@@ -187,7 +360,7 @@ sha1_k24:
 				sta	src1
 				lda	#(VD-BASE_START)
 				jsr	xor4
-sha1_shuffle:
+sha1_block_shuffle:
 			; T1 is src1 and dest here
 
 			lda	#(VE-BASE_START)
@@ -196,20 +369,24 @@ sha1_shuffle:
 			lda	k
 			jsr	add4	; T1=T1+k
 
-			; get data here
-			lda	#(ZERO-BASE_START)
+			jsr	get_w32_to_t2 	; W[i]->T2
+
+			lda	#(T1-BASE_START)
+			sta	dest
+			sta	src1
+			lda	#(T2-BASE_START)
 			jsr	add4	; T1=T1+data
 
 			lda	#(VD-BASE_START)
-			jsr 	copy_down	; E=D
+			jsr 	copy_up	; E=D
 			lda	#(VC-BASE_START)
-			jsr 	copy_down	; D=C
+			jsr 	copy_up	; D=C
 			lda	#(VB-BASE_START)	
-			jsr 	copy_down	; C=B
+			jsr 	copy_up	; C=B
 			lda	#30
 			jsr	rot_left	; C=ROTLEFT(C,30)
 			lda	#(VA-BASE_START)	
-			jsr 	copy_down	; B=A
+			jsr 	copy_up	; B=A
 
 			lda	#(VA-BASE_START)
 			sta	dest
@@ -224,29 +401,29 @@ sha1_shuffle:
 		lda	idx
 		add	#1
 		cmp 	#80
-		bhs	sha1_done
-		jmp 	sha1_loop
-sha1_done:
+		bhs	sha1_block_done
+		jmp 	sha1_block_loop
+sha1_block_done:
 
 
 		ldx	#0
-sha1_add0:
+sha1_block_add0:
 		clc
-sha1_add1:
+sha1_block_add1:
 		lda	VA,x
 		adc	H,x
-		sta	VA,x
+		sta	H,x
 		incx
 		txa	
 		and	#3
-		bne	sha1_add1
+		bne	sha1_block_add1
 		cpx	#20
-		bne	sha1_add0
+		bne	sha1_block_add0
 
 		rts
 
 
-copy_down:
+copy_up:
 		sta	src1
 		add	#4
 		sta	dest
@@ -312,7 +489,7 @@ op4_dst:
 
 rot_left:	
 		ldx dest
-		sta 	rot_cnt
+		sta 	j
 rot_loop:
 		lda	BASE_START,x
 		rola
@@ -328,7 +505,7 @@ rot_loop:
 		lda	BASE_START,x
 		rola
 		sta	BASE_START,x
-		dec	rot_cnt
+		dec	j
 		bne 	rot_loop
 		rts
 
@@ -336,14 +513,46 @@ rot_loop:
 BASE_START:
 
 KEY:
-		DW	$0e89
-		DW      $6e38
-		DW	$3cbb
-		DW	$e9cf
-		DW	$0000
-		DW	$e44f
+		DB	$89
+		DB	$0e
+		DB	$38
+		DB      $6e
+
+
+		DB	$bb
+		DB	$3c
+		DB	$cf
+		DB	$e9
+
+		DB	$00
+		DB	$00
+		DB	$4f
+		DB	$e4
+		
+IH:
+		DB      $01
+		DB      $23
+		DB      $45
+		DB      $67
+		DB      $89
+		DB      $AB
+		DB      $CD
+		DB      $EF
+		DB      $FE
+		DB      $DC
+		DB      $BA
+		DB      $98
+		DB      $76
+		DB      $54
+		DB      $32
+		DB      $10
+		DB      $F0
+		DB      $E1
+		DB      $D2
+		DB      $C3
 		
-IH:		DB      $01
+H:
+		DB      $01
 		DB      $23
 		DB      $45
 		DB      $67
@@ -395,6 +604,11 @@ FFFFFFFF:
 		DB 	$FF
 		DB 	$FF
 
+KEYOR:
+		DW	$3636
+		DW	$3636
+
+
 bss:
 
 VA:		DW	0
@@ -419,53 +633,52 @@ ZERO:		DW	0
 
 
 WBUF:		
+WBUF0:
 		DW	0
 		DW	0
+WBUF1:
 		DW	0
 		DW	0
+WBUF2:
 		DW	0
 		DW	0
+WBUF3:
 		DW	0
 		DW	0
+WBUF4:
 		DW	0
 		DW	0
+WBUF5:
 		DW	0
 		DW	0
+WBUF6:
 		DW	0
 		DW	0
+WBUF7:
 		DW	0
 		DW	0
+WBUF8:
 		DW	0
 		DW	0
+WBUF9:
 		DW	0
 		DW	0
+WBUF10:
 		DW	0
 		DW	0
+WBUF11:
 		DW	0
 		DW	0
+WBUF12:
 		DW	0
 		DW	0
+WBUF13:
 		DW	0
 		DW	0
+WBUF14:
 		DW	0
 		DW	0
-		DW	0
-		DW	0
-
-KEYOR:
-		DW	0
-		DW	0
-
-
-H:	
-		DW	0
-		DW	0
-		DW	0
-		DW	0
-		DW	0
-		DW	0
-		DW	0
-		DW	0
+WBUF15:
 		DW	0
 		DW	0
 
@@ -475,8 +688,10 @@ DATA1:
 		DW	0
 		DW	0
 		DW	0
-		DW	0
-		DW	0
+		DB	0
+		DB	0
+		DB	0
+		DB	$80
 		DW	0
 		DW	0
 		DW	0
@@ -484,8 +699,10 @@ DATA1:
 		DW	0
 		DW	0
 DATA2:	
-		DW	0
-		DW	0
+		DB	$40
+		DB	$02
+		DB	0
+		DB	0
 
 BASE_END: