changed arcfour api

[avr-crypto-lib.git] / arcfour-asm.S

diff --git a/arcfour-asm.S b/arcfour-asm.S
index d3ecadbf4857f6fc7e4cfd627c5b38a11120e094..05cac9255180b52b5dc5dbb7554a9936631d465e 100644 (file)
--- a/arcfour-asm.S
+++ b/arcfour-asm.S
@@ -26,68 +26,106 @@
  * 
  */
  
+#include <avr/io.h>
+
+
+.macro push_ p1:req, p2:vararg
+       push \p1
+.ifnb \p2      
+       push_ \p2
+.endif
+.endm
+
+.macro pop_ p1:req, p2:vararg
+       pop \p1
+.ifnb \p2      
+       pop_ \p2
+.endif
+.endm
+
+.macro push_range from:req, to:req
+       push \from
+.if     \to-\from
+       push_range "(\from+1)",\to
+.endif         
+.endm
+
+.macro pop_range from:req, to:req
+       pop \to
+.if     \to-\from
+       pop_range \from,"(\to-1)"       
+.endif
+.endm
+
+.macro stack_alloc size:req, reg1=r30, reg2=r31
+       in \reg1, _SFR_IO_ADDR(SPL)
+       in \reg2, _SFR_IO_ADDR(SPH)
+       sbiw r30, \size 
+       out  _SFR_IO_ADDR(SPH), \reg2
+       out  _SFR_IO_ADDR(SPL), \reg1
+.endm
+
+.macro stack_free size:req, reg1=r30, reg2=r31
+       in \reg1, _SFR_IO_ADDR(SPL)
+       in \reg2, _SFR_IO_ADDR(SPH)
+       adiw r30, \size 
+       out  _SFR_IO_ADDR(SPH), \reg2
+       out  _SFR_IO_ADDR(SPL), \reg1
+.endm
+ 
  /* +---+---+---------------------+
  *  | i | j | ......<256>........ |
  *  +---+---+---------------------+
  */
  
 .global arcfour_init
-
-;== arcfour_init ==
-;  this function initialises the context
-; param1: 16-bit pointer to a ctx struct
-;      given in r25,r24
-; param2: 16-bit pointer to a key
-;      given in r23,r22
-; param1: 8-bit integer indicating keylength in byte
-;      given in        r20
-
+/*
+ *== arcfour_init ==
+ *  this function initialises the context
+ * param1: 16-bit pointer to the key
+ *     given in r24:r25
+ * param2: 8-bit integer indicating keylength in byte
+ *     given in r22
+ * param3: 16-bit pointer to a ctx struct
+ *     given in r20:r21
+ */
 arcfour_init:
-       push r29
-       push r28
-       push r2
-       
-       movw r26, r24   /* X points to ctx */
-       movw r30, r22   /* Z points to key */
+       push_ r2, r28, r29
+       movw r26, r20   /* X points to ctx */
+       movw r30, r24   /* Z points to key */
        st X+, r1
-       st X+, r1               /* X points to S */
+       st X+, r1       /* X points to S */
+       movw r20, r26   /* store pointer to S in r21:r20 */
        
 1:             
        st X+, r1 
        inc r1
        brne 1b
        
-       adiw r24, 2             /* r24:r25 points to S */
-       clr r21                 /* r21 is j */
-       mov r18, r20            /* r18 is keyindex counter */
+       movw r26, r20
+       clr r18         /* r18 is keyindex counter */
        clr r0
+       clr r19
 2:
-       movw r26, r24
-       ld r19, Z+
-       add r21, r19            /* j+= key[i%length] */
-       
-       add r26, r1
-       adc r27, r0
-       ld r19, X
-       add r21, r19            /* j += S[i] */
-       
-       dec r18         /* check the key-index counter */
-       brne 3f
-       movw r30, r22
-       mov r18, r20
-3:     /* now swap(S[i], S[j]) */ /* r19 is still S[i] */
-       movw r28, r24 
-       add r28, r21
-       adc r29, r0             /* Y points to S[j]*/
+       ld r23, X
+       ld r2, Z+
+       add r19, r2
+       add r19, r23
+       movw r28, r20   /* load pointer to S in Y */
+       add r28, r19
+       adc r29, r1
        ld r2, Y
-       st Y, r19
-       st X, r2        
-       inc r1
+       st Y,  r23
+       st X+, r2 
+       inc r18
+       cp r18, r22
+       brne 3f
+       movw r30, r24
+       clr r18
+3:             
+       inc r0
        brne 2b 
-       
-       pop r2
-       pop r28
-       pop r29
+       pop_ r29, r28, r2
        ret
 
 /*