+RC5 +size-statistics tool +small modification to nessie_bc_test (optional free(...

[avr-crypto-lib.git] / noekeon_asm.S

diff --git a/noekeon_asm.S b/noekeon_asm.S
index f3f904dcc21303f2048185f024c0c8034b225e80..d880296389fdc63cdf2dd2adef5d57b77eba3792 100644 (file)
--- a/noekeon_asm.S
+++ b/noekeon_asm.S
@@ -83,7 +83,6 @@
        \op \p\()\d, \q\()\z
 .endm
 
-.global bigendian_rotl32
 ; === bigendian_rotl32 ===
 ; this function rotates a 32bit bigendian word n bits to the left
 ;  param1: the 32-bit value
@@ -94,7 +93,6 @@
 ;   given in r25,r24,r23,r22   
 
 bigendian_rotl32:
-       in r0, _SFR_IO_ADDR(SREG)
        /* copy high bit of r22 to carry */
        mov r1, r22
 2:
@@ -109,13 +107,11 @@ bigendian_rotl32:
        brne 2b
 bigendian_rotl32_exit:
        clr r1
-       out _SFR_IO_ADDR(SREG), r0
        ret
        
        
 /******************************************************************************/
 
-.global bigendian_rotr32
 ; === bigendian_rotl32 ===
 ; this function rotates a 32bit bigendian word n bits to the right
 ;  param1: the 32-bit value
@@ -126,7 +122,6 @@ bigendian_rotl32_exit:
 ;   given in r25,r24,r23,r22   
 
 bigendian_rotr32:
-       in r0, _SFR_IO_ADDR(SREG)
        /* copy high bit of r25 to carry */
 
        mov r1, r25
@@ -141,7 +136,6 @@ bigendian_rotr32:
        brne 2b
 bigendian_rotr32_exit:
        clr r1
-       out _SFR_IO_ADDR(SREG), r0
        ret
 
 /******************************************************************************/
@@ -457,7 +451,8 @@ round:
        push r25
        push r24
 pi_gamma_pi:
-       clc
+       ldi r30, lo8(bigendian_rotl32)
+       ldi r31, hi8(bigendian_rotl32)
        call pi
        /* pi1 done; now gamma */
        call gamma_1
@@ -485,7 +480,8 @@ pi_gamma_pi:
        eor state2_3, state3_3
 */
        call gamma_1
-       sec
+       ldi r30, lo8(bigendian_rotr32)
+       ldi r31, hi8(bigendian_rotr32)
        call pi
        ret
        
@@ -529,15 +525,7 @@ gamma_1:
        eor state0_3, r1
        ret
        
-pi:
-       brcs 1f
-       ldi r30, lo8(bigendian_rotl32)
-       ldi r31, hi8(bigendian_rotl32)
-       rjmp 2f
-1:
-       ldi r30, lo8(bigendian_rotr32)
-       ldi r31, hi8(bigendian_rotr32)
-2:     
+pi:    
        lsr r31
        ror r30
        /* a[1] <<<= 1*/
@@ -574,61 +562,60 @@ pi:
        mov state3_2, r24
        mov state3_3, r25
        ret
-       
-;------- trash follows --------
-
 
+/******************************************************************************/
+       
+/*
+void noekeon_init(void* key, noekeon_ctx_t* ctx){
+       uint8_t nullv[16];
+       
+       memset(nullv, 0, 16);
+       memcpy(ctx, key, 16);
+       noekeon_enc(ctx, nullv);
+}
+*/
+.global noekeon_init
+noekeon_init:
+; === noekeon_init ===
+;
+;  param1: pointer to key (r24,r25)
+;  param2: pointer to context  (r22,r23) 
+;
+       in r30, _SFR_IO_ADDR(SPL)
+       in r31, _SFR_IO_ADDR(SPH)
+       sbiw r30, 16 
+       out  _SFR_IO_ADDR(SPH), r31
+       out  _SFR_IO_ADDR(SPL), r30
 
-       /* load state */
        movw r26, r22
-       ldi r28, 2
-       clr r29 /* Y points at r2 aka state0_0 */
-       ldi r22, 16     
-1:     /* copy key to state */
-       ld r0, X+
-       st Y+, r0
-       dec r22
-       brne 1b
-
-       movw r26, r30
-       
-       clr r1
-       ldi r22, 16
-1:     /* set key to zero */
-       st Z+, r1
-       dec r22
+       adiw r30, 1
+       movw r22, r30
+       /* set nullv(stack) to zero */
+       ldi r20, 16
+1:     st Z+, r1
+       dec r20
        brne 1b
-               
-       call theta
 
-       ldi r22, 16
-1:     /* write key back */    
-       ld r0, -Y
-       st -Z, r0
-       dec r22
+       /* copy key data to ctx */
+       movw r30, r24
+       ldi r20, 16
+1:     ld r1, Z+
+       st X+, r1
+       dec r20
        brne 1b
-
-;      movw r26, r30 /* move keypointer to X */
-;      adiw r26, 1
-       movw r30, r24 /* Z points at state */
-       push r30 /* push state pointer */
-       push r31
+       clr r1
+       
+       sbiw r26, 16
+       movw r24, r26
+       call noekeon_enc
+       
+       in r30, _SFR_IO_ADDR(SPL)
+       in r31, _SFR_IO_ADDR(SPH)
+       adiw r30, 16 
+       out  _SFR_IO_ADDR(SPH), r31
+       out  _SFR_IO_ADDR(SPL), r30     
+       ret
        
-       ;--
-       clr r29
-       ldi r28, 2
-       ;--
-       ldi r22, 16
-       push r22 /* 16 is also the number of rounds and gets pushed here */
-       ldi r22, 16
-1:     /* load state */
-       ld r0, Z+
-       st Y+, r0
-       dec r22
-       brne 1b
-       /* state loaded */
-
 
-;------- ------------- --------