aesni-x86[_64].pl: optimize for Sandy Bridge and add XTS mode.

diff --git a/crypto/aes/asm/aesni-x86.pl b/crypto/aes/asm/aesni-x86.pl
index 1a1bf539cd45cfc1ee8a278d78c94506eed161c5..712149ab4b939e1af64ebd50c86c691be2d7e925 100644 (file)
--- a/crypto/aes/asm/aesni-x86.pl
+++ b/crypto/aes/asm/aesni-x86.pl
@@ -27,7 +27,21 @@
 # Lower ratios for smaller block sizes are perfectly understandable,
 # because function call overhead is higher in 32-bit mode. Largest
 # 8-KB block performance is virtually same: 32-bit code is less than
 # Lower ratios for smaller block sizes are perfectly understandable,
 # because function call overhead is higher in 32-bit mode. Largest
 # 8-KB block performance is virtually same: 32-bit code is less than

-# 1% slower for ECB, CBC and CCM, and ~3% slower otherwise. 
+# 1% slower for ECB, CBC and CCM, and ~3% slower otherwise.
+
+# January 2011
+#
+# See aesni-x86_64.pl for details. Unlike x86_64 version this module
+# interleaves at most 6 aes[enc|dec] instructions, because there are
+# not enough registers for 8x interleave [which should be optimal for
+# Sandy Bridge]. Actually, performance results for 6x interleave
+# factor presented in aesni-x86_64.pl (except for CTR) are for this
+# module.
+
+# April 2011
+#
+# Add aesni_xts_[en|de]crypt. Westmere spends 1.50 cycles processing
+# one byte out of 8KB with 128-bit key, Sandy Bridge - 1.09.

 
 $PREFIX="aesni";       # if $PREFIX is set to "AES", the script
                        # generates drop-in replacement for
 
 $PREFIX="aesni";       # if $PREFIX is set to "AES", the script
                        # generates drop-in replacement for


@@ -51,14 +65,14 @@ $out="edi";
 $rounds_="ebx";        # backup copy for $rounds
 $key_="ebp";   # backup copy for $key
 
 $rounds_="ebx";        # backup copy for $rounds
 $key_="ebp";   # backup copy for $key
 

-$inout0="xmm0";
-$inout1="xmm1";
-$inout2="xmm2";
-$rndkey0="xmm3";
-$rndkey1="xmm4";
-$ivec="xmm5";
-$in0="xmm6";
-$in1="xmm7";   $inout3="xmm7";
+$rndkey0="xmm0";
+$rndkey1="xmm1";
+$inout0="xmm2";
+$inout1="xmm3";
+$inout2="xmm4";
+$inout3="xmm5";        $in1="xmm5";
+$inout4="xmm6";        $in0="xmm6";
+$inout5="xmm7";        $ivec="xmm7";

 
 # AESNI extenstion
 sub aeskeygenassist
 
 # AESNI extenstion
 sub aeskeygenassist


@@ -80,13 +94,15 @@ sub aesdeclast      { aescommon(0xdf,@_); }
 # Inline version of internal aesni_[en|de]crypt1
 { my $sn;
 sub aesni_inline_generate1
 # Inline version of internal aesni_[en|de]crypt1
 { my $sn;
 sub aesni_inline_generate1

-{ my ($p,$inout)=@_; $inout=$inout0 if (!defined($inout));
+{ my ($p,$inout,$ivec)=@_; $inout=$inout0 if (!defined($inout));

   $sn++;
 
   $sn++;
 

-    &movdqu            ($rndkey0,&QWP(0,$key));
+    &$movekey          ($rndkey0,&QWP(0,$key));

     &$movekey          ($rndkey1,&QWP(16,$key));
     &$movekey          ($rndkey1,&QWP(16,$key));

+    &xorps             ($ivec,$rndkey0)        if (defined($ivec));

     &lea               ($key,&DWP(32,$key));
     &lea               ($key,&DWP(32,$key));

-    &pxor              ($inout,$rndkey0);
+    &xorps             ($inout,$ivec)          if (defined($ivec));
+    &xorps             ($inout,$rndkey0)       if (!defined($ivec));

     &set_label("${p}1_loop_$sn");
        eval"&aes${p}   ($inout,$rndkey1)";
        &dec            ($rounds);
     &set_label("${p}1_loop_$sn");
        eval"&aes${p}   ($inout,$rndkey1)";
        &dec            ($rounds);


@@ -100,9 +116,9 @@ sub aesni_generate1 # fully unrolled loop
 { my ($p,$inout)=@_; $inout=$inout0 if (!defined($inout));
 
     &function_begin_B("_aesni_${p}rypt1");
 { my ($p,$inout)=@_; $inout=$inout0 if (!defined($inout));
 
     &function_begin_B("_aesni_${p}rypt1");

-       &movdqu         ($rndkey0,&QWP(0,$key));
+       &movups         ($rndkey0,&QWP(0,$key));

        &$movekey       ($rndkey1,&QWP(0x10,$key));
        &$movekey       ($rndkey1,&QWP(0x10,$key));

-       &pxor           ($inout,$rndkey0);
+       &xorps          ($inout,$rndkey0);

        &$movekey       ($rndkey0,&QWP(0x20,$key));
        &lea            ($key,&DWP(0x30,$key));
        &cmp            ($rounds,11);
        &$movekey       ($rndkey0,&QWP(0x20,$key));
        &lea            ($key,&DWP(0x30,$key));
        &cmp            ($rounds,11);


@@ -147,7 +163,7 @@ sub aesni_generate1 # fully unrolled loop
 &function_begin_B("${PREFIX}_encrypt");
        &mov    ("eax",&wparam(0));
        &mov    ($key,&wparam(2));
 &function_begin_B("${PREFIX}_encrypt");
        &mov    ("eax",&wparam(0));
        &mov    ($key,&wparam(2));

-       &movdqu ($inout0,&QWP(0,"eax"));
+       &movups ($inout0,&QWP(0,"eax"));

        &mov    ($rounds,&DWP(240,$key));
        &mov    ("eax",&wparam(1));
        if ($inline)
        &mov    ($rounds,&DWP(240,$key));
        &mov    ("eax",&wparam(1));
        if ($inline)


@@ -163,7 +179,7 @@ sub aesni_generate1 # fully unrolled loop
 &function_begin_B("${PREFIX}_decrypt");
        &mov    ("eax",&wparam(0));
        &mov    ($key,&wparam(2));
 &function_begin_B("${PREFIX}_decrypt");
        &mov    ("eax",&wparam(0));
        &mov    ($key,&wparam(2));

-       &movdqu ($inout0,&QWP(0,"eax"));
+       &movups ($inout0,&QWP(0,"eax"));

        &mov    ($rounds,&DWP(240,$key));
        &mov    ("eax",&wparam(1));
        if ($inline)
        &mov    ($rounds,&DWP(240,$key));
        &mov    ("eax",&wparam(1));
        if ($inline)


@@ -174,16 +190,19 @@ sub aesni_generate1       # fully unrolled loop
        &ret    ();
 &function_end_B("${PREFIX}_decrypt");
 
        &ret    ();
 &function_end_B("${PREFIX}_decrypt");
 

-# _aesni_[en|de]crypt[34] are private interfaces, N denotes interleave
-# factor. Why 3x subroutine is used in loops? Even though aes[enc|dec]
-# latency is 6, it turned out that it can be scheduled only every
-# *second* cycle. Thus 3x interleave is the one providing optimal
+# _aesni_[en|de]cryptN are private interfaces, N denotes interleave
+# factor. Why 3x subroutine were originally used in loops? Even though
+# aes[enc|dec] latency was originally 6, it could be scheduled only
+# every *2nd* cycle. Thus 3x interleave was the one providing optimal

 # utilization, i.e. when subroutine's throughput is virtually same as
 # of non-interleaved subroutine [for number of input blocks up to 3].
 # utilization, i.e. when subroutine's throughput is virtually same as
 # of non-interleaved subroutine [for number of input blocks up to 3].

-# This is why it makes no sense to implement 2x subroutine. As soon
-# as/if Intel improves throughput by making it possible to schedule
-# the instructions in question *every* cycles I would have to
-# implement 6x interleave and use it in loop...
+# This is why it makes no sense to implement 2x subroutine.
+# aes[enc|dec] latency in next processor generation is 8, but the
+# instructions can be scheduled every cycle. Optimal interleave for
+# new processor is therefore 8x, but it's unfeasible to accommodate it
+# in XMM registers addreassable in 32-bit mode and therefore 6x is
+# used instead...
+

 sub aesni_generate3
 { my $p=shift;
 
 sub aesni_generate3
 { my $p=shift;
 


@@ -192,7 +211,7 @@ sub aesni_generate3
        &shr            ($rounds,1);
        &$movekey       ($rndkey1,&QWP(16,$key));
        &lea            ($key,&DWP(32,$key));
        &shr            ($rounds,1);
        &$movekey       ($rndkey1,&QWP(16,$key));
        &lea            ($key,&DWP(32,$key));

-       &pxor           ($inout0,$rndkey0);
+       &xorps          ($inout0,$rndkey0);

        &pxor           ($inout1,$rndkey0);
        &pxor           ($inout2,$rndkey0);
        &$movekey       ($rndkey0,&QWP(0,$key));
        &pxor           ($inout1,$rndkey0);
        &pxor           ($inout2,$rndkey0);
        &$movekey       ($rndkey0,&QWP(0,$key));


@@ -231,13 +250,13 @@ sub aesni_generate4
        &$movekey       ($rndkey1,&QWP(16,$key));
        &shr            ($rounds,1);
        &lea            ($key,&DWP(32,$key));
        &$movekey       ($rndkey1,&QWP(16,$key));
        &shr            ($rounds,1);
        &lea            ($key,&DWP(32,$key));

-       &pxor           ($inout0,$rndkey0);
+       &xorps          ($inout0,$rndkey0);

        &pxor           ($inout1,$rndkey0);
        &pxor           ($inout2,$rndkey0);
        &pxor           ($inout3,$rndkey0);
        &$movekey       ($rndkey0,&QWP(0,$key));
 
        &pxor           ($inout1,$rndkey0);
        &pxor           ($inout2,$rndkey0);
        &pxor           ($inout3,$rndkey0);
        &$movekey       ($rndkey0,&QWP(0,$key));
 

-    &set_label("${p}3_loop");
+    &set_label("${p}4_loop");

        eval"&aes${p}   ($inout0,$rndkey1)";
        eval"&aes${p}   ($inout1,$rndkey1)";
        &dec            ($rounds);
        eval"&aes${p}   ($inout0,$rndkey1)";
        eval"&aes${p}   ($inout1,$rndkey1)";
        &dec            ($rounds);


@@ -250,7 +269,7 @@ sub aesni_generate4
        eval"&aes${p}   ($inout2,$rndkey0)";
        eval"&aes${p}   ($inout3,$rndkey0)";
        &$movekey       ($rndkey0,&QWP(0,$key));
        eval"&aes${p}   ($inout2,$rndkey0)";
        eval"&aes${p}   ($inout3,$rndkey0)";
        &$movekey       ($rndkey0,&QWP(0,$key));

-    &jnz               (&label("${p}3_loop"));
+    &jnz               (&label("${p}4_loop"));

 
     eval"&aes${p}      ($inout0,$rndkey1)";
     eval"&aes${p}      ($inout1,$rndkey1)";
 
     eval"&aes${p}      ($inout0,$rndkey1)";
     eval"&aes${p}      ($inout1,$rndkey1)";


@@ -263,10 +282,73 @@ sub aesni_generate4
     &ret();
     &function_end_B("_aesni_${p}rypt4");
 }
     &ret();
     &function_end_B("_aesni_${p}rypt4");
 }

+
+sub aesni_generate6
+{ my $p=shift;
+
+    &function_begin_B("_aesni_${p}rypt6");
+    &static_label("_aesni_${p}rypt6_enter");
+       &$movekey       ($rndkey0,&QWP(0,$key));
+       &shr            ($rounds,1);
+       &$movekey       ($rndkey1,&QWP(16,$key));
+       &lea            ($key,&DWP(32,$key));
+       &xorps          ($inout0,$rndkey0);
+       &pxor           ($inout1,$rndkey0);     # pxor does better here
+       eval"&aes${p}   ($inout0,$rndkey1)";
+       &pxor           ($inout2,$rndkey0);
+       eval"&aes${p}   ($inout1,$rndkey1)";
+       &pxor           ($inout3,$rndkey0);
+       &dec            ($rounds);
+       eval"&aes${p}   ($inout2,$rndkey1)";
+       &pxor           ($inout4,$rndkey0);
+       eval"&aes${p}   ($inout3,$rndkey1)";
+       &pxor           ($inout5,$rndkey0);
+       eval"&aes${p}   ($inout4,$rndkey1)";
+       &$movekey       ($rndkey0,&QWP(0,$key));
+       eval"&aes${p}   ($inout5,$rndkey1)";
+       &jmp            (&label("_aesni_${p}rypt6_enter"));
+
+    &set_label("${p}6_loop",16);
+       eval"&aes${p}   ($inout0,$rndkey1)";
+       eval"&aes${p}   ($inout1,$rndkey1)";
+       &dec            ($rounds);
+       eval"&aes${p}   ($inout2,$rndkey1)";
+       eval"&aes${p}   ($inout3,$rndkey1)";
+       eval"&aes${p}   ($inout4,$rndkey1)";
+       eval"&aes${p}   ($inout5,$rndkey1)";
+    &set_label("_aesni_${p}rypt6_enter",16);
+       &$movekey       ($rndkey1,&QWP(16,$key));
+       eval"&aes${p}   ($inout0,$rndkey0)";
+       eval"&aes${p}   ($inout1,$rndkey0)";
+       &lea            ($key,&DWP(32,$key));
+       eval"&aes${p}   ($inout2,$rndkey0)";
+       eval"&aes${p}   ($inout3,$rndkey0)";
+       eval"&aes${p}   ($inout4,$rndkey0)";
+       eval"&aes${p}   ($inout5,$rndkey0)";
+       &$movekey       ($rndkey0,&QWP(0,$key));
+    &jnz               (&label("${p}6_loop"));
+
+    eval"&aes${p}      ($inout0,$rndkey1)";
+    eval"&aes${p}      ($inout1,$rndkey1)";
+    eval"&aes${p}      ($inout2,$rndkey1)";
+    eval"&aes${p}      ($inout3,$rndkey1)";
+    eval"&aes${p}      ($inout4,$rndkey1)";
+    eval"&aes${p}      ($inout5,$rndkey1)";
+    eval"&aes${p}last  ($inout0,$rndkey0)";
+    eval"&aes${p}last  ($inout1,$rndkey0)";
+    eval"&aes${p}last  ($inout2,$rndkey0)";
+    eval"&aes${p}last  ($inout3,$rndkey0)";
+    eval"&aes${p}last  ($inout4,$rndkey0)";
+    eval"&aes${p}last  ($inout5,$rndkey0)";
+    &ret();
+    &function_end_B("_aesni_${p}rypt6");
+}

 &aesni_generate3("enc") if ($PREFIX eq "aesni");
 &aesni_generate3("dec");
 &aesni_generate4("enc") if ($PREFIX eq "aesni");
 &aesni_generate4("dec");
 &aesni_generate3("enc") if ($PREFIX eq "aesni");
 &aesni_generate3("dec");
 &aesni_generate4("enc") if ($PREFIX eq "aesni");
 &aesni_generate4("dec");

+&aesni_generate6("enc") if ($PREFIX eq "aesni");
+&aesni_generate6("dec");

 \f
 if ($PREFIX eq "aesni") {
 ######################################################################
 \f
 if ($PREFIX eq "aesni") {
 ######################################################################


@@ -278,37 +360,62 @@ if ($PREFIX eq "aesni") {
        &mov    ($out,&wparam(1));
        &mov    ($len,&wparam(2));
        &mov    ($key,&wparam(3));
        &mov    ($out,&wparam(1));
        &mov    ($len,&wparam(2));
        &mov    ($key,&wparam(3));

-       &mov    ($rounds,&wparam(4));
-       &cmp    ($len,16);
-       &jb     (&label("ecb_ret"));
+       &mov    ($rounds_,&wparam(4));

        &and    ($len,-16);
        &and    ($len,-16);

-       &test   ($rounds,$rounds)
+       &jz     (&label("ecb_ret"));

        &mov    ($rounds,&DWP(240,$key));
        &mov    ($rounds,&DWP(240,$key));

+       &test   ($rounds_,$rounds_);
+       &jz     (&label("ecb_decrypt"));
+

        &mov    ($key_,$key);           # backup $key
        &mov    ($rounds_,$rounds);     # backup $rounds
        &mov    ($key_,$key);           # backup $key
        &mov    ($rounds_,$rounds);     # backup $rounds

-       &jz     (&label("ecb_decrypt"));
+       &cmp    ($len,0x60);
+       &jb     (&label("ecb_enc_tail"));
+
+       &movdqu ($inout0,&QWP(0,$inp));
+       &movdqu ($inout1,&QWP(0x10,$inp));
+       &movdqu ($inout2,&QWP(0x20,$inp));
+       &movdqu ($inout3,&QWP(0x30,$inp));
+       &movdqu ($inout4,&QWP(0x40,$inp));
+       &movdqu ($inout5,&QWP(0x50,$inp));
+       &lea    ($inp,&DWP(0x60,$inp));
+       &sub    ($len,0x60);
+       &jmp    (&label("ecb_enc_loop6_enter"));
+
+&set_label("ecb_enc_loop6",16);
+       &movups (&QWP(0,$out),$inout0);
+       &movdqu ($inout0,&QWP(0,$inp));
+       &movups (&QWP(0x10,$out),$inout1);
+       &movdqu ($inout1,&QWP(0x10,$inp));
+       &movups (&QWP(0x20,$out),$inout2);
+       &movdqu ($inout2,&QWP(0x20,$inp));
+       &movups (&QWP(0x30,$out),$inout3);
+       &movdqu ($inout3,&QWP(0x30,$inp));
+       &movups (&QWP(0x40,$out),$inout4);
+       &movdqu ($inout4,&QWP(0x40,$inp));
+       &movups (&QWP(0x50,$out),$inout5);
+       &lea    ($out,&DWP(0x60,$out));
+       &movdqu ($inout5,&QWP(0x50,$inp));
+       &lea    ($inp,&DWP(0x60,$inp));
+&set_label("ecb_enc_loop6_enter");

 
 

-       &cmp    ($len,0x40);
-       &jbe    (&label("ecb_enc_tail"));
-       &sub    ($len,0x40);
-       &jmp    (&label("ecb_enc_loop3"));
+       &call   ("_aesni_encrypt6");

 
 

-&set_label("ecb_enc_loop3",16);
-       &movups ($inout0,&QWP(0,$inp));
-       &movups ($inout1,&QWP(0x10,$inp));
-       &movups ($inout2,&QWP(0x20,$inp));
-       &call   ("_aesni_encrypt3");
-       &lea    ($inp,&DWP(0x30,$inp));
-       &movups (&QWP(0,$out),$inout0);

        &mov    ($key,$key_);           # restore $key
        &mov    ($key,$key_);           # restore $key

-       &movups (&QWP(0x10,$out),$inout1);

        &mov    ($rounds,$rounds_);     # restore $rounds
        &mov    ($rounds,$rounds_);     # restore $rounds

+       &sub    ($len,0x60);
+       &jnc    (&label("ecb_enc_loop6"));
+
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);

        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0x20,$out),$inout2);

-       &lea    ($out,&DWP(0x30,$out));
-       &sub    ($len,0x30);
-       &ja     (&label("ecb_enc_loop3"));
+       &movups (&QWP(0x30,$out),$inout3);
+       &movups (&QWP(0x40,$out),$inout4);
+       &movups (&QWP(0x50,$out),$inout5);
+       &lea    ($out,&DWP(0x60,$out));
+       &add    ($len,0x60);
+       &jz     (&label("ecb_ret"));

 
 

-       &add    ($len,0x40);

 &set_label("ecb_enc_tail");
        &movups ($inout0,&QWP(0,$inp));
        &cmp    ($len,0x20);
 &set_label("ecb_enc_tail");
        &movups ($inout0,&QWP(0,$inp));
        &cmp    ($len,0x20);


@@ -316,14 +423,18 @@ if ($PREFIX eq "aesni") {
        &movups ($inout1,&QWP(0x10,$inp));
        &je     (&label("ecb_enc_two"));
        &movups ($inout2,&QWP(0x20,$inp));
        &movups ($inout1,&QWP(0x10,$inp));
        &je     (&label("ecb_enc_two"));
        &movups ($inout2,&QWP(0x20,$inp));

-       &cmp    ($len,0x30);
-       &je     (&label("ecb_enc_three"));
+       &cmp    ($len,0x40);
+       &jb     (&label("ecb_enc_three"));

        &movups ($inout3,&QWP(0x30,$inp));
        &movups ($inout3,&QWP(0x30,$inp));

-       &call   ("_aesni_encrypt4");
+       &je     (&label("ecb_enc_four"));
+       &movups ($inout4,&QWP(0x40,$inp));
+       &xorps  ($inout5,$inout5);
+       &call   ("_aesni_encrypt6");

        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0x30,$out),$inout3);
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0x30,$out),$inout3);

+       &movups (&QWP(0x40,$out),$inout4);

        jmp     (&label("ecb_ret"));
 
 &set_label("ecb_enc_one",16);
        jmp     (&label("ecb_ret"));
 
 &set_label("ecb_enc_one",16);


@@ -335,7 +446,7 @@ if ($PREFIX eq "aesni") {
        &jmp    (&label("ecb_ret"));
 
 &set_label("ecb_enc_two",16);
        &jmp    (&label("ecb_ret"));
 
 &set_label("ecb_enc_two",16);

-       &pxor   ($inout2,$inout2);
+       &xorps  ($inout2,$inout2);

        &call   ("_aesni_encrypt3");
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &call   ("_aesni_encrypt3");
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);


@@ -347,29 +458,65 @@ if ($PREFIX eq "aesni") {
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &jmp    (&label("ecb_ret"));
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &jmp    (&label("ecb_ret"));

+
+&set_label("ecb_enc_four",16);
+       &call   ("_aesni_encrypt4");
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);
+       &movups (&QWP(0x20,$out),$inout2);
+       &movups (&QWP(0x30,$out),$inout3);
+       &jmp    (&label("ecb_ret"));

 ######################################################################
 &set_label("ecb_decrypt",16);
 ######################################################################
 &set_label("ecb_decrypt",16);

-       &cmp    ($len,0x40);
-       &jbe    (&label("ecb_dec_tail"));
-       &sub    ($len,0x40);
-       &jmp    (&label("ecb_dec_loop3"));
-
-&set_label("ecb_dec_loop3",16);
-       &movups ($inout0,&QWP(0,$inp));
-       &movups ($inout1,&QWP(0x10,$inp));
-       &movups ($inout2,&QWP(0x20,$inp));
-       &call   ("_aesni_decrypt3");
-       &lea    ($inp,&DWP(0x30,$inp));
+       &mov    ($key_,$key);           # backup $key
+       &mov    ($rounds_,$rounds);     # backup $rounds
+       &cmp    ($len,0x60);
+       &jb     (&label("ecb_dec_tail"));
+
+       &movdqu ($inout0,&QWP(0,$inp));
+       &movdqu ($inout1,&QWP(0x10,$inp));
+       &movdqu ($inout2,&QWP(0x20,$inp));
+       &movdqu ($inout3,&QWP(0x30,$inp));
+       &movdqu ($inout4,&QWP(0x40,$inp));
+       &movdqu ($inout5,&QWP(0x50,$inp));
+       &lea    ($inp,&DWP(0x60,$inp));
+       &sub    ($len,0x60);
+       &jmp    (&label("ecb_dec_loop6_enter"));
+
+&set_label("ecb_dec_loop6",16);

        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0,$out),$inout0);

-       &mov    ($key,$key_);           # restore $key
+       &movdqu ($inout0,&QWP(0,$inp));

        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x10,$out),$inout1);

+       &movdqu ($inout1,&QWP(0x10,$inp));
+       &movups (&QWP(0x20,$out),$inout2);
+       &movdqu ($inout2,&QWP(0x20,$inp));
+       &movups (&QWP(0x30,$out),$inout3);
+       &movdqu ($inout3,&QWP(0x30,$inp));
+       &movups (&QWP(0x40,$out),$inout4);
+       &movdqu ($inout4,&QWP(0x40,$inp));
+       &movups (&QWP(0x50,$out),$inout5);
+       &lea    ($out,&DWP(0x60,$out));
+       &movdqu ($inout5,&QWP(0x50,$inp));
+       &lea    ($inp,&DWP(0x60,$inp));
+&set_label("ecb_dec_loop6_enter");
+
+       &call   ("_aesni_decrypt6");
+
+       &mov    ($key,$key_);           # restore $key

        &mov    ($rounds,$rounds_);     # restore $rounds
        &mov    ($rounds,$rounds_);     # restore $rounds

+       &sub    ($len,0x60);
+       &jnc    (&label("ecb_dec_loop6"));
+
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);

        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0x20,$out),$inout2);

-       &lea    ($out,&DWP(0x30,$out));
-       &sub    ($len,0x30);
-       &ja     (&label("ecb_dec_loop3"));
+       &movups (&QWP(0x30,$out),$inout3);
+       &movups (&QWP(0x40,$out),$inout4);
+       &movups (&QWP(0x50,$out),$inout5);
+       &lea    ($out,&DWP(0x60,$out));
+       &add    ($len,0x60);
+       &jz     (&label("ecb_ret"));

 
 

-       &add    ($len,0x40);

 &set_label("ecb_dec_tail");
        &movups ($inout0,&QWP(0,$inp));
        &cmp    ($len,0x20);
 &set_label("ecb_dec_tail");
        &movups ($inout0,&QWP(0,$inp));
        &cmp    ($len,0x20);


@@ -377,14 +524,18 @@ if ($PREFIX eq "aesni") {
        &movups ($inout1,&QWP(0x10,$inp));
        &je     (&label("ecb_dec_two"));
        &movups ($inout2,&QWP(0x20,$inp));
        &movups ($inout1,&QWP(0x10,$inp));
        &je     (&label("ecb_dec_two"));
        &movups ($inout2,&QWP(0x20,$inp));

-       &cmp    ($len,0x30);
-       &je     (&label("ecb_dec_three"));
+       &cmp    ($len,0x40);
+       &jb     (&label("ecb_dec_three"));

        &movups ($inout3,&QWP(0x30,$inp));
        &movups ($inout3,&QWP(0x30,$inp));

-       &call   ("_aesni_decrypt4");
+       &je     (&label("ecb_dec_four"));
+       &movups ($inout4,&QWP(0x40,$inp));
+       &xorps  ($inout5,$inout5);
+       &call   ("_aesni_decrypt6");

        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0x30,$out),$inout3);
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0x30,$out),$inout3);

+       &movups (&QWP(0x40,$out),$inout4);

        &jmp    (&label("ecb_ret"));
 
 &set_label("ecb_dec_one",16);
        &jmp    (&label("ecb_ret"));
 
 &set_label("ecb_dec_one",16);


@@ -396,7 +547,7 @@ if ($PREFIX eq "aesni") {
        &jmp    (&label("ecb_ret"));
 
 &set_label("ecb_dec_two",16);
        &jmp    (&label("ecb_ret"));
 
 &set_label("ecb_dec_two",16);

-       &pxor   ($inout2,$inout2);
+       &xorps  ($inout2,$inout2);

        &call   ("_aesni_decrypt3");
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &call   ("_aesni_decrypt3");
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);


@@ -407,6 +558,14 @@ if ($PREFIX eq "aesni") {
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);
        &movups (&QWP(0,$out),$inout0);
        &movups (&QWP(0x10,$out),$inout1);
        &movups (&QWP(0x20,$out),$inout2);

+       &jmp    (&label("ecb_ret"));
+
+&set_label("ecb_dec_four",16);
+       &call   ("_aesni_decrypt4");
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);
+       &movups (&QWP(0x20,$out),$inout2);
+       &movups (&QWP(0x30,$out),$inout3);

 
 &set_label("ecb_ret");
 &function_end("aesni_ecb_encrypt");
 
 &set_label("ecb_ret");
 &function_end("aesni_ecb_encrypt");


@@ -420,6 +579,7 @@ if ($PREFIX eq "aesni") {
 # does not update *ivec! Nor does it finalize CMAC value
 # (see engine/eng_aesni.c for details)
 #
 # does not update *ivec! Nor does it finalize CMAC value
 # (see engine/eng_aesni.c for details)
 #

