Complete synchronization of aes-x86_64 with aes-586.
authorAndy Polyakov <appro@openssl.org>
Sat, 21 Jul 2007 14:20:46 +0000 (14:20 +0000)
committerAndy Polyakov <appro@openssl.org>
Sat, 21 Jul 2007 14:20:46 +0000 (14:20 +0000)
crypto/aes/asm/aes-x86_64.pl

index 91a19545de2507859229374f28a04b9dcce12e07..afd796b51772f254de554b6982f543388ccebd82 100755 (executable)
@@ -7,7 +7,7 @@
 # details see http://www.openssl.org/~appro/cryptogams/.
 # ====================================================================
 #
-# Version 2.0.
+# Version 2.1.
 #
 # aes-*-cbc benchmarks are improved by >70% [compared to gcc 3.3.2 on
 # Opteron 240 CPU] plus all the bells-n-whistles from 32-bit version
 #
 #              ECB encrypt     ECB decrypt     CBC large chunk
 # AMD64                33              41              13.0
-# EM64T                38              59              18.6
-# Core 2       30              43              14.5
+# EM64T                38              59              18.6(*)
+# Core 2       30              43              14.5(*)
+#
+# (*) with hyper-threading off
 
 $verticalspin=1;       # unlike 32-bit version $verticalspin performs
                        # ~15% better on both AMD and Intel cores
+$speed_limit=512;      # see aes-586.pl for details
 $output=shift;
 
 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
@@ -145,22 +148,17 @@ $code.=<<___;
        movzb   `&lo("$s0")`,$acc0
        movzb   `&lo("$s1")`,$acc1
        movzb   `&lo("$s2")`,$acc2
-       mov     2($sbox,$acc0,8),$t0
-       mov     2($sbox,$acc1,8),$t1
-       mov     2($sbox,$acc2,8),$t2
-
-       and     \$0x000000ff,$t0
-       and     \$0x000000ff,$t1
-       and     \$0x000000ff,$t2
+       movzb   2($sbox,$acc0,8),$t0
+       movzb   2($sbox,$acc1,8),$t1
+       movzb   2($sbox,$acc2,8),$t2
 
        movzb   `&lo("$s3")`,$acc0
        movzb   `&hi("$s1")`,$acc1
        movzb   `&hi("$s2")`,$acc2
-       mov     2($sbox,$acc0,8),$t3
+       movzb   2($sbox,$acc0,8),$t3
        mov     0($sbox,$acc1,8),$acc1  #$t0
        mov     0($sbox,$acc2,8),$acc2  #$t1
 
-       and     \$0x000000ff,$t3
        and     \$0x0000ff00,$acc1
        and     \$0x0000ff00,$acc2
 
@@ -621,7 +619,7 @@ AES_encrypt:
        # pick Te4 copy which can't "overlap" with stack frame or key schedule
        .picmeup        $sbox
        lea     AES_Te+2048-.($sbox),$sbox
-       lea     768-32(%rsp),%rbp
+       lea     768(%rsp),%rbp
        sub     $sbox,%rbp
        and     \$0x300,%rbp
        lea     ($sbox,%rbp),$sbox
@@ -1215,7 +1213,7 @@ AES_decrypt:
        # pick Td4 copy which can't "overlap" with stack frame or key schedule
        .picmeup        $sbox
        lea     AES_Td+2048-.($sbox),$sbox
-       lea     768-32(%rsp),%rbp
+       lea     768(%rsp),%rbp
        sub     $sbox,%rbp
        and     \$0x300,%rbp
        lea     ($sbox,%rbp),$sbox
@@ -1603,19 +1601,23 @@ ___
 {
 # stack frame layout
 # -8(%rsp)             return address
-my $_rsp="0(%rsp)";            # saved %rsp
-my $_len="8(%rsp)";            # copy of 3rd parameter, length
-my $_key="16(%rsp)";           # copy of 4th parameter, key
-my $_ivp="24(%rsp)";           # copy of 5th parameter, ivp
-my $keyp="32(%rsp)";           # one to pass as $key
-my $ivec="40(%rsp)";           # ivec[16]
-my $aes_key="56(%rsp)";                # copy of aes_key
-my $mark="56+240(%rsp)";       # copy of aes_key->rounds
+my $keyp="0(%rsp)";            # one to pass as $key
+my $keyend="8(%rsp)";          # &(keyp->rd_key[4*keyp->rounds])
+my $_rsp="16(%rsp)";           # saved %rsp
+my $_inp="24(%rsp)";           # copy of 1st parameter, inp
+my $_out="32(%rsp)";           # copy of 2nd parameter, out
+my $_len="40(%rsp)";           # copy of 3rd parameter, length
+my $_key="48(%rsp)";           # copy of 4th parameter, key
+my $_ivp="56(%rsp)";           # copy of 5th parameter, ivp
+my $ivec="64(%rsp)";           # ivec[16]
+my $aes_key="80(%rsp)";                # copy of aes_key
+my $mark="80+240(%rsp)";       # copy of aes_key->rounds
 
