This is a relatively significant edit that brings the Arduino copy of

optiboot up-to-date with the optiboot source repository as of Jun-2011
(the last changes made in the optiboot repository were in Oct-2010)

This adds support for several plaforms, fixes the "30k bug", and
refactors the source to have separate stk500.h, boot.h, and pin_defs.h

These are the arduino opticode issues fixed:
http://code.google.com/p/arduino/issues/detail?id=380
    optiboot has problems upload sketches bigger than 30 KB
http://code.google.com/p/arduino/issues/detail?id=556
    update optiboot to the point of the latest optiboot project sources.

These are issues that had been solved in the optiboot source aready:
http://code.google.com/p/arduino/issues/detail?id=364
   optiboot leaves timer1 configured when starting app, breaks PWM on
   pin 9 and 10.  (fixed with a workaround in arduino core.)
   aka http://code.google.com/p/optiboot/source/detail?r=c778fbe72df6ac13ef730c25283358c3c970f73e
   Support for ATmega8 and mega88.
   Fix fuse settings for mega168 _ISP targets
   Additional new platforms (mega, sanguino)

http://code.google.com/p/optiboot/issues/detail?id=26
   Set R1 to 0  (already in arduino code)
http://code.google.com/p/optiboot/issues/detail?id=36&can=1
   Fails to build correctly for mega88

After this commit, the only differences between the Arduino optiboot.c
and the optiboot repository optiboot.c are cosmetic.
(cherry picked from commit e2812ef91cc1489527827dcd14c843f3b3244f36)

1 files changed, 280 insertions, 259 deletions

diff --git a/bootloaders/optiboot/optiboot_lilypad.lst b/bootloaders/optiboot/optiboot_lilypad.lst
index 0e46bd1..cbae268 100644
--- a/bootloaders/optiboot/optiboot_lilypad.lst
+++ b/bootloaders/optiboot/optiboot_lilypad.lst
@@ -7,452 +7,473 @@ Idx Name          Size      VMA       LMA       File off  Algn
                   CONTENTS, ALLOC, LOAD, READONLY, CODE
   1 .debug_aranges 00000028  00000000  00000000  00000250  2**0
                   CONTENTS, READONLY, DEBUGGING
-  2 .debug_pubnames 00000078  00000000  00000000  00000278  2**0
+  2 .debug_pubnames 0000006a  00000000  00000000  00000278  2**0
                   CONTENTS, READONLY, DEBUGGING
-  3 .debug_info   00000277  00000000  00000000  000002f0  2**0
+  3 .debug_info   00000284  00000000  00000000  000002e2  2**0
                   CONTENTS, READONLY, DEBUGGING
-  4 .debug_abbrev 00000194  00000000  00000000  00000567  2**0
+  4 .debug_abbrev 000001ae  00000000  00000000  00000566  2**0
                   CONTENTS, READONLY, DEBUGGING
-  5 .debug_line   000003bb  00000000  00000000  000006fb  2**0
+  5 .debug_line   00000450  00000000  00000000  00000714  2**0
                   CONTENTS, READONLY, DEBUGGING
-  6 .debug_frame  000000a0  00000000  00000000  00000ab8  2**2
+  6 .debug_frame  00000090  00000000  00000000  00000b64  2**2
                   CONTENTS, READONLY, DEBUGGING
-  7 .debug_str    0000013f  00000000  00000000  00000b58  2**0
+  7 .debug_str    00000141  00000000  00000000  00000bf4  2**0
                   CONTENTS, READONLY, DEBUGGING
-  8 .debug_loc    000001a0  00000000  00000000  00000c97  2**0
+  8 .debug_loc    000001e1  00000000  00000000  00000d35  2**0
                   CONTENTS, READONLY, DEBUGGING
-  9 .debug_ranges 00000070  00000000  00000000  00000e37  2**0
+  9 .debug_ranges 00000068  00000000  00000000  00000f16  2**0
                   CONTENTS, READONLY, DEBUGGING
 
 Disassembly of section .text:
 
 00003e00 <main>:
-#ifdef VIRTUAL_BOOT_PARTITION
-#define rstVect (*(uint16_t*)(0x204))
-#define wdtVect (*(uint16_t*)(0x206))
+#define rstVect (*(uint16_t*)(RAMSTART+SPM_PAGESIZE*2+4))
+#define wdtVect (*(uint16_t*)(RAMSTART+SPM_PAGESIZE*2+6))
 #endif
+
 /* main program starts here */
 int main(void) {
-    3e00:	85 e0       	ldi	r24, 0x05	; 5
-    3e02:	80 93 81 00 	sts	0x0081, r24
-  UCSR0C = _BV(UCSZ00) | _BV(UCSZ01);
-  UBRR0L = (uint8_t)( (F_CPU + BAUD_RATE * 4L) / (BAUD_RATE * 8L) - 1 );
+    3e00:	11 24       	eor	r1, r1
+#ifdef __AVR_ATmega8__
+  SP=RAMEND;  // This is done by hardware reset
 #endif
 
