working ucoded alu

1 files changed, 216 insertions, 143 deletions

diff --git a/sha1/msauth.asm b/sha1/msauth.asm
index 397e2eb..fc5e291 100644
--- a/sha1/msauth.asm
+++ b/sha1/msauth.asm
@@ -3,12 +3,12 @@
 
 idx		EQU	$60
 lidx		EQU	$61
-k		EQU	$62
 src1		EQU	$63
 src2		EQU	$64
 dest		EQU	$65
 i		EQU	$66
 j		EQU	$67
+opidx		EQU	$62
 
 
 ; BYTES 60-67 are persisent RAM for our use
@@ -136,14 +136,14 @@ test_l2:
 		lda 	H+16
 		and	#$F
 		tax
-		add	#(H-BASE_START)
+		add	#(H-DATA_START)
 		sta	src1
 		lda	H+3,x
 		and	#$7F
 		sta	H+3,x
-		lda	#(T1-BASE_START)
+		lda	#(T1-DATA_START)
 		sta	dest
-		lda	#(ZERO-BASE_START)
+		lda	#(ZERO-DATA_START)
 		jsr 	add4
 		
 		
@@ -153,9 +153,9 @@ end:		bra end
 
 
 get_w32_015_to_t2:
-		lda	#(T2-BASE_START)
+		lda	#(T2-DATA_START)
 		sta	dest
-		lda	#(ZERO-BASE_START)
+		lda	#(ZERO-DATA_START)
 		sta	src1
 
 		lda	lidx
@@ -167,14 +167,14 @@ get_w32_015_to_t2:
 			bhs	get_w32_015_to_t2_z1
 				lsla
 				lsla
-				add	#(KEY-BASE_START)
+				add	#(KEY-DATA_START)
 				sta	src1
 get_w32_015_to_t2_z1:
-			lda	#(ZERO-BASE_START)
+			lda	#(ZERO-DATA_START)
 			jsr	add4
-			lda	#(T2-BASE_START)
+			lda	#(T2-DATA_START)
 			sta	src1
-			lda	#(KEYOR-BASE_START)
+			lda	#(KEYOR-DATA_START)
 			jsr 	xor4
 		rts
 get_w32_015_to_t2_l1:
@@ -183,15 +183,15 @@ get_w32_015_to_t2_l1:
 		bhs	get_w32_015_to_t2_l2
 			lsla
 			lsla
-			add	#(DATA1-BASE_START)
+			add	#(DATA1-DATA_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)
+			lda	#(DATA2-DATA_START)
 			bra	get_w32_015_to_t2_e
 get_w32_015_to_t2_l3:	
-			lda	#(ZERO-BASE_START)
+			lda	#(ZERO-DATA_START)
 get_w32_015_to_t2_e:
 		jsr	add4
 		rts
@@ -215,11 +215,11 @@ get_w32_to_t2_l1:
 		lda	j
 		lsla
 		lsla
-		add	#(WBUF-BASE_START)
+		add	#(WBUF-DATA_START)
 		sta 	dest
-		lda	#(T2-BASE_START)
+		lda	#(T2-DATA_START)
 		sta	src1
-		lda	#(ZERO-BASE_START)
+		lda	#(ZERO-DATA_START)
 		jsr 	add4
 
 	inc	j
@@ -229,17 +229,17 @@ get_w32_to_t2_l1:
 
 
 get_w32_to_t2_l2:
-		lda	#(T2-BASE_START)
+		lda	#(T2-DATA_START)
 		sta	dest
-		lda	#(WBUF13 - BASE_START)
+		lda	#(WBUF13 - DATA_START)
 		sta	src1
-		lda	#(WBUF8 - BASE_START)
+		lda	#(WBUF8 - DATA_START)
 		jsr	xor4		; T2 =W13 ^ W8
-		lda	#(T2-BASE_START)
+		lda	#(T2-DATA_START)
 		sta	src1
-		lda	#(WBUF2 - BASE_START)
+		lda	#(WBUF2 - DATA_START)
 		jsr	xor4		; T2 = T2 ^ W2
-		lda	#(WBUF0 - BASE_START)
+		lda	#(WBUF0 - DATA_START)
 		jsr	xor4		; T2= T2 ^ W0
 		lda	#1
 		jsr	rot_left	; T2= ROTLEFT(T2,1)
@@ -252,9 +252,9 @@ get_w32_to_t2_l3:
 			sta 	(WBUF14 + 3 - $FF),x
 		incx
 		bne	get_w32_to_t2_l3
-		lda	#(WBUF15-BASE_START)
+		lda	#(WBUF15-DATA_START)
 		sta	dest
-		lda	#(ZERO-BASE_START)
+		lda	#(ZERO-DATA_START)
 		jsr	add4		; WBUF15=T2
 		