 $code.=<<___;
 .globl AES_cbc_encrypt
 .type  AES_cbc_encrypt,\@function,6
 .align 16
+.extern        OPENSSL_ia32cap_P
 AES_cbc_encrypt:
        cmp     \$0,%rdx        # check length
        je      .Lcbc_just_ret
@@ -1629,20 +1631,27 @@ AES_cbc_encrypt:
        cld
 
        .picmeup $sbox
-.Lcbc_pic_point:
-
+       lea     AES_Te-.($sbox),$sbox
        cmp     \$0,%r9
-       je      .LDECRYPT
-
-       lea     AES_Te-.Lcbc_pic_point($sbox),$sbox
+       jne     .Lcbc_picked_te
+       lea     AES_Td-AES_Te($sbox),$sbox
+.Lcbc_picked_te:
+
+       mov     OPENSSL_ia32cap_P(%rip),%eax
+       cmp     \$$speed_limit,%rdx
+       jb      .Lcbc_slow_way
+       test    \$15,%rdx
+       jnz     .Lcbc_slow_way
+       bt      \$28,%eax
+       jc      .Lcbc_slow_way
 
        # allocate aligned stack frame...
-       lea     -64-248(%rsp),$key
+       lea     -88-248(%rsp),$key
        and     \$-64,$key
 
-       # ... and make it doesn't alias with AES_Te modulo 4096
+       # ... and make sure it doesn't alias with AES_T[ed] modulo 4096
        mov     $sbox,%r10
-       lea     2048($sbox),%r11
+       lea     2304($sbox),%r11
        mov     $key,%r12
        and     \$0xFFF,%r10    # s = $sbox&0xfff
        and     \$0xFFF,%r11    # e = ($sbox+2048)&0xfff
@@ -1664,20 +1673,24 @@ AES_cbc_encrypt:
        xchg    %rsp,$key
        add     \$8,%rsp        # reserve for return address!
        mov     $key,$_rsp      # save %rsp
+       mov     %rdi,$_inp      # save copy of inp
+       mov     %rsi,$_out      # save copy of out
        mov     %rdx,$_len      # save copy of len
        mov     %rcx,$_key      # save copy of key
        mov     %r8,$_ivp       # save copy of ivp
        movl    \$0,$mark       # copy of aes_key->rounds = 0;
        mov     %r8,%rbp        # rearrange input arguments
+       mov     %r9,%rbx
        mov     %rsi,$out
        mov     %rdi,$inp
        mov     %rcx,$key
 