   // Adaboot no-wait mod
   ch = MCUSR;
-    3e06:	84 b7       	in	r24, 0x34	; 52
+    3e02:	84 b7       	in	r24, 0x34	; 52
   MCUSR = 0;
-    3e08:	14 be       	out	0x34, r1	; 52
+    3e04:	14 be       	out	0x34, r1	; 52
   if (!(ch & _BV(EXTRF))) appStart();
-    3e0a:	81 ff       	sbrs	r24, 1
-    3e0c:	e4 d0       	rcall	.+456    	; 0x3fd6 <appStart>
+    3e06:	81 ff       	sbrs	r24, 1
+    3e08:	e6 d0       	rcall	.+460    	; 0x3fd6 <appStart>
+
+#if LED_START_FLASHES > 0
+  // Set up Timer 1 for timeout counter
+  TCCR1B = _BV(CS12) | _BV(CS10); // div 1024
+    3e0a:	85 e0       	ldi	r24, 0x05	; 5
+    3e0c:	80 93 81 00 	sts	0x0081, r24
+  UCSRA = _BV(U2X); //Double speed mode USART
+  UCSRB = _BV(RXEN) | _BV(TXEN);  // enable Rx & Tx
+  UCSRC = _BV(URSEL) | _BV(UCSZ1) | _BV(UCSZ0);  // config USART; 8N1
+  UBRRL = (uint8_t)( (F_CPU + BAUD_RATE * 4L) / (BAUD_RATE * 8L) - 1 );
+#else
+  UCSR0A = _BV(U2X0); //Double speed mode USART0
+    3e10:	82 e0       	ldi	r24, 0x02	; 2
+    3e12:	80 93 c0 00 	sts	0x00C0, r24
+  UCSR0B = _BV(RXEN0) | _BV(TXEN0);
+    3e16:	88 e1       	ldi	r24, 0x18	; 24
+    3e18:	80 93 c1 00 	sts	0x00C1, r24
+  UCSR0C = _BV(UCSZ00) | _BV(UCSZ01);
+    3e1c:	86 e0       	ldi	r24, 0x06	; 6
+    3e1e:	80 93 c2 00 	sts	0x00C2, r24
+  UBRR0L = (uint8_t)( (F_CPU + BAUD_RATE * 4L) / (BAUD_RATE * 8L) - 1 );
+    3e22:	88 e0       	ldi	r24, 0x08	; 8
+    3e24:	80 93 c4 00 	sts	0x00C4, r24
+#endif
+#endif
 
   // Set up watchdog to trigger after 500ms
-  watchdogConfig(WATCHDOG_500MS);
-    3e0e:	8d e0       	ldi	r24, 0x0D	; 13
-    3e10:	dc d0       	rcall	.+440    	; 0x3fca <watchdogConfig>
+  watchdogConfig(WATCHDOG_1S);
+    3e28:	8e e0       	ldi	r24, 0x0E	; 14
+    3e2a:	cf d0       	rcall	.+414    	; 0x3fca <watchdogConfig>
 
   /* Set LED pin as output */
   LED_DDR |= _BV(LED);
-    3e12:	25 9a       	sbi	0x04, 5	; 4
-
-#ifdef SOFT_UART
-  /* Set TX pin as output */
-  UART_DDR |= _BV(UART_TX_BIT);
-    3e14:	51 9a       	sbi	0x0a, 1	; 10
-    3e16:	86 e0       	ldi	r24, 0x06	; 6
+    3e2c:	25 9a       	sbi	0x04, 5	; 4
+    3e2e:	86 e0       	ldi	r24, 0x06	; 6
 }
 
 #if LED_START_FLASHES > 0
 void flash_led(uint8_t count) {
   do {
     TCNT1 = -(F_CPU/(1024*16));
-    3e18:	28 e1       	ldi	r18, 0x18	; 24
-    3e1a:	3e ef       	ldi	r19, 0xFE	; 254
+    3e30:	28 e1       	ldi	r18, 0x18	; 24
+    3e32:	3e ef       	ldi	r19, 0xFE	; 254
     TIFR1 = _BV(TOV1);
-    3e1c:	91 e0       	ldi	r25, 0x01	; 1
+    3e34:	91 e0       	ldi	r25, 0x01	; 1
 }
 
 #if LED_START_FLASHES > 0
 void flash_led(uint8_t count) {
   do {
     TCNT1 = -(F_CPU/(1024*16));
-    3e1e:	30 93 85 00 	sts	0x0085, r19
-    3e22:	20 93 84 00 	sts	0x0084, r18
+    3e36:	30 93 85 00 	sts	0x0085, r19
+    3e3a:	20 93 84 00 	sts	0x0084, r18
     TIFR1 = _BV(TOV1);
-    3e26:	96 bb       	out	0x16, r25	; 22
+    3e3e:	96 bb       	out	0x16, r25	; 22
     while(!(TIFR1 & _BV(TOV1)));
-    3e28:	b0 9b       	sbis	0x16, 0	; 22
-    3e2a:	fe cf       	rjmp	.-4      	; 0x3e28 <main+0x28>
+    3e40:	b0 9b       	sbis	0x16, 0	; 22
+    3e42:	fe cf       	rjmp	.-4      	; 0x3e40 <main+0x40>
+#ifdef __AVR_ATmega8__
+    LED_PORT ^= _BV(LED);
+#else
     LED_PIN |= _BV(LED);
-    3e2c:	1d 9a       	sbi	0x03, 5	; 3
+    3e44:	1d 9a       	sbi	0x03, 5	; 3
   return getch();
 }
 
 // Watchdog functions. These are only safe with interrupts turned off.
 void watchdogReset() {
   __asm__ __volatile__ (
-    3e2e:	a8 95       	wdr
-    TCNT1 = -(F_CPU/(1024*16));
-    TIFR1 = _BV(TOV1);
-    while(!(TIFR1 & _BV(TOV1)));
+    3e46:	a8 95       	wdr
+    LED_PORT ^= _BV(LED);
+#else
     LED_PIN |= _BV(LED);
+#endif
     watchdogReset();
   } while (--count);
-    3e30:	81 50       	subi	r24, 0x01	; 1
-    3e32:	a9 f7       	brne	.-22     	; 0x3e1e <main+0x1e>
+    3e48:	81 50       	subi	r24, 0x01	; 1
+    3e4a:	a9 f7       	brne	.-22     	; 0x3e36 <main+0x36>
     /* get character from UART */
     ch = getch();
 
     if(ch == STK_GET_PARAMETER) {
       // GET PARAMETER returns a generic 0x03 reply - enough to keep Avrdude happy
       getNch(1);
-    3e34:	dd 24       	eor	r13, r13
-    3e36:	d3 94       	inc	r13
-        boot_page_fill((uint16_t)(void*)addrPtr,a);
+    3e4c:	dd 24       	eor	r13, r13
+    3e4e:	d3 94       	inc	r13
+        __boot_page_fill_short((uint16_t)(void*)addrPtr,a);
         addrPtr += 2;
       } while (--ch);
       