+{ my $cmac=$inout1;

 &function_begin("aesni_ccm64_encrypt_blocks");
        &mov    ($inp,&wparam(0));
        &mov    ($out,&wparam(1));
 &function_begin("aesni_ccm64_encrypt_blocks");
        &mov    ($inp,&wparam(0));
        &mov    ($out,&wparam(1));


@@ -433,7 +593,7 @@ if ($PREFIX eq "aesni") {
        &mov    (&DWP(48,"esp"),$key_);
 
        &movdqu ($ivec,&QWP(0,$rounds_));       # load ivec
        &mov    (&DWP(48,"esp"),$key_);
 
        &movdqu ($ivec,&QWP(0,$rounds_));       # load ivec

-       &movdqu ($inout1,&QWP(0,$rounds));      # load cmac
+       &movdqu ($cmac,&QWP(0,$rounds));        # load cmac

 
        # compose byte-swap control mask for pshufb on stack
        &mov    (&DWP(0,"esp"),0x0c0d0e0f);
 
        # compose byte-swap control mask for pshufb on stack
        &mov    (&DWP(0,"esp"),0x0c0d0e0f);


@@ -458,27 +618,47 @@ if ($PREFIX eq "aesni") {
        &movdqa ($inout0,$ivec);
 
 &set_label("ccm64_enc_outer");
        &movdqa ($inout0,$ivec);
 
 &set_label("ccm64_enc_outer");

-       &movdqu ($in0,&QWP(0,$inp));
-       &pshufb ($inout0,$inout3);
-       &mov    ($key,$key_);
-       &mov    ($rounds,$rounds_);
-       &pxor   ($inout1,$in0);                 # cmac^=inp
-       &pxor   ($inout2,$inout2);
+       &movups         ($in0,&QWP(0,$inp));
+       &pshufb         ($inout0,$inout3);
+       &mov            ($key,$key_);
+       &mov            ($rounds,$rounds_);

 
 

-       &call   ("_aesni_encrypt3");
+       &$movekey       ($rndkey0,&QWP(0,$key));
+       &shr            ($rounds,1);
+       &$movekey       ($rndkey1,&QWP(16,$key));
+       &xorps          ($in0,$rndkey0);
+       &lea            ($key,&DWP(32,$key));
+       &xorps          ($inout0,$rndkey0);
+       &xorps          ($cmac,$in0);           # cmac^=inp
+       &$movekey       ($rndkey0,&QWP(0,$key));
+
+&set_label("ccm64_enc2_loop");
+       &aesenc         ($inout0,$rndkey1);
+       &dec            ($rounds);
+       &aesenc         ($cmac,$rndkey1);
+       &$movekey       ($rndkey1,&QWP(16,$key));
+       &aesenc         ($inout0,$rndkey0);
+       &lea            ($key,&DWP(32,$key));
+       &aesenc         ($cmac,$rndkey0);
+       &$movekey       ($rndkey0,&QWP(0,$key));
+       &jnz            (&label("ccm64_enc2_loop"));
+       &aesenc         ($inout0,$rndkey1);
+       &aesenc         ($cmac,$rndkey1);
+       &aesenclast     ($inout0,$rndkey0);
+       &aesenclast     ($cmac,$rndkey0);

 
        &paddq  ($ivec,&QWP(16,"esp"));
        &dec    ($len);
        &lea    ($inp,&DWP(16,$inp));
 
        &paddq  ($ivec,&QWP(16,"esp"));
        &dec    ($len);
        &lea    ($inp,&DWP(16,$inp));

-       &pxor   ($in0,$inout0);                 # inp^=E(ivec)
+       &xorps  ($in0,$inout0);                 # inp^=E(ivec)

        &movdqa ($inout0,$ivec);
        &movdqa ($inout0,$ivec);

-       &movdqu (&QWP(0,$out),$in0);
+       &movups (&QWP(0,$out),$in0);

        &lea    ($out,&DWP(16,$out));
        &jnz    (&label("ccm64_enc_outer"));
 
        &mov    ("esp",&DWP(48,"esp"));
        &mov    ($out,&wparam(5));
        &lea    ($out,&DWP(16,$out));
        &jnz    (&label("ccm64_enc_outer"));
 
        &mov    ("esp",&DWP(48,"esp"));
        &mov    ($out,&wparam(5));

-       &movdqu (&QWP(0,$out),$inout1);
+       &movups (&QWP(0,$out),$cmac);

 &function_end("aesni_ccm64_encrypt_blocks");
 
 &function_begin("aesni_ccm64_decrypt_blocks");
 &function_end("aesni_ccm64_encrypt_blocks");
 
 &function_begin("aesni_ccm64_decrypt_blocks");


@@ -494,7 +674,7 @@ if ($PREFIX eq "aesni") {
        &mov    (&DWP(48,"esp"),$key_);
 
        &movdqu ($ivec,&QWP(0,$rounds_));       # load ivec
        &mov    (&DWP(48,"esp"),$key_);
 
        &movdqu ($ivec,&QWP(0,$rounds_));       # load ivec

-       &movdqu ($inout1,&QWP(0,$rounds));      # load cmac
+       &movdqu ($cmac,&QWP(0,$rounds));        # load cmac

 
        # compose byte-swap control mask for pshufb on stack
        &mov    (&DWP(0,"esp"),0x0c0d0e0f);
 
        # compose byte-swap control mask for pshufb on stack
        &mov    (&DWP(0,"esp"),0x0c0d0e0f);


@@ -524,35 +704,56 @@ if ($PREFIX eq "aesni") {
        {   &call       ("_aesni_encrypt1");    }
 
 &set_label("ccm64_dec_outer");
        {   &call       ("_aesni_encrypt1");    }
 
 &set_label("ccm64_dec_outer");

-       &movdqu ($in0,&QWP(0,$inp));

        &paddq  ($ivec,&QWP(16,"esp"));
        &paddq  ($ivec,&QWP(16,"esp"));

-       &dec    ($len);
-       &lea    ($inp,&QWP(16,$inp));
-       &pxor   ($in0,$inout0);
+       &movups ($in0,&QWP(0,$inp));            # load inp
+       &xorps  ($in0,$inout0);

        &movdqa ($inout0,$ivec);
        &movdqa ($inout0,$ivec);

+       &lea    ($inp,&QWP(16,$inp));
+       &pshufb ($inout0,$inout3);

        &mov    ($key,$key_);
        &mov    ($rounds,$rounds_);
        &mov    ($key,$key_);
        &mov    ($rounds,$rounds_);

-       &pshufb ($inout0,$inout3);
-       &movdqu (&QWP(0,$out),$in0);
+       &movups (&QWP(0,$out),$in0);

        &lea    ($out,&DWP(16,$out));
 
        &lea    ($out,&DWP(16,$out));
 

+       &sub    ($len,1);

        &jz     (&label("ccm64_dec_break"));
 
        &jz     (&label("ccm64_dec_break"));
 

-       &pxor   ($inout2,$inout2);
-       &call   ("_aesni_encrypt3");
+       &$movekey       ($rndkey0,&QWP(0,$key));
+       &shr            ($rounds,1);
+       &$movekey       ($rndkey1,&QWP(16,$key));
+       &xorps          ($in0,$rndkey0);
+       &lea            ($key,&DWP(32,$key));
+       &xorps          ($inout0,$rndkey0);
+       &xorps          ($cmac,$in0);           # cmac^=out
+       &$movekey       ($rndkey0,&QWP(0,$key));

 
 

+&set_label("ccm64_dec2_loop");
+       &aesenc         ($inout0,$rndkey1);
+       &dec            ($rounds);
+       &aesenc         ($cmac,$rndkey1);
+       &$movekey       ($rndkey1,&QWP(16,$key));
+       &aesenc         ($inout0,$rndkey0);
+       &lea            ($key,&DWP(32,$key));
+       &aesenc         ($cmac,$rndkey0);
+       &$movekey       ($rndkey0,&QWP(0,$key));
+       &jnz            (&label("ccm64_dec2_loop"));
+       &aesenc         ($inout0,$rndkey1);
+       &aesenc         ($cmac,$rndkey1);
+       &aesenclast     ($inout0,$rndkey0);
+       &aesenclast     ($cmac,$rndkey0);

        &jmp    (&label("ccm64_dec_outer"));
 
 &set_label("ccm64_dec_break",16);
        if ($inline)
        &jmp    (&label("ccm64_dec_outer"));
 
 &set_label("ccm64_dec_break",16);
        if ($inline)

-       {   &aesni_inline_generate1("enc",$inout1);     }
+       {   &aesni_inline_generate1("enc",$cmac,$in0);  }

        else
        else

-       {   &call       ("_aesni_encrypt1",$inout1);    }
+       {   &call       ("_aesni_encrypt1",$cmac);      }

 
        &mov    ("esp",&DWP(48,"esp"));
        &mov    ($out,&wparam(5));
 
        &mov    ("esp",&DWP(48,"esp"));
        &mov    ($out,&wparam(5));

-       &movdqu (&QWP(0,$out),$inout1);
+       &movups (&QWP(0,$out),$cmac);

 &function_end("aesni_ccm64_decrypt_blocks");
 &function_end("aesni_ccm64_decrypt_blocks");

+}

 \f
 ######################################################################
 # void aesni_ctr32_encrypt_blocks (const void *in, void *out,
 \f
 ######################################################################
 # void aesni_ctr32_encrypt_blocks (const void *in, void *out,


@@ -562,6 +763,14 @@ if ($PREFIX eq "aesni") {
 # Handles only complete blocks, operates on 32-bit counter and
 # does not update *ivec! (see engine/eng_aesni.c for details)
 #
 # Handles only complete blocks, operates on 32-bit counter and
 # does not update *ivec! (see engine/eng_aesni.c for details)
 #

+# stack layout:
+#      0       pshufb mask
+#      16      vector addend: 0,6,6,6
+#      32      counter-less ivec
+#      48      1st triplet of counter vector
+#      64      2nd triplet of counter vector
+#      80      saved %esp
+

 &function_begin("aesni_ctr32_encrypt_blocks");
        &mov    ($inp,&wparam(0));
        &mov    ($out,&wparam(1));
 &function_begin("aesni_ctr32_encrypt_blocks");
        &mov    ($inp,&wparam(0));
        &mov    ($out,&wparam(1));


@@ -569,14 +778,14 @@ if ($PREFIX eq "aesni") {
        &mov    ($key,&wparam(3));
        &mov    ($rounds_,&wparam(4));
        &mov    ($key_,"esp");
        &mov    ($key,&wparam(3));
        &mov    ($rounds_,&wparam(4));
        &mov    ($key_,"esp");

-       &sub    ("esp",60);
+       &sub    ("esp",88);

        &and    ("esp",-16);                    # align stack
        &and    ("esp",-16);                    # align stack

-       &mov    (&DWP(48,"esp"),$key_);
+       &mov    (&DWP(80,"esp"),$key_);

 
        &cmp    ($len,1);
        &je     (&label("ctr32_one_shortcut"));
 
 
        &cmp    ($len,1);
        &je     (&label("ctr32_one_shortcut"));
 

-       &movups ($inout3,&QWP(0,$rounds_));     # load ivec
+       &movdqu ($inout5,&QWP(0,$rounds_));     # load ivec

 
        # compose byte-swap control mask for pshufb on stack
        &mov    (&DWP(0,"esp"),0x0c0d0e0f);
 
        # compose byte-swap control mask for pshufb on stack
        &mov    (&DWP(0,"esp"),0x0c0d0e0f);


@@ -585,139 +794,167 @@ if ($PREFIX eq "aesni") {
        &mov    (&DWP(12,"esp"),0x00010203);
 
        # compose counter increment vector on stack
        &mov    (&DWP(12,"esp"),0x00010203);
 
        # compose counter increment vector on stack

-       &mov    ($rounds,3);
+       &mov    ($rounds,6);

        &xor    ($key_,$key_);
        &mov    (&DWP(16,"esp"),$rounds);
        &mov    (&DWP(20,"esp"),$rounds);
        &mov    (&DWP(24,"esp"),$rounds);
        &mov    (&DWP(28,"esp"),$key_);
 
        &xor    ($key_,$key_);
        &mov    (&DWP(16,"esp"),$rounds);
        &mov    (&DWP(20,"esp"),$rounds);
        &mov    (&DWP(24,"esp"),$rounds);
        &mov    (&DWP(28,"esp"),$key_);
 

-       &pextrd ($rounds_,$inout3,3);           # pull 32-bit counter
-       &pinsrd ($inout3,$key_,3);              # wipe 32-bit counter
+       &pextrd ($rounds_,$inout5,3);           # pull 32-bit counter
+       &pinsrd ($inout5,$key_,3);              # wipe 32-bit counter

 
        &mov    ($rounds,&DWP(240,$key));       # key->rounds
 
        &mov    ($rounds,&DWP(240,$key));       # key->rounds

-       &movdqa ($rndkey0,&QWP(0,"esp"));       # load byte-swap mask

 
 

-       # $ivec is vector of 3 32-bit counters
-       &pxor   ($ivec,$ivec);
+       # compose 2 vectors of 3x32-bit counters

        &bswap  ($rounds_);
        &bswap  ($rounds_);

-       &pinsrd ($ivec,$rounds_,0);
+       &pxor   ($rndkey1,$rndkey1);
+       &pxor   ($rndkey0,$rndkey0);
+       &movdqa ($inout0,&QWP(0,"esp"));        # load byte-swap mask
+       &pinsrd ($rndkey1,$rounds_,0);
+       &lea    ($key_,&DWP(3,$rounds_));
+       &pinsrd ($rndkey0,$key_,0);

        &inc    ($rounds_);
        &inc    ($rounds_);

-       &pinsrd ($ivec,$rounds_,1);
+       &pinsrd ($rndkey1,$rounds_,1);
+       &inc    ($key_);
+       &pinsrd ($rndkey0,$key_,1);

        &inc    ($rounds_);
        &inc    ($rounds_);

-       &pinsrd ($ivec,$rounds_,2);
-       &pshufb ($ivec,$rndkey0);               # byte swap
-
-       &cmp    ($len,4);
-       &jbe    (&label("ctr32_tail"));
-       &movdqa (&QWP(32,"esp"),$inout3);       # save counter-less ivec
-       &mov    ($rounds_,$rounds);
-       &mov    ($key_,$key);
-       &sub    ($len,4);
-       &jmp    (&label("ctr32_loop3"));
-
-&set_label("ctr32_loop3",16);
-       &pshufd ($inout0,$ivec,3<<6);           # place counter to upper dword
-       &pshufd ($inout1,$ivec,2<<6);
-       &por    ($inout0,$inout3);              # merge counter-less ivec
-       &pshufd ($inout2,$ivec,1<<6);
-       &por    ($inout1,$inout3);
-       &por    ($inout2,$inout3);
-
-       # inline _aesni_encrypt3 and interleave last round
-       # with own code...
-
-       &$movekey       ($rndkey0,&QWP(0,$key));
-       &shr            ($rounds,1);
-       &$movekey       ($rndkey1,&QWP(16,$key));
-       &lea            ($key,&DWP(32,$key));
+       &pinsrd ($rndkey1,$rounds_,2);
+       &inc    ($key_);
+       &pinsrd ($rndkey0,$key_,2);
+       &movdqa (&QWP(48,"esp"),$rndkey1);      # save 1st triplet
+       &pshufb ($rndkey1,$inout0);             # byte swap
+       &movdqa (&QWP(64,"esp"),$rndkey0);      # save 2nd triplet
+       &pshufb ($rndkey0,$inout0);             # byte swap
+
+       &pshufd ($inout0,$rndkey1,3<<6);        # place counter to upper dword
+       &pshufd ($inout1,$rndkey1,2<<6);
+       &cmp    ($len,6);
+       &jb     (&label("ctr32_tail"));
+       &movdqa (&QWP(32,"esp"),$inout5);       # save counter-less ivec
+       &shr    ($rounds,1);
+       &mov    ($key_,$key);                   # backup $key
+       &mov    ($rounds_,$rounds);             # backup $rounds
+       &sub    ($len,6);
+       &jmp    (&label("ctr32_loop6"));
+
+&set_label("ctr32_loop6",16);
+       &pshufd ($inout2,$rndkey1,1<<6);
+       &movdqa ($rndkey1,&QWP(32,"esp"));      # pull counter-less ivec
+       &pshufd ($inout3,$rndkey0,3<<6);
+       &por    ($inout0,$rndkey1);             # merge counter-less ivec
+       &pshufd ($inout4,$rndkey0,2<<6);
+       &por    ($inout1,$rndkey1);
+       &pshufd ($inout5,$rndkey0,1<<6);
+       &por    ($inout2,$rndkey1);
+       &por    ($inout3,$rndkey1);
+       &por    ($inout4,$rndkey1);
+       &por    ($inout5,$rndkey1);
+
+       # inlining _aesni_encrypt6's prologue gives ~4% improvement...
+       &$movekey       ($rndkey0,&QWP(0,$key_));
+       &$movekey       ($rndkey1,&QWP(16,$key_));
+       &lea            ($key,&DWP(32,$key_));
+       &dec            ($rounds);

        &pxor           ($inout0,$rndkey0);
        &pxor           ($inout1,$rndkey0);
        &pxor           ($inout0,$rndkey0);
        &pxor           ($inout1,$rndkey0);

-       &pxor           ($inout2,$rndkey0);
-       &$movekey       ($rndkey0,&QWP(0,$key));
-
-&set_label("ctr32_enc_loop3");

        &aesenc         ($inout0,$rndkey1);
        &aesenc         ($inout0,$rndkey1);

+       &pxor           ($inout2,$rndkey0);

        &aesenc         ($inout1,$rndkey1);
        &aesenc         ($inout1,$rndkey1);

-       &dec            ($rounds);
+       &pxor           ($inout3,$rndkey0);

        &aesenc         ($inout2,$rndkey1);
        &aesenc         ($inout2,$rndkey1);

-       &$movekey       ($rndkey1,&QWP(16,$key));
-       &aesenc         ($inout0,$rndkey0);
-       &aesenc         ($inout1,$rndkey0);
-       &lea            ($key,&DWP(32,$key));
-       &aesenc         ($inout2,$rndkey0);
+       &pxor           ($inout4,$rndkey0);
+       &aesenc         ($inout3,$rndkey1);
+       &pxor           ($inout5,$rndkey0);
+       &aesenc         ($inout4,$rndkey1);

        &$movekey       ($rndkey0,&QWP(0,$key));
        &$movekey       ($rndkey0,&QWP(0,$key));

-       &jnz            (&label("ctr32_enc_loop3"));
+       &aesenc         ($inout5,$rndkey1);

 
 

-       &aesenc         ($inout0,$rndkey1);
-       &aesenc         ($inout1,$rndkey1);
-       &aesenc         ($inout2,$rndkey1);
-       &movdqa         ($rndkey1,&QWP(0,"esp"));       # load byte-swap mask
+       &call           (&label("_aesni_encrypt6_enter"));
+
+       &movups ($rndkey1,&QWP(0,$inp));
+       &movups ($rndkey0,&QWP(0x10,$inp));
+       &xorps  ($inout0,$rndkey1);
+       &movups ($rndkey1,&QWP(0x20,$inp));
+       &xorps  ($inout1,$rndkey0);
+       &movups (&QWP(0,$out),$inout0);
+       &movdqa ($rndkey0,&QWP(16,"esp"));      # load increment
+       &xorps  ($inout2,$rndkey1);
+       &movdqa ($rndkey1,&QWP(48,"esp"));      # load 1st triplet
+       &movups (&QWP(0x10,$out),$inout1);
+       &movups (&QWP(0x20,$out),$inout2);
+
+       &paddd  ($rndkey1,$rndkey0);            # 1st triplet increment
+       &paddd  ($rndkey0,&QWP(64,"esp"));      # 2nd triplet increment
+       &movdqa ($inout0,&QWP(0,"esp"));        # load byte swap mask
+
+       &movups ($inout1,&QWP(0x30,$inp));
+       &movups ($inout2,&QWP(0x40,$inp));
+       &xorps  ($inout3,$inout1);
+       &movups ($inout1,&QWP(0x50,$inp));
+       &lea    ($inp,&DWP(0x60,$inp));
+       &movdqa (&QWP(48,"esp"),$rndkey1);      # save 1st triplet
+       &pshufb ($rndkey1,$inout0);             # byte swap
+       &xorps  ($inout4,$inout2);
+       &movups (&QWP(0x30,$out),$inout3);
+       &xorps  ($inout5,$inout1);
+       &movdqa (&QWP(64,"esp"),$rndkey0);      # save 2nd triplet
+       &pshufb ($rndkey0,$inout0);             # byte swap
+       &movups (&QWP(0x40,$out),$inout4);
+       &pshufd ($inout0,$rndkey1,3<<6);
+       &movups (&QWP(0x50,$out),$inout5);
+       &lea    ($out,&DWP(0x60,$out));

 
 

-       &aesenclast     ($inout0,$rndkey0);
-       &pshufb         ($ivec,$rndkey1);               # byte swap
-       &movdqu         ($in0,&QWP(0,$inp));
-       &aesenclast     ($inout1,$rndkey0);
-       &paddd          ($ivec,&QWP(16,"esp"));         # counter increment
-       &movdqu         ($in1,&QWP(0x10,$inp));
-       &aesenclast     ($inout2,$rndkey0);
-       &pshufb         ($ivec,$rndkey1);               # byte swap
-       &movdqu         ($rndkey0,&QWP(0x20,$inp));
-       &lea            ($inp,&DWP(0x30,$inp));
-
-       &pxor   ($in0,$inout0);
-       &mov    ($key,$key_);
-       &pxor   ($in1,$inout1);
-       &movdqu (&QWP(0,$out),$in0);
-       &pxor   ($rndkey0,$inout2);
-       &movdqu (&QWP(0x10,$out),$in1);
-       &movdqu (&QWP(0x20,$out),$rndkey0);
-       &movdqa ($inout3,&QWP(32,"esp"));       # load counter-less ivec
-
-       &sub    ($len,3);
-       &lea    ($out,&DWP(0x30,$out));

        &mov    ($rounds,$rounds_);
        &mov    ($rounds,$rounds_);

-       &ja     (&label("ctr32_loop3"));
+       &pshufd ($inout1,$rndkey1,2<<6);
+       &sub    ($len,6);
+       &jnc    (&label("ctr32_loop6"));

 
 

-       &pextrd ($rounds_,$ivec,1);             # might need last counter value
-       &add    ($len,4);
-       &bswap  ($rounds_);
+       &add    ($len,6);
+       &jz     (&label("ctr32_ret"));
+       &mov    ($key,$key_);
+       &lea    ($rounds,&DWP(1,"",$rounds,2)); # restore $rounds
+       &movdqa ($inout5,&QWP(32,"esp"));       # pull count-less ivec

 
 &set_label("ctr32_tail");
 
 &set_label("ctr32_tail");

-       &pshufd ($inout0,$ivec,3<<6);
-       &pshufd ($inout1,$ivec,2<<6);
-       &por    ($inout0,$inout3);
+       &por    ($inout0,$inout5);

        &cmp    ($len,2);
        &jb     (&label("ctr32_one"));
        &cmp    ($len,2);
        &jb     (&label("ctr32_one"));

-       &lea    ($rounds_,&DWP(1,$rounds_));
-       &pshufd ($inout2,$ivec,1<<6);
-       &por    ($inout1,$inout3);
-       &je     (&label("ctr32_two"));
-       &bswap  ($rounds_);
-       &por    ($inout2,$inout3);
-       &cmp    ($len,3);
-       &je     (&label("ctr32_three"));
-
-       &pinsrd ($inout3,$rounds_,3);           # compose last counter value

 
 

-       &call   ("_aesni_encrypt4");
+       &pshufd ($inout2,$rndkey1,1<<6);
+       &por    ($inout1,$inout5);
+       &je     (&label("ctr32_two"));

 
 

-       &movdqu ($in0,&QWP(0,$inp));
-       &movdqu ($rndkey1,&QWP(0x10,$inp));
-       &pxor   ($in0,$inout0);
-       &movdqu ($rndkey0,&QWP(0x20,$inp));
-       &pxor   ($rndkey1,$inout1);
-       &movdqu ($ivec,&QWP(0x30,$inp));
-       &pxor   ($rndkey0,$inout2);
-       &movdqu (&QWP(0,$out),$in0);
-       &pxor   ($ivec,$inout3);
-       &movdqu (&QWP(0x10,$out),$rndkey1);
-       &movdqu (&QWP(0x20,$out),$rndkey0);
-       &movdqu (&QWP(0x30,$out),$ivec);
+       &pshufd ($inout3,$rndkey0,3<<6);
+       &por    ($inout2,$inout5);
+       &cmp    ($len,4);
+       &jb     (&label("ctr32_three"));
+
+       &pshufd ($inout4,$rndkey0,2<<6);
+       &por    ($inout3,$inout5);
+       &je     (&label("ctr32_four"));
+
+       &por    ($inout4,$inout5);
+       &call   ("_aesni_encrypt6");
+       &movups ($rndkey1,&QWP(0,$inp));
+       &movups ($rndkey0,&QWP(0x10,$inp));
+       &xorps  ($inout0,$rndkey1);
+       &movups ($rndkey1,&QWP(0x20,$inp));
+       &xorps  ($inout1,$rndkey0);
+       &movups ($rndkey0,&QWP(0x30,$inp));
+       &xorps  ($inout2,$rndkey1);
+       &movups ($rndkey1,&QWP(0x40,$inp));
+       &xorps  ($inout3,$rndkey0);
+       &movups (&QWP(0,$out),$inout0);
+       &xorps  ($inout4,$rndkey1);
+       &movups (&QWP(0x10,$out),$inout1);
+       &movups (&QWP(0x20,$out),$inout2);
+       &movups (&QWP(0x30,$out),$inout3);
+       &movups (&QWP(0x40,$out),$inout4);

        &jmp    (&label("ctr32_ret"));
 
 &set_label("ctr32_one_shortcut",16);
        &jmp    (&label("ctr32_ret"));
 
 &set_label("ctr32_one_shortcut",16);

-       &movdqu ($inout0,&QWP(0,$rounds_));     # load ivec
+       &movups ($inout0,&QWP(0,$rounds_));     # load ivec

        &mov    ($rounds,&DWP(240,$key));
        
 &set_label("ctr32_one");
        &mov    ($rounds,&DWP(240,$key));
        
 &set_label("ctr32_one");


@@ -725,37 +962,757 @@ if ($PREFIX eq "aesni") {
        {   &aesni_inline_generate1("enc");     }
        else
        {   &call       ("_aesni_encrypt1");    }
        {   &aesni_inline_generate1("enc");     }
        else
        {   &call       ("_aesni_encrypt1");    }

-       &movdqu ($in0,&QWP(0,$inp));
-       &pxor   ($in0,$inout0);
-       &movdqu (&QWP(0,$out),$in0);
+       &movups ($in0,&QWP(0,$inp));
+       &xorps  ($in0,$inout0);
+       &movups (&QWP(0,$out),$in0);

        &jmp    (&label("ctr32_ret"));
 
 &set_label("ctr32_two",16);
        &jmp    (&label("ctr32_ret"));
 
 &set_label("ctr32_two",16);

-       &pxor   ($inout2,$inout2);

        &call   ("_aesni_encrypt3");
        &call   ("_aesni_encrypt3");

-       &movdqu ($in0,&QWP(0,$inp));
-       &movdqu ($in1,&QWP(0x10,$inp));
-       &pxor   ($in0,$inout0);
-       &pxor   ($in1,$inout1);
-       &movdqu (&QWP(0,$out),$in0);
-       &movdqu (&QWP(0x10,$out),$in1);
+       &movups ($inout3,&QWP(0,$inp));
+       &movups ($inout4,&QWP(0x10,$inp));
+       &xorps  ($inout0,$inout3);
+       &xorps  ($inout1,$inout4);
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);

        &jmp    (&label("ctr32_ret"));
 
 &set_label("ctr32_three",16);
        &call   ("_aesni_encrypt3");
        &jmp    (&label("ctr32_ret"));
 
 &set_label("ctr32_three",16);
        &call   ("_aesni_encrypt3");

