diff --git a/Makefile b/Makefile
index b681f06..912e12e 100644
--- a/Makefile
+++ b/Makefile
@@ -16,7 +16,7 @@
 #-------------------------------------------------------------------------
 
 PROJNAME = main
-OBJS = main.o sha1.o hmac.o trunc.o usbdrv/oddebug.o usbdrv/usbdrvasm.o usbdrv/usbdrv.o
+OBJS = main.o sha1.o hmac.o trunc.o usbdrv/usbdrvasm.o usbdrv/usbdrv.o
 
 # target architecture
 MCU	= attiny45
@@ -60,17 +60,26 @@ OPTLEVEL=s
 
 # Flags
 CFLAGS=-I. $(INC) -g -mmcu=$(MCU) -O$(OPTLEVEL) \
-	-DF_CPU=16500000	\
-	-fpack-struct -fshort-enums             \
-	-funsigned-bitfields -funsigned-char    \
-	-Wall -Wstrict-prototypes               \
-	-Wa,-ahlms=$(firstword                  \
+	-DF_CPU=16500000	                        \
+    -mcall-prologues                            \
+    -mtiny-stack                                \
+    -fno-jump-tables                            \
+    -fno-inline-functions                       \
+    -fno-inline-small-functions                 \
+    -fno-split-wide-types                       \
+	-fpack-struct -fshort-enums                 \
+	-funsigned-bitfields -funsigned-char        \
+	-Wall -Wstrict-prototypes                   \
+	-Wa,-ahlms=$(firstword                      \
+    --gc-sections                               \
+    --print-gc-sections                         \
 	$(filter %.lst, $(<:.c=.lst)))
-ASMFLAGS =-I. $(INC) -mmcu=$(MCU)        \
-	-x assembler-with-cpp            \
-	-DF_CPU=16500000	\
-	-Wa,-gstabs,-ahlms=$(firstword   \
+ASMFLAGS =-I. $(INC) -mmcu=$(MCU)               \
+	-x assembler-with-cpp                       \
+	-DF_CPU=16500000	                        \
+	-Wa,-gstabs,-ahlms=$(firstword              \
 		$(<:.S=.lst) $(<.s=.lst))
+
 LDFLAGS	= -Wl,-Map=$*.map -mmcu=$(MCU)
 OCFLAGS	= -O $(BINFORMAT)
 # at16prog
diff --git a/main.c b/main.c
index 61dfd5a..4b90c9b 100644
--- a/main.c
+++ b/main.c
@@ -46,7 +46,8 @@ usbMsgLen_t usbFunctionSetup(uchar data[8]) {
 
 
 
-int main(void){
+int main(void) __attribute__((noreturn));
+int __attribute__((OS_main)) main(void) {
 	usbDeviceDisconnect();
 	_delay_ms(200);
 	usbDeviceConnect();
@@ -55,6 +56,5 @@ int main(void){
     for(;;){
 
 	}
-    return 0;
 }
 
diff --git a/sha1.c b/sha1.c
index f6b291f..7112dbf 100644
--- a/sha1.c
+++ b/sha1.c
@@ -20,17 +20,17 @@ static uint8_t chunk_cnt;
 
 void sha1_addbyte(uint8_t c){
 	uint8_t* p=(uint8_t*)chunk;
-	uint8_t i;
+	uint8_t i=(chunk_cnt & 0x03);
 
-	p[(chunk_cnt&0xfc)+3-(chunk_cnt & 0x03)]=c;
+	p[(chunk_cnt-i)+3-i]=c;
 	chunk_cnt++;
 	if(chunk_cnt==64){
 		/*chunk extension*/
 		uint32_t a,b,c,d,e,chunk_e[64];
 		for(i=0; i<64; i++){
-			chunk_e[i]= _LROT32( _W(i+16-3) ^ _W(i+16-8) ^ _W(i+16-14) ^ _W(i+16-16), 1 ); 
+			//chunk_e[i]= _LROT32( _W(i+16-3) ^ _W(i+16-8) ^ _W(i+16-14) ^ _W(i+16-16), 1 );
+            chunk_e[i]= _LROT32( _W(i+16-3) ^ _W(i+16-8) ^ _W(i+16-14) ^ _W(i+16-16), 1 );  
 		}
-
 		a=sha1_digest[4];
 		b=sha1_digest[3];
 		c=sha1_digest[2];
@@ -54,7 +54,7 @@ void sha1_addbyte(uint8_t c){
 				k=K4;
 				f=b ^ c ^ d;
 			}
-			temp = e + _LROT32(a,5) + f  + k + _W(i);
+			temp =  e + k + f + _LROT32(a,5) + _W(i);
 			e = d;
 			d = c;
 			c = _LROT32(b,30);
@@ -87,7 +87,7 @@ void sha1_fill(int32_t len){
 	}
 	{
 		uint64_t l=len*8;
-		/*pay attention to the endiness here*/
+		//pay attention to the endiness here
 		sha1_addbyte( ((char*)(&l))[7]);
 		sha1_addbyte( ((char*)(&l))[6]);
 		sha1_addbyte( ((char*)(&l))[5]);
diff --git a/test.c b/test.c
index a415b79..01166fd 100644
--- a/test.c
+++ b/test.c
@@ -4,6 +4,7 @@
 #include <string.h>
 
 #include "sha1.h"
+#include "hmac.h"
 
 char sha1sum[41];