       // Write from programming buffer
-      boot_page_write((uint16_t)(void*)address);
-    3e38:	a5 e0       	ldi	r26, 0x05	; 5
-    3e3a:	ea 2e       	mov	r14, r26
+      __boot_page_write_short((uint16_t)(void*)address);
+    3e50:	a5 e0       	ldi	r26, 0x05	; 5
+    3e52:	ea 2e       	mov	r14, r26
       boot_spm_busy_wait();
 
 #if defined(RWWSRE)
       // Reenable read access to flash
       boot_rww_enable();
-    3e3c:	f1 e1       	ldi	r31, 0x11	; 17
-    3e3e:	ff 2e       	mov	r15, r31
+    3e54:	f1 e1       	ldi	r31, 0x11	; 17
+    3e56:	ff 2e       	mov	r15, r31
 #endif
 
   /* Forever loop */
   for (;;) {
     /* get character from UART */
     ch = getch();
-    3e40:	ab d0       	rcall	.+342    	; 0x3f98 <getch>
+    3e58:	ab d0       	rcall	.+342    	; 0x3fb0 <getch>
 
     if(ch == STK_GET_PARAMETER) {
-    3e42:	81 34       	cpi	r24, 0x41	; 65
-    3e44:	21 f4       	brne	.+8      	; 0x3e4e <main+0x4e>
+    3e5a:	81 34       	cpi	r24, 0x41	; 65
+    3e5c:	21 f4       	brne	.+8      	; 0x3e66 <main+0x66>
       // GET PARAMETER returns a generic 0x03 reply - enough to keep Avrdude happy
       getNch(1);
-    3e46:	81 e0       	ldi	r24, 0x01	; 1
-    3e48:	d1 d0       	rcall	.+418    	; 0x3fec <verifySpace+0xc>
+    3e5e:	81 e0       	ldi	r24, 0x01	; 1
+    3e60:	c5 d0       	rcall	.+394    	; 0x3fec <getNch>
       putch(0x03);
-    3e4a:	83 e0       	ldi	r24, 0x03	; 3
-    3e4c:	24 c0       	rjmp	.+72     	; 0x3e96 <main+0x96>
+    3e62:	83 e0       	ldi	r24, 0x03	; 3
+    3e64:	20 c0       	rjmp	.+64     	; 0x3ea6 <main+0xa6>
     }
     else if(ch == STK_SET_DEVICE) {
-    3e4e:	82 34       	cpi	r24, 0x42	; 66
-    3e50:	11 f4       	brne	.+4      	; 0x3e56 <main+0x56>
+    3e66:	82 34       	cpi	r24, 0x42	; 66
+    3e68:	11 f4       	brne	.+4      	; 0x3e6e <main+0x6e>
       // SET DEVICE is ignored
       getNch(20);
-    3e52:	84 e1       	ldi	r24, 0x14	; 20
-    3e54:	03 c0       	rjmp	.+6      	; 0x3e5c <main+0x5c>
+    3e6a:	84 e1       	ldi	r24, 0x14	; 20
+    3e6c:	03 c0       	rjmp	.+6      	; 0x3e74 <main+0x74>
     }
     else if(ch == STK_SET_DEVICE_EXT) {
-    3e56:	85 34       	cpi	r24, 0x45	; 69
-    3e58:	19 f4       	brne	.+6      	; 0x3e60 <main+0x60>
+    3e6e:	85 34       	cpi	r24, 0x45	; 69
+    3e70:	19 f4       	brne	.+6      	; 0x3e78 <main+0x78>
       // SET DEVICE EXT is ignored
       getNch(5);
-    3e5a:	85 e0       	ldi	r24, 0x05	; 5
-    3e5c:	c7 d0       	rcall	.+398    	; 0x3fec <verifySpace+0xc>
-    3e5e:	8a c0       	rjmp	.+276    	; 0x3f74 <main+0x174>
+    3e72:	85 e0       	ldi	r24, 0x05	; 5
+    3e74:	bb d0       	rcall	.+374    	; 0x3fec <getNch>
+    3e76:	91 c0       	rjmp	.+290    	; 0x3f9a <main+0x19a>
     }
     else if(ch == STK_LOAD_ADDRESS) {
-    3e60:	85 35       	cpi	r24, 0x55	; 85
-    3e62:	a1 f4       	brne	.+40     	; 0x3e8c <main+0x8c>
+    3e78:	85 35       	cpi	r24, 0x55	; 85
+    3e7a:	81 f4       	brne	.+32     	; 0x3e9c <main+0x9c>
       // LOAD ADDRESS
-      address = getch();
-    3e64:	99 d0       	rcall	.+306    	; 0x3f98 <getch>
-    3e66:	08 2f       	mov	r16, r24
-    3e68:	10 e0       	ldi	r17, 0x00	; 0
-    3e6a:	10 93 01 02 	sts	0x0201, r17
-    3e6e:	00 93 00 02 	sts	0x0200, r16
-      address = (address & 0xff) | (getch() << 8);
-    3e72:	92 d0       	rcall	.+292    	; 0x3f98 <getch>
-    3e74:	90 e0       	ldi	r25, 0x00	; 0
-    3e76:	98 2f       	mov	r25, r24
-    3e78:	88 27       	eor	r24, r24
-    3e7a:	80 2b       	or	r24, r16
-    3e7c:	91 2b       	or	r25, r17
-      address += address; // Convert from word address to byte address
-    3e7e:	88 0f       	add	r24, r24
-    3e80:	99 1f       	adc	r25, r25
-    3e82:	90 93 01 02 	sts	0x0201, r25
-    3e86:	80 93 00 02 	sts	0x0200, r24
-    3e8a:	73 c0       	rjmp	.+230    	; 0x3f72 <main+0x172>
+      uint16_t newAddress;
+      newAddress = getch();
+    3e7c:	99 d0       	rcall	.+306    	; 0x3fb0 <getch>
+      newAddress = (newAddress & 0xff) | (getch() << 8);
+    3e7e:	08 2f       	mov	r16, r24
+    3e80:	10 e0       	ldi	r17, 0x00	; 0
+    3e82:	96 d0       	rcall	.+300    	; 0x3fb0 <getch>
+    3e84:	90 e0       	ldi	r25, 0x00	; 0
+    3e86:	98 2f       	mov	r25, r24
+    3e88:	88 27       	eor	r24, r24
+    3e8a:	80 2b       	or	r24, r16
+    3e8c:	91 2b       	or	r25, r17
+#ifdef RAMPZ
+      // Transfer top bit to RAMPZ
+      RAMPZ = (newAddress & 0x8000) ? 1 : 0;
+#endif
+      newAddress += newAddress; // Convert from word address to byte address
+    3e8e:	88 0f       	add	r24, r24
+    3e90:	99 1f       	adc	r25, r25
+      address = newAddress;
+    3e92:	90 93 01 02 	sts	0x0201, r25
+    3e96:	80 93 00 02 	sts	0x0200, r24
+    3e9a:	7e c0       	rjmp	.+252    	; 0x3f98 <main+0x198>
       verifySpace();
     }
     else if(ch == STK_UNIVERSAL) {
-    3e8c:	86 35       	cpi	r24, 0x56	; 86
-    3e8e:	29 f4       	brne	.+10     	; 0x3e9a <main+0x9a>
+    3e9c:	86 35       	cpi	r24, 0x56	; 86
+    3e9e:	29 f4       	brne	.+10     	; 0x3eaa <main+0xaa>
       // UNIVERSAL command is ignored
       getNch(4);
-    3e90:	84 e0       	ldi	r24, 0x04	; 4
-    3e92:	ac d0       	rcall	.+344    	; 0x3fec <verifySpace+0xc>
+    3ea0:	84 e0       	ldi	r24, 0x04	; 4
+    3ea2:	a4 d0       	rcall	.+328    	; 0x3fec <getNch>
       putch(0x00);
-    3e94:	80 e0       	ldi	r24, 0x00	; 0
-    3e96:	71 d0       	rcall	.+226    	; 0x3f7a <putch>
-    3e98:	6d c0       	rjmp	.+218    	; 0x3f74 <main+0x174>
+    3ea4:	80 e0       	ldi	r24, 0x00	; 0
+    3ea6:	7c d0       	rcall	.+248    	; 0x3fa0 <putch>
+    3ea8:	78 c0       	rjmp	.+240    	; 0x3f9a <main+0x19a>
     }
-    /* Write memory, length is big endian and is in bytes  */
+    /* Write memory, length is big endian and is in bytes */
     else if(ch == STK_PROG_PAGE) {
-    3e9a:	84 36       	cpi	r24, 0x64	; 100
-    3e9c:	09 f0       	breq	.+2      	; 0x3ea0 <main+0xa0>
-    3e9e:	43 c0       	rjmp	.+134    	; 0x3f26 <main+0x126>
+    3eaa:	84 36       	cpi	r24, 0x64	; 100
+    3eac:	09 f0       	breq	.+2      	; 0x3eb0 <main+0xb0>
+    3eae:	4e c0       	rjmp	.+156    	; 0x3f4c <main+0x14c>
       // PROGRAM PAGE - we support flash programming only, not EEPROM
       uint8_t *bufPtr;
       uint16_t addrPtr;
 