 	dec i
@@ -278,27 +278,17 @@ sha1_block_loop:
 			sta	idx
 			cmp	#20
 			bhs	sha1_block_k2
-				lda	#(K1-BASE_START)
+				lda	#(K1-DATA_START)
 				sta	k
 
-				; T1 = (B & C) | ((~B) & D)
-				lda	#(T2-BASE_START)
-				sta	dest
-				lda	#(VB-BASE_START)
-				sta	src1
-				lda	#(FFFFFFFF-BASE_START)
-				jsr	xor4	;T2=~B
-
-				lda	#(T2-BASE_START)
-				sta	src1
-				lda	#(VD-BASE_START)
-				jsr	and4	;T2=T2&D
+				ldx	#(T2eNotBandD-OPLIST_START)
+				jsr	op4_list
 
 				bra	sha1_block_t2_or_bandc ; T1=T2 | (B&C)
 sha1_block_k2:
 			cmp	#40
 			bhs	sha1_block_k3
-				lda	#(K2-BASE_START)
+				lda	#(K2-DATA_START)
 				sta	k
 				bra	sha1_block_k24 ; T1 = B ^ C ^ D
 
@@ -306,94 +296,32 @@ sha1_block_k3:
 			cmp	#60
 			bhs	sha1_block_k4
 				; T1 = (B & C) | (B & D) | (C & D)
-				lda	#(K3-BASE_START)
+				lda	#(K3-DATA_START)
 				sta	k
 
-				lda	#(T1-BASE_START)
-				sta	dest
-				lda	#(VD-BASE_START)
-				sta	src1
-				lda	#(VB-BASE_START)
-				jsr	and4	; T1=B&D
-
-				lda	#(T2-BASE_START)
-				sta	dest
-				lda	#(VC-BASE_START)
-				jsr	and4	; T2=C&D
-
-				lda	#(T2-BASE_START)
-				sta	src1
-				lda	#(T1-BASE_START)
-				jsr	or4	; T2=T1|T2
-
+				ldx	#(T2eCandDorBandD-OPLIST_START)
+				jsr	op4_list
+ 
 sha1_block_t2_or_bandc:
 
-				lda	#(T1-BASE_START)
-				sta	dest
-				lda	#(VB-BASE_START)
-				sta	src1
-				lda	#(VC-BASE_START)
-				jsr	and4	; T1 = B&C
-
-				lda	#(T1-BASE_START)
-				sta	src1
-				lda	#(T2-BASE_START)
-				jsr	or4	; T1= T1|T2
-
+				ldx	#(T1eT2orBandC-OPLIST_START)
+				jsr	op4_list
 				bra	sha1_block_shuffle
 
 sha1_block_k4:
-				lda	#(K4-BASE_START)
+				lda	#(K4-DATA_START)
 				sta	k
 sha1_block_k24:
-				; T1 = B ^ C ^ D
-				lda	#(T1-BASE_START)
-				sta	dest
-				lda	#(VB-BASE_START)
-				sta	src1
-				lda	#(VC-BASE_START)
-				jsr	xor4
-				lda	#(T1-BASE_START)
-				sta	src1
-				lda	#(VD-BASE_START)
-				jsr	xor4
-sha1_block_shuffle:
-			; T1 is src1 and dest here
 
-			lda	#(VE-BASE_START)
-			jsr	add4	; T1=T1+E
+				ldx 	#(T1eBxorCxorD-OPLIST_START)
+				jsr	op4_list
 
-			lda	k
-			jsr	add4	; T1=T1+k
+sha1_block_shuffle:
 