+       mov     240($key),%eax          # key->rounds
        # do we copy key schedule to stack?
        mov     $key,%r10
        sub     $sbox,%r10
        and     \$0xfff,%r10
-       cmp     \$2048,%r10
+       cmp     \$2304,%r10
        jb      .Lcbc_do_ecopy
        cmp     \$4096-248,%r10
        jb      .Lcbc_skip_ecopy
@@ -1688,12 +1701,11 @@ AES_cbc_encrypt:
                lea     $aes_key,$key
                mov     \$240/8,%ecx
                .long   0x90A548F3      # rep movsq
-               mov     (%rsi),%eax     # copy aes_key->rounds
-               mov     %eax,(%rdi)
+               mov     %eax,(%rdi)     # copy aes_key->rounds
 .Lcbc_skip_ecopy:
        mov     $key,$keyp      # save key pointer
 
-       mov     \$16,%ecx
+       mov     \$18,%ecx
 .align 4
 .Lcbc_prefetch_te:
                mov     0($sbox),%r10
@@ -1703,42 +1715,41 @@ AES_cbc_encrypt:
                lea     128($sbox),$sbox
                sub     \$1,%ecx
        jnz     .Lcbc_prefetch_te
-       sub     \$2048,$sbox
+       lea     -2304($sbox),$sbox
 
-       test    \$-16,%rdx              # check upon length
-       mov     %rdx,%r10
+       cmp     \$0,%rbx
+       je      .LFAST_DECRYPT
+
+#----------------------------- ENCRYPT -----------------------------#
        mov     0(%rbp),$s0             # load iv
        mov     4(%rbp),$s1
        mov     8(%rbp),$s2
        mov     12(%rbp),$s3
-       jz      .Lcbc_enc_tail          # short input...
 
 .align 4
-.Lcbc_enc_loop:
+.Lcbc_fast_enc_loop:
                xor     0($inp),$s0
                xor     4($inp),$s1
                xor     8($inp),$s2
                xor     12($inp),$s3
-               mov     $inp,$ivec      # if ($verticalspin) save inp
-
                mov     $keyp,$key      # restore key
+               mov     $inp,$_inp      # if ($verticalspin) save inp
+
                call    _x86_64_AES_encrypt
 
-               mov     $ivec,$inp      # if ($verticalspin) restore inp
+               mov     $_inp,$inp      # if ($verticalspin) restore inp
+               mov     $_len,%r10
                mov     $s0,0($out)
                mov     $s1,4($out)
                mov     $s2,8($out)
                mov     $s3,12($out)
 