       getLen();
-    3ea0:	8f d0       	rcall	.+286    	; 0x3fc0 <getLen>
-
-      // Immediately start page erase - this will 4.5ms
-      boot_page_erase((uint16_t)(void*)address);
-    3ea2:	e0 91 00 02 	lds	r30, 0x0200
-    3ea6:	f0 91 01 02 	lds	r31, 0x0201
-    3eaa:	83 e0       	ldi	r24, 0x03	; 3
-    3eac:	80 93 57 00 	sts	0x0057, r24
-    3eb0:	e8 95       	spm
-    3eb2:	c0 e0       	ldi	r28, 0x00	; 0
-    3eb4:	d1 e0       	ldi	r29, 0x01	; 1
-
+    3eb0:	87 d0       	rcall	.+270    	; 0x3fc0 <getLen>
+
+      // If we are in RWW section, immediately start page erase
+      if (address < NRWWSTART) __boot_page_erase_short((uint16_t)(void*)address);
+    3eb2:	e0 91 00 02 	lds	r30, 0x0200
+    3eb6:	f0 91 01 02 	lds	r31, 0x0201
+    3eba:	88 e3       	ldi	r24, 0x38	; 56
+    3ebc:	e0 30       	cpi	r30, 0x00	; 0
+    3ebe:	f8 07       	cpc	r31, r24
+    3ec0:	18 f4       	brcc	.+6      	; 0x3ec8 <main+0xc8>
+    3ec2:	83 e0       	ldi	r24, 0x03	; 3
+    3ec4:	87 bf       	out	0x37, r24	; 55
+    3ec6:	e8 95       	spm
+    3ec8:	c0 e0       	ldi	r28, 0x00	; 0
+    3eca:	d1 e0       	ldi	r29, 0x01	; 1
+      
       // While that is going on, read in page contents
       bufPtr = buff;
       do *bufPtr++ = getch();
-    3eb6:	70 d0       	rcall	.+224    	; 0x3f98 <getch>
-    3eb8:	89 93       	st	Y+, r24
+    3ecc:	71 d0       	rcall	.+226    	; 0x3fb0 <getch>
+    3ece:	89 93       	st	Y+, r24
       while (--length);
-    3eba:	80 91 02 02 	lds	r24, 0x0202
-    3ebe:	81 50       	subi	r24, 0x01	; 1
-    3ec0:	80 93 02 02 	sts	0x0202, r24
-    3ec4:	88 23       	and	r24, r24
-    3ec6:	b9 f7       	brne	.-18     	; 0x3eb6 <main+0xb6>
+    3ed0:	80 91 02 02 	lds	r24, 0x0202
+    3ed4:	81 50       	subi	r24, 0x01	; 1
+    3ed6:	80 93 02 02 	sts	0x0202, r24
+    3eda:	88 23       	and	r24, r24
+    3edc:	b9 f7       	brne	.-18     	; 0x3ecc <main+0xcc>
+
+      // If we are in NRWW section, page erase has to be delayed until now.
+      // Todo: Take RAMPZ into account
+      if (address >= NRWWSTART) __boot_page_erase_short((uint16_t)(void*)address);
+    3ede:	e0 91 00 02 	lds	r30, 0x0200
+    3ee2:	f0 91 01 02 	lds	r31, 0x0201
+    3ee6:	88 e3       	ldi	r24, 0x38	; 56
+    3ee8:	e0 30       	cpi	r30, 0x00	; 0
+    3eea:	f8 07       	cpc	r31, r24
+    3eec:	18 f0       	brcs	.+6      	; 0x3ef4 <main+0xf4>
+    3eee:	83 e0       	ldi	r24, 0x03	; 3
+    3ef0:	87 bf       	out	0x37, r24	; 55
+    3ef2:	e8 95       	spm
 