 			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_up	; E=D
-			lda	#(VC-BASE_START)
-			jsr 	copy_up	; D=C
-			lda	#(VB-BASE_START)	
-			jsr 	copy_up	; C=B
-			lda	#30
-			jsr	rot_left	; C=ROTLEFT(C,30)
-			lda	#(VA-BASE_START)	
-			jsr 	copy_up	; B=A
-
-			lda	#(VA-BASE_START)
-			sta	dest
-			sta	src1
-			lda	#5
-			jsr	rot_left	; A=ROTLEFT(A,5)
-
-			lda	#(T1-BASE_START)
-			jsr	add4		; A=A+T1
-
+			ldx	#(shuffleop-OPLIST_START)
+			jsr	op4_list
 
 		lda	idx
 		add	#1
@@ -417,14 +345,54 @@ sha1_block_add1:
 		cpx	#20
 		bne	sha1_block_add0
 
+rts1:
 		rts
 
-
-copy_up:
+op4_list:
+		lda	OPLIST_START,x
+		sta	op4_src2
+		incx	
+		lda	OPLIST_START,x
 		sta	src1
-		add	#4
+		incx	
+		lda	OPLIST_START,x
+		sta	src2
+		incx
+		lda	OPLIST_START,x
 		sta	dest
-		lda	#(ZERO-BASE_START)
+		incx	
+		stx	opidx
+		lda	op4_src2
+		tsta
+		beq	rts1
+		bmi     op4_list_l1
+
+			; rotates dest left src1 times
+			ldx dest
+op_list_l0:
+				lda	DATA_START,x
+				rola
+				lda	DATA_START+1,x
+				rola
+				sta	DATA_START+1,x
+				lda	DATA_START+2,x
+				rola
+				sta	DATA_START+2,x
+				lda	DATA_START+3,x
+				rola
+				sta	DATA_START+3,x
+				lda	DATA_START,x
+				rola
+				sta	DATA_START,x
+			dec	src1
+			bne 	op_list_l0
+		bra	op4_list_l2
+op4_list_l1:
+			jsr	op4_binary
+op4_list_l2:
+		ldx	opidx
+		bra	op4_list
+		
 
 add4:
 		sta	src2
@@ -432,42 +400,33 @@ add4:
 		sta 	op4_src2
 		bra 	op4_binary
 
-and4:
-		sta	src2
-		lda	#$d4	; and $ffff,a
-		sta 	op4_src2
-		bra 	op4_binary
-
 xor4:
 		sta	src2
 		lda	#$d8	; eor $ffff,a
 		sta 	op4_src2
 		bra	op4_binary
-or4:
-		sta	src2
-		lda	#$da	; or $ffff,a
-		sta 	op4_src2
+
 op4_binary:		
 		ldx	#0
-		lda	#BASE_START & $ff
+		lda	#DATA_START & $ff
 		add	src2
 		sta	op4_src2+2
 	        txa	
-		adc 	#BASE_START>>8
+		adc 	#DATA_START>>8
 		sta	op4_src2+1
 
-		lda	#BASE_START & $ff
+		lda	#DATA_START & $ff
 		add	src1
 		sta	op4_src1+2
 		txa 	
-		adc 	#BASE_START>>8
+		adc 	#DATA_START>>8
 		sta	op4_src1+1
 
-		lda	#BASE_START & $ff
+		lda	#DATA_START & $ff
 		add	dest
 		sta	op4_dst+2
 		txa 	
-		adc 	#BASE_START>>8
+		adc 	#DATA_START>>8
 		sta	op4_dst+1
 		 		
 		clc
@@ -488,26 +447,140 @@ rot_left:
 		ldx dest
 		sta 	j
 rot_loop:
-		lda	BASE_START,x
+		lda	DATA_START,x
 		rola
-		lda	BASE_START+1,x
+		lda	DATA_START+1,x
 		rola
-		sta	BASE_START+1,x
-		lda	BASE_START+2,x
+		sta	DATA_START+1,x
+		lda	DATA_START+2,x
 		rola
-		sta	BASE_START+2,x
-		lda	BASE_START+3,x
+		sta	DATA_START+2,x
+		lda	DATA_START+3,x
 		rola
-		sta	BASE_START+3,x
-		lda	BASE_START,x
+		sta	DATA_START+3,x
+		lda	DATA_START,x
 		rola
-		sta	BASE_START,x
+		sta	DATA_START,x
 		dec	j
 		bne 	rot_loop
 		rts
 