-               mov     $_len,%r10
                lea     16($inp),$inp
                lea     16($out),$out
                sub     \$16,%r10
                test    \$-16,%r10
                mov     %r10,$_len
-       jnz     .Lcbc_enc_loop
-       test    \$15,%r10
-       jnz     .Lcbc_enc_tail
+       jnz     .Lcbc_fast_enc_loop
        mov     $_ivp,%rbp      # restore ivp
        mov     $s0,0(%rbp)     # save ivec
        mov     $s1,4(%rbp)
@@ -1749,12 +1760,12 @@ AES_cbc_encrypt:
 .Lcbc_cleanup:
        cmpl    \$0,$mark       # was the key schedule copied?
        lea     $aes_key,%rdi
-       mov     $_rsp,%rsp
        je      .Lcbc_exit
                mov     \$240/8,%ecx
                xor     %rax,%rax
                .long   0x90AB48F3      # rep stosq
 .Lcbc_exit:
+       mov     $_rsp,%rsp
        popfq
        pop     %r15
        pop     %r14
@@ -1764,122 +1775,35 @@ AES_cbc_encrypt:
        pop     %rbx
 .Lcbc_just_ret:
        ret
-.align 4
-.Lcbc_enc_tail:
-       cmp     $inp,$out
-       je      .Lcbc_enc_in_place
-       mov     %r10,%rcx
-       mov     $inp,%rsi
-       mov     $out,%rdi
-       .long   0xF689A4F3              # rep movsb
-.Lcbc_enc_in_place:
-       mov     \$16,%rcx               # zero tail
-       sub     %r10,%rcx
-       xor     %rax,%rax
-       .long   0xF689AAF3              # rep stosb
-       mov     $out,$inp               # this is not a mistake!
-       movq    \$16,$_len              # len=16
-       jmp     .Lcbc_enc_loop          # one more spin...
+
 #----------------------------- DECRYPT -----------------------------#
 .align 16
-.LDECRYPT:
-       lea     AES_Td-.Lcbc_pic_point($sbox),$sbox
-
-       # allocate aligned stack frame...
-       lea     -64-248(%rsp),$key
-       and     \$-64,$key
-
-       # ... and make it doesn't alias with AES_Td modulo 4096
-       mov     $sbox,%r10
-       lea     2304($sbox),%r11
-       mov     $key,%r12
-       and     \$0xFFF,%r10    # s = $sbox&0xfff
-       and     \$0xFFF,%r11    # e = ($sbox+2048+256)&0xfff
-       and     \$0xFFF,%r12    # p = %rsp&0xfff
-
-       cmp     %r11,%r12       # if (p=>e) %rsp =- (p-e);
-       jb      .Lcbc_td_break_out
-       sub     %r11,%r12
-       sub     %r12,$key
-       jmp     .Lcbc_td_ok
-.Lcbc_td_break_out:            # else %rsp -= (p-s)&0xfff + framesz
-       sub     %r10,%r12
-       and     \$0xFFF,%r12
-       add     \$320,%r12
-       sub     %r12,$key
-.align 4
-.Lcbc_td_ok:
-
-       xchg    %rsp,$key
-       add     \$8,%rsp        # reserve for return address!
-       mov     $key,$_rsp      # save %rsp
-       mov     %rdx,$_len      # save copy of len
-       mov     %rcx,$_key      # save copy of key
-       mov     %r8,$_ivp       # save copy of ivp
-       movl    \$0,$mark       # copy of aes_key->rounds = 0;
-       mov     %r8,%rbp        # rearrange input arguments
-       mov     %rsi,$out
-       mov     %rdi,$inp
-       mov     %rcx,$key
-
-       # do we copy key schedule to stack?
-       mov     $key,%r10
-       sub     $sbox,%r10
-       and     \$0xfff,%r10
-       cmp     \$2304,%r10
-       jb      .Lcbc_do_dcopy
-       cmp     \$4096-248,%r10
-       jb      .Lcbc_skip_dcopy
-.align 4
-.Lcbc_do_dcopy:
-               mov     $key,%rsi
-               lea     $aes_key,%rdi
-               lea     $aes_key,$key
-               mov     \$240/8,%ecx
-               .long   0x90A548F3      # rep movsq
-               mov     (%rsi),%eax     # copy aes_key->rounds
-               mov     %eax,(%rdi)
-.Lcbc_skip_dcopy:
-       mov     $key,$keyp      # save key pointer
-
-       mov     \$18,%ecx
-.align 4
-.Lcbc_prefetch_td:
-               mov     0($sbox),%r10
-               mov     32($sbox),%r11
-               mov     64($sbox),%r12
-               mov     96($sbox),%r13
-               lea     128($sbox),$sbox
-               sub     \$1,%ecx
-       jnz     .Lcbc_prefetch_td
-       sub     \$2304,$sbox
-
+.LFAST_DECRYPT:
        cmp     $inp,$out
-       je      .Lcbc_dec_in_place
+       je      .Lcbc_fast_dec_in_place
 
        mov     %rbp,$ivec
 .align 4
-.Lcbc_dec_loop:
-               mov     0($inp),$s0             # read input
+.Lcbc_fast_dec_loop:
+               mov     0($inp),$s0     # read input
                mov     4($inp),$s1
                mov     8($inp),$s2
                mov     12($inp),$s3
-               mov     $inp,8+$ivec    # if ($verticalspin) save inp
-
                mov     $keyp,$key      # restore key
+               mov     $inp,$_inp      # if ($verticalspin) save inp
+
                call    _x86_64_AES_decrypt
 
                mov     $ivec,%rbp      # load ivp
-               mov     8+$ivec,$inp    # if ($verticalspin) restore inp
+               mov     $_inp,$inp      # if ($verticalspin) restore inp
+               mov     $_len,%r10      # load len
                xor     0(%rbp),$s0     # xor iv
                xor     4(%rbp),$s1
                xor     8(%rbp),$s2
                xor     12(%rbp),$s3
                mov     $inp,%rbp       # current input, next iv
 