       // Read command terminator, start reply
       verifySpace();
-    3ec8:	8b d0       	rcall	.+278    	; 0x3fe0 <verifySpace>
+    3ef4:	75 d0       	rcall	.+234    	; 0x3fe0 <verifySpace>
       
       // If only a partial page is to be programmed, the erase might not be complete.
       // So check that here
       boot_spm_busy_wait();
-    3eca:	07 b6       	in	r0, 0x37	; 55
-    3ecc:	00 fc       	sbrc	r0, 0
-    3ece:	fd cf       	rjmp	.-6      	; 0x3eca <main+0xca>
+    3ef6:	07 b6       	in	r0, 0x37	; 55
+    3ef8:	00 fc       	sbrc	r0, 0
+    3efa:	fd cf       	rjmp	.-6      	; 0x3ef6 <main+0xf6>
       }
 #endif
 
       // Copy buffer into programming buffer
       bufPtr = buff;
       addrPtr = (uint16_t)(void*)address;
-    3ed0:	40 91 00 02 	lds	r20, 0x0200
-    3ed4:	50 91 01 02 	lds	r21, 0x0201
-    3ed8:	a0 e0       	ldi	r26, 0x00	; 0
-    3eda:	b1 e0       	ldi	r27, 0x01	; 1
+    3efc:	40 91 00 02 	lds	r20, 0x0200
+    3f00:	50 91 01 02 	lds	r21, 0x0201
+    3f04:	a0 e0       	ldi	r26, 0x00	; 0
+    3f06:	b1 e0       	ldi	r27, 0x01	; 1
       ch = SPM_PAGESIZE / 2;
       do {
         uint16_t a;
         a = *bufPtr++;
-    3edc:	2c 91       	ld	r18, X
-    3ede:	30 e0       	ldi	r19, 0x00	; 0
+    3f08:	2c 91       	ld	r18, X
+    3f0a:	30 e0       	ldi	r19, 0x00	; 0
         a |= (*bufPtr++) << 8;
-    3ee0:	11 96       	adiw	r26, 0x01	; 1
-    3ee2:	8c 91       	ld	r24, X
-    3ee4:	11 97       	sbiw	r26, 0x01	; 1
-    3ee6:	90 e0       	ldi	r25, 0x00	; 0
-    3ee8:	98 2f       	mov	r25, r24
-    3eea:	88 27       	eor	r24, r24
-    3eec:	82 2b       	or	r24, r18
-    3eee:	93 2b       	or	r25, r19
-#ifdef VIRTUAL_BOOT_PARTITION
-#define rstVect (*(uint16_t*)(0x204))
-#define wdtVect (*(uint16_t*)(0x206))
+    3f0c:	11 96       	adiw	r26, 0x01	; 1
+    3f0e:	8c 91       	ld	r24, X
+    3f10:	11 97       	sbiw	r26, 0x01	; 1
+    3f12:	90 e0       	ldi	r25, 0x00	; 0
+    3f14:	98 2f       	mov	r25, r24
+    3f16:	88 27       	eor	r24, r24
+    3f18:	82 2b       	or	r24, r18
+    3f1a:	93 2b       	or	r25, r19
+#define rstVect (*(uint16_t*)(RAMSTART+SPM_PAGESIZE*2+4))
+#define wdtVect (*(uint16_t*)(RAMSTART+SPM_PAGESIZE*2+6))
 #endif
+
 /* main program starts here */
 int main(void) {
-    3ef0:	12 96       	adiw	r26, 0x02	; 2
+    3f1c:	12 96       	adiw	r26, 0x02	; 2
       ch = SPM_PAGESIZE / 2;
       do {
         uint16_t a;
         a = *bufPtr++;
         a |= (*bufPtr++) << 8;
-        boot_page_fill((uint16_t)(void*)addrPtr,a);
-    3ef2:	fa 01       	movw	r30, r20
-    3ef4:	0c 01       	movw	r0, r24
-    3ef6:	d0 92 57 00 	sts	0x0057, r13
-    3efa:	e8 95       	spm
-    3efc:	11 24       	eor	r1, r1
+        __boot_page_fill_short((uint16_t)(void*)addrPtr,a);
+    3f1e:	fa 01       	movw	r30, r20
+    3f20:	0c 01       	movw	r0, r24
+    3f22:	d7 be       	out	0x37, r13	; 55
+    3f24:	e8 95       	spm
+    3f26:	11 24       	eor	r1, r1
         addrPtr += 2;
-    3efe:	4e 5f       	subi	r20, 0xFE	; 254
-    3f00:	5f 4f       	sbci	r21, 0xFF	; 255
+    3f28:	4e 5f       	subi	r20, 0xFE	; 254
+    3f2a:	5f 4f       	sbci	r21, 0xFF	; 255
       } while (--ch);
-    3f02:	f1 e0       	ldi	r31, 0x01	; 1
-    3f04:	a0 38       	cpi	r26, 0x80	; 128
-    3f06:	bf 07       	cpc	r27, r31
-    3f08:	49 f7       	brne	.-46     	; 0x3edc <main+0xdc>
+    3f2c:	f1 e0       	ldi	r31, 0x01	; 1
+    3f2e:	a0 38       	cpi	r26, 0x80	; 128
+    3f30:	bf 07       	cpc	r27, r31
+    3f32:	51 f7       	brne	.-44     	; 0x3f08 <main+0x108>
       
