; aes_gf2_mul(uint8_t *out, uint8_t *op1, uint8_t *op2);
_aes_gf2_mul:
; Galois-Field GF(2^128) multiplication routine
; little endian fields expected
	ld hl, -16
	call ti._frameset
	lea de, ix - 16		; stack mem?
	ld hl, (ix + 9)		; op1 (save a copy)
	ld bc, 16
	ldir				; ix - 32 = tmp = op1

	; zero out output
	ld de, (ix + 6)		; op 1
	xor a
	ld (de), a
	inc de
	ld hl, (ix + 6)
	ld bc, 15
	ldir

	ld hl, (ix + 12)		; op2 = for bit in bits
	ld bc, 0
	ld c, 16
	;add hl, bc				; comment out for little endian
	;inc c
.loop_op2:
	ld a, (hl)
	push hl
		ld b, 8
.loop_bits_in_byte:
		rra
		push af
			sbc a,a
			push bc
				ld c,a

				; add op1 (res) + tmp
				ld hl, (ix + 6)		; hl = (dest)
				lea de, ix - 16		; de = tmp (src)
				ld b, 16
.loop_add:
				ld a, (de)
				and a, c
				xor a, (hl)
				ld (hl), a
				inc hl
				inc de
				djnz .loop_add

               ; now double tmp
				lea hl, ix - 16		; tmp in hl	little endian
                ;lea hl, ix - 1			; tmp in hl big endian
                ld b, 16
                or a                ; reset carry
.loop_mul2:
                rl (hl)
                inc hl		; little endian
                ;dec hl		; big endian
                djnz .loop_mul2

                ; now xor with polynomial x^128 + x^7 + x^2 + x + 1
                ; if bit 128 set, xor least-significant byte with 10000111b

                sbc a, a
                and a, 10000111b
                xor a, (ix - 16)		; little endian
                ld (ix - 16), a
                ;xor a, (ix - 1)			; big endian
                ;ld (ix - 1), a

.no_xor_poly:
			pop bc
		pop af
		djnz .loop_bits_in_byte
	pop hl
	inc hl		; little endian
	;dec hl			; big endian
	dec c
	jr nz, .loop_op2
	ld sp, ix
	pop ix
	ret