-       &movdqu ($in0,&QWP(0,$inp));
-       &movdqu ($in1,&QWP(0x10,$inp));
-       &movdqu ($rndkey1,&QWP(0x20,$inp));
-       &pxor   ($in0,$inout0);
-       &pxor   ($in1,$inout1);
-       &movdqu (&QWP(0,$out),$in0);
-       &pxor   ($rndkey1,$inout2);
-       &movdqu (&QWP(0x10,$out),$in1);
-       &movdqu (&QWP(0x20,$out),$rndkey1);
+       &movups ($inout3,&QWP(0,$inp));
+       &movups ($inout4,&QWP(0x10,$inp));
+       &xorps  ($inout0,$inout3);
+       &movups ($inout5,&QWP(0x20,$inp));
+       &xorps  ($inout1,$inout4);
+       &movups (&QWP(0,$out),$inout0);
+       &xorps  ($inout2,$inout5);
+       &movups (&QWP(0x10,$out),$inout1);
+       &movups (&QWP(0x20,$out),$inout2);
+       &jmp    (&label("ctr32_ret"));
+
+&set_label("ctr32_four",16);
+       &call   ("_aesni_encrypt4");
+       &movups ($inout4,&QWP(0,$inp));
+       &movups ($inout5,&QWP(0x10,$inp));
+       &movups ($rndkey1,&QWP(0x20,$inp));
+       &xorps  ($inout0,$inout4);
+       &movups ($rndkey0,&QWP(0x30,$inp));
+       &xorps  ($inout1,$inout5);
+       &movups (&QWP(0,$out),$inout0);
+       &xorps  ($inout2,$rndkey1);
+       &movups (&QWP(0x10,$out),$inout1);
+       &xorps  ($inout3,$rndkey0);
+       &movups (&QWP(0x20,$out),$inout2);
+       &movups (&QWP(0x30,$out),$inout3);

 
 &set_label("ctr32_ret");
 
 &set_label("ctr32_ret");

-       &mov    ("esp",&DWP(48,"esp"));
+       &mov    ("esp",&DWP(80,"esp"));

 &function_end("aesni_ctr32_encrypt_blocks");
 &function_end("aesni_ctr32_encrypt_blocks");

+\f
+######################################################################
+# void aesni_xts_[en|de]crypt(const char *inp,char *out,size_t len,
+#      const AES_KEY *key1, const AES_KEY *key2
+#      const unsigned char iv[16]);
+#
+{ my ($tweak,$twtmp,$twres,$twmask)=($rndkey1,$rndkey0,$inout0,$inout1);
+
+&function_begin("aesni_xts_encrypt");
+       &mov    ($key,&wparam(4));              # key2
+       &mov    ($inp,&wparam(5));              # clear-text tweak
+
+       &mov    ($rounds,&DWP(240,$key));       # key2->rounds
+       &movups ($inout0,&QWP(0,$inp));
+       if ($inline)
+       {   &aesni_inline_generate1("enc");     }
+       else
+       {   &call       ("_aesni_encrypt1");    }
+
+       &mov    ($inp,&wparam(0));
+       &mov    ($out,&wparam(1));
+       &mov    ($len,&wparam(2));
+       &mov    ($key,&wparam(3));              # key1
+
+       &mov    ($key_,"esp");
+       &sub    ("esp",16*7+8);
+       &mov    ($rounds,&DWP(240,$key));       # key1->rounds
+       &and    ("esp",-16);                    # align stack
+
+       &mov    (&DWP(16*6+0,"esp"),0x87);      # compose the magic constant
+       &mov    (&DWP(16*6+4,"esp"),0);
+       &mov    (&DWP(16*6+8,"esp"),1);
+       &mov    (&DWP(16*6+12,"esp"),0);
+       &mov    (&DWP(16*7+0,"esp"),$len);      # save original $len
+       &mov    (&DWP(16*7+4,"esp"),$key_);     # save original %esp
+
+       &movdqa ($tweak,$inout0);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($twmask,&QWP(6*16,"esp"));     # 0x0...010...87
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+
+       &and    ($len,-16);
+       &mov    ($key_,$key);                   # backup $key
+       &mov    ($rounds_,$rounds);             # backup $rounds
+       &sub    ($len,16*6);
+       &jc     (&label("xts_enc_short"));
+
+       &shr    ($rounds,1);
+       &mov    ($rounds_,$rounds);
+       &jmp    (&label("xts_enc_loop6"));
+
+&set_label("xts_enc_loop6",16);
+       for ($i=0;$i<4;$i++) {
+           &pshufd     ($twres,$twtmp,0x13);
+           &pxor       ($twtmp,$twtmp);
+           &movdqa     (&QWP(16*$i,"esp"),$tweak);
+           &paddq      ($tweak,$tweak);        # &psllq($tweak,1);
+           &pand       ($twres,$twmask);       # isolate carry and residue
+           &pcmpgtd    ($twtmp,$tweak);        # broadcast upper bits
+           &pxor       ($tweak,$twres);
+       }
+       &pshufd ($inout5,$twtmp,0x13);
+       &movdqa (&QWP(16*$i++,"esp"),$tweak);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+        &$movekey      ($rndkey0,&QWP(0,$key_));
+       &pand   ($inout5,$twmask);              # isolate carry and residue
+        &movups        ($inout0,&QWP(0,$inp)); # load input
+       &pxor   ($inout5,$tweak);
+
+       # inline _aesni_encrypt6 prologue and flip xor with tweak and key[0]
+       &movdqu ($inout1,&QWP(16*1,$inp));
+        &xorps         ($inout0,$rndkey0);     # input^=rndkey[0]
+       &movdqu ($inout2,&QWP(16*2,$inp));
+        &pxor          ($inout1,$rndkey0);
+       &movdqu ($inout3,&QWP(16*3,$inp));
+        &pxor          ($inout2,$rndkey0);
+       &movdqu ($inout4,&QWP(16*4,$inp));
+        &pxor          ($inout3,$rndkey0);
+       &movdqu ($rndkey1,&QWP(16*5,$inp));
+        &pxor          ($inout4,$rndkey0);
+       &lea    ($inp,&DWP(16*6,$inp));
+       &pxor   ($inout0,&QWP(16*0,"esp"));     # input^=tweak
+       &movdqa (&QWP(16*$i,"esp"),$inout5);    # save last tweak
+       &pxor   ($inout5,$rndkey1);
+
+        &$movekey      ($rndkey1,&QWP(16,$key_));
+        &lea           ($key,&DWP(32,$key_));
+       &pxor   ($inout1,&QWP(16*1,"esp"));
+        &aesenc        ($inout0,$rndkey1);
+       &pxor   ($inout2,&QWP(16*2,"esp"));
+        &aesenc        ($inout1,$rndkey1);
+       &pxor   ($inout3,&QWP(16*3,"esp"));
+        &dec           ($rounds);
+        &aesenc        ($inout2,$rndkey1);
+       &pxor   ($inout4,&QWP(16*4,"esp"));
+        &aesenc        ($inout3,$rndkey1);
+       &pxor           ($inout5,$rndkey0);
+        &aesenc        ($inout4,$rndkey1);
+        &$movekey      ($rndkey0,&QWP(0,$key));
+        &aesenc        ($inout5,$rndkey1);
+       &call           (&label("_aesni_encrypt6_enter"));
+
+       &movdqa ($tweak,&QWP(16*5,"esp"));      # last tweak
+       &pxor   ($twtmp,$twtmp);
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # output^=tweak
+       &pcmpgtd        ($twtmp,$tweak);                # broadcast upper bits
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &xorps  ($inout2,&QWP(16*2,"esp"));
+       &movups (&QWP(16*1,$out),$inout1);
+       &xorps  ($inout3,&QWP(16*3,"esp"));
+       &movups (&QWP(16*2,$out),$inout2);
+       &xorps  ($inout4,&QWP(16*4,"esp"));
+       &movups (&QWP(16*3,$out),$inout3);
+       &xorps  ($inout5,$tweak);
+       &movups (&QWP(16*4,$out),$inout4);
+       &pshufd ($twres,$twtmp,0x13);
+       &movups (&QWP(16*5,$out),$inout5);
+       &lea    ($out,&DWP(16*6,$out));
+       &movdqa ($twmask,&QWP(16*6,"esp"));     # 0x0...010...87
+
+       &pxor   ($twtmp,$twtmp);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &mov    ($rounds,$rounds_);             # restore $rounds
+       &pxor   ($tweak,$twres);
+
+       &sub    ($len,16*6);
+       &jnc    (&label("xts_enc_loop6"));
+
+       &lea    ($rounds,&DWP(1,"",$rounds,2)); # restore $rounds
+       &mov    ($key,$key_);                   # restore $key
+       &mov    ($rounds_,$rounds);
+
+&set_label("xts_enc_short");
+       &add    ($len,16*6);
+       &jz     (&label("xts_enc_done6x"));
+
+       &movdqa ($inout3,$tweak);               # put aside previous tweak
+       &cmp    ($len,0x20);
+       &jb     (&label("xts_enc_one"));
+
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+       &je     (&label("xts_enc_two"));
+
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($inout4,$tweak);               # put aside previous tweak
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+       &cmp    ($len,0x40);
+       &jb     (&label("xts_enc_three"));
+
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($inout5,$tweak);               # put aside previous tweak
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+       &movdqa (&QWP(16*0,"esp"),$inout3);
+       &movdqa (&QWP(16*1,"esp"),$inout4);
+       &je     (&label("xts_enc_four"));
+
+       &movdqa (&QWP(16*2,"esp"),$inout5);
+       &pshufd ($inout5,$twtmp,0x13);
+       &movdqa (&QWP(16*3,"esp"),$tweak);
+       &paddq  ($tweak,$tweak);                # &psllq($inout0,1);
+       &pand   ($inout5,$twmask);              # isolate carry and residue
+       &pxor   ($inout5,$tweak);
+
+       &movdqu ($inout0,&QWP(16*0,$inp));      # load input
+       &movdqu ($inout1,&QWP(16*1,$inp));
+       &movdqu ($inout2,&QWP(16*2,$inp));
+       &pxor   ($inout0,&QWP(16*0,"esp"));     # input^=tweak
+       &movdqu ($inout3,&QWP(16*3,$inp));
+       &pxor   ($inout1,&QWP(16*1,"esp"));
+       &movdqu ($inout4,&QWP(16*4,$inp));
+       &pxor   ($inout2,&QWP(16*2,"esp"));
+       &lea    ($inp,&DWP(16*5,$inp));
+       &pxor   ($inout3,&QWP(16*3,"esp"));
+       &movdqa (&QWP(16*4,"esp"),$inout5);     # save last tweak
+       &pxor   ($inout4,$inout5);
+
+       &call   ("_aesni_encrypt6");
+
+       &movaps ($tweak,&QWP(16*4,"esp"));      # last tweak
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # output^=tweak
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &xorps  ($inout2,&QWP(16*2,"esp"));
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &xorps  ($inout3,&QWP(16*3,"esp"));
+       &movups (&QWP(16*1,$out),$inout1);
+       &xorps  ($inout4,$tweak);
+       &movups (&QWP(16*2,$out),$inout2);
+       &movups (&QWP(16*3,$out),$inout3);
+       &movups (&QWP(16*4,$out),$inout4);
+       &lea    ($out,&DWP(16*5,$out));
+       &jmp    (&label("xts_enc_done"));
+
+&set_label("xts_enc_one",16);
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &lea    ($inp,&DWP(16*1,$inp));
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       if ($inline)
+       {   &aesni_inline_generate1("enc");     }
+       else
+       {   &call       ("_aesni_encrypt1");    }
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &lea    ($out,&DWP(16*1,$out));
+
+       &movdqa ($tweak,$inout3);               # last tweak
+       &jmp    (&label("xts_enc_done"));
+
+&set_label("xts_enc_two",16);
+       &movaps ($inout4,$tweak);               # put aside last tweak
+
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &movups ($inout1,&QWP(16*1,$inp));
+       &lea    ($inp,&DWP(16*2,$inp));
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       &xorps  ($inout1,$inout4);
+       &xorps  ($inout2,$inout2);
+
+       &call   ("_aesni_encrypt3");
+
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &xorps  ($inout1,$inout4);
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &movups (&QWP(16*1,$out),$inout1);
+       &lea    ($out,&DWP(16*2,$out));
+
+       &movdqa ($tweak,$inout4);               # last tweak
+       &jmp    (&label("xts_enc_done"));
+
+&set_label("xts_enc_three",16);
+       &movaps ($inout5,$tweak);               # put aside last tweak
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &movups ($inout1,&QWP(16*1,$inp));
+       &movups ($inout2,&QWP(16*2,$inp));
+       &lea    ($inp,&DWP(16*3,$inp));
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       &xorps  ($inout1,$inout4);
+       &xorps  ($inout2,$inout5);
+
+       &call   ("_aesni_encrypt3");
+
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &xorps  ($inout1,$inout4);
+       &xorps  ($inout2,$inout5);
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &movups (&QWP(16*1,$out),$inout1);
+       &movups (&QWP(16*2,$out),$inout2);
+       &lea    ($out,&DWP(16*3,$out));
+
+       &movdqa ($tweak,$inout5);               # last tweak
+       &jmp    (&label("xts_enc_done"));
+
+&set_label("xts_enc_four",16);
+       &movaps ($inout4,$tweak);               # put aside last tweak
+
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &movups ($inout1,&QWP(16*1,$inp));
+       &movups ($inout2,&QWP(16*2,$inp));
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # input^=tweak
+       &movups ($inout3,&QWP(16*3,$inp));
+       &lea    ($inp,&DWP(16*4,$inp));
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &xorps  ($inout2,$inout5);
+       &xorps  ($inout3,$inout4);
+
+       &call   ("_aesni_encrypt4");
+
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # output^=tweak
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &xorps  ($inout2,$inout5);
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &xorps  ($inout3,$inout4);
+       &movups (&QWP(16*1,$out),$inout1);
+       &movups (&QWP(16*2,$out),$inout2);
+       &movups (&QWP(16*3,$out),$inout3);
+       &lea    ($out,&DWP(16*4,$out));
+
+       &movdqa ($tweak,$inout4);               # last tweak
+       &jmp    (&label("xts_enc_done"));
+
+&set_label("xts_enc_done6x",16);               # $tweak is pre-calculated
+       &mov    ($len,&DWP(16*7+0,"esp"));      # restore original $len
+       &and    ($len,15);
+       &jz     (&label("xts_enc_ret"));
+       &movdqa ($inout3,$tweak);
+       &mov    (&DWP(16*7+0,"esp"),$len);      # save $len%16
+       &jmp    (&label("xts_enc_steal"));
+
+&set_label("xts_enc_done",16);
+       &mov    ($len,&DWP(16*7+0,"esp"));      # restore original $len
+       &pxor   ($twtmp,$twtmp);
+       &and    ($len,15);
+       &jz     (&label("xts_enc_ret"));
+
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &mov    (&DWP(16*7+0,"esp"),$len);      # save $len%16
+       &pshufd ($inout3,$twtmp,0x13);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($inout3,&QWP(16*6,"esp"));     # isolate carry and residue
+       &pxor   ($inout3,$tweak);
+
+&set_label("xts_enc_steal");
+       &movz   ($rounds,&BP(0,$inp));
+       &movz   ($key,&BP(-16,$out));
+       &lea    ($inp,&DWP(1,$inp));
+       &mov    (&BP(-16,$out),&LB($rounds));
+       &mov    (&BP(0,$out),&LB($key));
+       &lea    ($out,&DWP(1,$out));
+       &sub    ($len,1);
+       &jnz    (&label("xts_enc_steal"));
+
+       &sub    ($out,&DWP(16*7+0,"esp"));      # rewind $out
+       &mov    ($key,$key_);                   # restore $key
+       &mov    ($rounds,$rounds_);             # restore $rounds
+
+       &movups ($inout0,&QWP(-16,$out));       # load input
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       if ($inline)
+       {   &aesni_inline_generate1("enc");     }
+       else
+       {   &call       ("_aesni_encrypt1");    }
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &movups (&QWP(-16,$out),$inout0);       # write output
+
+&set_label("xts_enc_ret");
+       &mov    ("esp",&DWP(16*7+4,"esp"));     # restore %esp
+&function_end("aesni_xts_encrypt");
+
+&function_begin("aesni_xts_decrypt");
+       &mov    ($key,&wparam(4));              # key2
+       &mov    ($inp,&wparam(5));              # clear-text tweak
+
+       &mov    ($rounds,&DWP(240,$key));       # key2->rounds
+       &movups ($inout0,&QWP(0,$inp));
+       if ($inline)
+       {   &aesni_inline_generate1("enc");     }
+       else
+       {   &call       ("_aesni_encrypt1");    }
+
+       &mov    ($inp,&wparam(0));
+       &mov    ($out,&wparam(1));
+       &mov    ($len,&wparam(2));
+       &mov    ($key,&wparam(3));              # key1
+
+       &mov    ($key_,"esp");
+       &sub    ("esp",16*7+8);
+       &and    ("esp",-16);                    # align stack
+
+       &xor    ($rounds_,$rounds_);            # if(len%16) len-=16;
+       &test   ($len,15);
+       &setnz  (&LB($rounds_));
+       &shl    ($rounds_,4);
+       &sub    ($len,$rounds_);
+
+       &mov    (&DWP(16*6+0,"esp"),0x87);      # compose the magic constant
+       &mov    (&DWP(16*6+4,"esp"),0);
+       &mov    (&DWP(16*6+8,"esp"),1);
+       &mov    (&DWP(16*6+12,"esp"),0);
+       &mov    (&DWP(16*7+0,"esp"),$len);      # save original $len
+       &mov    (&DWP(16*7+4,"esp"),$key_);     # save original %esp
+
+       &mov    ($rounds,&DWP(240,$key));       # key1->rounds
+       &mov    ($key_,$key);                   # backup $key
+       &mov    ($rounds_,$rounds);             # backup $rounds
+
+       &movdqa ($tweak,$inout0);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($twmask,&QWP(6*16,"esp"));     # 0x0...010...87
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+
+       &and    ($len,-16);
+       &sub    ($len,16*6);
+       &jc     (&label("xts_dec_short"));
+
+       &shr    ($rounds,1);
+       &mov    ($rounds_,$rounds);
+       &jmp    (&label("xts_dec_loop6"));
+
+&set_label("xts_dec_loop6",16);
+       for ($i=0;$i<4;$i++) {
+           &pshufd     ($twres,$twtmp,0x13);
+           &pxor       ($twtmp,$twtmp);
+           &movdqa     (&QWP(16*$i,"esp"),$tweak);
+           &paddq      ($tweak,$tweak);        # &psllq($tweak,1);
+           &pand       ($twres,$twmask);       # isolate carry and residue
+           &pcmpgtd    ($twtmp,$tweak);        # broadcast upper bits
+           &pxor       ($tweak,$twres);
+       }
+       &pshufd ($inout5,$twtmp,0x13);
+       &movdqa (&QWP(16*$i++,"esp"),$tweak);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+        &$movekey      ($rndkey0,&QWP(0,$key_));
+       &pand   ($inout5,$twmask);              # isolate carry and residue
+        &movups        ($inout0,&QWP(0,$inp)); # load input
+       &pxor   ($inout5,$tweak);
+
+       # inline _aesni_encrypt6 prologue and flip xor with tweak and key[0]
+       &movdqu ($inout1,&QWP(16*1,$inp));
+        &xorps         ($inout0,$rndkey0);     # input^=rndkey[0]
+       &movdqu ($inout2,&QWP(16*2,$inp));
+        &pxor          ($inout1,$rndkey0);
+       &movdqu ($inout3,&QWP(16*3,$inp));
+        &pxor          ($inout2,$rndkey0);
+       &movdqu ($inout4,&QWP(16*4,$inp));
+        &pxor          ($inout3,$rndkey0);
+       &movdqu ($rndkey1,&QWP(16*5,$inp));
+        &pxor          ($inout4,$rndkey0);
+       &lea    ($inp,&DWP(16*6,$inp));
+       &pxor   ($inout0,&QWP(16*0,"esp"));     # input^=tweak
+       &movdqa (&QWP(16*$i,"esp"),$inout5);    # save last tweak
+       &pxor   ($inout5,$rndkey1);
+
+        &$movekey      ($rndkey1,&QWP(16,$key_));
+        &lea           ($key,&DWP(32,$key_));
+       &pxor   ($inout1,&QWP(16*1,"esp"));
+        &aesdec        ($inout0,$rndkey1);
+       &pxor   ($inout2,&QWP(16*2,"esp"));
+        &aesdec        ($inout1,$rndkey1);
+       &pxor   ($inout3,&QWP(16*3,"esp"));
+        &dec           ($rounds);
+        &aesdec        ($inout2,$rndkey1);
+       &pxor   ($inout4,&QWP(16*4,"esp"));
+        &aesdec        ($inout3,$rndkey1);
+       &pxor           ($inout5,$rndkey0);
+        &aesdec        ($inout4,$rndkey1);
+        &$movekey      ($rndkey0,&QWP(0,$key));
+        &aesdec        ($inout5,$rndkey1);
+       &call           (&label("_aesni_decrypt6_enter"));
+
+       &movdqa ($tweak,&QWP(16*5,"esp"));      # last tweak
+       &pxor   ($twtmp,$twtmp);
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # output^=tweak
+       &pcmpgtd        ($twtmp,$tweak);                # broadcast upper bits
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &xorps  ($inout2,&QWP(16*2,"esp"));
+       &movups (&QWP(16*1,$out),$inout1);
+       &xorps  ($inout3,&QWP(16*3,"esp"));
+       &movups (&QWP(16*2,$out),$inout2);
+       &xorps  ($inout4,&QWP(16*4,"esp"));
+       &movups (&QWP(16*3,$out),$inout3);
+       &xorps  ($inout5,$tweak);
+       &movups (&QWP(16*4,$out),$inout4);
+       &pshufd ($twres,$twtmp,0x13);
+       &movups (&QWP(16*5,$out),$inout5);
+       &lea    ($out,&DWP(16*6,$out));
+       &movdqa ($twmask,&QWP(16*6,"esp"));     # 0x0...010...87
+
+       &pxor   ($twtmp,$twtmp);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &mov    ($rounds,$rounds_);             # restore $rounds
+       &pxor   ($tweak,$twres);
+
+       &sub    ($len,16*6);
+       &jnc    (&label("xts_dec_loop6"));
+
+       &lea    ($rounds,&DWP(1,"",$rounds,2)); # restore $rounds
+       &mov    ($key,$key_);                   # restore $key
+       &mov    ($rounds_,$rounds);
+
+&set_label("xts_dec_short");
+       &add    ($len,16*6);
+       &jz     (&label("xts_dec_done6x"));
+
+       &movdqa ($inout3,$tweak);               # put aside previous tweak
+       &cmp    ($len,0x20);
+       &jb     (&label("xts_dec_one"));
+
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+       &je     (&label("xts_dec_two"));
+
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($inout4,$tweak);               # put aside previous tweak
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+       &cmp    ($len,0x40);
+       &jb     (&label("xts_dec_three"));
+
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($inout5,$tweak);               # put aside previous tweak
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+       &movdqa (&QWP(16*0,"esp"),$inout3);
+       &movdqa (&QWP(16*1,"esp"),$inout4);
+       &je     (&label("xts_dec_four"));
+
+       &movdqa (&QWP(16*2,"esp"),$inout5);
+       &pshufd ($inout5,$twtmp,0x13);
+       &movdqa (&QWP(16*3,"esp"),$tweak);
+       &paddq  ($tweak,$tweak);                # &psllq($inout0,1);
+       &pand   ($inout5,$twmask);              # isolate carry and residue
+       &pxor   ($inout5,$tweak);
+
+       &movdqu ($inout0,&QWP(16*0,$inp));      # load input
+       &movdqu ($inout1,&QWP(16*1,$inp));
+       &movdqu ($inout2,&QWP(16*2,$inp));
+       &pxor   ($inout0,&QWP(16*0,"esp"));     # input^=tweak
+       &movdqu ($inout3,&QWP(16*3,$inp));
+       &pxor   ($inout1,&QWP(16*1,"esp"));
+       &movdqu ($inout4,&QWP(16*4,$inp));
+       &pxor   ($inout2,&QWP(16*2,"esp"));
+       &lea    ($inp,&DWP(16*5,$inp));
+       &pxor   ($inout3,&QWP(16*3,"esp"));
+       &movdqa (&QWP(16*4,"esp"),$inout5);     # save last tweak
+       &pxor   ($inout4,$inout5);
+
+       &call   ("_aesni_decrypt6");
+
+       &movaps ($tweak,&QWP(16*4,"esp"));      # last tweak
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # output^=tweak
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &xorps  ($inout2,&QWP(16*2,"esp"));
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &xorps  ($inout3,&QWP(16*3,"esp"));
+       &movups (&QWP(16*1,$out),$inout1);
+       &xorps  ($inout4,$tweak);
+       &movups (&QWP(16*2,$out),$inout2);
+       &movups (&QWP(16*3,$out),$inout3);
+       &movups (&QWP(16*4,$out),$inout4);
+       &lea    ($out,&DWP(16*5,$out));
+       &jmp    (&label("xts_dec_done"));
+
+&set_label("xts_dec_one",16);
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &lea    ($inp,&DWP(16*1,$inp));
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       if ($inline)
+       {   &aesni_inline_generate1("dec");     }
+       else
+       {   &call       ("_aesni_decrypt1");    }
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &lea    ($out,&DWP(16*1,$out));
+
+       &movdqa ($tweak,$inout3);               # last tweak
+       &jmp    (&label("xts_dec_done"));
+
+&set_label("xts_dec_two",16);
+       &movaps ($inout4,$tweak);               # put aside last tweak
+
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &movups ($inout1,&QWP(16*1,$inp));
+       &lea    ($inp,&DWP(16*2,$inp));
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       &xorps  ($inout1,$inout4);
+
+       &call   ("_aesni_decrypt3");
+
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &xorps  ($inout1,$inout4);
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &movups (&QWP(16*1,$out),$inout1);
+       &lea    ($out,&DWP(16*2,$out));
+
+       &movdqa ($tweak,$inout4);               # last tweak
+       &jmp    (&label("xts_dec_done"));
+
+&set_label("xts_dec_three",16);
+       &movaps ($inout5,$tweak);               # put aside last tweak
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &movups ($inout1,&QWP(16*1,$inp));
+       &movups ($inout2,&QWP(16*2,$inp));
+       &lea    ($inp,&DWP(16*3,$inp));
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       &xorps  ($inout1,$inout4);
+       &xorps  ($inout2,$inout5);
+
+       &call   ("_aesni_decrypt3");
+
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &xorps  ($inout1,$inout4);
+       &xorps  ($inout2,$inout5);
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &movups (&QWP(16*1,$out),$inout1);
+       &movups (&QWP(16*2,$out),$inout2);
+       &lea    ($out,&DWP(16*3,$out));
+
+       &movdqa ($tweak,$inout5);               # last tweak
+       &jmp    (&label("xts_dec_done"));
+
+&set_label("xts_dec_four",16);
+       &movaps ($inout4,$tweak);               # put aside last tweak
+
+       &movups ($inout0,&QWP(16*0,$inp));      # load input
+       &movups ($inout1,&QWP(16*1,$inp));
+       &movups ($inout2,&QWP(16*2,$inp));
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # input^=tweak
+       &movups ($inout3,&QWP(16*3,$inp));
+       &lea    ($inp,&DWP(16*4,$inp));
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &xorps  ($inout2,$inout5);
+       &xorps  ($inout3,$inout4);
+
+       &call   ("_aesni_decrypt4");
+
+       &xorps  ($inout0,&QWP(16*0,"esp"));     # output^=tweak
+       &xorps  ($inout1,&QWP(16*1,"esp"));
+       &xorps  ($inout2,$inout5);
+       &movups (&QWP(16*0,$out),$inout0);      # write output
+       &xorps  ($inout3,$inout4);
+       &movups (&QWP(16*1,$out),$inout1);
+       &movups (&QWP(16*2,$out),$inout2);
+       &movups (&QWP(16*3,$out),$inout3);
+       &lea    ($out,&DWP(16*4,$out));
+
+       &movdqa ($tweak,$inout4);               # last tweak
+       &jmp    (&label("xts_dec_done"));
+
+&set_label("xts_dec_done6x",16);               # $tweak is pre-calculated
+       &mov    ($len,&DWP(16*7+0,"esp"));      # restore original $len
+       &and    ($len,15);
+       &jz     (&label("xts_dec_ret"));
+       &mov    (&DWP(16*7+0,"esp"),$len);      # save $len%16
+       &jmp    (&label("xts_dec_only_one_more"));
+
+&set_label("xts_dec_done",16);
+       &mov    ($len,&DWP(16*7+0,"esp"));      # restore original $len
+       &pxor   ($twtmp,$twtmp);
+       &and    ($len,15);
+       &jz     (&label("xts_dec_ret"));
+
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &mov    (&DWP(16*7+0,"esp"),$len);      # save $len%16
+       &pshufd ($twres,$twtmp,0x13);
+       &pxor   ($twtmp,$twtmp);
+       &movdqa ($twmask,&QWP(16*6,"esp"));
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($twres,$twmask);               # isolate carry and residue
+       &pcmpgtd($twtmp,$tweak);                # broadcast upper bits
+       &pxor   ($tweak,$twres);
+
+&set_label("xts_dec_only_one_more");
+       &pshufd ($inout3,$twtmp,0x13);
+       &movdqa ($inout4,$tweak);               # put aside previous tweak
+       &paddq  ($tweak,$tweak);                # &psllq($tweak,1);
+       &pand   ($inout3,$twmask);              # isolate carry and residue
+       &pxor   ($inout3,$tweak);
+
+       &mov    ($key,$key_);                   # restore $key
+       &mov    ($rounds,$rounds_);             # restore $rounds
+
+       &movups ($inout0,&QWP(0,$inp));         # load input
+       &xorps  ($inout0,$inout3);              # input^=tweak
+       if ($inline)
+       {   &aesni_inline_generate1("dec");     }
+       else
+       {   &call       ("_aesni_decrypt1");    }
+       &xorps  ($inout0,$inout3);              # output^=tweak
+       &movups (&QWP(0,$out),$inout0);         # write output
+
+&set_label("xts_dec_steal");
+       &movz   ($rounds,&BP(16,$inp));
+       &movz   ($key,&BP(0,$out));
+       &lea    ($inp,&DWP(1,$inp));
+       &mov    (&BP(0,$out),&LB($rounds));
+       &mov    (&BP(16,$out),&LB($key));
+       &lea    ($out,&DWP(1,$out));
+       &sub    ($len,1);
+       &jnz    (&label("xts_dec_steal"));
+
+       &sub    ($out,&DWP(16*7+0,"esp"));      # rewind $out
+       &mov    ($key,$key_);                   # restore $key
+       &mov    ($rounds,$rounds_);             # restore $rounds
+
+       &movups ($inout0,&QWP(0,$out));         # load input
+       &xorps  ($inout0,$inout4);              # input^=tweak
+       if ($inline)
+       {   &aesni_inline_generate1("dec");     }
+       else
+       {   &call       ("_aesni_decrypt1");    }
+       &xorps  ($inout0,$inout4);              # output^=tweak
+       &movups (&QWP(0,$out),$inout0);         # write output
+
+&set_label("xts_dec_ret");
+       &mov    ("esp",&DWP(16*7+4,"esp"));     # restore %esp
+&function_end("aesni_xts_decrypt");
+}