       // Write from programming buffer
-      boot_page_write((uint16_t)(void*)address);
-    3f0a:	e0 91 00 02 	lds	r30, 0x0200
-    3f0e:	f0 91 01 02 	lds	r31, 0x0201
-    3f12:	e0 92 57 00 	sts	0x0057, r14
-    3f16:	e8 95       	spm
+      __boot_page_write_short((uint16_t)(void*)address);
+    3f34:	e0 91 00 02 	lds	r30, 0x0200
+    3f38:	f0 91 01 02 	lds	r31, 0x0201
+    3f3c:	e7 be       	out	0x37, r14	; 55
+    3f3e:	e8 95       	spm
       boot_spm_busy_wait();
-    3f18:	07 b6       	in	r0, 0x37	; 55
-    3f1a:	00 fc       	sbrc	r0, 0
-    3f1c:	fd cf       	rjmp	.-6      	; 0x3f18 <main+0x118>
+    3f40:	07 b6       	in	r0, 0x37	; 55
+    3f42:	00 fc       	sbrc	r0, 0
+    3f44:	fd cf       	rjmp	.-6      	; 0x3f40 <main+0x140>
 
 #if defined(RWWSRE)
       // Reenable read access to flash
       boot_rww_enable();
-    3f1e:	f0 92 57 00 	sts	0x0057, r15
-    3f22:	e8 95       	spm
-    3f24:	27 c0       	rjmp	.+78     	; 0x3f74 <main+0x174>
+    3f46:	f7 be       	out	0x37, r15	; 55
+    3f48:	e8 95       	spm
+    3f4a:	27 c0       	rjmp	.+78     	; 0x3f9a <main+0x19a>
 #endif
 
     }
     /* Read memory block mode, length is big endian.  */
     else if(ch == STK_READ_PAGE) {
-    3f26:	84 37       	cpi	r24, 0x74	; 116
-    3f28:	b9 f4       	brne	.+46     	; 0x3f58 <main+0x158>
+    3f4c:	84 37       	cpi	r24, 0x74	; 116
+    3f4e:	b9 f4       	brne	.+46     	; 0x3f7e <main+0x17e>
       // READ PAGE - we only read flash
       getLen();
-    3f2a:	4a d0       	rcall	.+148    	; 0x3fc0 <getLen>
+    3f50:	37 d0       	rcall	.+110    	; 0x3fc0 <getLen>
       verifySpace();
-    3f2c:	59 d0       	rcall	.+178    	; 0x3fe0 <verifySpace>
-        else ch = pgm_read_byte_near(address);
+    3f52:	46 d0       	rcall	.+140    	; 0x3fe0 <verifySpace>
+        putch(result);
         address++;
-        putch(ch);
-      } while (--length);
+      }
+      while (--length);
 #else
       do putch(pgm_read_byte_near(address++));
-    3f2e:	e0 91 00 02 	lds	r30, 0x0200
-    3f32:	f0 91 01 02 	lds	r31, 0x0201
-    3f36:	31 96       	adiw	r30, 0x01	; 1
-    3f38:	f0 93 01 02 	sts	0x0201, r31
-    3f3c:	e0 93 00 02 	sts	0x0200, r30
-    3f40:	31 97       	sbiw	r30, 0x01	; 1
-    3f42:	e4 91       	lpm	r30, Z+
-    3f44:	8e 2f       	mov	r24, r30
-    3f46:	19 d0       	rcall	.+50     	; 0x3f7a <putch>
+    3f54:	e0 91 00 02 	lds	r30, 0x0200
+    3f58:	f0 91 01 02 	lds	r31, 0x0201
+    3f5c:	31 96       	adiw	r30, 0x01	; 1
+    3f5e:	f0 93 01 02 	sts	0x0201, r31
+    3f62:	e0 93 00 02 	sts	0x0200, r30
+    3f66:	31 97       	sbiw	r30, 0x01	; 1
+    3f68:	e4 91       	lpm	r30, Z+
+    3f6a:	8e 2f       	mov	r24, r30
+    3f6c:	19 d0       	rcall	.+50     	; 0x3fa0 <putch>
       while (--length);
-    3f48:	80 91 02 02 	lds	r24, 0x0202
-    3f4c:	81 50       	subi	r24, 0x01	; 1
-    3f4e:	80 93 02 02 	sts	0x0202, r24
-    3f52:	88 23       	and	r24, r24
-    3f54:	61 f7       	brne	.-40     	; 0x3f2e <main+0x12e>
-    3f56:	0e c0       	rjmp	.+28     	; 0x3f74 <main+0x174>
+    3f6e:	80 91 02 02 	lds	r24, 0x0202
+    3f72:	81 50       	subi	r24, 0x01	; 1
+    3f74:	80 93 02 02 	sts	0x0202, r24
+    3f78:	88 23       	and	r24, r24
+    3f7a:	61 f7       	brne	.-40     	; 0x3f54 <main+0x154>
+    3f7c:	0e c0       	rjmp	.+28     	; 0x3f9a <main+0x19a>
+#endif
 #endif
     }
 