-               mov     $_len,%r10      # load len
                sub     \$16,%r10
-               jc      .Lcbc_dec_partial
                mov     %r10,$_len      # update len
                mov     %rbp,$ivec      # update ivp
 
@@ -1890,8 +1814,7 @@ AES_cbc_encrypt:
 
                lea     16($inp),$inp
                lea     16($out),$out
-       jnz     .Lcbc_dec_loop
-.Lcbc_dec_end:
+       jnz     .Lcbc_fast_dec_loop
        mov     $_ivp,%r12              # load user ivp
        mov     0(%rbp),%r10            # load iv
        mov     8(%rbp),%r11
@@ -1899,64 +1822,239 @@ AES_cbc_encrypt:
        mov     %r11,8(%r12)
        jmp     .Lcbc_cleanup
 
-.align 4
-.Lcbc_dec_partial:
-       mov     $s0,0+$ivec             # dump output to stack
-       mov     $s1,4+$ivec
-       mov     $s2,8+$ivec
-       mov     $s3,12+$ivec
-       mov     $out,%rdi
-       lea     $ivec,%rsi
-       mov     \$16,%rcx
-       add     %r10,%rcx               # number of bytes to copy
-       .long   0xF689A4F3              # rep movsb
-       jmp     .Lcbc_dec_end
-
 .align 16
-.Lcbc_dec_in_place:
+.Lcbc_fast_dec_in_place:
+       mov     0(%rbp),%r10            # copy iv to stack
+       mov     8(%rbp),%r11
+       mov     %r10,0+$ivec
+       mov     %r11,8+$ivec
+.align 4
+.Lcbc_fast_dec_in_place_loop:
                mov     0($inp),$s0     # load input
                mov     4($inp),$s1
                mov     8($inp),$s2
                mov     12($inp),$s3
+               mov     $keyp,$key      # restore key
+               mov     $inp,$_inp      # if ($verticalspin) save inp
 
-               mov     $inp,$ivec      # if ($verticalspin) save inp
-               mov     $keyp,$key
                call    _x86_64_AES_decrypt
 
-               mov     $ivec,$inp      # if ($verticalspin) restore inp
-               mov     $_ivp,%rbp
-               xor     0(%rbp),$s0
-               xor     4(%rbp),$s1
-               xor     8(%rbp),$s2
-               xor     12(%rbp),$s3
+               mov     $_inp,$inp      # if ($verticalspin) restore inp
+               mov     $_len,%r10
+               xor     0+$ivec,$s0
+               xor     4+$ivec,$s1
+               xor     8+$ivec,$s2
+               xor     12+$ivec,$s3
+
+               mov     0($inp),%r11    # load input
+               mov     8($inp),%r12
+               sub     \$16,%r10
+               jz      .Lcbc_fast_dec_in_place_done
 
-               mov     0($inp),%r10    # copy input to iv
-               mov     8($inp),%r11
-               mov     %r10,0(%rbp)
-               mov     %r11,8(%rbp)
+               mov     %r11,0+$ivec    # copy input to iv
+               mov     %r12,8+$ivec
 
                mov     $s0,0($out)     # save output [zaps input]
                mov     $s1,4($out)
                mov     $s2,8($out)
                mov     $s3,12($out)
 
