faster and faster

diff --git a/aes_enc-asm.S b/aes_enc-asm.S
index a2235cf2cef7f48f1a478646c4e4b76545258a68..d8cec74b60cf292195924b4e8a759cc9d4542a16 100644 (file)
--- a/aes_enc-asm.S
+++ b/aes_enc-asm.S
@@ -185,28 +185,25 @@ aes_encrypt_core:
        brtc 2f
        rjmp 1b
 2:     
-       
-       mov r1, ST00
-       eor r1, ST01
-       eor r1, ST02
+ /* mixrow 1*/
+       mov r1, ST02
        eor r1, ST03
                
        mov T0, ST00
-       eor T0, ST01
-       lsl T0
+       eor ST00, ST01
+       eor r1, ST00
+       lsl ST00
        brcc 3f
-       eor T0, xREDUCER
-3:     
-       eor T0, r1
-       eor T0, ST00
+       eor ST00, xREDUCER
+3:     eor ST00, r1
+       eor ST00, T0
        
        mov T1, ST01
        eor T1, ST02
        lsl T1
        brcc 3f
        eor T1, xREDUCER
-3:                     
-       eor T1, r1
+3:     eor T1, r1
        eor ST01, T1
        
        mov T2, ST02
@@ -214,42 +211,36 @@ aes_encrypt_core:
        lsl T2
        brcc 3f
        eor T2, xREDUCER
-3:
-       eor T2, r1
+3:  eor T2, r1
        eor ST02, T2
        
        mov T3, ST03
-       eor T3, ST00
+       eor T3, T0
        lsl T3
        brcc 3f
        eor T3, xREDUCER
-3:     
-       eor T3, r1
+3:     eor T3, r1
        eor ST03, T3
                
-       mov ST00, T0
-
-       mov r1, ST10
-       eor r1, ST11
-       eor r1, ST12
+ /* mixrow 2 */
+       mov r1, ST12
        eor r1, ST13
                
        mov T0, ST10
-       eor T0, ST11
-       lsl T0
+       eor ST10, ST11
+       eor r1, ST10
+       lsl ST10
        brcc 3f
-       eor T0, xREDUCER
-3:     
-       eor T0, r1
-       eor T0, ST10
+       eor ST10, xREDUCER
+3:     eor ST10, r1
+       eor ST10, T0
        
        mov T1, ST11
        eor T1, ST12
        lsl T1
        brcc 3f
        eor T1, xREDUCER
-3:                     
-       eor T1, r1
+3:     eor T1, r1
        eor ST11, T1
        
        mov T2, ST12
@@ -257,42 +248,36 @@ aes_encrypt_core:
        lsl T2
        brcc 3f
        eor T2, xREDUCER
-3:
-       eor T2, r1
+3:  eor T2, r1
        eor ST12, T2
        
        mov T3, ST13
-       eor T3, ST10
+       eor T3, T0
        lsl T3
        brcc 3f
        eor T3, xREDUCER
-3:     
-       eor T3, r1
+3:     eor T3, r1
        eor ST13, T3
                
-       mov ST10, T0
-
-       mov r1, ST20
-       eor r1, ST21
-       eor r1, ST22
+ /* mixrow 3*/
+       mov r1, ST22
        eor r1, ST23
                
        mov T0, ST20
-       eor T0, ST21
-       lsl T0
+       eor ST20, ST21
+       eor r1, ST20
+       lsl ST20
        brcc 3f
-       eor T0, xREDUCER
-3:     
-       eor T0, r1
-       eor T0, ST20
+       eor ST20, xREDUCER
+3:     eor ST20, r1
+       eor ST20, T0
        
        mov T1, ST21
        eor T1, ST22
        lsl T1
        brcc 3f
        eor T1, xREDUCER
-3:                     
-       eor T1, r1
+3:     eor T1, r1
        eor ST21, T1
        
        mov T2, ST22
@@ -300,42 +285,36 @@ aes_encrypt_core:
        lsl T2
        brcc 3f
        eor T2, xREDUCER
-3:
-       eor T2, r1
+3:  eor T2, r1
        eor ST22, T2
        
        mov T3, ST23
-       eor T3, ST20
+       eor T3, T0
        lsl T3
        brcc 3f
        eor T3, xREDUCER
-3:     
-       eor T3, r1
+3:     eor T3, r1
        eor ST23, T3
                
-       mov ST20, T0
-
-       mov r1, ST30
-       eor r1, ST31
-       eor r1, ST32
+ /* mixrow 4 */
+       mov r1, ST32
        eor r1, ST33
                
        mov T0, ST30
-       eor T0, ST31
-       lsl T0
+       eor ST30, ST31
+       eor r1, ST30
+       lsl ST30
        brcc 3f
-       eor T0, xREDUCER
-3:     
-       eor T0, r1
-       eor T0, ST30
+       eor ST30, xREDUCER
+3:     eor ST30, r1
+       eor ST30, T0
        
        mov T1, ST31
        eor T1, ST32
        lsl T1
        brcc 3f
        eor T1, xREDUCER
-3:                     
-       eor T1, r1
+3:     eor T1, r1
        eor ST31, T1
        
        mov T2, ST32
@@ -343,20 +322,17 @@ aes_encrypt_core:
        lsl T2
        brcc 3f
        eor T2, xREDUCER
-3:
-       eor T2, r1
+3:  eor T2, r1
        eor ST32, T2
        
        mov T3, ST33
-       eor T3, ST30
+       eor T3, T0
        lsl T3
        brcc 3f
        eor T3, xREDUCER
-3:     
-       eor T3, r1
+3:     eor T3, r1
        eor ST33, T3
                
-       mov ST30, T0
 
        /* mix colums (rows) done */