     /* Get device signature bytes  */
     else if(ch == STK_READ_SIGN) {
-    3f58:	85 37       	cpi	r24, 0x75	; 117
-    3f5a:	39 f4       	brne	.+14     	; 0x3f6a <main+0x16a>
+    3f7e:	85 37       	cpi	r24, 0x75	; 117
+    3f80:	39 f4       	brne	.+14     	; 0x3f90 <main+0x190>
       // READ SIGN - return what Avrdude wants to hear
       verifySpace();
-    3f5c:	41 d0       	rcall	.+130    	; 0x3fe0 <verifySpace>
+    3f82:	2e d0       	rcall	.+92     	; 0x3fe0 <verifySpace>
       putch(SIGNATURE_0);
-    3f5e:	8e e1       	ldi	r24, 0x1E	; 30
-    3f60:	0c d0       	rcall	.+24     	; 0x3f7a <putch>
+    3f84:	8e e1       	ldi	r24, 0x1E	; 30
+    3f86:	0c d0       	rcall	.+24     	; 0x3fa0 <putch>
       putch(SIGNATURE_1);
-    3f62:	84 e9       	ldi	r24, 0x94	; 148
-    3f64:	0a d0       	rcall	.+20     	; 0x3f7a <putch>
+    3f88:	84 e9       	ldi	r24, 0x94	; 148
+    3f8a:	0a d0       	rcall	.+20     	; 0x3fa0 <putch>
       putch(SIGNATURE_2);
-    3f66:	86 e0       	ldi	r24, 0x06	; 6
-    3f68:	96 cf       	rjmp	.-212    	; 0x3e96 <main+0x96>
+    3f8c:	86 e0       	ldi	r24, 0x06	; 6
+    3f8e:	8b cf       	rjmp	.-234    	; 0x3ea6 <main+0xa6>
     }
     else if (ch == 'Q') {
-    3f6a:	81 35       	cpi	r24, 0x51	; 81
-    3f6c:	11 f4       	brne	.+4      	; 0x3f72 <main+0x172>
+    3f90:	81 35       	cpi	r24, 0x51	; 81
+    3f92:	11 f4       	brne	.+4      	; 0x3f98 <main+0x198>
       // Adaboot no-wait mod
       watchdogConfig(WATCHDOG_16MS);
-    3f6e:	88 e0       	ldi	r24, 0x08	; 8
-    3f70:	2c d0       	rcall	.+88     	; 0x3fca <watchdogConfig>
+    3f94:	88 e0       	ldi	r24, 0x08	; 8
+    3f96:	19 d0       	rcall	.+50     	; 0x3fca <watchdogConfig>
       verifySpace();
     }
     else {
       // This covers the response to commands like STK_ENTER_PROGMODE
       verifySpace();
-    3f72:	36 d0       	rcall	.+108    	; 0x3fe0 <verifySpace>
+    3f98:	23 d0       	rcall	.+70     	; 0x3fe0 <verifySpace>
     }
     putch(STK_OK);
-    3f74:	80 e1       	ldi	r24, 0x10	; 16
-    3f76:	01 d0       	rcall	.+2      	; 0x3f7a <putch>
-    3f78:	63 cf       	rjmp	.-314    	; 0x3e40 <main+0x40>
+    3f9a:	80 e1       	ldi	r24, 0x10	; 16
+    3f9c:	01 d0       	rcall	.+2      	; 0x3fa0 <putch>
+    3f9e:	5c cf       	rjmp	.-328    	; 0x3e58 <main+0x58>
+
+00003fa0 <putch>:
+  }
+}
 
-00003f7a <putch>:
 void putch(char ch) {
+    3fa0:	98 2f       	mov	r25, r24
 #ifndef SOFT_UART
   while (!(UCSR0A & _BV(UDRE0)));
+    3fa2:	80 91 c0 00 	lds	r24, 0x00C0
+    3fa6:	85 ff       	sbrs	r24, 5
+    3fa8:	fc cf       	rjmp	.-8      	; 0x3fa2 <putch+0x2>
   UDR0 = ch;
-#else
-  __asm__ __volatile__ (
-    3f7a:	2a e0       	ldi	r18, 0x0A	; 10
-    3f7c:	30 e0       	ldi	r19, 0x00	; 0
-    3f7e:	80 95       	com	r24
-    3f80:	08 94       	sec
-    3f82:	10 f4       	brcc	.+4      	; 0x3f88 <putch+0xe>
-    3f84:	59 98       	cbi	0x0b, 1	; 11
-    3f86:	02 c0       	rjmp	.+4      	; 0x3f8c <putch+0x12>
-    3f88:	59 9a       	sbi	0x0b, 1	; 11
-    3f8a:	00 00       	nop
-    3f8c:	15 d0       	rcall	.+42     	; 0x3fb8 <uartDelay>
-    3f8e:	14 d0       	rcall	.+40     	; 0x3fb8 <uartDelay>
-    3f90:	86 95       	lsr	r24
-    3f92:	2a 95       	dec	r18
-    3f94:	b1 f7       	brne	.-20     	; 0x3f82 <putch+0x8>
+    3faa:	90 93 c6 00 	sts	0x00C6, r25
       [uartBit] "I" (UART_TX_BIT)
     :
       "r25"
   );
 #endif
 }
-    3f96:	08 95       	ret
+    3fae:	08 95       	ret
 