 }
 \f
 ######################################################################
 }
 \f
 ######################################################################


@@ -764,34 +1721,38 @@ if ($PREFIX eq "aesni") {
 #                           unsigned char *ivp,const int enc);
 &function_begin("${PREFIX}_cbc_encrypt");
        &mov    ($inp,&wparam(0));
 #                           unsigned char *ivp,const int enc);
 &function_begin("${PREFIX}_cbc_encrypt");
        &mov    ($inp,&wparam(0));

+       &mov    ($rounds_,"esp");

        &mov    ($out,&wparam(1));
        &mov    ($out,&wparam(1));

+       &sub    ($rounds_,24);

        &mov    ($len,&wparam(2));
        &mov    ($len,&wparam(2));

+       &and    ($rounds_,-16);

        &mov    ($key,&wparam(3));
        &mov    ($key_,&wparam(4));
        &test   ($len,$len);
        &mov    ($key,&wparam(3));
        &mov    ($key_,&wparam(4));
        &test   ($len,$len);

-       &jz     (&label("cbc_ret"));
+       &jz     (&label("cbc_abort"));

 
        &cmp    (&wparam(5),0);
 
        &cmp    (&wparam(5),0);

-       &movdqu ($ivec,&QWP(0,$key_));  # load IV
+       &xchg   ($rounds_,"esp");               # alloca
+       &movups ($ivec,&QWP(0,$key_));          # load IV

        &mov    ($rounds,&DWP(240,$key));
        &mov    ($rounds,&DWP(240,$key));

-       &mov    ($key_,$key);           # backup $key
-       &mov    ($rounds_,$rounds);     # backup $rounds
+       &mov    ($key_,$key);                   # backup $key
+       &mov    (&DWP(16,"esp"),$rounds_);      # save original %esp
+       &mov    ($rounds_,$rounds);             # backup $rounds

        &je     (&label("cbc_decrypt"));
 
        &je     (&label("cbc_decrypt"));
 

-       &movdqa ($inout0,$ivec);
+       &movaps ($inout0,$ivec);

        &cmp    ($len,16);
        &jb     (&label("cbc_enc_tail"));
        &sub    ($len,16);
        &jmp    (&label("cbc_enc_loop"));
 
 &set_label("cbc_enc_loop",16);
        &cmp    ($len,16);
        &jb     (&label("cbc_enc_tail"));
        &sub    ($len,16);
        &jmp    (&label("cbc_enc_loop"));
 
 &set_label("cbc_enc_loop",16);

-       &movdqu ($ivec,&QWP(0,$inp));
+       &movups ($ivec,&QWP(0,$inp));           # input actually

        &lea    ($inp,&DWP(16,$inp));
        &lea    ($inp,&DWP(16,$inp));

-       &pxor   ($inout0,$ivec);

        if ($inline)
        if ($inline)

-       {   &aesni_inline_generate1("enc");     }
+       {   &aesni_inline_generate1("enc",$inout0,$ivec);       }

        else
        else

-       {   &call       ("_aesni_encrypt1");    }
+       {   &xorps($inout0,$ivec); &call("_aesni_encrypt1");    }

        &mov    ($rounds,$rounds_);     # restore $rounds
        &mov    ($key,$key_);           # restore $key
        &movups (&QWP(0,$out),$inout0); # store output
        &mov    ($rounds,$rounds_);     # restore $rounds
        &mov    ($key,$key_);           # restore $key
        &movups (&QWP(0,$out),$inout0); # store output


@@ -817,61 +1778,97 @@ if ($PREFIX eq "aesni") {
        &jmp    (&label("cbc_enc_loop"));
 ######################################################################
 &set_label("cbc_decrypt",16);
        &jmp    (&label("cbc_enc_loop"));
 ######################################################################
 &set_label("cbc_decrypt",16);

-       &cmp    ($len,0x40);
+       &cmp    ($len,0x50);

        &jbe    (&label("cbc_dec_tail"));
        &jbe    (&label("cbc_dec_tail"));

-       &sub    ($len,0x40);
-       &jmp    (&label("cbc_dec_loop3"));
+       &movaps (&QWP(0,"esp"),$ivec);          # save IV
+       &sub    ($len,0x50);
+       &jmp    (&label("cbc_dec_loop6_enter"));

 
 

-&set_label("cbc_dec_loop3",16);
-       &movups ($inout0,&QWP(0,$inp));
-       &movups ($inout1,&QWP(0x10,$inp));
-       &movups ($inout2,&QWP(0x20,$inp));
-       &movaps ($in0,$inout0);
-       &movaps ($in1,$inout1);
-
-       &call   ("_aesni_decrypt3");
-
-       &pxor   ($inout0,$ivec);
-       &pxor   ($inout1,$in0);
-       &movdqu ($ivec,&QWP(0x20,$inp));
-       &lea    ($inp,&DWP(0x30,$inp));
-       &pxor   ($inout2,$in1);
-       &movdqu (&QWP(0,$out),$inout0);
-       &mov    ($rounds,$rounds_)      # restore $rounds
-       &movdqu (&QWP(0x10,$out),$inout1);
-       &mov    ($key,$key_);           # restore $key
-       &movdqu (&QWP(0x20,$out),$inout2);
-       &lea    ($out,&DWP(0x30,$out));
-       &sub    ($len,0x30);
-       &ja     (&label("cbc_dec_loop3"));
-
-       &add    ($len,0x40);
+&set_label("cbc_dec_loop6",16);
+       &movaps (&QWP(0,"esp"),$rndkey0);       # save IV
+       &movups (&QWP(0,$out),$inout5);
+       &lea    ($out,&DWP(0x10,$out));
+&set_label("cbc_dec_loop6_enter");
+       &movdqu ($inout0,&QWP(0,$inp));
+       &movdqu ($inout1,&QWP(0x10,$inp));
+       &movdqu ($inout2,&QWP(0x20,$inp));
+       &movdqu ($inout3,&QWP(0x30,$inp));
+       &movdqu ($inout4,&QWP(0x40,$inp));
+       &movdqu ($inout5,&QWP(0x50,$inp));
+
+       &call   ("_aesni_decrypt6");
+
+       &movups ($rndkey1,&QWP(0,$inp));
+       &movups ($rndkey0,&QWP(0x10,$inp));
+       &xorps  ($inout0,&QWP(0,"esp"));        # ^=IV
+       &xorps  ($inout1,$rndkey1);
+       &movups ($rndkey1,&QWP(0x20,$inp));
+       &xorps  ($inout2,$rndkey0);
+       &movups ($rndkey0,&QWP(0x30,$inp));
+       &xorps  ($inout3,$rndkey1);
+       &movups ($rndkey1,&QWP(0x40,$inp));
+       &xorps  ($inout4,$rndkey0);
+       &movups ($rndkey0,&QWP(0x50,$inp));     # IV
+       &xorps  ($inout5,$rndkey1);
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);
+       &lea    ($inp,&DWP(0x60,$inp));
+       &movups (&QWP(0x20,$out),$inout2);
+       &mov    ($rounds,$rounds_)              # restore $rounds
+       &movups (&QWP(0x30,$out),$inout3);
+       &mov    ($key,$key_);                   # restore $key
+       &movups (&QWP(0x40,$out),$inout4);
+       &lea    ($out,&DWP(0x50,$out));
+       &sub    ($len,0x60);
+       &ja     (&label("cbc_dec_loop6"));
+
+       &movaps ($inout0,$inout5);
+       &movaps ($ivec,$rndkey0);
+       &add    ($len,0x50);
+       &jle    (&label("cbc_dec_tail_collected"));
+       &movups (&QWP(0,$out),$inout0);
+       &lea    ($out,&DWP(0x10,$out));

 &set_label("cbc_dec_tail");
        &movups ($inout0,&QWP(0,$inp));
        &movaps ($in0,$inout0);
        &cmp    ($len,0x10);
        &jbe    (&label("cbc_dec_one"));
 &set_label("cbc_dec_tail");
        &movups ($inout0,&QWP(0,$inp));
        &movaps ($in0,$inout0);
        &cmp    ($len,0x10);
        &jbe    (&label("cbc_dec_one"));

+

        &movups ($inout1,&QWP(0x10,$inp));
        &movaps ($in1,$inout1);
        &cmp    ($len,0x20);
        &jbe    (&label("cbc_dec_two"));
        &movups ($inout1,&QWP(0x10,$inp));
        &movaps ($in1,$inout1);
        &cmp    ($len,0x20);
        &jbe    (&label("cbc_dec_two"));

+

        &movups ($inout2,&QWP(0x20,$inp));
        &cmp    ($len,0x30);
        &jbe    (&label("cbc_dec_three"));
        &movups ($inout2,&QWP(0x20,$inp));
        &cmp    ($len,0x30);
        &jbe    (&label("cbc_dec_three"));

+

        &movups ($inout3,&QWP(0x30,$inp));
        &movups ($inout3,&QWP(0x30,$inp));

-       &call   ("_aesni_decrypt4");
-       &movdqu ($rndkey0,&QWP(0x10,$inp));
-       &movdqu ($rndkey1,&QWP(0x20,$inp));
-       &pxor   ($inout0,$ivec);
-       &pxor   ($inout1,$in0);
-       &movdqu ($ivec,&QWP(0x30,$inp));
-       &movdqu (&QWP(0,$out),$inout0);
-       &pxor   ($inout2,$rndkey0);
-       &pxor   ($inout3,$rndkey1);
-       &movdqu (&QWP(0x10,$out),$inout1);
-       &movdqu (&QWP(0x20,$out),$inout2);
-       &movdqa ($inout0,$inout3);
-       &lea    ($out,&DWP(0x30,$out));
+       &cmp    ($len,0x40);
+       &jbe    (&label("cbc_dec_four"));
+
+       &movups ($inout4,&QWP(0x40,$inp));
+       &movaps (&QWP(0,"esp"),$ivec);          # save IV
+       &movups ($inout0,&QWP(0,$inp));
+       &xorps  ($inout5,$inout5);
+       &call   ("_aesni_decrypt6");
+       &movups ($rndkey1,&QWP(0,$inp));
+       &movups ($rndkey0,&QWP(0x10,$inp));
+       &xorps  ($inout0,&QWP(0,"esp"));        # ^= IV
+       &xorps  ($inout1,$rndkey1);
+       &movups ($rndkey1,&QWP(0x20,$inp));
+       &xorps  ($inout2,$rndkey0);
+       &movups ($rndkey0,&QWP(0x30,$inp));
+       &xorps  ($inout3,$rndkey1);
+       &movups ($ivec,&QWP(0x40,$inp));        # IV
+       &xorps  ($inout4,$rndkey0);
+       &movups (&QWP(0,$out),$inout0);
+       &movups (&QWP(0x10,$out),$inout1);
+       &movups (&QWP(0x20,$out),$inout2);
+       &movups (&QWP(0x30,$out),$inout3);
+       &lea    ($out,&DWP(0x40,$out));
+       &movaps ($inout0,$inout4);
+       &sub    ($len,0x50);

        &jmp    (&label("cbc_dec_tail_collected"));
 
 &set_label("cbc_dec_one",16);
        &jmp    (&label("cbc_dec_tail_collected"));
 
 &set_label("cbc_dec_one",16);


@@ -879,51 +1876,70 @@ if ($PREFIX eq "aesni") {
        {   &aesni_inline_generate1("dec");     }
        else
        {   &call       ("_aesni_decrypt1");    }
        {   &aesni_inline_generate1("dec");     }
        else
        {   &call       ("_aesni_decrypt1");    }

-       &pxor   ($inout0,$ivec);
-       &movdqa ($ivec,$in0);
+       &xorps  ($inout0,$ivec);
+       &movaps ($ivec,$in0);
+       &sub    ($len,0x10);

        &jmp    (&label("cbc_dec_tail_collected"));
 
 &set_label("cbc_dec_two",16);
        &jmp    (&label("cbc_dec_tail_collected"));
 
 &set_label("cbc_dec_two",16);

-       &pxor   ($inout2,$inout2);
+       &xorps  ($inout2,$inout2);

        &call   ("_aesni_decrypt3");
        &call   ("_aesni_decrypt3");

-       &pxor   ($inout0,$ivec);
-       &pxor   ($inout1,$in0);
-       &movdqu (&QWP(0,$out),$inout0);
-       &movdqa ($inout0,$inout1);
-       &movdqa ($ivec,$in1);
+       &xorps  ($inout0,$ivec);
+       &xorps  ($inout1,$in0);
+       &movups (&QWP(0,$out),$inout0);
+       &movaps ($inout0,$inout1);

        &lea    ($out,&DWP(0x10,$out));
        &lea    ($out,&DWP(0x10,$out));

+       &movaps ($ivec,$in1);
+       &sub    ($len,0x20);

        &jmp    (&label("cbc_dec_tail_collected"));
 
 &set_label("cbc_dec_three",16);
        &call   ("_aesni_decrypt3");
        &jmp    (&label("cbc_dec_tail_collected"));
 
 &set_label("cbc_dec_three",16);
        &call   ("_aesni_decrypt3");

-       &pxor   ($inout0,$ivec);
-       &pxor   ($inout1,$in0);
-       &pxor   ($inout2,$in1);
-       &movdqu (&QWP(0,$out),$inout0);
-       &movdqu (&QWP(0x10,$out),$inout1);
-       &movdqa ($inout0,$inout2);
-       &movdqu ($ivec,&QWP(0x20,$inp));
+       &xorps  ($inout0,$ivec);
+       &xorps  ($inout1,$in0);
+       &xorps  ($inout2,$in1);
+       &movups (&QWP(0,$out),$inout0);
+       &movaps ($inout0,$inout2);
+       &movups (&QWP(0x10,$out),$inout1);

        &lea    ($out,&DWP(0x20,$out));
        &lea    ($out,&DWP(0x20,$out));

+       &movups ($ivec,&QWP(0x20,$inp));
+       &sub    ($len,0x30);
+       &jmp    (&label("cbc_dec_tail_collected"));
+
+&set_label("cbc_dec_four",16);
+       &call   ("_aesni_decrypt4");
+       &movups ($rndkey1,&QWP(0x10,$inp));
+       &movups ($rndkey0,&QWP(0x20,$inp));
+       &xorps  ($inout0,$ivec);
+       &movups ($ivec,&QWP(0x30,$inp));
+       &xorps  ($inout1,$in0);
+       &movups (&QWP(0,$out),$inout0);
+       &xorps  ($inout2,$rndkey1);
+       &movups (&QWP(0x10,$out),$inout1);
+       &xorps  ($inout3,$rndkey0);
+       &movups (&QWP(0x20,$out),$inout2);
+       &lea    ($out,&DWP(0x30,$out));
+       &movaps ($inout0,$inout3);
+       &sub    ($len,0x40);

 
 &set_label("cbc_dec_tail_collected");
        &and    ($len,15);
        &jnz    (&label("cbc_dec_tail_partial"));
 
 &set_label("cbc_dec_tail_collected");
        &and    ($len,15);
        &jnz    (&label("cbc_dec_tail_partial"));

-       &movdqu (&QWP(0,$out),$inout0);
+       &movups (&QWP(0,$out),$inout0);

        &jmp    (&label("cbc_ret"));
 
 &set_label("cbc_dec_tail_partial",16);
        &jmp    (&label("cbc_ret"));
 
 &set_label("cbc_dec_tail_partial",16);

-       &mov    ($key_,"esp");
-       &sub    ("esp",16);
-       &and    ("esp",-16);
-       &movdqa (&QWP(0,"esp"),$inout0);
+       &movaps (&QWP(0,"esp"),$inout0);
+       &mov    ("ecx",16);

        &mov    ($inp,"esp");
        &mov    ($inp,"esp");

-       &mov    ("ecx",$len);
+       &sub    ("ecx",$len);

        &data_word(0xA4F3F689);         # rep movsb
        &data_word(0xA4F3F689);         # rep movsb

-       &mov    ("esp",$key_);

 
 &set_label("cbc_ret");
 
 &set_label("cbc_ret");

+       &mov    ("esp",&DWP(16,"esp")); # pull original %esp

        &mov    ($key_,&wparam(4));
        &movups (&QWP(0,$key_),$ivec);  # output IV
        &mov    ($key_,&wparam(4));
        &movups (&QWP(0,$key_),$ivec);  # output IV

+&set_label("cbc_abort");

 &function_end("${PREFIX}_cbc_encrypt");
 \f
 ######################################################################
 &function_end("${PREFIX}_cbc_encrypt");
 \f
 ######################################################################


@@ -945,7 +1961,7 @@ if ($PREFIX eq "aesni") {
        &jz     (&label("bad_pointer"));
 
        &movups ("xmm0",&QWP(0,"eax")); # pull first 128 bits of *userKey
        &jz     (&label("bad_pointer"));
 
        &movups ("xmm0",&QWP(0,"eax")); # pull first 128 bits of *userKey

-       &pxor   ("xmm4","xmm4");        # low dword of xmm4 is assumed 0
+       &xorps  ("xmm4","xmm4");        # low dword of xmm4 is assumed 0

        &lea    ($key,&DWP(16,$key));
        &cmp    ($rounds,256);
        &je     (&label("14rounds"));
        &lea    ($key,&DWP(16,$key));
        &cmp    ($rounds,256);
        &je     (&label("14rounds"));


@@ -987,11 +2003,11 @@ if ($PREFIX eq "aesni") {
        &lea            ($key,&DWP(16,$key));
 &set_label("key_128_cold");
        &shufps         ("xmm4","xmm0",0b00010000);
        &lea            ($key,&DWP(16,$key));
 &set_label("key_128_cold");
        &shufps         ("xmm4","xmm0",0b00010000);

-       &pxor           ("xmm0","xmm4");
-       &shufps         ("xmm4","xmm0",0b10001100,);
-       &pxor           ("xmm0","xmm4");
-       &pshufd         ("xmm1","xmm1",0b11111111);     # critical path
-       &pxor           ("xmm0","xmm1");
+       &xorps          ("xmm0","xmm4");
+       &shufps         ("xmm4","xmm0",0b10001100);
+       &xorps          ("xmm0","xmm4");
+       &shufps         ("xmm1","xmm1",0b11111111);     # critical path
+       &xorps          ("xmm0","xmm1");

        &ret();
 
 &set_label("12rounds",16);
        &ret();
 
 &set_label("12rounds",16);


@@ -1026,11 +2042,11 @@ if ($PREFIX eq "aesni") {
        &movaps         ("xmm5","xmm2");
 &set_label("key_192b_warm");
        &shufps         ("xmm4","xmm0",0b00010000);
        &movaps         ("xmm5","xmm2");
 &set_label("key_192b_warm");
        &shufps         ("xmm4","xmm0",0b00010000);

-       &movaps         ("xmm3","xmm2");
-       &pxor           ("xmm0","xmm4");
+       &movdqa         ("xmm3","xmm2");
+       &xorps          ("xmm0","xmm4");

        &shufps         ("xmm4","xmm0",0b10001100);
        &pslldq         ("xmm3",4);
        &shufps         ("xmm4","xmm0",0b10001100);
        &pslldq         ("xmm3",4);

-       &pxor           ("xmm0","xmm4");
+       &xorps          ("xmm0","xmm4");

        &pshufd         ("xmm1","xmm1",0b01010101);     # critical path
        &pxor           ("xmm2","xmm3");
        &pxor           ("xmm0","xmm1");
        &pshufd         ("xmm1","xmm1",0b01010101);     # critical path
        &pxor           ("xmm2","xmm3");
        &pxor           ("xmm0","xmm1");


@@ -1089,11 +2105,11 @@ if ($PREFIX eq "aesni") {
        &lea            ($key,&DWP(16,$key));
 &set_label("key_256a_cold");
        &shufps         ("xmm4","xmm0",0b00010000);
        &lea            ($key,&DWP(16,$key));
 &set_label("key_256a_cold");
        &shufps         ("xmm4","xmm0",0b00010000);

-       &pxor           ("xmm0","xmm4");
+       &xorps          ("xmm0","xmm4");

        &shufps         ("xmm4","xmm0",0b10001100);
        &shufps         ("xmm4","xmm0",0b10001100);

-       &pxor           ("xmm0","xmm4");
-       &pshufd         ("xmm1","xmm1",0b11111111);     # critical path
-       &pxor           ("xmm0","xmm1");
+       &xorps          ("xmm0","xmm4");
+       &shufps         ("xmm1","xmm1",0b11111111);     # critical path
+       &xorps          ("xmm0","xmm1");

        &ret();
 
 &set_label("key_256b",16);
        &ret();
 
 &set_label("key_256b",16);


@@ -1101,11 +2117,11 @@ if ($PREFIX eq "aesni") {
        &lea            ($key,&DWP(16,$key));
 
        &shufps         ("xmm4","xmm2",0b00010000);
        &lea            ($key,&DWP(16,$key));
 
        &shufps         ("xmm4","xmm2",0b00010000);

-       &pxor           ("xmm2","xmm4");
+       &xorps          ("xmm2","xmm4");

        &shufps         ("xmm4","xmm2",0b10001100);
        &shufps         ("xmm4","xmm2",0b10001100);

-       &pxor           ("xmm2","xmm4");
-       &pshufd         ("xmm1","xmm1",0b10101010);     # critical path
-       &pxor           ("xmm2","xmm1");
+       &xorps          ("xmm2","xmm4");
+       &shufps         ("xmm1","xmm1",0b10101010);     # critical path
+       &xorps          ("xmm2","xmm1");

        &ret();
 
 &set_label("bad_pointer",4);
        &ret();
 
 &set_label("bad_pointer",4);

diff --git a/crypto/aes/asm/aesni-x86_64.pl b/crypto/aes/asm/aesni-x86_64.pl
index 4602068ffcf176a91e66ec99730f3763cce0d4b1..ea9cf9404d7bde451b210b72bc105ce1ae0d578e 100644 (file)
--- a/crypto/aes/asm/aesni-x86_64.pl
+++ b/crypto/aes/asm/aesni-x86_64.pl
@@ -18,7 +18,7 @@
 # non-parallelizable modes such as CBC encrypt is 3.75 cycles per byte
 # processed with 128-bit key. And given their throughput asymptotic
 # performance for parallelizable modes is 1.25 cycles per byte. Being
 # non-parallelizable modes such as CBC encrypt is 3.75 cycles per byte
 # processed with 128-bit key. And given their throughput asymptotic
 # performance for parallelizable modes is 1.25 cycles per byte. Being

-# asymptotic limit is not something you commonly achieve in reality,
+# asymptotic limit it's not something you commonly achieve in reality,

 # but how close does one get? Below are results collected for
 # different modes and block sized. Pairs of numbers are for en-/
 # decryption.
 # but how close does one get? Below are results collected for
 # different modes and block sized. Pairs of numbers are for en-/
 # decryption.


@@ -79,6 +79,84 @@
 # 10/11.7 and 10/13.4 (as opposite to 10/12.0 and 10/14.0), but one
 # observe even 10/11.2 and 10/12.4 (CTR, OFB, CFB)...
 
 # 10/11.7 and 10/13.4 (as opposite to 10/12.0 and 10/14.0), but one
 # observe even 10/11.2 and 10/12.4 (CTR, OFB, CFB)...
 

