squeezing out a little more speed

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

diff --git a/aes_enc-asm.S b/aes_enc-asm.S
index 72b9e6ecbb5d3cbe0b5856436beae9e6015a8147..a2235cf2cef7f48f1a478646c4e4b76545258a68 100644 (file)
--- a/aes_enc-asm.S
+++ b/aes_enc-asm.S
@@ -207,7 +207,7 @@ aes_encrypt_core:
        eor T1, xREDUCER
 3:                     
        eor T1, r1
-       eor T1, ST01
+       eor ST01, T1
        
        mov T2, ST02
        eor T2, ST03
@@ -216,7 +216,7 @@ aes_encrypt_core:
        eor T2, xREDUCER
 3:
        eor T2, r1
-       eor T2, ST02
+       eor ST02, T2
        
        mov T3, ST03
        eor T3, ST00
@@ -225,13 +225,9 @@ aes_encrypt_core:
        eor T3, xREDUCER
 3:     
        eor T3, r1
-       eor T3, ST03
+       eor ST03, T3
                
        mov ST00, T0
-       mov ST01, T1
-       mov ST02, T2
-       mov ST03, T3
-
 
        mov r1, ST10
        eor r1, ST11
@@ -254,7 +250,7 @@ aes_encrypt_core:
        eor T1, xREDUCER
 3:                     
        eor T1, r1
-       eor T1, ST11
+       eor ST11, T1
        
        mov T2, ST12
        eor T2, ST13
@@ -263,7 +259,7 @@ aes_encrypt_core:
        eor T2, xREDUCER
 3:
        eor T2, r1
-       eor T2, ST12
+       eor ST12, T2
        
        mov T3, ST13
        eor T3, ST10
@@ -272,12 +268,9 @@ aes_encrypt_core:
        eor T3, xREDUCER
 3:     
        eor T3, r1
-       eor T3, ST13
+       eor ST13, T3
                
        mov ST10, T0
-       mov ST11, T1
-       mov ST12, T2
-       mov ST13, T3
 
        mov r1, ST20
        eor r1, ST21
@@ -300,7 +293,7 @@ aes_encrypt_core:
        eor T1, xREDUCER
 3:                     
        eor T1, r1
-       eor T1, ST21
+       eor ST21, T1
        
        mov T2, ST22
        eor T2, ST23
@@ -309,7 +302,7 @@ aes_encrypt_core:
        eor T2, xREDUCER
 3:
        eor T2, r1
-       eor T2, ST22
+       eor ST22, T2
        
        mov T3, ST23
        eor T3, ST20
@@ -318,12 +311,9 @@ aes_encrypt_core:
        eor T3, xREDUCER
 3:     
        eor T3, r1
-       eor T3, ST23
+       eor ST23, T3
                
        mov ST20, T0
-       mov ST21, T1
-       mov ST22, T2
-       mov ST23, T3
 
        mov r1, ST30
        eor r1, ST31
@@ -346,7 +336,7 @@ aes_encrypt_core:
        eor T1, xREDUCER
 3:                     
        eor T1, r1
-       eor T1, ST31
+       eor ST31, T1
        
        mov T2, ST32
        eor T2, ST33
@@ -355,7 +345,7 @@ aes_encrypt_core:
        eor T2, xREDUCER
 3:
        eor T2, r1
-       eor T2, ST32
+       eor ST32, T2
        
        mov T3, ST33
        eor T3, ST30
@@ -364,12 +354,9 @@ aes_encrypt_core:
        eor T3, xREDUCER
 3:     
        eor T3, r1
-       eor T3, ST33
+       eor ST33, T3
                
        mov ST30, T0
-       mov ST31, T1
-       mov ST32, T2
-       mov ST33, T3
 
        /* mix colums (rows) done */