10% performance tweak in 64-bit mode.
authorAndy Polyakov <appro@openssl.org>
Thu, 27 Sep 2007 06:19:25 +0000 (06:19 +0000)
committerAndy Polyakov <appro@openssl.org>
Thu, 27 Sep 2007 06:19:25 +0000 (06:19 +0000)
crypto/aes/asm/aes-ppc.pl

index 8a7a221..903075e 100644 (file)
@@ -880,6 +880,12 @@ Lppc_AES_decrypt_compact:
        addi    $key,$key,16
        ori     $mask80,$mask80,0x8080
        ori     $mask1b,$mask1b,0x1b1b
+___
+$code.=<<___ if ($SIZE_T==8);
+       insrdi  $mask80,$mask80,32,0
+       insrdi  $mask1b,$mask1b,32,0
+___
+$code.=<<___;
        mtctr   $acc00
 .align 4
 Ldec_compact_loop:
@@ -942,7 +948,70 @@ Ldec_compact_loop:
 
        addi    $key,$key,16
        bdz     Ldec_compact_done
+___
+$code.=<<___ if ($SIZE_T==8);
+       # vectorized permutation improves decrypt performance by 10%
+       insrdi  $s0,$s1,32,0
+       insrdi  $s2,$s3,32,0
 
+       and     $acc00,$s0,$mask80      # r1=r0&0x80808080
+       and     $acc02,$s2,$mask80
+       srdi    $acc04,$acc00,7         # r1>>7
+       srdi    $acc06,$acc02,7
+       andc    $acc08,$s0,$mask80      # r0&0x7f7f7f7f
+       andc    $acc10,$s2,$mask80
+       sub     $acc00,$acc00,$acc04    # r1-(r1>>7)
+       sub     $acc02,$acc02,$acc06
+       add     $acc08,$acc08,$acc08    # (r0&0x7f7f7f7f)<<1
+       add     $acc10,$acc10,$acc10
+       and     $acc00,$acc00,$mask1b   # (r1-(r1>>7))&0x1b1b1b1b
+       and     $acc02,$acc02,$mask1b
+       xor     $acc00,$acc00,$acc08    # r2
+       xor     $acc02,$acc02,$acc10
+
+       and     $acc04,$acc00,$mask80   # r1=r2&0x80808080
+       and     $acc06,$acc02,$mask80
+       srdi    $acc08,$acc04,7         # r1>>7
+       srdi    $acc10,$acc06,7
+       andc    $acc12,$acc00,$mask80   # r2&0x7f7f7f7f
+       andc    $acc14,$acc02,$mask80
+       sub     $acc04,$acc04,$acc08    # r1-(r1>>7)
+       sub     $acc06,$acc06,$acc10
+       add     $acc12,$acc12,$acc12    # (r2&0x7f7f7f7f)<<1
+       add     $acc14,$acc14,$acc14
+       and     $acc04,$acc04,$mask1b   # (r1-(r1>>7))&0x1b1b1b1b
+       and     $acc06,$acc06,$mask1b
+       xor     $acc04,$acc04,$acc12    # r4
+       xor     $acc06,$acc06,$acc14
+
+       and     $acc08,$acc04,$mask80   # r1=r4&0x80808080
+       and     $acc10,$acc06,$mask80
+       srdi    $acc12,$acc08,7         # r1>>7
+       srdi    $acc14,$acc10,7
+       sub     $acc08,$acc08,$acc12    # r1-(r1>>7)
+       sub     $acc10,$acc10,$acc14
+       andc    $acc12,$acc04,$mask80   # r4&0x7f7f7f7f
+       andc    $acc14,$acc06,$mask80
+       add     $acc12,$acc12,$acc12    # (r4&0x7f7f7f7f)<<1
+       add     $acc14,$acc14,$acc14
+       and     $acc08,$acc08,$mask1b   # (r1-(r1>>7))&0x1b1b1b1b
+       and     $acc10,$acc10,$mask1b
+       xor     $acc08,$acc08,$acc12    # r8
+       xor     $acc10,$acc10,$acc14
+
+       xor     $acc00,$acc00,$s0       # r2^r0
+       xor     $acc02,$acc02,$s2
+       xor     $acc04,$acc04,$s0       # r4^r0
+       xor     $acc06,$acc06,$s2
+
+       extrdi  $acc01,$acc00,0,32
+       extrdi  $acc03,$acc02,0,32
+       extrdi  $acc05,$acc04,0,32
+       extrdi  $acc07,$acc06,0,32
+       extrdi  $acc09,$acc08,0,32
+       extrdi  $acc11,$acc10,0,32
+___
+$code.=<<___ if ($SIZE_T==4);
        and     $acc00,$s0,$mask80      # r1=r0&0x80808080
        and     $acc01,$s1,$mask80
        and     $acc02,$s2,$mask80
@@ -1038,6 +1107,8 @@ Ldec_compact_loop:
        xor     $acc05,$acc05,$s1
        xor     $acc06,$acc06,$s2
        xor     $acc07,$acc07,$s3
+___
+$code.=<<___;
        rotrwi  $s0,$s0,8               # = ROTATE(r0,8)
        rotrwi  $s1,$s1,8
        rotrwi  $s2,$s2,8