+# January 2011
+#
+# While Westmere processor features 6 cycles latency for aes[enc|dec]
+# instructions, which can be scheduled every second cycle, Sandy
+# Bridge spends 8 cycles per instruction, but it can schedule them
+# every cycle. This means that code targeting Westmere would perform
+# suboptimally on Sandy Bridge. Therefore this update.
+#
+# In addition, non-parallelizable CBC encrypt (as well as CCM) is
+# optimized. Relative improvement might appear modest, 8% on Westmere,
+# but in absolute terms it's 3.77 cycles per byte encrypted with
+# 128-bit key on Westmere, and 5.07 - on Sandy Bridge. These numbers
+# should be compared to asymptotic limits of 3.75 for Westmere and
+# 5.00 for Sandy Bridge. Actually, the fact that they get this close
+# to asymptotic limits is quite amazing. Indeed, the limit is
+# calculated as latency times number of rounds, 10 for 128-bit key,
+# and divided by 16, the number of bytes in block, or in other words
+# it accounts *solely* for aesenc instructions. But there are extra
+# instructions, and numbers so close to the asymptotic limits mean
+# that it's as if it takes as little as *one* additional cycle to
+# execute all of them. How is it possible? It is possible thanks to
+# out-of-order execution logic, which manages to overlap post-
+# processing of previous block, things like saving the output, with
+# actual encryption of current block, as well as pre-processing of
+# current block, things like fetching input and xor-ing it with
+# 0-round element of the key schedule, with actual encryption of
+# previous block. Keep this in mind...
+#
+# For parallelizable modes, such as ECB, CBC decrypt, CTR, higher
+# performance is achieved by interleaving instructions working on
+# independent blocks. In which case asymptotic limit for such modes
+# can be obtained by dividing above mentioned numbers by AES
+# instructions' interleave factor. Westmere can execute at most 3 
+# instructions at a time, meaning that optimal interleave factor is 3,
+# and that's where the "magic" number of 1.25 come from. "Optimal
+# interleave factor" means that increase of interleave factor does
+# not improve performance. The formula has proven to reflect reality
+# pretty well on Westmere... Sandy Bridge on the other hand can
+# execute up to 8 AES instructions at a time, so how does varying
+# interleave factor affect the performance? Here is table for ECB
+# (numbers are cycles per byte processed with 128-bit key):
+#
+# instruction interleave factor                3x      6x      8x
+# theoretical asymptotic limit         1.67    0.83    0.625
+# measured performance for 8KB block   1.05    0.86    0.84
+#
+# "as if" interleave factor            4.7x    5.8x    6.0x
+#
+# Further data for other parallelizable modes:
+#
+# CBC decrypt                          1.16    0.93    0.93
+# CTR                                  1.14    0.91    n/a
+#
+# Well, given 3x column it's probably inappropriate to call the limit
+# asymptotic, if it can be surpassed, isn't it? What happens there?
+# Rewind to CBC paragraph for the answer. Yes, out-of-order execution
+# magic is responsible for this. Processor overlaps not only the
+# additional instructions with AES ones, but even AES instuctions
+# processing adjacent triplets of independent blocks. In the 6x case
+# additional instructions  still claim disproportionally small amount
+# of additional cycles, but in 8x case number of instructions must be
+# a tad too high for out-of-order logic to cope with, and AES unit
+# remains underutilized... As you can see 8x interleave is hardly
+# justifiable, so there no need to feel bad that 32-bit aesni-x86.pl
+# utilizies 6x interleave because of limited register bank capacity.
+#
+# Higher interleave factors do have negative impact on Westmere
+# performance. While for ECB mode it's negligible ~1.5%, other
+# parallelizables perform ~5% worse, which is outweighed by ~25%
+# improvement on Sandy Bridge. To balance regression on Westmere
+# CTR mode was implemented with 6x aesenc interleave factor.
+
+# April 2011
+#
+# Add aesni_xts_[en|de]crypt. Westmere spends 1.33 cycles processing
+# one byte out of 8KB with 128-bit key, Sandy Bridge - 0.97. Just like
+# in CTR mode AES instruction interleave factor was chosen to be 6x.
+

 $PREFIX="aesni";       # if $PREFIX is set to "AES", the script
                        # generates drop-in replacement for
                        # crypto/aes/asm/aes-x86_64.pl:-)
 $PREFIX="aesni";       # if $PREFIX is set to "AES", the script
                        # generates drop-in replacement for
                        # crypto/aes/asm/aes-x86_64.pl:-)


@@ -114,12 +192,14 @@ $rnds_="%r10d";   # backup copy for $rounds
 $key_="%r11";  # backup copy for $key
 
 # %xmm register layout
 $key_="%r11";  # backup copy for $key
 
 # %xmm register layout

-$inout0="%xmm0";       $inout1="%xmm1";
-$inout2="%xmm2";       $inout3="%xmm3";
-$rndkey0="%xmm4";      $rndkey1="%xmm5";
-
-$iv="%xmm6";           $in0="%xmm7";   # used in CBC decrypt, CTR, ...
-$in1="%xmm8";          $in2="%xmm9";
+$rndkey0="%xmm0";      $rndkey1="%xmm1";
+$inout0="%xmm2";       $inout1="%xmm3";
+$inout2="%xmm4";       $inout3="%xmm5";
+$inout4="%xmm6";       $inout5="%xmm7";
+$inout6="%xmm8";       $inout7="%xmm9";
+
+$in2="%xmm6";          $in1="%xmm7";   # used in CBC decrypt, CTR, ...
+$in0="%xmm8";          $iv="%xmm9";

 \f
 # Inline version of internal aesni_[en|de]crypt1.
 #
 \f
 # Inline version of internal aesni_[en|de]crypt1.
 #


@@ -127,13 +207,22 @@ $in1="%xmm8";             $in2="%xmm9";
 # cycles which take care of loop variables...
 { my $sn;
 sub aesni_generate1 {
 # cycles which take care of loop variables...
 { my $sn;
 sub aesni_generate1 {

-my ($p,$key,$rounds,$inout)=@_;        $inout=$inout0 if (!defined($inout));
+my ($p,$key,$rounds,$inout,$ivec)=@_;  $inout=$inout0 if (!defined($inout));

 ++$sn;
 $code.=<<___;
 ++$sn;
 $code.=<<___;

-       movdqu  ($key),$rndkey0
+       $movkey ($key),$rndkey0

        $movkey 16($key),$rndkey1
        $movkey 16($key),$rndkey1

+___
+$code.=<<___ if (defined($ivec));
+       xorps   $rndkey0,$ivec
+       lea     32($key),$key
+       xorps   $ivec,$inout
+___
+$code.=<<___ if (!defined($ivec));

        lea     32($key),$key
        lea     32($key),$key

-       pxor    $rndkey0,$inout
+       xorps   $rndkey0,$inout
+___
+$code.=<<___;

 .Loop_${p}1_$sn:
        aes${p} $rndkey1,$inout
        dec     $rounds
 .Loop_${p}1_$sn:
        aes${p} $rndkey1,$inout
        dec     $rounds


@@ -152,8 +241,8 @@ $code.=<<___;
 .type  ${PREFIX}_encrypt,\@abi-omnipotent
 .align 16
 ${PREFIX}_encrypt:
 .type  ${PREFIX}_encrypt,\@abi-omnipotent
 .align 16
 ${PREFIX}_encrypt:

-       movdqu  ($inp),$inout0          # load input
-       mov     240($key),$rounds       # pull $rounds
+       movups  ($inp),$inout0          # load input
+       mov     240($key),$rounds       # key->rounds

 ___
        &aesni_generate1("enc",$key,$rounds);
 $code.=<<___;
 ___
        &aesni_generate1("enc",$key,$rounds);
 $code.=<<___;


@@ -165,8 +254,8 @@ $code.=<<___;
 .type  ${PREFIX}_decrypt,\@abi-omnipotent
 .align 16
 ${PREFIX}_decrypt:
 .type  ${PREFIX}_decrypt,\@abi-omnipotent
 .align 16
 ${PREFIX}_decrypt:

-       movdqu  ($inp),$inout0          # load input
-       mov     240($key),$rounds       # pull $rounds
+       movups  ($inp),$inout0          # load input
+       mov     240($key),$rounds       # key->rounds

 ___
        &aesni_generate1("dec",$key,$rounds);
 $code.=<<___;
 ___
        &aesni_generate1("dec",$key,$rounds);
 $code.=<<___;


@@ -176,16 +265,16 @@ $code.=<<___;
 ___
 }
 \f
 ___
 }
 \f

-# _aesni_[en|de]crypt[34] are private interfaces, N denotes interleave
-# factor. Why 3x subroutine is used in loops? Even though aes[enc|dec]
-# latency is 6, it turned out that it can be scheduled only every
-# *second* cycle. Thus 3x interleave is the one providing optimal
+# _aesni_[en|de]cryptN are private interfaces, N denotes interleave
+# factor. Why 3x subroutine were originally used in loops? Even though
+# aes[enc|dec] latency was originally 6, it could be scheduled only
+# every *2nd* cycle. Thus 3x interleave was the one providing optimal

 # utilization, i.e. when subroutine's throughput is virtually same as
 # of non-interleaved subroutine [for number of input blocks up to 3].
 # utilization, i.e. when subroutine's throughput is virtually same as
 # of non-interleaved subroutine [for number of input blocks up to 3].

-# This is why it makes no sense to implement 2x subroutine. As soon
-# as/if Intel improves throughput by making it possible to schedule
-# the instructions in question *every* cycles I would have to
-# implement 6x interleave and use it in loop...
+# This is why it makes no sense to implement 2x subroutine.
+# aes[enc|dec] latency in next processor generation is 8, but the
+# instructions can be scheduled every cycle. Optimal interleave for
+# new processor is therefore 8x...

 sub aesni_generate3 {
 my $dir=shift;
 # As already mentioned it takes in $key and $rounds, which are *not*
 sub aesni_generate3 {
 my $dir=shift;
 # As already mentioned it takes in $key and $rounds, which are *not*


@@ -198,9 +287,9 @@ _aesni_${dir}rypt3:
        shr     \$1,$rounds
        $movkey 16($key),$rndkey1
        lea     32($key),$key
        shr     \$1,$rounds
        $movkey 16($key),$rndkey1
        lea     32($key),$key

-       pxor    $rndkey0,$inout0
-       pxor    $rndkey0,$inout1
-       pxor    $rndkey0,$inout2
+       xorps   $rndkey0,$inout0
+       xorps   $rndkey0,$inout1
+       xorps   $rndkey0,$inout2

        $movkey         ($key),$rndkey0
 
 .L${dir}_loop3:
        $movkey         ($key),$rndkey0
 
 .L${dir}_loop3:


@@ -242,11 +331,11 @@ _aesni_${dir}rypt4:
        shr     \$1,$rounds
        $movkey 16($key),$rndkey1
        lea     32($key),$key
        shr     \$1,$rounds
        $movkey 16($key),$rndkey1
        lea     32($key),$key

-       pxor    $rndkey0,$inout0
-       pxor    $rndkey0,$inout1
-       pxor    $rndkey0,$inout2
-       pxor    $rndkey0,$inout3
-       $movkey         ($key),$rndkey0
+       xorps   $rndkey0,$inout0
+       xorps   $rndkey0,$inout1
+       xorps   $rndkey0,$inout2
+       xorps   $rndkey0,$inout3
+       $movkey ($key),$rndkey0

 
 .L${dir}_loop4:
        aes${dir}       $rndkey1,$inout0
 
 .L${dir}_loop4:
        aes${dir}       $rndkey1,$inout0


@@ -275,10 +364,155 @@ _aesni_${dir}rypt4:
 .size  _aesni_${dir}rypt4,.-_aesni_${dir}rypt4
 ___
 }
 .size  _aesni_${dir}rypt4,.-_aesni_${dir}rypt4
 ___
 }

+sub aesni_generate6 {
+my $dir=shift;
+# As already mentioned it takes in $key and $rounds, which are *not*
+# preserved. $inout[0-5] is cipher/clear text...
+$code.=<<___;
+.type  _aesni_${dir}rypt6,\@abi-omnipotent
+.align 16
+_aesni_${dir}rypt6:
+       $movkey         ($key),$rndkey0
+       shr             \$1,$rounds
+       $movkey         16($key),$rndkey1
+       lea             32($key),$key
+       xorps           $rndkey0,$inout0
+       pxor            $rndkey0,$inout1
+       aes${dir}       $rndkey1,$inout0
+       pxor            $rndkey0,$inout2
+       aes${dir}       $rndkey1,$inout1
+       pxor            $rndkey0,$inout3
+       aes${dir}       $rndkey1,$inout2
+       pxor            $rndkey0,$inout4
+       aes${dir}       $rndkey1,$inout3
+       pxor            $rndkey0,$inout5
+       dec             $rounds
+       aes${dir}       $rndkey1,$inout4
+       $movkey         ($key),$rndkey0
+       aes${dir}       $rndkey1,$inout5
+       jmp             .L${dir}_loop6_enter
+.align 16
+.L${dir}_loop6:
+       aes${dir}       $rndkey1,$inout0
+       aes${dir}       $rndkey1,$inout1
+       dec             $rounds
+       aes${dir}       $rndkey1,$inout2
+       aes${dir}       $rndkey1,$inout3
+       aes${dir}       $rndkey1,$inout4
+       aes${dir}       $rndkey1,$inout5
+.L${dir}_loop6_enter:                          # happens to be 16-byte aligned
+       $movkey         16($key),$rndkey1
+       aes${dir}       $rndkey0,$inout0
+       aes${dir}       $rndkey0,$inout1
+       lea             32($key),$key
+       aes${dir}       $rndkey0,$inout2
+       aes${dir}       $rndkey0,$inout3
+       aes${dir}       $rndkey0,$inout4
+       aes${dir}       $rndkey0,$inout5
+       $movkey         ($key),$rndkey0
+       jnz             .L${dir}_loop6
+
+       aes${dir}       $rndkey1,$inout0
+       aes${dir}       $rndkey1,$inout1
+       aes${dir}       $rndkey1,$inout2
+       aes${dir}       $rndkey1,$inout3
+       aes${dir}       $rndkey1,$inout4
+       aes${dir}       $rndkey1,$inout5
+       aes${dir}last   $rndkey0,$inout0
+       aes${dir}last   $rndkey0,$inout1
+       aes${dir}last   $rndkey0,$inout2
+       aes${dir}last   $rndkey0,$inout3
+       aes${dir}last   $rndkey0,$inout4
+       aes${dir}last   $rndkey0,$inout5
+       ret
+.size  _aesni_${dir}rypt6,.-_aesni_${dir}rypt6
+___
+}
+sub aesni_generate8 {
+my $dir=shift;
+# As already mentioned it takes in $key and $rounds, which are *not*
+# preserved. $inout[0-7] is cipher/clear text...
+$code.=<<___;
+.type  _aesni_${dir}rypt8,\@abi-omnipotent
+.align 16
+_aesni_${dir}rypt8:
+       $movkey         ($key),$rndkey0
+       shr             \$1,$rounds
+       $movkey         16($key),$rndkey1
+       lea             32($key),$key
+       xorps           $rndkey0,$inout0
+       xorps           $rndkey0,$inout1
+       aes${dir}       $rndkey1,$inout0
+       pxor            $rndkey0,$inout2
+       aes${dir}       $rndkey1,$inout1
+       pxor            $rndkey0,$inout3
+       aes${dir}       $rndkey1,$inout2
+       pxor            $rndkey0,$inout4
+       aes${dir}       $rndkey1,$inout3
+       pxor            $rndkey0,$inout5
+       dec             $rounds
+       aes${dir}       $rndkey1,$inout4
+       pxor            $rndkey0,$inout6
+       aes${dir}       $rndkey1,$inout5
+       pxor            $rndkey0,$inout7
+       $movkey         ($key),$rndkey0
+       aes${dir}       $rndkey1,$inout6
+       aes${dir}       $rndkey1,$inout7
+       $movkey         16($key),$rndkey1
+       jmp             .L${dir}_loop8_enter
+.align 16
+.L${dir}_loop8:
+       aes${dir}       $rndkey1,$inout0
+       aes${dir}       $rndkey1,$inout1
+       dec             $rounds
+       aes${dir}       $rndkey1,$inout2
+       aes${dir}       $rndkey1,$inout3
+       aes${dir}       $rndkey1,$inout4
+       aes${dir}       $rndkey1,$inout5
+       aes${dir}       $rndkey1,$inout6
+       aes${dir}       $rndkey1,$inout7
+       $movkey         16($key),$rndkey1
+.L${dir}_loop8_enter:                          # happens to be 16-byte aligned
+       aes${dir}       $rndkey0,$inout0
+       aes${dir}       $rndkey0,$inout1
+       lea             32($key),$key
+       aes${dir}       $rndkey0,$inout2
+       aes${dir}       $rndkey0,$inout3
+       aes${dir}       $rndkey0,$inout4
+       aes${dir}       $rndkey0,$inout5
+       aes${dir}       $rndkey0,$inout6
+       aes${dir}       $rndkey0,$inout7
+       $movkey         ($key),$rndkey0
+       jnz             .L${dir}_loop8
+
+       aes${dir}       $rndkey1,$inout0
+       aes${dir}       $rndkey1,$inout1
+       aes${dir}       $rndkey1,$inout2
+       aes${dir}       $rndkey1,$inout3
+       aes${dir}       $rndkey1,$inout4
+       aes${dir}       $rndkey1,$inout5
+       aes${dir}       $rndkey1,$inout6
+       aes${dir}       $rndkey1,$inout7
+       aes${dir}last   $rndkey0,$inout0
+       aes${dir}last   $rndkey0,$inout1
+       aes${dir}last   $rndkey0,$inout2
+       aes${dir}last   $rndkey0,$inout3
+       aes${dir}last   $rndkey0,$inout4
+       aes${dir}last   $rndkey0,$inout5
+       aes${dir}last   $rndkey0,$inout6
+       aes${dir}last   $rndkey0,$inout7
+       ret
+.size  _aesni_${dir}rypt8,.-_aesni_${dir}rypt8
+___
+}

 &aesni_generate3("enc") if ($PREFIX eq "aesni");
 &aesni_generate3("dec");
 &aesni_generate4("enc") if ($PREFIX eq "aesni");
 &aesni_generate4("dec");
 &aesni_generate3("enc") if ($PREFIX eq "aesni");
 &aesni_generate3("dec");
 &aesni_generate4("enc") if ($PREFIX eq "aesni");
 &aesni_generate4("dec");

+&aesni_generate6("enc") if ($PREFIX eq "aesni");
+&aesni_generate6("dec");
+&aesni_generate8("enc") if ($PREFIX eq "aesni");
+&aesni_generate8("dec");

 \f
 if ($PREFIX eq "aesni") {
 ########################################################################
 \f
 if ($PREFIX eq "aesni") {
 ########################################################################


@@ -290,37 +524,73 @@ $code.=<<___;
 .type  aesni_ecb_encrypt,\@function,5
 .align 16
 aesni_ecb_encrypt:
 .type  aesni_ecb_encrypt,\@function,5
 .align 16
 aesni_ecb_encrypt:

-       cmp     \$16,$len               # check length
-       jb      .Lecb_ret
-
-       mov     240($key),$rounds       # pull $rounds

        and     \$-16,$len
        and     \$-16,$len

+       jz      .Lecb_ret
+
+       mov     240($key),$rounds       # key->rounds
+       $movkey ($key),$rndkey0

        mov     $key,$key_              # backup $key
        mov     $rounds,$rnds_          # backup $rounds
        test    %r8d,%r8d               # 5th argument
        jz      .Lecb_decrypt
 #--------------------------- ECB ENCRYPT ------------------------------#
        mov     $key,$key_              # backup $key
        mov     $rounds,$rnds_          # backup $rounds
        test    %r8d,%r8d               # 5th argument
        jz      .Lecb_decrypt
 #--------------------------- ECB ENCRYPT ------------------------------#

-       cmp     \$0x40,$len
-       jbe     .Lecb_enc_tail
-       sub     \$0x40,$len
-       jmp     .Lecb_enc_loop3
+       cmp     \$0x80,$len
+       jb      .Lecb_enc_tail
+
+       movdqu  ($inp),$inout0
+       movdqu  0x10($inp),$inout1
+       movdqu  0x20($inp),$inout2
+       movdqu  0x30($inp),$inout3
+       movdqu  0x40($inp),$inout4
+       movdqu  0x50($inp),$inout5
+       movdqu  0x60($inp),$inout6
+       movdqu  0x70($inp),$inout7
+       lea     0x80($inp),$inp
+       sub     \$0x80,$len
+       jmp     .Lecb_enc_loop8_enter

 .align 16
 .align 16

-.Lecb_enc_loop3:
-       movups  ($inp),$inout0
-       movups  0x10($inp),$inout1
-       movups  0x20($inp),$inout2
-       call    _aesni_encrypt3
-       lea     0x30($inp),$inp
+.Lecb_enc_loop8:

        movups  $inout0,($out)
        movups  $inout0,($out)

+       mov     $key_,$key              # restore $key
+       movdqu  ($inp),$inout0

        mov     $rnds_,$rounds          # restore $rounds
        movups  $inout1,0x10($out)
        mov     $rnds_,$rounds          # restore $rounds
        movups  $inout1,0x10($out)

+       movdqu  0x10($inp),$inout1
+       movups  $inout2,0x20($out)
+       movdqu  0x20($inp),$inout2
+       movups  $inout3,0x30($out)
+       movdqu  0x30($inp),$inout3
+       movups  $inout4,0x40($out)
+       movdqu  0x40($inp),$inout4
+       movups  $inout5,0x50($out)
+       movdqu  0x50($inp),$inout5
+       movups  $inout6,0x60($out)
+       movdqu  0x60($inp),$inout6
+       movups  $inout7,0x70($out)
+       lea     0x80($out),$out
+       movdqu  0x70($inp),$inout7
+       lea     0x80($inp),$inp
+.Lecb_enc_loop8_enter:
+
+       call    _aesni_encrypt8
+
+       sub     \$0x80,$len
+       jnc     .Lecb_enc_loop8
+
+       movups  $inout0,($out)

        mov     $key_,$key              # restore $key
        mov     $key_,$key              # restore $key

+       movups  $inout1,0x10($out)
+       mov     $rnds_,$rounds          # restore $rounds

        movups  $inout2,0x20($out)
        movups  $inout2,0x20($out)

-       lea     0x30($out),$out
-       sub     \$0x30,$len
-       ja      .Lecb_enc_loop3
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)
+       movups  $inout6,0x60($out)
+       movups  $inout7,0x70($out)
+       lea     0x80($out),$out
+       add     \$0x80,$len
+       jz      .Lecb_ret

 
 

-       add     \$0x40,$len

 .Lecb_enc_tail:
        movups  ($inp),$inout0
        cmp     \$0x20,$len
 .Lecb_enc_tail:
        movups  ($inp),$inout0
        cmp     \$0x20,$len


@@ -328,14 +598,24 @@ aesni_ecb_encrypt:
        movups  0x10($inp),$inout1
        je      .Lecb_enc_two
        movups  0x20($inp),$inout2
        movups  0x10($inp),$inout1
        je      .Lecb_enc_two
        movups  0x20($inp),$inout2

-       cmp     \$0x30,$len
-       je      .Lecb_enc_three
+       cmp     \$0x40,$len
+       jb      .Lecb_enc_three

        movups  0x30($inp),$inout3
        movups  0x30($inp),$inout3

-       call    _aesni_encrypt4
+       je      .Lecb_enc_four
+       movups  0x40($inp),$inout4
+       cmp     \$0x60,$len
+       jb      .Lecb_enc_five
+       movups  0x50($inp),$inout5
+       je      .Lecb_enc_six
+       movdqu  0x60($inp),$inout6
+       call    _aesni_encrypt8

        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        movups  $inout3,0x30($out)
        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        movups  $inout3,0x30($out)

+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)
+       movups  $inout6,0x60($out)

        jmp     .Lecb_ret
 .align 16
 .Lecb_enc_one:
        jmp     .Lecb_ret
 .align 16
 .Lecb_enc_one:


@@ -346,7 +626,7 @@ $code.=<<___;
        jmp     .Lecb_ret
 .align 16
 .Lecb_enc_two:
        jmp     .Lecb_ret
 .align 16
 .Lecb_enc_two:

-       pxor    $inout2,$inout2
+       xorps   $inout2,$inout2

        call    _aesni_encrypt3
        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        call    _aesni_encrypt3
        movups  $inout0,($out)
        movups  $inout1,0x10($out)


@@ -358,30 +638,95 @@ $code.=<<___;
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        jmp     .Lecb_ret
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        jmp     .Lecb_ret

+.align 16
+.Lecb_enc_four:
+       call    _aesni_encrypt4
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       jmp     .Lecb_ret
+.align 16
+.Lecb_enc_five:
+       xorps   $inout5,$inout5
+       call    _aesni_encrypt6
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       jmp     .Lecb_ret
+.align 16
+.Lecb_enc_six:
+       call    _aesni_encrypt6
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)
+       jmp     .Lecb_ret

 \f#--------------------------- ECB DECRYPT ------------------------------#
 .align 16
 .Lecb_decrypt:
 \f#--------------------------- ECB DECRYPT ------------------------------#
 .align 16
 .Lecb_decrypt:

-       cmp     \$0x40,$len
-       jbe     .Lecb_dec_tail
-       sub     \$0x40,$len
-       jmp     .Lecb_dec_loop3
+       cmp     \$0x80,$len
+       jb      .Lecb_dec_tail
+
+       movdqu  ($inp),$inout0
+       movdqu  0x10($inp),$inout1
+       movdqu  0x20($inp),$inout2
+       movdqu  0x30($inp),$inout3
+       movdqu  0x40($inp),$inout4
+       movdqu  0x50($inp),$inout5
+       movdqu  0x60($inp),$inout6
+       movdqu  0x70($inp),$inout7
+       lea     0x80($inp),$inp
+       sub     \$0x80,$len
+       jmp     .Lecb_dec_loop8_enter

 .align 16
 .align 16

-.Lecb_dec_loop3:
-       movups  ($inp),$inout0
-       movups  0x10($inp),$inout1
-       movups  0x20($inp),$inout2
-       call    _aesni_decrypt3
-       lea     0x30($inp),$inp
+.Lecb_dec_loop8:

        movups  $inout0,($out)
        movups  $inout0,($out)

+       mov     $key_,$key              # restore $key
+       movdqu  ($inp),$inout0

        mov     $rnds_,$rounds          # restore $rounds
        movups  $inout1,0x10($out)
        mov     $rnds_,$rounds          # restore $rounds
        movups  $inout1,0x10($out)

+       movdqu  0x10($inp),$inout1
+       movups  $inout2,0x20($out)
+       movdqu  0x20($inp),$inout2
+       movups  $inout3,0x30($out)
+       movdqu  0x30($inp),$inout3
+       movups  $inout4,0x40($out)
+       movdqu  0x40($inp),$inout4
+       movups  $inout5,0x50($out)
+       movdqu  0x50($inp),$inout5
+       movups  $inout6,0x60($out)
+       movdqu  0x60($inp),$inout6
+       movups  $inout7,0x70($out)
+       lea     0x80($out),$out
+       movdqu  0x70($inp),$inout7
+       lea     0x80($inp),$inp
+.Lecb_dec_loop8_enter:
+
+       call    _aesni_decrypt8
+
+       $movkey ($key_),$rndkey0
+       sub     \$0x80,$len
+       jnc     .Lecb_dec_loop8
+
+       movups  $inout0,($out)

        mov     $key_,$key              # restore $key
        mov     $key_,$key              # restore $key

+       movups  $inout1,0x10($out)
+       mov     $rnds_,$rounds          # restore $rounds

        movups  $inout2,0x20($out)
        movups  $inout2,0x20($out)

-       lea     0x30($out),$out
-       sub     \$0x30,$len
-       ja      .Lecb_dec_loop3
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)
+       movups  $inout6,0x60($out)
+       movups  $inout7,0x70($out)
+       lea     0x80($out),$out
+       add     \$0x80,$len
+       jz      .Lecb_ret

 
 

-       add     \$0x40,$len

 .Lecb_dec_tail:
        movups  ($inp),$inout0
        cmp     \$0x20,$len
 .Lecb_dec_tail:
        movups  ($inp),$inout0
        cmp     \$0x20,$len


@@ -389,14 +734,25 @@ $code.=<<___;
        movups  0x10($inp),$inout1
        je      .Lecb_dec_two
        movups  0x20($inp),$inout2
        movups  0x10($inp),$inout1
        je      .Lecb_dec_two
        movups  0x20($inp),$inout2

-       cmp     \$0x30,$len
-       je      .Lecb_dec_three
+       cmp     \$0x40,$len
+       jb      .Lecb_dec_three

        movups  0x30($inp),$inout3
        movups  0x30($inp),$inout3

-       call    _aesni_decrypt4
+       je      .Lecb_dec_four
+       movups  0x40($inp),$inout4
+       cmp     \$0x60,$len
+       jb      .Lecb_dec_five
+       movups  0x50($inp),$inout5
+       je      .Lecb_dec_six
+       movups  0x60($inp),$inout6
+       $movkey ($key),$rndkey0
+       call    _aesni_decrypt8

        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        movups  $inout3,0x30($out)
        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        movups  $inout3,0x30($out)

+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)
+       movups  $inout6,0x60($out)

        jmp     .Lecb_ret
 .align 16
 .Lecb_dec_one:
        jmp     .Lecb_ret
 .align 16
 .Lecb_dec_one:


@@ -407,7 +763,7 @@ $code.=<<___;
        jmp     .Lecb_ret
 .align 16
 .Lecb_dec_two:
        jmp     .Lecb_ret
 .align 16
 .Lecb_dec_two:

-       pxor    $inout2,$inout2
+       xorps   $inout2,$inout2

        call    _aesni_decrypt3
        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        call    _aesni_decrypt3
        movups  $inout0,($out)
        movups  $inout1,0x10($out)


@@ -418,6 +774,34 @@ $code.=<<___;
        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)
        movups  $inout0,($out)
        movups  $inout1,0x10($out)
        movups  $inout2,0x20($out)

+       jmp     .Lecb_ret
+.align 16
+.Lecb_dec_four:
+       call    _aesni_decrypt4
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       jmp     .Lecb_ret
+.align 16
+.Lecb_dec_five:
+       xorps   $inout5,$inout5
+       call    _aesni_decrypt6
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       jmp     .Lecb_ret
+.align 16
+.Lecb_dec_six:
+       call    _aesni_decrypt6
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)

 
 .Lecb_ret:
        ret
 
 .Lecb_ret:
        ret


@@ -467,25 +851,45 @@ $code.=<<___;
        movdqa  $iv,$inout0
 
 .Lccm64_enc_outer:
        movdqa  $iv,$inout0
 
 .Lccm64_enc_outer:

-       movdqu  ($inp),$in0                     # load inp
+       movups  ($inp),$in0                     # load inp

        pshufb  $bswap_mask,$inout0
        mov     $key_,$key
        mov     $rnds_,$rounds
        pshufb  $bswap_mask,$inout0
        mov     $key_,$key
        mov     $rnds_,$rounds

-       pxor    $in0,$inout1                    # cmac^=inp
-       pxor    $inout2,$inout2

 
 

