#include <common.h>

#define _INASM_
#include "usb_table.h"

#define PIN_OUT        0x60000300
#define PIN_OUT_SET    0x60000304
#define PIN_OUT_CLEAR  0x60000308
#define PIN_DIR        0x6000030C
#define PIN_DIR_OUTPUT 0x60000310
#define PIN_DIR_INPUT  0x60000314
#define PIN_IN         0x60000318
#define _BV(x) ((1)<<(x))


#define DEBUG_HIGH    s32i.n	a13, a12, 0
#define DEBUG_LOW     s32i.n	a13, a12, 4
//(PIN_OUT_CLEAR - PIN_OUT_SET)

//Long versions of the commands to help with byte-alignment? Maybe?
#define DEBUG_HIGH_L  .byte 0xd2; .byte 0x6c; .byte 0x00
#define DEBUG_LOW_L   .byte 0xd2; .byte 0x6c; .byte 0x01
#define NOP_L			.byte 0xf0;	.byte 0x20;	.byte 0x00

//4010113f:	1cd9         	1.n	a13, a12, 4
//40101141:	0cd9      	s32i.n	a13, a12, 0


//Detailed analysis of some useful stuff and performance tweaking: http://naberius.de/2015/05/14/esp8266-gpio-output-performance/
//Reverse engineerd boot room can be helpful, too: http://cholla.mmto.org/esp8266/bootrom/boot.txt
//USB Protocol read from wikipedia: https://en.wikipedia.org/wiki/USB

#define SIZE_OF_BUFFER 24
.global usb_buffer
#define usb_buffer_end (usb_buffer + SIZE_OF_BUFFER)

.global usb_buff_data
.global gpio_intr

gpio_intr:
	addi	a1, a1, -60  //Must be 4 higher than highest pusher
	s32i.n	a0, a1, 52
	s32i.n	a2, a1, 48
	s32i.n	a3, a1, 44
	s32i.n	a8, a1, 24
	s32i.n	a9, a1, 20
	s32i.n	a10, a1, 16
	s32i.n	a11, a1, 12 
	s32i.n	a12, a1, 8    //debug pin address
	s32i.n	a13, a1, 4    //debug pin
	s32i.n	a14, a1, 0    //usb_ramtable

	movi    a14, usb_ramtable	//This is actually very slow.
	movi    a13, 1<<DEBUGPIN

	DISABLE_INTERRUPTS_LCL

	l32i    a12, a14, PIN_OUT_SET_OFFSET

	//Acknowledge interrupt
	l32i.n a2, a14, GPIO_STATUS_OFFSET
	l32i.n a4, a2, 0
	s32i.n a4, a2, (GPIO_STATUS_W1TC - GPIO_STATUS)

	movi a9, 512
rkloop:
	nop
	DEBUG_HIGH
	nop
	addi a9, a9, -1
	nop
	DEBUG_LOW
	nop
	bnez a9, rkloop

/*
	movi a9, 1023 #timeout
find_high:
    l32i.n a2, a11, 0
	bbsi a2, DPLUS, done_high
	addi.n a9, a9, -1
	bnez a9, find_high
	j end_gpio_intr



	DEBUG_LOW

	//TODO: Make a mode that waits for data to finish out, despite there being an error.
	movi a0, usb_buff_data
	movi a2, 0xffffffff
	s32i a0, a2, 0
end_intr_with_extra_vars:
	l32i.n	a4, a1, 44
	l32i.n	a5, a1, 40
	l32i.n	a6, a1, 36
	l32i.n	a7, a1, 32
	l32i.n	a8, a1, 28

end_gpio_intr:
*/

	DEBUG_LOW
	DEBUG_LOW
	DEBUG_LOW
	DEBUG_LOW

	ENABLE_INTERRUPTS_LCL

	l32i.n	a0, a1, 52
	l32i.n	a2, a1, 48
	l32i.n	a3, a1, 44
	l32i.n	a8, a1, 24
	l32i.n	a9, a1, 20
	l32i.n	a10, a1, 16
	l32i.n	a11, a1, 12
	l32i.n	a12, a1, 8
	l32i.n	a13, a1, 4
	l32i.n	a14, a1, 0
	addi	a1, a1, 60
	ret.n




.global my_table
.global my_func
my_func:
	addi a1, a1, -24
	s32i.n a0, a1, 0
	s32i.n a2, a1, 4
	s32i.n a3, a1, 8
	s32i.n a12, a1, 12
	s32i.n a13, a1, 16
	s32i.n a14, a1, 20

	movi    a14, usb_ramtable	//This is actually very slow.
	movi    a13, 1<<DEBUGPIN
	l32i    a12, a14, PIN_OUT_SET_OFFSET

//At 80
//49 = 1179
//50 = 1203  (12 cycles) (DBH + DBL)
//49 = 591 (6 cycles)
//50 = 603 (DBL only)
//49 = 425 (NOTHING 5 cycles)
//At 160
//49 = 1138 / 50 = 1162 (DBL only -> 12 cycles)
//Both
//49 = 2314 / 50 = 2362 (DBL+DBH = 24 cycles
	movi a3, 50
	slli a3, a3, 1


	nop
	rsr a9, ccount
	nop
retry:
	DEBUG_HIGH
	nop; nop; nop; nop; nop; nop;
//	nop; nop; nop; nop; nop; nop;

	//Just checking:  Yep!  This is 6 cycles.
	add a11, a14, a3
	l8ui a2, a11, GPIO_STATUS_OFFSET
	nop;
	bbsi a2, 0, skipper
	nop;
	nop;
skipper:

	DEBUG_LOW
	nop; nop; nop; nop; nop; nop;
	nop; nop; nop; nop; nop; nop;

	DEBUG_HIGH
	nop; nop; nop; nop; nop; nop;
	nop; nop; nop; nop; nop; nop;

	DEBUG_LOW
	nop; nop; nop; nop; nop; nop;
	nop; nop; nop; nop; nop; nop;

	DEBUG_HIGH
	nop; nop; nop; nop; nop; nop;
	nop; nop; nop; nop; nop; nop

	DEBUG_LOW
	nop; nop; nop; nop; nop; nop;
	nop; nop; nop; nop; 

	addi a3, a3, -1
	bnez a3, retry

rsr a11, ccount

	l32i.n a14, a1, 20
	l32i.n a13, a1, 16
	l32i.n a12, a1, 12
	l32i.n a3, a1, 8
	l32i.n a2, a1, 4
	l32i.n a0, a1, 0
	addi a1, a1, 24
	ret.n