-00003f98 <getch>:
+00003fb0 <getch>:
   return getch();
 }
 
 // Watchdog functions. These are only safe with interrupts turned off.
 void watchdogReset() {
   __asm__ __volatile__ (
-    3f98:	a8 95       	wdr
-#ifdef LED_DATA_FLASH
+    3fb0:	a8 95       	wdr
+      [uartBit] "I" (UART_RX_BIT)
+    :
+      "r25"
+);
+#else
+  while(!(UCSR0A & _BV(RXC0)));
+    3fb2:	80 91 c0 00 	lds	r24, 0x00C0
+    3fb6:	87 ff       	sbrs	r24, 7
+    3fb8:	fc cf       	rjmp	.-8      	; 0x3fb2 <getch+0x2>
+  ch = UDR0;
+    3fba:	80 91 c6 00 	lds	r24, 0x00C6
   LED_PIN |= _BV(LED);
 #endif
+#endif
 
   return ch;
 }
-    3f9a:	29 e0       	ldi	r18, 0x09	; 9
-    3f9c:	30 e0       	ldi	r19, 0x00	; 0
-    3f9e:	48 99       	sbic	0x09, 0	; 9
-    3fa0:	fe cf       	rjmp	.-4      	; 0x3f9e <getch+0x6>
-    3fa2:	0a d0       	rcall	.+20     	; 0x3fb8 <uartDelay>
-    3fa4:	09 d0       	rcall	.+18     	; 0x3fb8 <uartDelay>
-    3fa6:	08 d0       	rcall	.+16     	; 0x3fb8 <uartDelay>
-    3fa8:	88 94       	clc
-    3faa:	48 99       	sbic	0x09, 0	; 9
-    3fac:	08 94       	sec
-    3fae:	2a 95       	dec	r18
-    3fb0:	11 f0       	breq	.+4      	; 0x3fb6 <getch+0x1e>
-    3fb2:	87 95       	ror	r24
-    3fb4:	f7 cf       	rjmp	.-18     	; 0x3fa4 <getch+0xc>
-    3fb6:	08 95       	ret
-
-00003fb8 <uartDelay>:
-#if UART_B_VALUE > 255
-#error Baud rate too slow for soft UART
-#endif
-
-void uartDelay() {
-  __asm__ __volatile__ (
-    3fb8:	98 e0       	ldi	r25, 0x08	; 8
-    3fba:	9a 95       	dec	r25
-    3fbc:	f1 f7       	brne	.-4      	; 0x3fba <uartDelay+0x2>
     3fbe:	08 95       	ret
 
 00003fc0 <getLen>:
@@ -462,13 +483,13 @@ void uartDelay() {
 
 uint8_t getLen() {
   getch();
-    3fc0:	eb df       	rcall	.-42     	; 0x3f98 <getch>
+    3fc0:	f7 df       	rcall	.-18     	; 0x3fb0 <getch>
   length = getch();
-    3fc2:	ea df       	rcall	.-44     	; 0x3f98 <getch>
+    3fc2:	f6 df       	rcall	.-20     	; 0x3fb0 <getch>
     3fc4:	80 93 02 02 	sts	0x0202, r24
   return getch();
 }
-    3fc8:	e7 cf       	rjmp	.-50     	; 0x3f98 <getch>
+    3fc8:	f3 cf       	rjmp	.-26     	; 0x3fb0 <getch>
 
 00003fca <watchdogConfig>:
     "wdr\n"
@@ -504,14 +525,16 @@ void appStart() {
 
 void verifySpace() {
   if (getch() != CRC_EOP) appStart();
-    3fe0:	db df       	rcall	.-74     	; 0x3f98 <getch>
+    3fe0:	e7 df       	rcall	.-50     	; 0x3fb0 <getch>
     3fe2:	80 32       	cpi	r24, 0x20	; 32
     3fe4:	09 f0       	breq	.+2      	; 0x3fe8 <verifySpace+0x8>
     3fe6:	f7 df       	rcall	.-18     	; 0x3fd6 <appStart>
   putch(STK_INSYNC);
     3fe8:	84 e1       	ldi	r24, 0x14	; 20
 }
-    3fea:	c7 cf       	rjmp	.-114    	; 0x3f7a <putch>
+    3fea:	da cf       	rjmp	.-76     	; 0x3fa0 <putch>
+
+00003fec <getNch>:
     ::[count] "M" (UART_B_VALUE)
   );
 }
@@ -520,12 +543,10 @@ void verifySpace() {
 void getNch(uint8_t count) {
     3fec:	1f 93       	push	r17
     3fee:	18 2f       	mov	r17, r24
-
-00003ff0 <getNch>:
   do getch(); while (--count);
-    3ff0:	d3 df       	rcall	.-90     	; 0x3f98 <getch>
+    3ff0:	df df       	rcall	.-66     	; 0x3fb0 <getch>
     3ff2:	11 50       	subi	r17, 0x01	; 1
-    3ff4:	e9 f7       	brne	.-6      	; 0x3ff0 <getNch>
+    3ff4:	e9 f7       	brne	.-6      	; 0x3ff0 <getNch+0x4>
   verifySpace();
     3ff6:	f4 df       	rcall	.-24     	; 0x3fe0 <verifySpace>
 }