-       call    _aesni_encrypt3
+       $movkey ($key),$rndkey0
+       shr     \$1,$rounds
+       $movkey 16($key),$rndkey1
+       xorps   $rndkey0,$in0
+       lea     32($key),$key
+       xorps   $rndkey0,$inout0
+       xorps   $inout1,$in0                    # cmac^=inp
+       $movkey ($key),$rndkey0
+
+.Lccm64_enc2_loop:
+       aesenc  $rndkey1,$inout0
+       dec     $rounds
+       aesenc  $rndkey1,$inout1
+       $movkey 16($key),$rndkey1
+       aesenc  $rndkey0,$inout0
+       lea     32($key),$key
+       aesenc  $rndkey0,$inout1
+       $movkey 0($key),$rndkey0
+       jnz     .Lccm64_enc2_loop
+       aesenc  $rndkey1,$inout0
+       aesenc  $rndkey1,$inout1
+       aesenclast      $rndkey0,$inout0
+       aesenclast      $rndkey0,$inout1

 
        paddq   $increment,$iv
        dec     $len
        lea     16($inp),$inp
 
        paddq   $increment,$iv
        dec     $len
        lea     16($inp),$inp

-       pxor    $inout0,$in0                    # inp ^= E(iv)
+       xorps   $inout0,$in0                    # inp ^= E(iv)

        movdqa  $iv,$inout0
        movdqa  $iv,$inout0

-       movdqu  $in0,($out)                     # save output
+       movups  $in0,($out)                     # save output

        lea     16($out),$out
        jnz     .Lccm64_enc_outer
 
        lea     16($out),$out
        jnz     .Lccm64_enc_outer
 

-       movdqu  $inout1,($cmac)
+       movups  $inout1,($cmac)

 ___
 $code.=<<___ if ($win64);
        movaps  (%rsp),%xmm6
 ___
 $code.=<<___ if ($win64);
        movaps  (%rsp),%xmm6


@@ -529,24 +933,42 @@ ___
        &aesni_generate1("enc",$key,$rounds);
 $code.=<<___;
 .Lccm64_dec_outer:
        &aesni_generate1("enc",$key,$rounds);
 $code.=<<___;
 .Lccm64_dec_outer:

-       movdqu  ($inp),$in0                     # load inp

        paddq   $increment,$iv
        paddq   $increment,$iv

-       dec     $len
-       lea     16($inp),$inp
-       pxor    $inout0,$in0
+       movups  ($inp),$in0                     # load inp
+       xorps   $inout0,$in0

        movdqa  $iv,$inout0
        movdqa  $iv,$inout0

+       lea     16($inp),$inp
+       pshufb  $bswap_mask,$inout0

        mov     $key_,$key
        mov     $rnds_,$rounds
        mov     $key_,$key
        mov     $rnds_,$rounds

-       pshufb  $bswap_mask,$inout0
-       movdqu  $in0,($out)
+       movups  $in0,($out)

        lea     16($out),$out
        lea     16($out),$out

-       pxor    $in0,$inout1                    # cmac^=out

 
 

+       sub     \$1,$len

        jz      .Lccm64_dec_break
 
        jz      .Lccm64_dec_break
 

-       pxor    $inout2,$inout2
-       call    _aesni_encrypt3
+       $movkey ($key),$rndkey0
+       shr     \$1,$rounds
+       $movkey 16($key),$rndkey1
+       xorps   $rndkey0,$in0
+       lea     32($key),$key
+       xorps   $rndkey0,$inout0
+       xorps   $in0,$inout1                    # cmac^=out
+       $movkey ($key),$rndkey0

 
 

+.Lccm64_dec2_loop:
+       aesenc  $rndkey1,$inout0
+       dec     $rounds
+       aesenc  $rndkey1,$inout1
+       $movkey 16($key),$rndkey1
+       aesenc  $rndkey0,$inout0
+       lea     32($key),$key
+       aesenc  $rndkey0,$inout1
+       $movkey 0($key),$rndkey0
+       jnz     .Lccm64_dec2_loop
+       aesenc  $rndkey1,$inout0
+       aesenc  $rndkey1,$inout1
+       aesenclast      $rndkey0,$inout0

        jmp     .Lccm64_dec_outer
 
 .align 16
        jmp     .Lccm64_dec_outer
 
 .align 16


@@ -554,7 +976,7 @@ $code.=<<___;
 ___
        &aesni_generate1("enc",$key,$rounds,$inout1);
 $code.=<<___;
 ___
        &aesni_generate1("enc",$key,$rounds,$inout1);
 $code.=<<___;

-       movdqu  $inout1,($cmac)
+       movups  $inout1,($cmac)

 ___
 $code.=<<___ if ($win64);
        movaps  (%rsp),%xmm6
 ___
 $code.=<<___ if ($win64);
        movaps  (%rsp),%xmm6


@@ -566,219 +988,1133 @@ $code.=<<___ if ($win64);
 ___
 $code.=<<___;
        ret
 ___
 $code.=<<___;
        ret

-.size  aesni_ccm64_decrypt_blocks,.-aesni_ccm64_decrypt_blocks
+.size  aesni_ccm64_decrypt_blocks,.-aesni_ccm64_decrypt_blocks
+___
+}\f
+######################################################################
+# void aesni_ctr32_encrypt_blocks (const void *in, void *out,
+#                         size_t blocks, const AES_KEY *key,
+#                         const char *ivec);
+#
+# Handles only complete blocks, operates on 32-bit counter and
+# does not update *ivec! (see engine/eng_aesni.c for details)
+#
+{
+my $reserved = $win64?0:-0x28;
+my ($in0,$in1,$in2,$in3)=map("%xmm$_",(8..11));
+my ($iv0,$iv1,$ivec)=("%xmm12","%xmm13","%xmm14");
+my $bswap_mask="%xmm15";
+
+$code.=<<___;
+.globl aesni_ctr32_encrypt_blocks
+.type  aesni_ctr32_encrypt_blocks,\@function,5
+.align 16
+aesni_ctr32_encrypt_blocks:
+___
+$code.=<<___ if ($win64);
+       lea     -0xc8(%rsp),%rsp
+       movaps  %xmm6,0x20(%rsp)
+       movaps  %xmm7,0x30(%rsp)
+       movaps  %xmm8,0x40(%rsp)
+       movaps  %xmm9,0x50(%rsp)
+       movaps  %xmm10,0x60(%rsp)
+       movaps  %xmm11,0x70(%rsp)
+       movaps  %xmm12,0x80(%rsp)
+       movaps  %xmm13,0x90(%rsp)
+       movaps  %xmm14,0xa0(%rsp)
+       movaps  %xmm15,0xb0(%rsp)
+.Lctr32_body:
+___
+$code.=<<___;
+       cmp     \$1,$len
+       je      .Lctr32_one_shortcut
+
+       movdqu  ($ivp),$ivec
+       movdqa  .Lbswap_mask(%rip),$bswap_mask
+       xor     $rounds,$rounds
+       pextrd  \$3,$ivec,$rnds_                # pull 32-bit counter
+       pinsrd  \$3,$rounds,$ivec               # wipe 32-bit counter
+
+       mov     240($key),$rounds               # key->rounds
+       bswap   $rnds_
+       pxor    $iv0,$iv0                       # vector of 3 32-bit counters
+       pxor    $iv1,$iv1                       # vector of 3 32-bit counters
+       pinsrd  \$0,$rnds_,$iv0
+       lea     3($rnds_),$key_
+       pinsrd  \$0,$key_,$iv1
+       inc     $rnds_
+       pinsrd  \$1,$rnds_,$iv0
+       inc     $key_
+       pinsrd  \$1,$key_,$iv1
+       inc     $rnds_
+       pinsrd  \$2,$rnds_,$iv0
+       inc     $key_
+       pinsrd  \$2,$key_,$iv1
+       movdqa  $iv0,$reserved(%rsp)
+       pshufb  $bswap_mask,$iv0
+       movdqa  $iv1,`$reserved+0x10`(%rsp)
+       pshufb  $bswap_mask,$iv1
+
+       pshufd  \$`3<<6`,$iv0,$inout0           # place counter to upper dword
+       pshufd  \$`2<<6`,$iv0,$inout1
+       pshufd  \$`1<<6`,$iv0,$inout2
+       cmp     \$6,$len
+       jb      .Lctr32_tail
+       shr     \$1,$rounds
+       mov     $key,$key_                      # backup $key
+       mov     $rounds,$rnds_                  # backup $rounds
+       sub     \$6,$len
+       jmp     .Lctr32_loop6
+
+.align 16
+.Lctr32_loop6:
+       pshufd  \$`3<<6`,$iv1,$inout3
+       por     $ivec,$inout0                   # merge counter-less ivec
+        $movkey        ($key_),$rndkey0
+       pshufd  \$`2<<6`,$iv1,$inout4
+       por     $ivec,$inout1
+        $movkey        16($key_),$rndkey1
+       pshufd  \$`1<<6`,$iv1,$inout5
+       por     $ivec,$inout2
+       por     $ivec,$inout3
+        xorps          $rndkey0,$inout0
+       por     $ivec,$inout4
+       por     $ivec,$inout5
+
+       # inline _aesni_encrypt6 and interleave last rounds
+       # with own code...
+
+       pxor            $rndkey0,$inout1
+       aesenc          $rndkey1,$inout0
+       lea             32($key_),$key
+       pxor            $rndkey0,$inout2
+       aesenc          $rndkey1,$inout1
+        movdqa         .Lincrement32(%rip),$iv1
+       pxor            $rndkey0,$inout3
+       aesenc          $rndkey1,$inout2
+        movdqa         $reserved(%rsp),$iv0
+       pxor            $rndkey0,$inout4
+       aesenc          $rndkey1,$inout3
+       pxor            $rndkey0,$inout5
+       $movkey         ($key),$rndkey0
+       dec             $rounds
+       aesenc          $rndkey1,$inout4
+       aesenc          $rndkey1,$inout5
+       jmp             .Lctr32_enc_loop6_enter
+.align 16
+.Lctr32_enc_loop6:
+       aesenc          $rndkey1,$inout0
+       aesenc          $rndkey1,$inout1
+       dec             $rounds
+       aesenc          $rndkey1,$inout2
+       aesenc          $rndkey1,$inout3
+       aesenc          $rndkey1,$inout4
+       aesenc          $rndkey1,$inout5
+.Lctr32_enc_loop6_enter:
+       $movkey         16($key),$rndkey1
+       aesenc          $rndkey0,$inout0
+       aesenc          $rndkey0,$inout1
+       lea             32($key),$key
+       aesenc          $rndkey0,$inout2
+       aesenc          $rndkey0,$inout3
+       aesenc          $rndkey0,$inout4
+       aesenc          $rndkey0,$inout5
+       $movkey         ($key),$rndkey0
+       jnz             .Lctr32_enc_loop6
+
+       aesenc          $rndkey1,$inout0
+        paddd          $iv1,$iv0               # increment counter vector
+       aesenc          $rndkey1,$inout1
+        paddd          `$reserved+0x10`(%rsp),$iv1
+       aesenc          $rndkey1,$inout2
+        movdqa         $iv0,$reserved(%rsp)    # save counter vector
+       aesenc          $rndkey1,$inout3
+        movdqa         $iv1,`$reserved+0x10`(%rsp)
+       aesenc          $rndkey1,$inout4
+        pshufb         $bswap_mask,$iv0        # byte swap
+       aesenc          $rndkey1,$inout5
+        pshufb         $bswap_mask,$iv1
+
+       aesenclast      $rndkey0,$inout0
+        movups         ($inp),$in0             # load input
+       aesenclast      $rndkey0,$inout1
+        movups         0x10($inp),$in1
+       aesenclast      $rndkey0,$inout2
+        movups         0x20($inp),$in2
+       aesenclast      $rndkey0,$inout3
+        movups         0x30($inp),$in3
+       aesenclast      $rndkey0,$inout4
+        movups         0x40($inp),$rndkey1
+       aesenclast      $rndkey0,$inout5
+        movups         0x50($inp),$rndkey0
+        lea    0x60($inp),$inp
+
+       xorps   $inout0,$in0                    # xor
+        pshufd \$`3<<6`,$iv0,$inout0
+       xorps   $inout1,$in1
+        pshufd \$`2<<6`,$iv0,$inout1
+       movups  $in0,($out)                     # store output
+       xorps   $inout2,$in2
+        pshufd \$`1<<6`,$iv0,$inout2
+       movups  $in1,0x10($out)
+       xorps   $inout3,$in3
+       movups  $in2,0x20($out)
+       xorps   $inout4,$rndkey1
+       movups  $in3,0x30($out)
+       xorps   $inout5,$rndkey0
+       movups  $rndkey1,0x40($out)
+       movups  $rndkey0,0x50($out)
+       lea     0x60($out),$out
+       mov     $rnds_,$rounds
+       sub     \$6,$len
+       jnc     .Lctr32_loop6
+
+       add     \$6,$len
+       jz      .Lctr32_done
+       mov     $key_,$key                      # restore $key
+       lea     1($rounds,$rounds),$rounds      # restore original value
+
+.Lctr32_tail:
+       por     $ivec,$inout0
+       movups  ($inp),$in0
+       cmp     \$2,$len
+       jb      .Lctr32_one
+
+       por     $ivec,$inout1
+       movups  0x10($inp),$in1
+       je      .Lctr32_two
+
+       pshufd  \$`3<<6`,$iv1,$inout3
+       por     $ivec,$inout2
+       movups  0x20($inp),$in2
+       cmp     \$4,$len
+       jb      .Lctr32_three
+
+       pshufd  \$`2<<6`,$iv1,$inout4
+       por     $ivec,$inout3
+       movups  0x30($inp),$in3
+       je      .Lctr32_four
+
+       por     $ivec,$inout4
+       xorps   $inout5,$inout5
+
+       call    _aesni_encrypt6
+
+       movups  0x40($inp),$rndkey1
+       xorps   $inout0,$in0
+       xorps   $inout1,$in1
+       movups  $in0,($out)
+       xorps   $inout2,$in2
+       movups  $in1,0x10($out)
+       xorps   $inout3,$in3
+       movups  $in2,0x20($out)
+       xorps   $inout4,$rndkey1
+       movups  $in3,0x30($out)
+       movups  $rndkey1,0x40($out)
+       jmp     .Lctr32_done
+
+.align 16
+.Lctr32_one_shortcut:
+       movups  ($ivp),$inout0
+       movups  ($inp),$in0
+       mov     240($key),$rounds               # key->rounds
+.Lctr32_one:
+___
+       &aesni_generate1("enc",$key,$rounds);
+$code.=<<___;
+       xorps   $inout0,$in0
+       movups  $in0,($out)
+       jmp     .Lctr32_done
+
+.align 16
+.Lctr32_two:
+       xorps   $inout2,$inout2
+       call    _aesni_encrypt3
+       xorps   $inout0,$in0
+       xorps   $inout1,$in1
+       movups  $in0,($out)
+       movups  $in1,0x10($out)
+       jmp     .Lctr32_done
+
+.align 16
+.Lctr32_three:
+       call    _aesni_encrypt3
+       xorps   $inout0,$in0
+       xorps   $inout1,$in1
+       movups  $in0,($out)
+       xorps   $inout2,$in2
+       movups  $in1,0x10($out)
+       movups  $in2,0x20($out)
+       jmp     .Lctr32_done
+
+.align 16
+.Lctr32_four:
+       call    _aesni_encrypt4
+       xorps   $inout0,$in0
+       xorps   $inout1,$in1
+       movups  $in0,($out)
+       xorps   $inout2,$in2
+       movups  $in1,0x10($out)
+       xorps   $inout3,$in3
+       movups  $in2,0x20($out)
+       movups  $in3,0x30($out)
+
+.Lctr32_done:
+___
+$code.=<<___ if ($win64);
+       movaps  0x20(%rsp),%xmm6
+       movaps  0x30(%rsp),%xmm7
+       movaps  0x40(%rsp),%xmm8
+       movaps  0x50(%rsp),%xmm9
+       movaps  0x60(%rsp),%xmm10
+       movaps  0x70(%rsp),%xmm11
+       movaps  0x80(%rsp),%xmm12
+       movaps  0x90(%rsp),%xmm13
+       movaps  0xa0(%rsp),%xmm14
+       movaps  0xb0(%rsp),%xmm15
+       lea     0xc8(%rsp),%rsp
+.Lctr32_ret:
+___
+$code.=<<___;
+       ret
+.size  aesni_ctr32_encrypt_blocks,.-aesni_ctr32_encrypt_blocks
+___
+}
+\f
+######################################################################
+# void aesni_xts_[en|de]crypt(const char *inp,char *out,size_t len,
+#      const AES_KEY *key1, const AES_KEY *key2
+#      const unsigned char iv[16]);
+#
+{
+my @tweak=map("%xmm$_",(10..15));
+my ($twmask,$twres,$twtmp)=("%xmm8","%xmm9",@tweak[4]);
+my ($key2,$ivp,$len_)=("%r8","%r9","%r9");
+my $frame_size = 0x68 + ($win64?160:0);
+
+$code.=<<___;
+.globl aesni_xts_encrypt
+.type  aesni_xts_encrypt,\@function,6
+.align 16
+aesni_xts_encrypt:
+       lea     -$frame_size(%rsp),%rsp
+___
+$code.=<<___ if ($win64);
+       movaps  %xmm6,0x60(%rsp)
+       movaps  %xmm7,0x70(%rsp)
+       movaps  %xmm8,0x80(%rsp)
+       movaps  %xmm9,0x90(%rsp)
+       movaps  %xmm10,0xa0(%rsp)
+       movaps  %xmm11,0xb0(%rsp)
+       movaps  %xmm12,0xc0(%rsp)
+       movaps  %xmm13,0xd0(%rsp)
+       movaps  %xmm14,0xe0(%rsp)
+       movaps  %xmm15,0xf0(%rsp)
+.Lxts_enc_body:
+___
+$code.=<<___;
+       movups  ($ivp),@tweak[5]                # load clear-text tweak
+       mov     240(%r8),$rounds                # key2->rounds
+       mov     240($key),$rnds_                # key1->rounds
+___
+       # generate the tweak
+       &aesni_generate1("enc",$key2,$rounds,@tweak[5]);
+$code.=<<___;
+       mov     $key,$key_                      # backup $key
+       mov     $rnds_,$rounds                  # backup $rounds
+       mov     $len,$len_                      # backup $len
+       and     \$-16,$len
+
+       movdqa  .Lxts_magic(%rip),$twmask
+       pxor    $twtmp,$twtmp
+       pcmpgtd @tweak[5],$twtmp                # broadcast upper bits
+___
+    for ($i=0;$i<4;$i++) {
+    $code.=<<___;
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[$i]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+       pand    $twmask,$twres                  # isolate carry and residue
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+       pxor    $twres,@tweak[5]
+___
+    }
+$code.=<<___;
+       sub     \$16*6,$len
+       jc      .Lxts_enc_short
+
+       shr     \$1,$rounds
+       sub     \$1,$rounds
+       mov     $rounds,$rnds_
+       jmp     .Lxts_enc_grandloop
+
+.align 16
+.Lxts_enc_grandloop:
+       pshufd  \$0x13,$twtmp,$twres
+       movdqa  @tweak[5],@tweak[4]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+       movdqu  `16*0`($inp),$inout0            # load input
+       pand    $twmask,$twres                  # isolate carry and residue
+       movdqu  `16*1`($inp),$inout1
+       pxor    $twres,@tweak[5]
+
+       movdqu  `16*2`($inp),$inout2
+       pxor    @tweak[0],$inout0               # input^=tweak
+       movdqu  `16*3`($inp),$inout3
+       pxor    @tweak[1],$inout1
+       movdqu  `16*4`($inp),$inout4
+       pxor    @tweak[2],$inout2
+       movdqu  `16*5`($inp),$inout5
+       lea     `16*6`($inp),$inp
+       pxor    @tweak[3],$inout3
+       $movkey         ($key_),$rndkey0
+       pxor    @tweak[4],$inout4
+       pxor    @tweak[5],$inout5
+
+       # inline _aesni_encrypt6 and interleave first and last rounds
+       # with own code...
+       $movkey         16($key_),$rndkey1
+       pxor            $rndkey0,$inout0
+       pxor            $rndkey0,$inout1
+        movdqa @tweak[0],`16*0`(%rsp)          # put aside tweaks
+       aesenc          $rndkey1,$inout0
+       lea             32($key_),$key
+       pxor            $rndkey0,$inout2
+        movdqa @tweak[1],`16*1`(%rsp)
+       aesenc          $rndkey1,$inout1
+       pxor            $rndkey0,$inout3
+        movdqa @tweak[2],`16*2`(%rsp)
+       aesenc          $rndkey1,$inout2
+       pxor            $rndkey0,$inout4
+        movdqa @tweak[3],`16*3`(%rsp)
+       aesenc          $rndkey1,$inout3
+       pxor            $rndkey0,$inout5
+       $movkey         ($key),$rndkey0
+       dec             $rounds
+        movdqa @tweak[4],`16*4`(%rsp)
+       aesenc          $rndkey1,$inout4
+        movdqa @tweak[5],`16*5`(%rsp)
+       aesenc          $rndkey1,$inout5
+       pxor    $twtmp,$twtmp
+       pcmpgtd @tweak[5],$twtmp
+       jmp             .Lxts_enc_loop6_enter
+
+.align 16
+.Lxts_enc_loop6:
+       aesenc          $rndkey1,$inout0
+       aesenc          $rndkey1,$inout1
+       dec             $rounds
+       aesenc          $rndkey1,$inout2
+       aesenc          $rndkey1,$inout3
+       aesenc          $rndkey1,$inout4
+       aesenc          $rndkey1,$inout5
+.Lxts_enc_loop6_enter:
+       $movkey         16($key),$rndkey1
+       aesenc          $rndkey0,$inout0
+       aesenc          $rndkey0,$inout1
+       lea             32($key),$key
+       aesenc          $rndkey0,$inout2
+       aesenc          $rndkey0,$inout3
+       aesenc          $rndkey0,$inout4
+       aesenc          $rndkey0,$inout5
+       $movkey         ($key),$rndkey0
+       jnz             .Lxts_enc_loop6
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesenc         $rndkey1,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesenc         $rndkey1,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcast upper bits
+        aesenc         $rndkey1,$inout2
+       pxor    $twres,@tweak[5]
+        aesenc         $rndkey1,$inout3
+        aesenc         $rndkey1,$inout4
+        aesenc         $rndkey1,$inout5
+        $movkey        16($key),$rndkey1
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[0]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesenc         $rndkey0,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesenc         $rndkey0,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+        aesenc         $rndkey0,$inout2
+       pxor    $twres,@tweak[5]
+        aesenc         $rndkey0,$inout3
+        aesenc         $rndkey0,$inout4
+        aesenc         $rndkey0,$inout5
+        $movkey        32($key),$rndkey0
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[1]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesenc         $rndkey1,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesenc         $rndkey1,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+        aesenc         $rndkey1,$inout2
+       pxor    $twres,@tweak[5]
+        aesenc         $rndkey1,$inout3
+        aesenc         $rndkey1,$inout4
+        aesenc         $rndkey1,$inout5
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[2]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesenclast     $rndkey0,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesenclast     $rndkey0,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+        aesenclast     $rndkey0,$inout2
+       pxor    $twres,@tweak[5]
+        aesenclast     $rndkey0,$inout3
+        aesenclast     $rndkey0,$inout4
+        aesenclast     $rndkey0,$inout5
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[3]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        xorps  `16*0`(%rsp),$inout0            # output^=tweak
+       pand    $twmask,$twres                  # isolate carry and residue
+        xorps  `16*1`(%rsp),$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+       pxor    $twres,@tweak[5]
+
+       xorps   `16*2`(%rsp),$inout2
+       movups  $inout0,`16*0`($out)            # write output
+       xorps   `16*3`(%rsp),$inout3
+       movups  $inout1,`16*1`($out)
+       xorps   `16*4`(%rsp),$inout4
+       movups  $inout2,`16*2`($out)
+       xorps   `16*5`(%rsp),$inout5
+       movups  $inout3,`16*3`($out)
+       mov     $rnds_,$rounds                  # restore $rounds
+       movups  $inout4,`16*4`($out)
+       movups  $inout5,`16*5`($out)
+       lea     `16*6`($out),$out
+       sub     \$16*6,$len
+       jnc     .Lxts_enc_grandloop
+
+       lea     3($rounds,$rounds),$rounds      # restore original value
+       mov     $key_,$key                      # restore $key
+       mov     $rounds,$rnds_                  # backup $rounds
+
+.Lxts_enc_short:
+       add     \$16*6,$len
+       jz      .Lxts_enc_done
+
+       cmp     \$0x20,$len
+       jb      .Lxts_enc_one
+       je      .Lxts_enc_two
+
+       cmp     \$0x40,$len
+       jb      .Lxts_enc_three
+       je      .Lxts_enc_four
+
+       pshufd  \$0x13,$twtmp,$twres
+       movdqa  @tweak[5],@tweak[4]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        movdqu ($inp),$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        movdqu 16*1($inp),$inout1
+       pxor    $twres,@tweak[5]
+
+       movdqu  16*2($inp),$inout2
+       pxor    @tweak[0],$inout0
+       movdqu  16*3($inp),$inout3
+       pxor    @tweak[1],$inout1
+       movdqu  16*4($inp),$inout4
+       lea     16*5($inp),$inp
+       pxor    @tweak[2],$inout2
+       pxor    @tweak[3],$inout3
+       pxor    @tweak[4],$inout4
+
+       call    _aesni_encrypt6
+
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[5],@tweak[0]
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+       movdqu  $inout0,($out)
+       xorps   @tweak[3],$inout3
+       movdqu  $inout1,16*1($out)
+       xorps   @tweak[4],$inout4
+       movdqu  $inout2,16*2($out)
+       movdqu  $inout3,16*3($out)
+       movdqu  $inout4,16*4($out)
+       lea     16*5($out),$out
+       jmp     .Lxts_enc_done
+
+.align 16
+.Lxts_enc_one:
+       movups  ($inp),$inout0
+       lea     16*1($inp),$inp
+       xorps   @tweak[0],$inout0
+___
+       &aesni_generate1("enc",$key,$rounds);
+$code.=<<___;
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[1],@tweak[0]
+       movups  $inout0,($out)
+       lea     16*1($out),$out
+       jmp     .Lxts_enc_done
+
+.align 16
+.Lxts_enc_two:
+       movups  ($inp),$inout0
+       movups  16($inp),$inout1
+       lea     32($inp),$inp
+       xorps   @tweak[0],$inout0
+       xorps   @tweak[1],$inout1
+
+       call    _aesni_encrypt3
+
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[2],@tweak[0]
+       xorps   @tweak[1],$inout1
+       movups  $inout0,($out)
+       movups  $inout1,16*1($out)
+       lea     16*2($out),$out
+       jmp     .Lxts_enc_done
+
+.align 16
+.Lxts_enc_three:
+       movups  ($inp),$inout0
+       movups  16*1($inp),$inout1
+       movups  16*2($inp),$inout2
+       lea     16*3($inp),$inp
+       xorps   @tweak[0],$inout0
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+
+       call    _aesni_encrypt3
+
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[3],@tweak[0]
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+       movups  $inout0,($out)
+       movups  $inout1,16*1($out)
+       movups  $inout2,16*2($out)
+       lea     16*3($out),$out
+       jmp     .Lxts_enc_done
+
+.align 16
+.Lxts_enc_four:
+       movups  ($inp),$inout0
+       movups  16*1($inp),$inout1
+       movups  16*2($inp),$inout2
+       xorps   @tweak[0],$inout0
+       movups  16*3($inp),$inout3
+       lea     16*4($inp),$inp
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+       xorps   @tweak[3],$inout3
+
+       call    _aesni_encrypt4
+
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[5],@tweak[0]
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+       movups  $inout0,($out)
+       xorps   @tweak[3],$inout3
+       movups  $inout1,16*1($out)
+       movups  $inout2,16*2($out)
+       movups  $inout3,16*3($out)
+       lea     16*4($out),$out
+       jmp     .Lxts_enc_done
+
+.align 16
+.Lxts_enc_done:
+       and     \$15,$len_
+       jz      .Lxts_enc_ret
+       mov     $len_,$len
+
+.Lxts_enc_steal:
+       movzb   ($inp),%eax                     # borrow $rounds ...
+       movzb   -16($out),%ecx                  # ... and $key
+       lea     1($inp),$inp
+       mov     %al,-16($out)
+       mov     %cl,0($out)
+       lea     1($out),$out
+       sub     \$1,$len
+       jnz     .Lxts_enc_steal
+
+       sub     $len_,$out                      # rewind $out
+       mov     $key_,$key                      # restore $key
+       mov     $rnds_,$rounds                  # restore $rounds
+
+       movups  -16($out),$inout0
+       xorps   @tweak[0],$inout0
+___
+       &aesni_generate1("enc",$key,$rounds);
+$code.=<<___;
+       xorps   @tweak[0],$inout0
+       movups  $inout0,-16($out)
+
+.Lxts_enc_ret:
+___
+$code.=<<___ if ($win64);
+       movaps  0x60(%rsp),%xmm6
+       movaps  0x70(%rsp),%xmm7
+       movaps  0x80(%rsp),%xmm8
+       movaps  0x90(%rsp),%xmm9
+       movaps  0xa0(%rsp),%xmm10
+       movaps  0xb0(%rsp),%xmm11
+       movaps  0xc0(%rsp),%xmm12
+       movaps  0xd0(%rsp),%xmm13
+       movaps  0xe0(%rsp),%xmm14
+       movaps  0xf0(%rsp),%xmm15
+___
+$code.=<<___;
+       lea     $frame_size(%rsp),%rsp
+.Lxts_enc_epilogue:
+       ret
+.size  aesni_xts_encrypt,.-aesni_xts_encrypt

