!to "shasum.o", plain
	* = $6000
	jmp main

	;; clear addresses:
	;; (http://apple2.org.za/gswv/a2zine/faqs/csa2pfaq.html#017)
	;; 06-09
	;; EB-EF
	;; FA-FD
	
	!addr	SRC = $06
	!addr	DEST = $08
	!addr   INPUT = $eb
	!addr   LENGTH = $ee

	!addr PRBYTE = $FDDA
	!addr COUT = $FDED
	
h0:	!32 0			; return value (hash)
h1:	!32 0
h2:	!32 0
h3:	!32 0
h4:	!32 0
ml:	!32 0			; message length
w:	!fill 64, 0
w_next:	!fill 64, 0
kh0:	!be32 $67452301		; initial values for h0..h4
kh1:	!be32 $EFCDAB89
kh2:	!be32 $98BADCFE
kh3:	!be32 $10325476
kh4:	!be32 $C3D2E1F0
k1:	!be32 $5A827999		; k = constants
k2:	!be32 $6ED9EBA1
k3:	!be32 $8F1BBCDC
k4:	!be32 $CA62C1D6	

;;; Print a string of bytes, as hex.
;;; Address in SRC, count in A.
;;; Burns Y, A.
prbytes:
	ldy #0
-	pha
	lda (SRC),y
	jsr PRBYTE
	iny
	pla
	adc #$ff
	bne -
	rts
	
main:
	;; Test PRBYTES
	;; lda #<kh0
	;; sta SRC
	;; lda #>kh0
	;; sta SRC+1
	;; lda #8
	;; jsr prbytes
	;; LDA #$8D
	;; jsr COUT

	;; Test shasum
	lda #0
	sta INPUT
	lda #$ff
	sta INPUT+1
	lda #0
	sta LENGTH+1
	lda #$3b
	sta LENGTH
	jsr shasum
	
	rts

shasum:
	;; Initialize h0..h4
	ldy #(ml-h0-1)
-	lda kh0,y
	sta h0,y
	dey
	bpl -
	;; Initialize message length (ml)
	lda #0
	sta ml
	sta ml+1
	lda LENGTH
	sta ml+3
	lda LENGTH+1
	sta ml+2

	;; Message length is in bits
	ldy #3
-	clc
	rol ml+3
	rol ml+2
	rol ml+1
	dey
	bne -

	;; Initialize chunk counter
	;; ldy #0			; already zero

	;; Invert length so we can inc instead of dec
	lda LENGTH
	sec
	lda #0
	sbc LENGTH
	sta LENGTH
	lda #0
	sbc LENGTH+1
	sta LENGTH+1
	ora LENGTH
	beq .msgdone
	
.loop	lda (INPUT),y
	sta w,y
	iny
	cpy #$40
	bne +

	;; Call do_chunk
	jsr do_chunk
	ldy #0

	clc
	lda INPUT
	adc #$40
	sta INPUT
	bcc +
	inc INPUT+1
	
+	inc LENGTH
	bne .loop
	inc LENGTH+1
	bne .loop
	
.msgdone:
	lda #$80
	sta w,y
	iny
	cpy #$40
	bne .zeros
	jsr do_chunk
	ldy #0

.zeros
	cpy #$3C
	beq .length
	lda #0
	sta w,y
	iny
	cpy #$40
	bne .zeros
	jsr do_chunk
	ldy #0
	jmp .zeros
.length
	ldy #3
-	lda ml,y
	sta w+$3c,y
	dey
	bpl -
	jsr do_chunk
	rts

do_chunk:
	lda #<w
	sta SRC
	lda #>w
	sta SRC+1
	lda #$40
	jsr prbytes
	lda #$8D
	jsr COUT
	rts
	
	!eof
TODOs
[X] Routine to print n hex bytes (address, length (byte))
[ ] Routine to get the next 16 values (64 bytes)
[ ] Routine to get w[i] one i at a time, and rebuild next 16 values
[ ] Routine to get the next 16 values (64 bytes) of input

Chacha20 code isn't so useful, because it's little-endian.

611B ADD32
612B XOR32
613A ROL16
6152 RORX
6165 ROL12
616E ROL8
6180 ROL7
6189 PR32

Needed sha1sum:
- leftrotate 1  --> ?
- leftrotate 5  --> ROL8, RORX 3
- leftrotate 30 --> RORX 2