-               mov     $_len,%rcx
                lea     16($inp),$inp
                lea     16($out),$out
-               sub     \$16,%rcx
-               jc      .Lcbc_dec_in_place_partial
-               mov     %rcx,$_len
-       jnz     .Lcbc_dec_in_place
+               mov     %r10,$_len
+       jmp     .Lcbc_fast_dec_in_place_loop
+.Lcbc_fast_dec_in_place_done:
+       mov     $_ivp,%rdi
+       mov     %r11,0(%rdi)    # copy iv back to user
+       mov     %r12,8(%rdi)
+
+       mov     $s0,0($out)     # save output [zaps input]
+       mov     $s1,4($out)
+       mov     $s2,8($out)
+       mov     $s3,12($out)
+
        jmp     .Lcbc_cleanup
 
+#--------------------------- SLOW ROUTINE ---------------------------#
+.align 16
+.Lcbc_slow_way:
+       # allocate aligned stack frame...
+       lea     -80(%rsp),%rbp
+       and     \$-64,%rbp
+       # ... just "above" key schedule
+       lea     -80-63(%rcx),%rax
+       sub     %rbp,%rax
+       neg     %rax
+       and     \$0x3c0,%rax
+       sub     %rax,%rbp
+
+       xchg    %rsp,%rbp
+       add     \$8,%rsp        # reserve for return address!
+       mov     %rbp,$_rsp      # save %rsp
+       mov     %rdi,$_inp      # save copy of inp
+       mov     %rsi,$_out      # save copy of out
+       mov     %rdx,$_len      # save copy of len
+       #mov    %rcx,$_key      # save copy of key
+       mov     %r8,$_ivp       # save copy of ivp
+       mov     %r8,%rbp        # rearrange input arguments
+       mov     %r9,%rbx
+       mov     %rsi,$out
+       mov     %rdi,$inp
+       mov     %rcx,$key
+
+       mov     240($key),%eax
+       mov     $key,$keyp      # save key pointer
+       shl     \$4,%eax
+       lea     ($key,%rax),%rax
+       mov     %rax,$keyend
+
+       # pick Te4 copy which can't "overlap" with stack frame or key scdedule
+       lea     2048($sbox),$sbox
+       lea     768-8(%rsp),%rax
+       sub     $sbox,%rax
+       and     \$0x300,%rax
+       lea     ($sbox,%rax),$sbox
+
+       cmp     \$0,%rbx
+       je      .LSLOW_DECRYPT
+
+#--------------------------- SLOW ENCRYPT ---------------------------#
+       test    \$-16,%rdx              # check upon length
+       mov     %rdx,%r10
+       mov     0(%rbp),$s0             # load iv
+       mov     4(%rbp),$s1
+       mov     8(%rbp),$s2
+       mov     12(%rbp),$s3
+       jz      .Lcbc_slow_enc_tail     # short input...
+
 .align 4
-.Lcbc_dec_in_place_partial:
-       # one can argue if this is actually required
-       lea     ($out,%rcx),%rdi
-       lea     (%rbp,%rcx),%rsi
-       neg     %rcx
-       .long   0xF689A4F3      # rep movsb     # restore tail
-       jmp     .Lcbc_cleanup
+.Lcbc_slow_enc_loop:
+               xor     0($inp),$s0
+               xor     4($inp),$s1
+               xor     8($inp),$s2
+               xor     12($inp),$s3
+               mov     $keyp,$key      # restore key
+               mov     $inp,$_inp      # save inp
+               mov     $out,$_out      # save out
+
+               call    _x86_64_AES_encrypt_compact
+
+               mov     $_inp,$inp      # restore inp
+               mov     $_out,$out      # restore out
+               mov     $_len,%r10
+               mov     $s0,0($out)
+               mov     $s1,4($out)
+               mov     $s2,8($out)
+               mov     $s3,12($out)
+
+               lea     16($inp),$inp
+               lea     16($out),$out
+               sub     \$16,%r10
+               test    \$-16,%r10
+               mov     %r10,$_len
+       jnz     .Lcbc_slow_enc_loop
+       test    \$15,%r10