 ___
 ___

-}\f
-######################################################################
-# void aesni_ctr32_encrypt_blocks (const void *in, void *out,
-#                         size_t blocks, const AES_KEY *key,
-#                         const char *ivec);
-#
-# Handles only complete blocks, operates on 32-bit counter and
-# does not update *ivec! (see engine/eng_aesni.c for details)
-#
-my $increment="%xmm10";
-my $bswap_mask="%xmm11";

 
 $code.=<<___;
 
 $code.=<<___;

-.globl aesni_ctr32_encrypt_blocks
-.type  aesni_ctr32_encrypt_blocks,\@function,5
+.globl aesni_xts_decrypt
+.type  aesni_xts_decrypt,\@function,6

 .align 16
 .align 16

-aesni_ctr32_encrypt_blocks:
+aesni_xts_decrypt:
+       lea     -$frame_size(%rsp),%rsp

 ___
 $code.=<<___ if ($win64);
 ___
 $code.=<<___ if ($win64);

-       lea     -0x68(%rsp),%rsp
-       movaps  %xmm6,(%rsp)
-       movaps  %xmm7,0x10(%rsp)
-       movaps  %xmm8,0x20(%rsp)
-       movaps  %xmm9,0x30(%rsp)
-       movaps  %xmm10,0x40(%rsp)
-       movaps  %xmm11,0x50(%rsp)
-
-.Lctr32_body:
+       movaps  %xmm6,0x60(%rsp)
+       movaps  %xmm7,0x70(%rsp)
+       movaps  %xmm8,0x80(%rsp)
+       movaps  %xmm9,0x90(%rsp)
+       movaps  %xmm10,0xa0(%rsp)
+       movaps  %xmm11,0xb0(%rsp)
+       movaps  %xmm12,0xc0(%rsp)
+       movaps  %xmm13,0xd0(%rsp)
+       movaps  %xmm14,0xe0(%rsp)
+       movaps  %xmm15,0xf0(%rsp)
+.Lxts_dec_body:

 ___
 $code.=<<___;
 ___
 $code.=<<___;

-       cmp     \$1,$len
-       je      .Lctr32_one_shortcut
-
-       movdqu  ($ivp),$inout3
-       movdqa  .Lincrement32(%rip),$increment
-       movdqa  .Lbswap_mask(%rip),$bswap_mask
-       xor     $rounds,$rounds
-       pextrd  \$3,$inout3,$rnds_              # pull 32-bit counter
-       pinsrd  \$3,$rounds,$inout3             # wipe 32-bit counter
+       movups  ($ivp),@tweak[5]                # load clear-text tweak
+       mov     240($key2),$rounds              # key2->rounds
+       mov     240($key),$rnds_                # key1->rounds
+___
+       # generate the tweak
+       &aesni_generate1("enc",$key2,$rounds,@tweak[5]);
+$code.=<<___;
+       xor     %eax,%eax                       # if ($len%16) len-=16;
+       test    \$15,$len
+       setnz   %al
+       shl     \$4,%rax
+       sub     %rax,$len
+
+       mov     $key,$key_                      # backup $key
+       mov     $rnds_,$rounds                  # backup $rounds
+       mov     $len,$len_                      # backup $len
+       and     \$-16,$len

 
 

-       mov     240($key),$rounds               # key->rounds
-       pxor    $iv,$iv                         # vector of 3 32-bit counters
-       bswap   $rnds_
-       pinsrd  \$0,$rnds_,$iv
-       inc     $rnds_
-       pinsrd  \$1,$rnds_,$iv
-       inc     $rnds_
-       pinsrd  \$2,$rnds_,$iv
-       pshufb  $bswap_mask,$iv
+       movdqa  .Lxts_magic(%rip),$twmask
+       pxor    $twtmp,$twtmp
+       pcmpgtd @tweak[5],$twtmp                # broadcast upper bits
+___
+    for ($i=0;$i<4;$i++) {
+    $code.=<<___;
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[$i]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+       pand    $twmask,$twres                  # isolate carry and residue
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+       pxor    $twres,@tweak[5]
+___
+    }
+$code.=<<___;
+       sub     \$16*6,$len
+       jc      .Lxts_dec_short

 
 

-       cmp     \$4,$len
-       jbe     .Lctr32_tail
+       shr     \$1,$rounds
+       sub     \$1,$rounds

        mov     $rounds,$rnds_
        mov     $rounds,$rnds_

-       mov     $key,$key_
-       sub     \$4,$len
-
-.Lctr32_loop3:
-       pshufd  \$`3<<6`,$iv,$inout0            # place counter to upper dword
-       pshufd  \$`2<<6`,$iv,$inout1
-       por     $inout3,$inout0                 # merge counter-less ivec
-       pshufd  \$`1<<6`,$iv,$inout2
-       por     $inout3,$inout1
-       por     $inout3,$inout2
+       jmp     .Lxts_dec_grandloop

 
 

-       # inline _aesni_encrypt3 and interleave last round
+.align 16
+.Lxts_dec_grandloop:
+       pshufd  \$0x13,$twtmp,$twres
+       movdqa  @tweak[5],@tweak[4]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+       movdqu  `16*0`($inp),$inout0            # load input
+       pand    $twmask,$twres                  # isolate carry and residue
+       movdqu  `16*1`($inp),$inout1
+       pxor    $twres,@tweak[5]
+
+       movdqu  `16*2`($inp),$inout2
+       pxor    @tweak[0],$inout0               # input^=tweak
+       movdqu  `16*3`($inp),$inout3
+       pxor    @tweak[1],$inout1
+       movdqu  `16*4`($inp),$inout4
+       pxor    @tweak[2],$inout2
+       movdqu  `16*5`($inp),$inout5
+       lea     `16*6`($inp),$inp
+       pxor    @tweak[3],$inout3
+       $movkey         ($key_),$rndkey0
+       pxor    @tweak[4],$inout4
+       pxor    @tweak[5],$inout5
+
+       # inline _aesni_decrypt6 and interleave first and last rounds

        # with own code...
        # with own code...

+       $movkey         16($key_),$rndkey1
+       pxor            $rndkey0,$inout0
+       pxor            $rndkey0,$inout1
+        movdqa @tweak[0],`16*0`(%rsp)          # put aside tweaks
+       aesdec          $rndkey1,$inout0
+       lea             32($key_),$key
+       pxor            $rndkey0,$inout2
+        movdqa @tweak[1],`16*1`(%rsp)
+       aesdec          $rndkey1,$inout1
+       pxor            $rndkey0,$inout3
+        movdqa @tweak[2],`16*2`(%rsp)
+       aesdec          $rndkey1,$inout2
+       pxor            $rndkey0,$inout4
+        movdqa @tweak[3],`16*3`(%rsp)
+       aesdec          $rndkey1,$inout3
+       pxor            $rndkey0,$inout5
+       $movkey         ($key),$rndkey0
+       dec             $rounds
+        movdqa @tweak[4],`16*4`(%rsp)
+       aesdec          $rndkey1,$inout4
+        movdqa @tweak[5],`16*5`(%rsp)
+       aesdec          $rndkey1,$inout5
+       pxor    $twtmp,$twtmp
+       pcmpgtd @tweak[5],$twtmp
+       jmp             .Lxts_dec_loop6_enter

 
 

-       $movkey ($key),$rndkey0
-       shr     \$1,$rounds
-       $movkey 16($key),$rndkey1
-       lea     32($key),$key
-       pxor    $rndkey0,$inout0
-       pxor    $rndkey0,$inout1
-       pxor    $rndkey0,$inout2
-       $movkey ($key),$rndkey0
-       jmp     .Lctr32_enc_loop3

 .align 16
 .align 16

-.Lctr32_enc_loop3:
-       aesenc          $rndkey1,$inout0
-       aesenc          $rndkey1,$inout1
+.Lxts_dec_loop6:
+       aesdec          $rndkey1,$inout0
+       aesdec          $rndkey1,$inout1

        dec             $rounds
        dec             $rounds

-       aesenc          $rndkey1,$inout2
+       aesdec          $rndkey1,$inout2
+       aesdec          $rndkey1,$inout3
+       aesdec          $rndkey1,$inout4
+       aesdec          $rndkey1,$inout5
+.Lxts_dec_loop6_enter:

        $movkey         16($key),$rndkey1
        $movkey         16($key),$rndkey1

-       aesenc          $rndkey0,$inout0
-       aesenc          $rndkey0,$inout1
+       aesdec          $rndkey0,$inout0
+       aesdec          $rndkey0,$inout1

        lea             32($key),$key
        lea             32($key),$key

-       aesenc          $rndkey0,$inout2
+       aesdec          $rndkey0,$inout2
+       aesdec          $rndkey0,$inout3
+       aesdec          $rndkey0,$inout4
+       aesdec          $rndkey0,$inout5

        $movkey         ($key),$rndkey0
        $movkey         ($key),$rndkey0

-       jnz             .Lctr32_enc_loop3
+       jnz             .Lxts_dec_loop6
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesdec         $rndkey1,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesdec         $rndkey1,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcast upper bits
+        aesdec         $rndkey1,$inout2
+       pxor    $twres,@tweak[5]
+        aesdec         $rndkey1,$inout3
+        aesdec         $rndkey1,$inout4
+        aesdec         $rndkey1,$inout5
+        $movkey        16($key),$rndkey1
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[0]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesdec         $rndkey0,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesdec         $rndkey0,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+        aesdec         $rndkey0,$inout2
+       pxor    $twres,@tweak[5]
+        aesdec         $rndkey0,$inout3
+        aesdec         $rndkey0,$inout4
+        aesdec         $rndkey0,$inout5
+        $movkey        32($key),$rndkey0
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[1]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesdec         $rndkey1,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesdec         $rndkey1,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+        aesdec         $rndkey1,$inout2
+       pxor    $twres,@tweak[5]
+        aesdec         $rndkey1,$inout3
+        aesdec         $rndkey1,$inout4
+        aesdec         $rndkey1,$inout5
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[2]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        aesdeclast     $rndkey0,$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        aesdeclast     $rndkey0,$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+        aesdeclast     $rndkey0,$inout2
+       pxor    $twres,@tweak[5]
+        aesdeclast     $rndkey0,$inout3
+        aesdeclast     $rndkey0,$inout4
+        aesdeclast     $rndkey0,$inout5
+
+       pshufd  \$0x13,$twtmp,$twres
+       pxor    $twtmp,$twtmp
+       movdqa  @tweak[5],@tweak[3]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        xorps  `16*0`(%rsp),$inout0            # output^=tweak
+       pand    $twmask,$twres                  # isolate carry and residue
+        xorps  `16*1`(%rsp),$inout1
+       pcmpgtd @tweak[5],$twtmp                # broadcat upper bits
+       pxor    $twres,@tweak[5]
+
+       xorps   `16*2`(%rsp),$inout2
+       movups  $inout0,`16*0`($out)            # write output
+       xorps   `16*3`(%rsp),$inout3
+       movups  $inout1,`16*1`($out)
+       xorps   `16*4`(%rsp),$inout4
+       movups  $inout2,`16*2`($out)
+       xorps   `16*5`(%rsp),$inout5
+       movups  $inout3,`16*3`($out)
+       mov     $rnds_,$rounds                  # restore $rounds
+       movups  $inout4,`16*4`($out)
+       movups  $inout5,`16*5`($out)
+       lea     `16*6`($out),$out
+       sub     \$16*6,$len
+       jnc     .Lxts_dec_grandloop
+
+       lea     3($rounds,$rounds),$rounds      # restore original value
+       mov     $key_,$key                      # restore $key
+       mov     $rounds,$rnds_                  # backup $rounds
+
+.Lxts_dec_short:
+       add     \$16*6,$len
+       jz      .Lxts_dec_done

 
 

-       aesenc          $rndkey1,$inout0
-       aesenc          $rndkey1,$inout1
-       aesenc          $rndkey1,$inout2
+       cmp     \$0x20,$len
+       jb      .Lxts_dec_one
+       je      .Lxts_dec_two

 
 

-       pshufb          $bswap_mask,$iv
-       movdqu          ($inp),$in0
-       aesenclast      $rndkey0,$inout0
-       movdqu          0x10($inp),$in1
-       paddd           $increment,$iv
-       aesenclast      $rndkey0,$inout1
-       movdqu          0x20($inp),$in2
-       pshufb          $bswap_mask,$iv
-       aesenclast      $rndkey0,$inout2
-       lea             0x30($inp),$inp
+       cmp     \$0x40,$len
+       jb      .Lxts_dec_three
+       je      .Lxts_dec_four
+
+       pshufd  \$0x13,$twtmp,$twres
+       movdqa  @tweak[5],@tweak[4]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        movdqu ($inp),$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        movdqu 16*1($inp),$inout1
+       pxor    $twres,@tweak[5]
+
+       movdqu  16*2($inp),$inout2
+       pxor    @tweak[0],$inout0
+       movdqu  16*3($inp),$inout3
+       pxor    @tweak[1],$inout1
+       movdqu  16*4($inp),$inout4
+       lea     16*5($inp),$inp
+       pxor    @tweak[2],$inout2
+       pxor    @tweak[3],$inout3
+       pxor    @tweak[4],$inout4
+
+       call    _aesni_decrypt6
+
+       xorps   @tweak[0],$inout0
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+       movdqu  $inout0,($out)
+       xorps   @tweak[3],$inout3
+       movdqu  $inout1,16*1($out)
+       xorps   @tweak[4],$inout4
+       movdqu  $inout2,16*2($out)
+        pxor           $twtmp,$twtmp
+       movdqu  $inout3,16*3($out)
+        pcmpgtd        @tweak[5],$twtmp
+       movdqu  $inout4,16*4($out)
+       lea     16*5($out),$out
+        pshufd         \$0x13,$twtmp,@tweak[1] # $twres
+       and     \$15,$len_
+       jz      .Lxts_dec_ret
+
+       movdqa  @tweak[5],@tweak[0]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+       pand    $twmask,@tweak[1]               # isolate carry and residue
+       pxor    @tweak[5],@tweak[1]
+       jmp     .Lxts_dec_done2

 
 

-       mov     $key_,$key
-       pxor    $inout0,$in0
-       sub     \$3,$len
-       mov     $rnds_,$rounds
-       pxor    $inout1,$in1
-       movdqu  $in0,($out)
-       pxor    $inout2,$in2
-       movdqu  $in1,0x10($out)
-       movdqu  $in2,0x20($out)
-       lea     0x30($out),$out
-       ja      .Lctr32_loop3
+.align 16
+.Lxts_dec_one:
+       movups  ($inp),$inout0
+       lea     16*1($inp),$inp
+       xorps   @tweak[0],$inout0
+___
+       &aesni_generate1("dec",$key,$rounds);
+$code.=<<___;
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[1],@tweak[0]
+       movups  $inout0,($out)
+       movdqa  @tweak[2],@tweak[1]
+       lea     16*1($out),$out
+       jmp     .Lxts_dec_done

 
 

-       pextrd  \$1,$iv,$rnds_                  # might need last counter value
-       add     \$4,$len
-       bswap   $rnds_
+.align 16
+.Lxts_dec_two:
+       movups  ($inp),$inout0
+       movups  16($inp),$inout1
+       lea     32($inp),$inp
+       xorps   @tweak[0],$inout0
+       xorps   @tweak[1],$inout1

 
 

-.Lctr32_tail:
-       pshufd  \$`3<<6`,$iv,$inout0
-       pshufd  \$`2<<6`,$iv,$inout1
-       por     $inout3,$inout0
-       movdqu  ($inp),$in0
-       cmp     \$2,$len
-       jb      .Lctr32_one
-       lea     1($rnds_),$rnds_
-       pshufd  \$`1<<6`,$iv,$inout2
-       por     $inout3,$inout1
-       movdqu  0x10($inp),$in1
-       je      .Lctr32_two
-       bswap   $rnds_
-       por     $inout3,$inout2
-       movdqu  0x20($inp),$in2
-       cmp     \$3,$len
-       je      .Lctr32_three
+       call    _aesni_decrypt3

 
 

-       pinsrd  \$3,$rnds_,$inout3              # compose last counter value
-       movdqu  0x30($inp),$iv
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[2],@tweak[0]
+       xorps   @tweak[1],$inout1
+       movdqa  @tweak[3],@tweak[1]
+       movups  $inout0,($out)
+       movups  $inout1,16*1($out)
+       lea     16*2($out),$out
+       jmp     .Lxts_dec_done

 
 

-       call    _aesni_encrypt4
+.align 16
+.Lxts_dec_three:
+       movups  ($inp),$inout0
+       movups  16*1($inp),$inout1
+       movups  16*2($inp),$inout2
+       lea     16*3($inp),$inp
+       xorps   @tweak[0],$inout0
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2

 
 

-       pxor    $inout0,$in0
-       pxor    $inout1,$in1
-       pxor    $inout2,$in2
-       movdqu  $in0,($out)
-       pxor    $inout3,$iv
-       movdqu  $in1,0x10($out)
-       movdqu  $in2,0x20($out)
-       movdqu  $iv,0x30($out)
-       jmp     .Lctr32_done
+       call    _aesni_decrypt3

 
 

-.align 16
-.Lctr32_one_shortcut:
-       movdqu  ($ivp),$inout0
-       movdqu  ($inp),$in0
-       mov     240($key),$rounds               # key->rounds
-.Lctr32_one:
-___
-       &aesni_generate1("enc",$key,$rounds);
-$code.=<<___;
-       pxor    $inout0,$in0
-       movdqu  $in0,($out)
-       jmp     .Lctr32_done
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[3],@tweak[0]
+       xorps   @tweak[1],$inout1
+       movdqa  @tweak[5],@tweak[1]
+       xorps   @tweak[2],$inout2
+       movups  $inout0,($out)
+       movups  $inout1,16*1($out)
+       movups  $inout2,16*2($out)
+       lea     16*3($out),$out
+       jmp     .Lxts_dec_done

 
 .align 16
 
 .align 16

-.Lctr32_two:
-       pxor    $inout2,$inout2
-       call    _aesni_encrypt3
-       pxor    $inout0,$in0
-       pxor    $inout1,$in1
-       movdqu  $in0,($out)
-       movdqu  $in1,0x10($out)
-       jmp     .Lctr32_done
+.Lxts_dec_four:
+       pshufd  \$0x13,$twtmp,$twres
+       movdqa  @tweak[5],@tweak[4]
+       paddq   @tweak[5],@tweak[5]             # psllq 1,$tweak
+        movups ($inp),$inout0
+       pand    $twmask,$twres                  # isolate carry and residue
+        movups 16*1($inp),$inout1
+       pxor    $twres,@tweak[5]
+
+       movups  16*2($inp),$inout2
+       xorps   @tweak[0],$inout0
+       movups  16*3($inp),$inout3
+       lea     16*4($inp),$inp
+       xorps   @tweak[1],$inout1
+       xorps   @tweak[2],$inout2
+       xorps   @tweak[3],$inout3
+
+       call    _aesni_decrypt4
+
+       xorps   @tweak[0],$inout0
+       movdqa  @tweak[4],@tweak[0]
+       xorps   @tweak[1],$inout1
+       movdqa  @tweak[5],@tweak[1]
+       xorps   @tweak[2],$inout2
+       movups  $inout0,($out)
+       xorps   @tweak[3],$inout3
+       movups  $inout1,16*1($out)
+       movups  $inout2,16*2($out)
+       movups  $inout3,16*3($out)
+       lea     16*4($out),$out
+       jmp     .Lxts_dec_done

 
 .align 16
 
 .align 16

-.Lctr32_three:
-       call    _aesni_encrypt3
-       pxor    $inout0,$in0
-       pxor    $inout1,$in1
-       pxor    $inout2,$in2
-       movdqu  $in0,($out)
-       movdqu  $in1,0x10($out)
-       movdqu  $in2,0x20($out)
+.Lxts_dec_done:
+       and     \$15,$len_
+       jz      .Lxts_dec_ret
+.Lxts_dec_done2:
+       mov     $len_,$len
+       mov     $key_,$key                      # restore $key
+       mov     $rnds_,$rounds                  # restore $rounds

 
 

-.Lctr32_done:
+       movups  ($inp),$inout0
+       xorps   @tweak[1],$inout0
+___
+       &aesni_generate1("dec",$key,$rounds);
+$code.=<<___;
+       xorps   @tweak[1],$inout0
+       movups  $inout0,($out)
+
+.Lxts_dec_steal:
+       movzb   16($inp),%eax                   # borrow $rounds ...
+       movzb   ($out),%ecx                     # ... and $key
+       lea     1($inp),$inp
+       mov     %al,($out)
+       mov     %cl,16($out)
+       lea     1($out),$out
+       sub     \$1,$len
+       jnz     .Lxts_dec_steal
+
+       sub     $len_,$out                      # rewind $out
+       mov     $key_,$key                      # restore $key
+       mov     $rnds_,$rounds                  # restore $rounds
+
+       movups  ($out),$inout0
+       xorps   @tweak[0],$inout0

 ___
 ___

+       &aesni_generate1("dec",$key,$rounds);
+$code.=<<___;
+       xorps   @tweak[0],$inout0
+       movups  $inout0,($out)

 
 

+.Lxts_dec_ret:
+___

 $code.=<<___ if ($win64);
 $code.=<<___ if ($win64);

-       movaps  (%rsp),%xmm6
-       movaps  0x10(%rsp),%xmm7
-       movaps  0x20(%rsp),%xmm8
-       movaps  0x30(%rsp),%xmm9
-       movaps  0x40(%rsp),%xmm10
-       movaps  0x50(%rsp),%xmm11
-       lea     0x68(%rsp),%rsp
-.Lctr32_ret:
+       movaps  0x60(%rsp),%xmm6
+       movaps  0x70(%rsp),%xmm7
+       movaps  0x80(%rsp),%xmm8
+       movaps  0x90(%rsp),%xmm9
+       movaps  0xa0(%rsp),%xmm10
+       movaps  0xb0(%rsp),%xmm11
+       movaps  0xc0(%rsp),%xmm12
+       movaps  0xd0(%rsp),%xmm13
+       movaps  0xe0(%rsp),%xmm14
+       movaps  0xf0(%rsp),%xmm15

 ___
 $code.=<<___;
 ___
 $code.=<<___;

+       lea     $frame_size(%rsp),%rsp
+.Lxts_dec_epilogue:

        ret
        ret

-.size  aesni_ctr32_encrypt_blocks,.-aesni_ctr32_encrypt_blocks
+.size  aesni_xts_decrypt,.-aesni_xts_decrypt

 ___
 ___

-}}
+} }}

 \f
 ########################################################################
 # void $PREFIX_cbc_encrypt (const void *inp, void *out,
 #                          size_t length, const AES_KEY *key,
 #                          unsigned char *ivp,const int enc);
 \f
 ########################################################################
 # void $PREFIX_cbc_encrypt (const void *inp, void *out,
 #                          size_t length, const AES_KEY *key,
 #                          unsigned char *ivp,const int enc);

-$reserved = $win64?0x40:-0x18; # used in decrypt
+{
+my $reserved = $win64?0x40:-0x18;      # used in decrypt

 $code.=<<___;
 .globl ${PREFIX}_cbc_encrypt
 .type  ${PREFIX}_cbc_encrypt,\@function,6
 $code.=<<___;
 .globl ${PREFIX}_cbc_encrypt
 .type  ${PREFIX}_cbc_encrypt,\@function,6


@@ -787,12 +2123,12 @@ ${PREFIX}_cbc_encrypt:
        test    $len,$len               # check length
        jz      .Lcbc_ret
 
        test    $len,$len               # check length
        jz      .Lcbc_ret
 

-       mov     240($key),$rnds_        # pull $rounds
+       mov     240($key),$rnds_        # key->rounds

        mov     $key,$key_              # backup $key
        test    %r9d,%r9d               # 6th argument
        jz      .Lcbc_decrypt
 #--------------------------- CBC ENCRYPT ------------------------------#
        mov     $key,$key_              # backup $key
        test    %r9d,%r9d               # 6th argument
        jz      .Lcbc_decrypt
 #--------------------------- CBC ENCRYPT ------------------------------#

-       movdqu  ($ivp),$inout0          # load iv as initial state
+       movups  ($ivp),$inout0          # load iv as initial state

        mov     $rnds_,$rounds
        cmp     \$16,$len
        jb      .Lcbc_enc_tail
        mov     $rnds_,$rounds
        cmp     \$16,$len
        jb      .Lcbc_enc_tail


@@ -800,11 +2136,11 @@ ${PREFIX}_cbc_encrypt:
        jmp     .Lcbc_enc_loop
 .align 16
 .Lcbc_enc_loop:
        jmp     .Lcbc_enc_loop
 .align 16
 .Lcbc_enc_loop:

-       movdqu  ($inp),$inout1          # load input
+       movups  ($inp),$inout1          # load input

        lea     16($inp),$inp
        lea     16($inp),$inp

-       pxor    $inout1,$inout0
+       #xorps  $inout1,$inout0

 ___
 ___

-       &aesni_generate1("enc",$key,$rounds);
+       &aesni_generate1("enc",$key,$rounds,$inout0,$inout1);

 $code.=<<___;
        mov     $rnds_,$rounds          # restore $rounds
        mov     $key_,$key              # restore $key
 $code.=<<___;
        mov     $rnds_,$rounds          # restore $rounds
        mov     $key_,$key              # restore $key


@@ -846,106 +2182,251 @@ ___
 $code.=<<___;
        movups  ($ivp),$iv
        mov     $rnds_,$rounds
 $code.=<<___;
        movups  ($ivp),$iv
        mov     $rnds_,$rounds

-       cmp     \$0x40,$len
+       cmp     \$0x70,$len

        jbe     .Lcbc_dec_tail
        jbe     .Lcbc_dec_tail

-       sub     \$0x40,$len
-       jmp     .Lcbc_dec_loop3
+       shr     \$1,$rnds_
+       sub     \$0x70,$len
+       mov     $rnds_,$rounds
+       movaps  $iv,$reserved(%rsp)
+       jmp     .Lcbc_dec_loop8_enter

 .align 16
 .align 16

-.Lcbc_dec_loop3:
-       movups  ($inp),$inout0
+.Lcbc_dec_loop8:
+       movaps  $rndkey0,$reserved(%rsp)        # save IV
+       movups  $inout7,($out)
+       lea     0x10($out),$out
+.Lcbc_dec_loop8_enter:
+       $movkey         ($key),$rndkey0
+       movups  ($inp),$inout0                  # load input

        movups  0x10($inp),$inout1
        movups  0x10($inp),$inout1

-       movups  0x20($inp),$inout2
-       movaps  $inout0,$in0
-       movaps  $inout1,$in1
-       movaps  $inout2,$in2
-
-       call    _aesni_decrypt3
+       $movkey         16($key),$rndkey1

 
 

