Work in progress, source loop unrolling.

Rev: src/nettle/sparc/aes.asm:1.115

diff --git a/sparc/aes.asm b/sparc/aes.asm
index cb63816e445654fa777fa1b61e30f5c14b7550ee..347cac3ff477cc4c71ea589f11f96d930d52991c 100644 (file)
--- a/sparc/aes.asm
+++ b/sparc/aes.asm
@@ -277,47 +277,49 @@ C .Lsource_loop:
        C add   src, 4, src
        
        C i = 2
-       ldub    [src+3], t3
-       ldub    [src+2], t2
-       sll     t3, 24, t3
-       ldub    [src+1], t1
-       
-       sll     t2, 16, t2
-       or      t3, t2, t3
-       ldub    [src], t0
-       sll     t1, 8, t1
-       
-       ! Get subkey
-       ld      [src+%g2], t2
-       or      t3, t1, t3
-       or      t3, t0, t3
-       xor     t3, t2, t3
-       
-       C cmp   src, %g1
-       st      t3, [src+%g3]
-       C bleu  .Lsource_loop
-       add     src, 4, src
+       AES_LOAD(8)
+       C ldub  [src+3], t3
+       C ldub  [src+2], t2
+       C sll   t3, 24, t3
+       C ldub  [src+1], t1
+       C 
+       C sll   t2, 16, t2
+       C or    t3, t2, t3
+       C ldub  [src], t0
+       C sll   t1, 8, t1
+       C 
+       C ! Get subkey
+       C ld    [src+%g2], t2
+       C or    t3, t1, t3
+       C or    t3, t0, t3
+       C xor   t3, t2, t3
+       C 
+       C C cmp src, %g1
+       C st    t3, [src+%g3]
+       C C bleu        .Lsource_loop
+       C add   src, 4, src
        C i = 3
-       ldub    [src+3], t3
-       ldub    [src+2], t2
-       sll     t3, 24, t3
-       ldub    [src+1], t1
-       
-       sll     t2, 16, t2
-       or      t3, t2, t3
-       ldub    [src], t0
-       sll     t1, 8, t1
-       
-       ! Get subkey
-       ld      [src+%g2], t2
-       or      t3, t1, t3
-       or      t3, t0, t3
-       xor     t3, t2, t3
-       
-       C cmp   src, %g1
-       st      t3, [src+%g3]
-       C bleu  .Lsource_loop
-       add     src, 4, src
+       AES_LOAD(12)
+       C ldub  [src+3], t3
+       C ldub  [src+2], t2
+       C sll   t3, 24, t3
+       C ldub  [src+1], t1
+       C 
+       C sll   t2, 16, t2
+       C or    t3, t2, t3
+       C ldub  [src], t0
+       C sll   t1, 8, t1
+       C 
+       C ! Get subkey
+       C ld    [src+%g2], t2
+       C or    t3, t1, t3
+       C or    t3, t0, t3
+       C xor   t3, t2, t3
+       C 
+       C C cmp src, %g1
+       C st    t3, [src+%g3]
+       C C bleu        .Lsource_loop
+       C add   src, 4, src
 
        C End loop