+       jnz     .Lcbc_slow_enc_tail
+       mov     $_ivp,%rbp      # restore ivp
+       mov     $s0,0(%rbp)     # save ivec
+       mov     $s1,4(%rbp)
+       mov     $s2,8(%rbp)
+       mov     $s3,12(%rbp)
+
+       jmp     .Lcbc_exit
+.align 4
+.Lcbc_slow_enc_tail:
+       cmp     $inp,$out
+       je      .Lcbc_slow_enc_in_place
+       mov     %r10,%rcx
+       mov     $inp,%rsi
+       mov     $out,%rdi
+       .long   0xF689A4F3              # rep movsb
+.Lcbc_slow_enc_in_place:
+       mov     \$16,%rcx               # zero tail
+       sub     %r10,%rcx
+       xor     %rax,%rax
+       .long   0xF689AAF3              # rep stosb
+       mov     $out,$inp               # this is not a mistake!
+       movq    \$16,$_len              # len=16
+       jmp     .Lcbc_slow_enc_loop     # one more spin...
+#--------------------------- SLOW DECRYPT ---------------------------#
+.align 16
+.LSLOW_DECRYPT:
+       shr     \$3,%rax
+       add     %rax,$sbox              # recall "magic" constants!
+
+       mov     0(%rbp),%r10            # copy iv to stack
+       mov     8(%rbp),%r11
+       mov     %r10,0+$ivec
+       mov     %r11,8+$ivec
+
+.align 4
+.Lcbc_slow_dec_loop:
+               mov     0($inp),$s0     # load input
+               mov     4($inp),$s1
+               mov     8($inp),$s2
+               mov     12($inp),$s3
+               mov     $keyp,$key      # restore key
+               mov     $inp,$_inp      # save inp
+               mov     $out,$_out      # save out
+
+               call    _x86_64_AES_decrypt_compact
+
+               mov     $_inp,$inp      # restore inp
+               mov     $_out,$out      # restore out
+               mov     $_len,%r10
+               xor     0+$ivec,$s0
+               xor     4+$ivec,$s1
+               xor     8+$ivec,$s2
+               xor     12+$ivec,$s3
+
+               mov     0($inp),%r11    # load input
+               mov     8($inp),%r12
+               sub     \$16,%r10
+               jc      .Lcbc_slow_dec_partial
+               jz      .Lcbc_slow_dec_done
+
+               mov     %r11,0+$ivec    # copy input to iv
+               mov     %r12,8+$ivec
+
+               mov     $s0,0($out)     # save output [can zap input]
+               mov     $s1,4($out)
+               mov     $s2,8($out)
+               mov     $s3,12($out)
+
+               lea     16($inp),$inp
+               lea     16($out),$out
+               mov     %r10,$_len
+       jmp     .Lcbc_slow_dec_loop
+.Lcbc_slow_dec_done:
+       mov     $_ivp,%rdi
+       mov     %r11,0(%rdi)            # copy iv back to user
+       mov     %r12,8(%rdi)
+
+       mov     $s0,0($out)             # save output [can zap input]
+       mov     $s1,4($out)
+       mov     $s2,8($out)
+       mov     $s3,12($out)
+
+       jmp     .Lcbc_exit
+
+.align 4
+.Lcbc_slow_dec_partial:
+       mov     $_ivp,%rdi
+       mov     %r11,0(%rdi)            # copy iv back to user
+       mov     %r12,8(%rdi)
+
+       mov     $s0,0+$ivec             # save output to stack
+       mov     $s1,4+$ivec
+       mov     $s2,8+$ivec
+       mov     $s3,12+$ivec
+
+       mov     $out,%rdi
+       lea     $ivec,%rsi
+       lea     16(%r10),%rcx
+       .long   0xF689A4F3      # rep movsb
+       jmp     .Lcbc_exit
 .size  AES_cbc_encrypt,.-AES_cbc_encrypt
 ___
 }
@@ -2385,6 +2483,7 @@ $code.=<<___;
        .long   0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
        .long   0x1b1b1b1b, 0x1b1b1b1b, 0, 0
 .asciz  "AES for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+.align 64
 ___
 
 $code =~ s/\`([^\`]*)\`/eval($1)/gem;