-       sub     \$0x30,$len
-       lea     0x30($inp),$inp
-       lea     0x30($out),$out
-       pxor    $iv,$inout0
-       pxor    $in0,$inout1
-       movaps  $in2,$iv
-       pxor    $in1,$inout2
-       movdqu  $inout0,-0x30($out)
-       mov     $rnds_,$rounds  # restore $rounds
-       movdqu  $inout1,-0x20($out)
-       mov     $key_,$key      # restore $key
-       movdqu  $inout2,-0x10($out)
-       ja      .Lcbc_dec_loop3
+       lea             32($key),$key
+       movdqu  0x20($inp),$inout2
+       xorps           $rndkey0,$inout0
+       movdqu  0x30($inp),$inout3
+       xorps           $rndkey0,$inout1
+       movdqu  0x40($inp),$inout4
+       aesdec          $rndkey1,$inout0
+       pxor            $rndkey0,$inout2
+       movdqu  0x50($inp),$inout5
+       aesdec          $rndkey1,$inout1
+       pxor            $rndkey0,$inout3
+       movdqu  0x60($inp),$inout6
+       aesdec          $rndkey1,$inout2
+       pxor            $rndkey0,$inout4
+       movdqu  0x70($inp),$inout7
+       aesdec          $rndkey1,$inout3
+       pxor            $rndkey0,$inout5
+       dec             $rounds
+       aesdec          $rndkey1,$inout4
+       pxor            $rndkey0,$inout6
+       aesdec          $rndkey1,$inout5
+       pxor            $rndkey0,$inout7
+       $movkey         ($key),$rndkey0
+       aesdec          $rndkey1,$inout6
+       aesdec          $rndkey1,$inout7
+       $movkey         16($key),$rndkey1

 
 

-       add     \$0x40,$len
-       movups  $iv,($ivp)
+       call            .Ldec_loop8_enter
+
+       movups  ($inp),$rndkey1         # re-load input
+       movups  0x10($inp),$rndkey0
+       xorps   $reserved(%rsp),$inout0 # ^= IV
+       xorps   $rndkey1,$inout1
+       movups  0x20($inp),$rndkey1
+       xorps   $rndkey0,$inout2
+       movups  0x30($inp),$rndkey0
+       xorps   $rndkey1,$inout3
+       movups  0x40($inp),$rndkey1
+       xorps   $rndkey0,$inout4
+       movups  0x50($inp),$rndkey0
+       xorps   $rndkey1,$inout5
+       movups  0x60($inp),$rndkey1
+       xorps   $rndkey0,$inout6
+       movups  0x70($inp),$rndkey0     # IV
+       xorps   $rndkey1,$inout7
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       mov     $rnds_,$rounds          # restore $rounds
+       movups  $inout4,0x40($out)
+       mov     $key_,$key              # restore $key
+       movups  $inout5,0x50($out)
+       lea     0x80($inp),$inp
+       movups  $inout6,0x60($out)
+       lea     0x70($out),$out
+       sub     \$0x80,$len
+       ja      .Lcbc_dec_loop8
+
+       movaps  $inout7,$inout0
+       movaps  $rndkey0,$iv
+       add     \$0x70,$len
+       jle     .Lcbc_dec_tail_collected
+       movups  $inout0,($out)
+       lea     1($rnds_,$rnds_),$rounds
+       lea     0x10($out),$out

 .Lcbc_dec_tail:
        movups  ($inp),$inout0
        movaps  $inout0,$in0
        cmp     \$0x10,$len
        jbe     .Lcbc_dec_one
 .Lcbc_dec_tail:
        movups  ($inp),$inout0
        movaps  $inout0,$in0
        cmp     \$0x10,$len
        jbe     .Lcbc_dec_one

+

        movups  0x10($inp),$inout1
        movaps  $inout1,$in1
        cmp     \$0x20,$len
        jbe     .Lcbc_dec_two
        movups  0x10($inp),$inout1
        movaps  $inout1,$in1
        cmp     \$0x20,$len
        jbe     .Lcbc_dec_two

+

        movups  0x20($inp),$inout2
        movaps  $inout2,$in2
        cmp     \$0x30,$len
        jbe     .Lcbc_dec_three
        movups  0x20($inp),$inout2
        movaps  $inout2,$in2
        cmp     \$0x30,$len
        jbe     .Lcbc_dec_three

+

        movups  0x30($inp),$inout3
        movups  0x30($inp),$inout3

-       call    _aesni_decrypt4
-       pxor    $iv,$inout0
-       movups  0x30($inp),$iv
-       pxor    $in0,$inout1
-       movdqu  $inout0,($out)
-       pxor    $in1,$inout2
-       movdqu  $inout1,0x10($out)
-       pxor    $in2,$inout3
-       movdqu  $inout2,0x20($out)
-       movdqa  $inout3,$inout0
-       lea     0x30($out),$out
+       cmp     \$0x40,$len
+       jbe     .Lcbc_dec_four
+
+       movups  0x40($inp),$inout4
+       cmp     \$0x50,$len
+       jbe     .Lcbc_dec_five
+
+       movups  0x50($inp),$inout5
+       cmp     \$0x60,$len
+       jbe     .Lcbc_dec_six
+
+       movups  0x60($inp),$inout6
+       movaps  $iv,$reserved(%rsp)     # save IV
+       call    _aesni_decrypt8
+       movups  ($inp),$rndkey1
+       movups  0x10($inp),$rndkey0
+       xorps   $reserved(%rsp),$inout0 # ^= IV
+       xorps   $rndkey1,$inout1
+       movups  0x20($inp),$rndkey1
+       xorps   $rndkey0,$inout2
+       movups  0x30($inp),$rndkey0
+       xorps   $rndkey1,$inout3
+       movups  0x40($inp),$rndkey1
+       xorps   $rndkey0,$inout4
+       movups  0x50($inp),$rndkey0
+       xorps   $rndkey1,$inout5
+       movups  0x60($inp),$iv          # IV
+       xorps   $rndkey0,$inout6
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       movups  $inout5,0x50($out)
+       lea     0x60($out),$out
+       movaps  $inout6,$inout0
+       sub     \$0x70,$len

        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_one:
 ___
        &aesni_generate1("dec",$key,$rounds);
 $code.=<<___;
        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_one:
 ___
        &aesni_generate1("dec",$key,$rounds);
 $code.=<<___;

-       pxor    $iv,$inout0
+       xorps   $iv,$inout0

        movaps  $in0,$iv
        movaps  $in0,$iv

+       sub     \$0x10,$len

        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_two:
        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_two:

-       pxor    $inout2,$inout2
+       xorps   $inout2,$inout2

        call    _aesni_decrypt3
        call    _aesni_decrypt3

-       pxor    $iv,$inout0
-       pxor    $in0,$inout1
-       movdqu  $inout0,($out)
+       xorps   $iv,$inout0
+       xorps   $in0,$inout1
+       movups  $inout0,($out)

        movaps  $in1,$iv
        movaps  $in1,$iv

-       movdqa  $inout1,$inout0
+       movaps  $inout1,$inout0

        lea     0x10($out),$out
        lea     0x10($out),$out

+       sub     \$0x20,$len

        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_three:
        call    _aesni_decrypt3
        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_three:
        call    _aesni_decrypt3

-       pxor    $iv,$inout0
-       pxor    $in0,$inout1
-       movdqu  $inout0,($out)
-       pxor    $in1,$inout2
-       movdqu  $inout1,0x10($out)
+       xorps   $iv,$inout0
+       xorps   $in0,$inout1
+       movups  $inout0,($out)
+       xorps   $in1,$inout2
+       movups  $inout1,0x10($out)

        movaps  $in2,$iv
        movaps  $in2,$iv

-       movdqa  $inout2,$inout0
+       movaps  $inout2,$inout0

        lea     0x20($out),$out
        lea     0x20($out),$out

+       sub     \$0x30,$len
+       jmp     .Lcbc_dec_tail_collected
+.align 16
+.Lcbc_dec_four:
+       call    _aesni_decrypt4
+       xorps   $iv,$inout0
+       movups  0x30($inp),$iv
+       xorps   $in0,$inout1
+       movups  $inout0,($out)
+       xorps   $in1,$inout2
+       movups  $inout1,0x10($out)
+       xorps   $in2,$inout3
+       movups  $inout2,0x20($out)
+       movaps  $inout3,$inout0
+       lea     0x30($out),$out
+       sub     \$0x40,$len
+       jmp     .Lcbc_dec_tail_collected
+.align 16
+.Lcbc_dec_five:
+       xorps   $inout5,$inout5
+       call    _aesni_decrypt6
+       movups  0x10($inp),$rndkey1
+       movups  0x20($inp),$rndkey0
+       xorps   $iv,$inout0
+       xorps   $in0,$inout1
+       xorps   $rndkey1,$inout2
+       movups  0x30($inp),$rndkey1
+       xorps   $rndkey0,$inout3
+       movups  0x40($inp),$iv
+       xorps   $rndkey1,$inout4
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       lea     0x40($out),$out
+       movaps  $inout4,$inout0
+       sub     \$0x50,$len
+       jmp     .Lcbc_dec_tail_collected
+.align 16
+.Lcbc_dec_six:
+       call    _aesni_decrypt6
+       movups  0x10($inp),$rndkey1
+       movups  0x20($inp),$rndkey0
+       xorps   $iv,$inout0
+       xorps   $in0,$inout1
+       xorps   $rndkey1,$inout2
+       movups  0x30($inp),$rndkey1
+       xorps   $rndkey0,$inout3
+       movups  0x40($inp),$rndkey0
+       xorps   $rndkey1,$inout4
+       movups  0x50($inp),$iv
+       xorps   $rndkey0,$inout5
+       movups  $inout0,($out)
+       movups  $inout1,0x10($out)
+       movups  $inout2,0x20($out)
+       movups  $inout3,0x30($out)
+       movups  $inout4,0x40($out)
+       lea     0x50($out),$out
+       movaps  $inout5,$inout0
+       sub     \$0x60,$len

        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_tail_collected:
        and     \$15,$len
        movups  $iv,($ivp)
        jnz     .Lcbc_dec_tail_partial
        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_tail_collected:
        and     \$15,$len
        movups  $iv,($ivp)
        jnz     .Lcbc_dec_tail_partial

-       movdqu  $inout0,($out)
+       movups  $inout0,($out)

        jmp     .Lcbc_dec_ret
 .align 16
 .Lcbc_dec_tail_partial:
        movaps  $inout0,$reserved(%rsp)
        jmp     .Lcbc_dec_ret
 .align 16
 .Lcbc_dec_tail_partial:
        movaps  $inout0,$reserved(%rsp)

+       mov     \$16,%rcx

        mov     $out,%rdi
        mov     $out,%rdi

-       mov     $len,%rcx
+       sub     $len,%rcx

        lea     $reserved(%rsp),%rsi
        .long   0x9066A4F3      # rep movsb
 
        lea     $reserved(%rsp),%rsi
        .long   0x9066A4F3      # rep movsb
 


@@ -963,7 +2444,7 @@ $code.=<<___;
        ret
 .size  ${PREFIX}_cbc_encrypt,.-${PREFIX}_cbc_encrypt
 ___
        ret
 .size  ${PREFIX}_cbc_encrypt,.-${PREFIX}_cbc_encrypt
 ___

-\f
+} \f

 # int $PREFIX_set_[en|de]crypt_key (const unsigned char *userKey,
 #                              int bits, AES_KEY *key)
 { my ($inp,$bits,$key) = @_4args;
 # int $PREFIX_set_[en|de]crypt_key (const unsigned char *userKey,
 #                              int bits, AES_KEY *key)
 { my ($inp,$bits,$key) = @_4args;


@@ -1033,7 +2514,7 @@ __aesni_set_encrypt_key:
        jz      .Lenc_key_ret
 
        movups  ($inp),%xmm0            # pull first 128 bits of *userKey
        jz      .Lenc_key_ret
 
        movups  ($inp),%xmm0            # pull first 128 bits of *userKey

-       pxor    %xmm4,%xmm4             # low dword of xmm4 is assumed 0
+       xorps   %xmm4,%xmm4             # low dword of xmm4 is assumed 0

        lea     16($key),%rax
        cmp     \$256,$bits
        je      .L14rounds
        lea     16($key),%rax
        cmp     \$256,$bits
        je      .L14rounds


@@ -1148,11 +2629,11 @@ __aesni_set_encrypt_key:
        lea     16(%rax),%rax
 .Lkey_expansion_128_cold:
        shufps  \$0b00010000,%xmm0,%xmm4
        lea     16(%rax),%rax
 .Lkey_expansion_128_cold:
        shufps  \$0b00010000,%xmm0,%xmm4

-       pxor    %xmm4, %xmm0
+       xorps   %xmm4, %xmm0

        shufps  \$0b10001100,%xmm0,%xmm4
        shufps  \$0b10001100,%xmm0,%xmm4

-       pxor    %xmm4, %xmm0
-       pshufd  \$0b11111111,%xmm1,%xmm1        # critical path
-       pxor    %xmm1,%xmm0
+       xorps   %xmm4, %xmm0
+       shufps  \$0b11111111,%xmm1,%xmm1        # critical path
+       xorps   %xmm1,%xmm0

        ret
 
 .align 16
        ret
 
 .align 16


@@ -1163,11 +2644,11 @@ __aesni_set_encrypt_key:
        movaps  %xmm2, %xmm5
 .Lkey_expansion_192b_warm:
        shufps  \$0b00010000,%xmm0,%xmm4
        movaps  %xmm2, %xmm5
 .Lkey_expansion_192b_warm:
        shufps  \$0b00010000,%xmm0,%xmm4

-       movaps  %xmm2,%xmm3
-       pxor    %xmm4,%xmm0
+       movdqa  %xmm2,%xmm3
+       xorps   %xmm4,%xmm0

        shufps  \$0b10001100,%xmm0,%xmm4
        pslldq  \$4,%xmm3
        shufps  \$0b10001100,%xmm0,%xmm4
        pslldq  \$4,%xmm3

-       pxor    %xmm4,%xmm0
+       xorps   %xmm4,%xmm0

        pshufd  \$0b01010101,%xmm1,%xmm1        # critical path
        pxor    %xmm3,%xmm2
        pxor    %xmm1,%xmm0
        pshufd  \$0b01010101,%xmm1,%xmm1        # critical path
        pxor    %xmm3,%xmm2
        pxor    %xmm1,%xmm0


@@ -1191,11 +2672,11 @@ __aesni_set_encrypt_key:
        lea     16(%rax),%rax
 .Lkey_expansion_256a_cold:
        shufps  \$0b00010000,%xmm0,%xmm4
        lea     16(%rax),%rax
 .Lkey_expansion_256a_cold:
        shufps  \$0b00010000,%xmm0,%xmm4

-       pxor    %xmm4,%xmm0
+       xorps   %xmm4,%xmm0

        shufps  \$0b10001100,%xmm0,%xmm4
        shufps  \$0b10001100,%xmm0,%xmm4

-       pxor    %xmm4,%xmm0
-       pshufd  \$0b11111111,%xmm1,%xmm1        # critical path
-       pxor    %xmm1,%xmm0
+       xorps   %xmm4,%xmm0
+       shufps  \$0b11111111,%xmm1,%xmm1        # critical path
+       xorps   %xmm1,%xmm0

        ret
 
 .align 16
        ret
 
 .align 16


@@ -1204,13 +2685,14 @@ __aesni_set_encrypt_key:
        lea     16(%rax),%rax
 
        shufps  \$0b00010000,%xmm2,%xmm4
        lea     16(%rax),%rax
 
        shufps  \$0b00010000,%xmm2,%xmm4

-       pxor    %xmm4,%xmm2
+       xorps   %xmm4,%xmm2

        shufps  \$0b10001100,%xmm2,%xmm4
        shufps  \$0b10001100,%xmm2,%xmm4

-       pxor    %xmm4,%xmm2
-       pshufd  \$0b10101010,%xmm1,%xmm1        # critical path
-       pxor    %xmm1,%xmm2
+       xorps   %xmm4,%xmm2
+       shufps  \$0b10101010,%xmm1,%xmm1        # critical path
+       xorps   %xmm1,%xmm2

        ret
 .size  ${PREFIX}_set_encrypt_key,.-${PREFIX}_set_encrypt_key
        ret
 .size  ${PREFIX}_set_encrypt_key,.-${PREFIX}_set_encrypt_key

+.size  __aesni_set_encrypt_key,.-__aesni_set_encrypt_key

 ___
 }
 \f
 ___
 }
 \f


@@ -1219,9 +2701,12 @@ $code.=<<___;
 .Lbswap_mask:
        .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
 .Lincrement32:
 .Lbswap_mask:
        .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
 .Lincrement32:

-       .long   3,3,3,0
+       .long   6,6,6,0

 .Lincrement64:
        .long   1,0,0,0
 .Lincrement64:
        .long   1,0,0,0

+.Lxts_magic:
+       .long   0x87,0,1,0
+

 .asciz  "AES for Intel AES-NI, CRYPTOGAMS by <appro\@openssl.org>"
 .align 64
 ___
 .asciz  "AES for Intel AES-NI, CRYPTOGAMS by <appro\@openssl.org>"
 .align 64
 ___


@@ -1253,12 +2738,8 @@ ecb_se_handler:
        sub     \$64,%rsp
 
        mov     152($context),%rax      # pull context->Rsp
        sub     \$64,%rsp
 
        mov     152($context),%rax      # pull context->Rsp

-       mov     8(%rax),%rdi
-       mov     16(%rax),%rsi
-       mov     %rsi,168($context)      # restore context->Rsi
-       mov     %rdi,176($context)      # restore context->Rdi

 
 

-       jmp     .Lcommon_seh_exit
+       jmp     .Lcommon_seh_tail

 .size  ecb_se_handler,.-ecb_se_handler
 
 .type  ccm64_se_handler,\@abi-omnipotent
 .size  ecb_se_handler,.-ecb_se_handler
 
 .type  ccm64_se_handler,\@abi-omnipotent


@@ -1284,29 +2765,22 @@ ccm64_se_handler:
        mov     0(%r11),%r10d           # HandlerData[0]
        lea     (%rsi,%r10),%r10        # prologue label
        cmp     %r10,%rbx               # context->Rip<prologue label
        mov     0(%r11),%r10d           # HandlerData[0]
        lea     (%rsi,%r10),%r10        # prologue label
        cmp     %r10,%rbx               # context->Rip<prologue label

-       jb      .Lin_ccm64_prologue
+       jb      .Lcommon_seh_tail

 
        mov     152($context),%rax      # pull context->Rsp
 
        mov     4(%r11),%r10d           # HandlerData[1]
        lea     (%rsi,%r10),%r10        # epilogue label
        cmp     %r10,%rbx               # context->Rip>=epilogue label
 
        mov     152($context),%rax      # pull context->Rsp
 
        mov     4(%r11),%r10d           # HandlerData[1]
        lea     (%rsi,%r10),%r10        # epilogue label
        cmp     %r10,%rbx               # context->Rip>=epilogue label

-       jae     .Lin_ccm64_prologue
+       jae     .Lcommon_seh_tail

 
 

-       lea     0(%rax),%rsi            # top of stack
+       lea     0(%rax),%rsi            # %xmm save area

        lea     512($context),%rdi      # &context.Xmm6
        mov     \$8,%ecx                # 4*sizeof(%xmm0)/sizeof(%rax)
        .long   0xa548f3fc              # cld; rep movsq
        lea     0x58(%rax),%rax         # adjust stack pointer
 
        lea     512($context),%rdi      # &context.Xmm6
        mov     \$8,%ecx                # 4*sizeof(%xmm0)/sizeof(%rax)
        .long   0xa548f3fc              # cld; rep movsq
        lea     0x58(%rax),%rax         # adjust stack pointer
 

-.Lin_ccm64_prologue:
-       mov     8(%rax),%rdi
-       mov     16(%rax),%rsi
-       mov     %rax,152($context)      # restore context->Rsp
-       mov     %rsi,168($context)      # restore context->Rsi
-       mov     %rdi,176($context)      # restore context->Rdi
-
-       jmp     .Lcommon_seh_exit
+       jmp     .Lcommon_seh_tail

 .size  ccm64_se_handler,.-ccm64_se_handler
 
 .type  ctr32_se_handler,\@abi-omnipotent
 .size  ccm64_se_handler,.-ccm64_se_handler
 
 .type  ctr32_se_handler,\@abi-omnipotent


@@ -1328,29 +2802,63 @@ ctr32_se_handler:
 
        lea     .Lctr32_body(%rip),%r10
        cmp     %r10,%rbx               # context->Rip<"prologue" label
 
        lea     .Lctr32_body(%rip),%r10
        cmp     %r10,%rbx               # context->Rip<"prologue" label

-       jb      .Lin_ctr32_prologue
+       jb      .Lcommon_seh_tail

 
        mov     152($context),%rax      # pull context->Rsp
 
        lea     .Lctr32_ret(%rip),%r10
        cmp     %r10,%rbx
 
        mov     152($context),%rax      # pull context->Rsp
 
        lea     .Lctr32_ret(%rip),%r10
        cmp     %r10,%rbx

-       jae     .Lin_ctr32_prologue
+       jae     .Lcommon_seh_tail

 
 

-       lea     0(%rax),%rsi            # top of stack
+       lea     0x20(%rax),%rsi         # %xmm save area

        lea     512($context),%rdi      # &context.Xmm6
        lea     512($context),%rdi      # &context.Xmm6

-       mov     \$12,%ecx               # 6*sizeof(%xmm0)/sizeof(%rax)
+       mov     \$20,%ecx               # 10*sizeof(%xmm0)/sizeof(%rax)

        .long   0xa548f3fc              # cld; rep movsq
        .long   0xa548f3fc              # cld; rep movsq

-       lea     0x68(%rax),%rax         # adjust stack pointer
-
-.Lin_ctr32_prologue:
-       mov     8(%rax),%rdi
-       mov     16(%rax),%rsi
-       mov     %rax,152($context)      # restore context->Rsp
-       mov     %rsi,168($context)      # restore context->Rsi
-       mov     %rdi,176($context)      # restore context->Rdi
+       lea     0xc8(%rax),%rax         # adjust stack pointer

 
 

-       jmp     .Lcommon_seh_exit
+       jmp     .Lcommon_seh_tail

 .size  ctr32_se_handler,.-ctr32_se_handler
 .size  ctr32_se_handler,.-ctr32_se_handler

+
+.type  xts_se_handler,\@abi-omnipotent
+.align 16
+xts_se_handler:
+       push    %rsi
+       push    %rdi
+       push    %rbx
+       push    %rbp
+       push    %r12
+       push    %r13
+       push    %r14
+       push    %r15
+       pushfq
+       sub     \$64,%rsp
+
+       mov     120($context),%rax      # pull context->Rax
+       mov     248($context),%rbx      # pull context->Rip
+
+       mov     8($disp),%rsi           # disp->ImageBase
+       mov     56($disp),%r11          # disp->HandlerData
+
+       mov     0(%r11),%r10d           # HandlerData[0]
+       lea     (%rsi,%r10),%r10        # prologue lable
+       cmp     %r10,%rbx               # context->Rip<prologue label
+       jb      .Lcommon_seh_tail
+
+       mov     152($context),%rax      # pull context->Rsp
+
+       mov     4(%r11),%r10d           # HandlerData[1]
+       lea     (%rsi,%r10),%r10        # epilogue label
+       cmp     %r10,%rbx               # context->Rip>=epilogue label
+       jae     .Lcommon_seh_tail
+
+       lea     0x60(%rax),%rsi         # %xmm save area
+       lea     512($context),%rdi      # & context.Xmm6
+       mov     \$20,%ecx               # 10*sizeof(%xmm0)/sizeof(%rax)
+       .long   0xa548f3fc              # cld; rep movsq
+       lea     0x68+160(%rax),%rax     # adjust stack pointer
+
+       jmp     .Lcommon_seh_tail
+.size  xts_se_handler,.-xts_se_handler

 ___
 $code.=<<___;
 .type  cbc_se_handler,\@abi-omnipotent
 ___
 $code.=<<___;
 .type  cbc_se_handler,\@abi-omnipotent


@@ -1372,7 +2880,7 @@ cbc_se_handler:
 
        lea     .Lcbc_decrypt(%rip),%r10
        cmp     %r10,%rbx               # context->Rip<"prologue" label
 
        lea     .Lcbc_decrypt(%rip),%r10
        cmp     %r10,%rbx               # context->Rip<"prologue" label

-       jb      .Lin_cbc_prologue
+       jb      .Lcommon_seh_tail

 
        lea     .Lcbc_decrypt_body(%rip),%r10
        cmp     %r10,%rbx               # context->Rip<cbc_decrypt_body
 
        lea     .Lcbc_decrypt_body(%rip),%r10
        cmp     %r10,%rbx               # context->Rip<cbc_decrypt_body


@@ -1380,26 +2888,25 @@ cbc_se_handler:
 
        lea     .Lcbc_ret(%rip),%r10
        cmp     %r10,%rbx               # context->Rip>="epilogue" label
 
        lea     .Lcbc_ret(%rip),%r10
        cmp     %r10,%rbx               # context->Rip>="epilogue" label

-       jae     .Lin_cbc_prologue
+       jae     .Lcommon_seh_tail

 
        lea     0(%rax),%rsi            # top of stack
        lea     512($context),%rdi      # &context.Xmm6
        mov     \$8,%ecx                # 4*sizeof(%xmm0)/sizeof(%rax)
        .long   0xa548f3fc              # cld; rep movsq
        lea     0x58(%rax),%rax         # adjust stack pointer
 
        lea     0(%rax),%rsi            # top of stack
        lea     512($context),%rdi      # &context.Xmm6
        mov     \$8,%ecx                # 4*sizeof(%xmm0)/sizeof(%rax)
        .long   0xa548f3fc              # cld; rep movsq
        lea     0x58(%rax),%rax         # adjust stack pointer

-       jmp     .Lin_cbc_prologue
+       jmp     .Lcommon_seh_tail

 
 .Lrestore_cbc_rax:
        mov     120($context),%rax
 
 .Lrestore_cbc_rax:
        mov     120($context),%rax

-.Lin_cbc_prologue:
+
+.Lcommon_seh_tail:

        mov     8(%rax),%rdi
        mov     16(%rax),%rsi
        mov     %rax,152($context)      # restore context->Rsp
        mov     %rsi,168($context)      # restore context->Rsi
        mov     %rdi,176($context)      # restore context->Rdi
 
        mov     8(%rax),%rdi
        mov     16(%rax),%rsi
        mov     %rax,152($context)      # restore context->Rsp
        mov     %rsi,168($context)      # restore context->Rsi
        mov     %rdi,176($context)      # restore context->Rdi
 

-.Lcommon_seh_exit:
-

        mov     40($disp),%rdi          # disp->ContextRecord
        mov     $context,%rsi           # context
        mov     \$154,%ecx              # sizeof(CONTEXT)
        mov     40($disp),%rdi          # disp->ContextRecord
        mov     $context,%rsi           # context
        mov     \$154,%ecx              # sizeof(CONTEXT)


@@ -1452,6 +2959,14 @@ $code.=<<___ if ($PREFIX eq "aesni");
        .rva    .LSEH_begin_aesni_ctr32_encrypt_blocks
        .rva    .LSEH_end_aesni_ctr32_encrypt_blocks
        .rva    .LSEH_info_ctr32
        .rva    .LSEH_begin_aesni_ctr32_encrypt_blocks
        .rva    .LSEH_end_aesni_ctr32_encrypt_blocks
        .rva    .LSEH_info_ctr32

+
+       .rva    .LSEH_begin_aesni_xts_encrypt
+       .rva    .LSEH_end_aesni_xts_encrypt
+       .rva    .LSEH_info_xts_enc
+
+       .rva    .LSEH_begin_aesni_xts_decrypt
+       .rva    .LSEH_end_aesni_xts_decrypt
+       .rva    .LSEH_info_xts_dec

 ___
 $code.=<<___;
        .rva    .LSEH_begin_${PREFIX}_cbc_encrypt
 ___
 $code.=<<___;
        .rva    .LSEH_begin_${PREFIX}_cbc_encrypt


@@ -1483,6 +2998,14 @@ $code.=<<___ if ($PREFIX eq "aesni");
 .LSEH_info_ctr32:
        .byte   9,0,0,0
        .rva    ctr32_se_handler
 .LSEH_info_ctr32:
        .byte   9,0,0,0
        .rva    ctr32_se_handler

+.LSEH_info_xts_enc:
+       .byte   9,0,0,0
+       .rva    xts_se_handler
+       .rva    .Lxts_enc_body,.Lxts_enc_epilogue       # HandlerData[]
+.LSEH_info_xts_dec:
+       .byte   9,0,0,0
+       .rva    xts_se_handler
+       .rva    .Lxts_dec_body,.Lxts_dec_epilogue       # HandlerData[]

 ___
 $code.=<<___;
 .LSEH_info_cbc:
 ___
 $code.=<<___;
 .LSEH_info_cbc: