; 6502 mode
.p02

; define where we want to be loaded
.global __LOADADDR__
__LOADADDR__ = $7000
; and where data should reside
.global __DATAADDR__
__DATAADDR__ = $7200

; load useful register/memory locations
.include "pet.inc"

; for some reason, the PIA registers are missing in pet.inc...
PIA1 := $E810
PIA1_PA  := PIA1+$0 ; PORT A or DDR A: Data Direction Register A
PIA1_CRA := PIA1+$1 ; CRA: Control Register A
PIA1_PB  := PIA1+$2 ; PORT B or DDR B: Data Direction Register B
PIA1_CRB := PIA1+$3 ; CRB: Control Register B
PIA2 := $E820
PIA2_PA  := PIA2+$0
PIA2_CRA := PIA2+$1
PIA2_PB  := PIA2+$2
PIA2_CRB := PIA2+$3

; 1MHz phase2 clock
PHI2_CLOCK = 1000000
; don't set the baud rate too high, or other operations will be starved
BAUD_RATE = 300

; parameters and return values can reside in an unused area of the zero page
.zeropage
	; available bytes to read or write
	.export rs_available
	rs_available: .byte 0
	; single byte transfer from input / to output fifo
	.export rs_data
	rs_data: .byte 0
	; status of operation
	rs_status_ok = 0
	rs_status_read_buffer_empty = 1
	rs_status_write_buffer_full = 2
	rs_status_device_not_initialized = 3
	rs_status_device_already_initialized = 4
	.export rs_status
	rs_status: .byte 0

; specify the load address, so code and data will be located there
;.org __LOADADDR__

; entry points
.code
	; these are convenience entry points right at the beginning of the page,
	; to reduce dependency on code size.
	; relocatable code would be perfect, but that's a lot more work.
	.export rs_install
	rs_install:
		jmp install
	.export rs_uninstall
	rs_uninstall:
		jmp uninstall
	.export rs_read
	rs_read:
		jmp read
	.export rs_write
	rs_write:
		jmp write

; driver state data
.data
	; 1=driver initialized
	initialized: .byte 0
	; saved IRQ vector
	oldvector: .word 0
	; state of the CA1 input: bit2=0 low level, bit2=1 high level
	ca1state: .byte 0
	; output buffer
	; this is 16bit, because we need to process start and stop bits as well
	outbuf: .word 0
	; current shifted bit (if 0, no data is being transferred)
	outshift: .byte 0
	; output queue (16 bytes)
	; this is a push from the front - pop from the tail fifo
	outqlen: .byte 0
	outq: .res 16, 0
	; input buffer
	inbuf: .word 0
	; current shifted bit (if 0, no data is being transferred)
	inshift: .byte 0
	; input queue (16 bytes)
	; this is a push from the tail - pop from the front fifo
	inqlen: .byte 0
	inq: .res 16, 0

; main code follows
.code

	; driver installation, must be called once to set up IRQs, etc.
	install:
		; check if we're already initialized
		lda initialized
		beq @hwinit
		; signal error
		lda #rs_status_device_already_initialized
		sta rs_status
		; and return
		rts

	@hwinit:
		; disable interrupts, so we're not disturbed
		cli
		
		; save the previous handler
		lda IRQVec
		sta oldvector
		lda IRQVec+1
		sta oldvector+1
		; assign our custom interrupt handler to the IRQ vector
		lda #>irqhandler
		sta IRQVec
		lda #<irqhandler
		sta IRQVec+1

		; VIA timer 2 cannot be enabled and disabled on the fly, so we'll
		; configure it here and arm it later
		; one-shot mode
		lda VIA_CR
		and #%11011111
		sta VIA_CR
		; enable interrupt
		; interrupt enable is done by setting bit7 and the desired interrupt
		; bit in IER to 1 (no need to read-modify-write)
		lda #%10100000
		sta VIA_IER

		; set up the PIA1 and VIA to process I/O
		; TXD: VIA PB3, output
		lda VIA_DDRB
		ora #%00001000
		sta VIA_DDRB
		; RXD: PIA1 CA1, input, interrupt
		lda PIA1_CRA
		; we start with the negative transition (RS232 start bit: H->L)
		and #%11111101
		; IRQ enable
		ora #%00000001
		sta PIA1_CRA
		; CTS: PIA1 CB2, output (optional)
		; RTS: PIA1 PA4, input (optional)
		
		; clear some of the state variables
		; the buffers are controlled by other variables anyway and don't need clearing
		lda #0
		sta ca1state
		sta outshift
		sta outqlen
		sta inshift
		sta inqlen
		
		; we're ready
		lda #1
		sta initialized
		; and ok
		lda #rs_status_ok
		sta rs_status
		; fire away
		sei
		; return
		rts

	; driver uninstallation, must be called once to set up IRQs, etc.
	.export uninstall
	uninstall:
		; check if we're already initialized
		lda initialized
		bne @hwuninit
		; signal error
		lda #rs_status_device_not_initialized
		sta rs_status
		; and return
		rts

	@hwuninit:
		; disable interrupts, so we're not disturbed
		cli

		; restore the previous handler
		lda oldvector
		ldx oldvector+1
		sta IRQVec
		stx IRQVec+1

		; disable all interrupts
		; RXD: PIA1 CA1
		lda PIA1_CRA
		; IRQ disable
		and #%11111110
		sta PIA1_CRA
		; and the VIA timer 2
		; clearing is done by setting bit 7 to 0 and the desired interrupt to 1
		lda #%00100000
		sta VIA_IER

		; we leave the I/O pins and the state alone

		; but we'll go back to uninitalized state
		lda #0
		sta initialized

		; and ok
		lda #rs_status_ok
		sta rs_status
		; the rest of the system will probably want interrupts enabled
		sei
		; return
		rts
		

	; write one byte to FIFO
	.export write
	write:
		; check if we're already initialized
		lda initialized
		bne @dowrite

		; signal error
		lda #rs_status_device_not_initialized
		sta rs_status
		; and return
		rts

	@dowrite:
		; disable interrupts while we write into the buffer
		cli

		; test if there's space in the buffer first
		lda #16
		cmp outqlen
		beq @wrnodata

		; move data up first
		ldy outqlen
		; check if we need to move data at all
		bne @wrbufend

		; move all existing elements up, starting from the top
	@wrbufloop:
		lda outq-1,y
		sta outq,y
		dey
		bne @wrbufloop
	@wrbufend:

		; write data to head
		lda rs_data
		sta outq
		; update queue length
		ldy outqlen
		iny
		sty outqlen
		; also store to return value
		sty rs_available

		; all right
		lda #rs_status_ok
		sta rs_status

		; an unconditional branch would be useful...
		clc
		bcc @wrdone

	@wrnodata:
		; return error
		lda #rs_status_write_buffer_full
		sta rs_status

	@wrdone:
		; and re-enable
		sei

		; return
		rts

	; read one byte from FIFO
	.export read
	read:
		; check if we're already initialized
		lda initialized
		bne @doread

		; signal error
		lda #rs_status_device_not_initialized
		sta rs_status
		; and return
		rts

	@doread:
		; disable interrupts while we read the buffer
		cli

		; test if we have any data first
		lda inqlen
		beq @rdnodata

		; return data from head
		lda inq
		sta rs_data

		; update the queue length
		ldy inqlen
		dey
		sty inqlen
		; also store to return value
		sty rs_available
		; check if we need to move any data
		bne @rdbufend

		; move all other data elements down
		ldx #0
	@rdbufloop:
		lda inq+1,x
		sta inq,x
		inx
		dey
		bne @rdbufloop
	@rdbufend:

		; all right
		lda #rs_status_ok
		sta rs_status

		; an unconditional branch would be useful...
		clc
		bcc @rddone

	@rdnodata:
		; return error
		lda #rs_status_read_buffer_empty
		sta rs_status

	@rddone:
		; and re-enable
		sei

		; return
		rts

	; start the timer
	starttimer:
		; arm VIA timer 2 by latching the counter
		period = PHI2_CLOCK/BAUD_RATE
		lda #>period
		sta VIA_T1CL
		lda #<period
		sta VIA_T1CH
		; return
		rts
	
	; stop the timer
	@stoptimer:
		; VIA timer 1 disabled
		lda VIA_CR
		and #%10111111
		sta VIA_CR
		; return
		rts

	; IRQ handler
	.interruptor irqhandler
	irqhandler:
		; disable interrupts
		cli
		; clear decimal flag to avoid unexpected behavior
		cld
		; save registers
		pha
		txa
		pha
		tya
		pha

		; test for the interrupts we're expecting

	@handleflip:
		; read PIA1 CRA, interrupt flag is automatically cleared
		lda PIA1_CRA
		; save for later, so we don't need to fetch it again
		tax
		; bit 7 is the CA1 interrupt flag - fired?
		and #%10000000
		; nope, skip flip handling
		beq @handletimer

		; handle CA1 interrupt
		; fetch saved CRA
		txa
		; save current value for later use:
		; bit 2 is contains the active value of CA1, which is what we want to know
		sta ca1state
		; invert transition (H->L <-> L->H)
		eor #%00000010
		; mask out interrupt flag - we don't want another IRQ right away
		and #%01111111
		sta PIA1_CRA

	@handletimer:
		; VIA timer 2 fired?
		lda VIA_IFR
		and #%00100000
		; zero, no interrupt
		beq @irqreturn

		; timer 2 fired: rearm (we're in one-shot mode!), and clear the interrupt flag
		; FIXME this should be done earlier, for more accurate timing
		; FIXME we could skip writing the low byte here, writing the high byte is enough
		jsr starttimer

	@loadout:
		; do we have more bits to send?
		lda outshift
		; yes, skip loading next byte
		bne @shiftout

		; load next byte from FIFO

		; do we have more bytes?
		lda outqlen
		; nope, skip ahead to input processing
		beq @shiftin

		; load length
		ldx outqlen
		; decrement, also gives the index
		dex
		; load next byte from fifo
		lda outq,x
		; store new queue length
		stx outqlen

		; prepare shift register
		; layout: 000000E7 6543210S
		; bit0 = start bit = 0
		; bit1..7 = data0..6
		clc
		rol
		sta outbuf
		; bit8 = data7
		; bit9 = stop bit = 1
		lda #%00000001
		rol
		sta outbuf+1
		; we're shifting 10 bits out
		lda #10
		sta outshift

	@shiftout:
		; make sure overflow flag is clear, so we can do unconditional branches
		clv
		; load shift register bit 0
		lda outbuf
		and #%00000001
		; set or clear?
		beq @shiftoutclear
		; set output
		lda #%00001000
		ora VIA_PB
		bvc @shiftoutstore
	@shiftoutclear:
		; clear output
		lda #%11110111
		and VIA_PB
	@shiftoutstore:
		sta VIA_PB

		; shift the next bit in
		; 0 -> high byte -> carry
		lsr outbuf+1
		; carry -> low byte
		ror outbuf

		; decrement counter
		dec outshift

	@shiftin:
		; TODO process input bit-bang
	@storein:
		; TODO process input fifo

	@irqreturn:
		; restore registers
		pla
		tay
		pla
		tax
		pla
		; enable interrupts
		; FIXME should we do this here? what's common practice? is this compatible with other interrupt handlers
		sei
		; jump to previous handler
		; we don't need to return or restore flags, this will be done by the chained interrupt handler
		jmp (oldvector)