+
+OPLIST_START:
+
+T2eNotBandD:
+				DB	$d8		;eor
+				DB	(VB-DATA_START)	;src1
+				DB	(FFFFFFFF-DATA_START)	;src2
+				DB	(T2-DATA_START)	;dest
+				DB	$d4		;and
+				DB	(VD-DATA_START)	;src1
+				DB	(T2-DATA_START)	;src2
+				DB	(T2-DATA_START)	;dest
+				DB	0		;stop
+				
+
+T2eCandDorBandD:		; T1 = C&D, T2=B&D, T2=T1|T2
+				DB	$d4		;and
+				DB	(VC-DATA_START) ;src1
+				DB	(VD-DATA_START)	;src2
+				DB	(T1-DATA_START)	;dest
+				DB	$d4		;and
+				DB	(VB-DATA_START) ;src1
+				DB	(VD-DATA_START)	;src2
+				DB	(T2-DATA_START)	;dest
+				DB	$da		;or
+				DB	(T1-DATA_START)	;src1
+				DB	(T2-DATA_START)	;src2
+				DB	(T2-DATA_START)	;dest
+				DB	0		;stop
+
+T1eT2orBandC:			; T1=B&C; T1=T1|T2
+				DB	$d4		;and
+				DB	(VB-DATA_START) ;src1
+				DB	(VC-DATA_START)	;src2
+				DB	(T1-DATA_START)	;dest
+				DB	$da		;or
+				DB	(T1-DATA_START)	;src1
+				DB	(T2-DATA_START)	;src2
+				DB	(T1-DATA_START)	;dest
+				DB	0		;stop
+			
+				
+
+T1eBxorCxorD:
+			DB	$d8		;xor
+			DB	(VB-DATA_START)	;src1
+			DB	(VC-DATA_START)	;src2
+			DB	(T1-DATA_START)	;dst
+			DB	$d8		;xor
+			DB	(T1-DATA_START)	;src1
+			DB	(VD-DATA_START) ;src2
+			DB	(T1-DATA_START)	;dst
+			DB	0		;stop
+	
+shuffleop:
+
+
+			DB	$d9		;add
+			DB	(T1-DATA_START)	;src1
+			DB	(VE-DATA_START) ;src2
+			DB	(T1-DATA_START)	;dest
+
+			DB	$d9		;add
+			DB	(T1-DATA_START)	;src1
+k:
+			DB	0		;src2
+			DB	(T1-DATA_START)	;dest
+
+			DB	$d9		;add
+			DB	(T1-DATA_START)	;src1
+			DB	(T2-DATA_START)	;src2
+			DB	(T1-DATA_START)	;dest
+
+			DB	$d9		;add
+			DB	(VD-DATA_START) ;src1
+			DB	(ZERO-DATA_START) ;src2
+			DB	(VE-DATA_START) ;dest
+
+			DB	$d9		;add
+			DB	(VC-DATA_START) ;src1
+			DB	(ZERO-DATA_START) ;src2
+			DB	(VD-DATA_START) ;dest
+
+			DB	$d9		;add
+			DB	(VB-DATA_START) ;src1
+			DB	(ZERO-DATA_START) ;src2
+			DB	(VC-DATA_START) ;dest
+
+			DB	$d9		;add
+			DB	(VA-DATA_START) ;src1
+			DB	(ZERO-DATA_START) ;src2
+			DB	(VB-DATA_START) ;dest
+
+			DB	$1		;rot left
+			DB	30		;30 times
+			DB	0		; N/A
+			DB	(VC-DATA_START) ;dest
+
+			DB	$1		;rot left
+			DB	5		;5 times
+			DB	0		; N/A
+			DB	(VA-DATA_START) ;dest
+
+			DB	$d9		;add
+			DB	(VA-DATA_START) ;src1
+			DB	(T1-DATA_START) ;src2
+			DB	(VA-DATA_START) ;dest
+
+			DB 	$0
+
+
+OPLIST_END:
+
+
 		
-BASE_START:
+DATA_START:
 
 KEY:
 		DB	$89
@@ -701,5 +774,5 @@ DATA2:
 		DB	0
 		DB	0
 
-BASE_END:
+DATA_END: