bn/asm/rsax-avx2.pl: minor optimization [for Decoded ICache].
authorAndy Polyakov <appro@openssl.org>
Thu, 10 Oct 2013 21:06:43 +0000 (23:06 +0200)
committerAndy Polyakov <appro@openssl.org>
Thu, 10 Oct 2013 21:06:43 +0000 (23:06 +0200)
crypto/bn/asm/rsaz-avx2.pl

index 2d9e2399125d1e44092933623cb4b80bc95a6a7e..09c45b0ec2d2af75be1611aa929303cd9801738e 100755 (executable)
@@ -48,7 +48,7 @@
 # +10% improvement by <appro@openssl.org>
 #
 # rsa2048 sign/sec     OpenSSL 1.0.1   scalar(*)       this
-# 2GHz Haswell         544             632/+16%        947/+74%
+# 2.3GHz Haswell       621             732/+18%        1112/+79%
 #
 # (*)  if system doesn't support AVX2, for reference purposes;
 
@@ -125,6 +125,7 @@ my $FrameSize=32*18+32*8;   # place for A^2 and 2*A
 my $aap=$r0;
 my $tp0="%rbx";
 my $tp1=$r3;
+my $tpa=$tmp;
 
 $np="%r13";                    # reassigned argument
 
@@ -263,7 +264,7 @@ $code.=<<___;
         vpbroadcastq   32*2-128($ap), $B1
         vmovdqu        $ACC9, 32*17-448($tp1)
 
-       xor     $tmp, $tmp
+       mov     $ap, $tpa
        mov     \$4, $i
        jmp     .Lsqr_entry_1024
 ___
@@ -272,38 +273,29 @@ $TEMP2=$Y2;
 $code.=<<___;
 .align 32
 .LOOP_SQR_1024:
-       vmovdqu         32*0(%rsp,$tmp), $TEMP0 # 32*0-192($tp0,$tmp)
-       vmovdqu         32*1(%rsp,$tmp), $TEMP1 # 32*1-192($tp0,$tmp)
-        vpbroadcastq   32*1-128($ap,$tmp), $B2
+        vpbroadcastq   32*1-128($tpa), $B2
        vpmuludq        32*0-128($ap), $B1, $ACC0
-       vmovdqu         32*2-192($tp0,$tmp), $TEMP2
-       vpaddq          $TEMP0, $ACC0, $ACC0
+       vpaddq          32*0-192($tp0), $ACC0, $ACC0
        vpmuludq        32*0-128($aap), $B1, $ACC1
-       vmovdqu         32*3-192($tp0,$tmp), $TEMP0
-       vpaddq          $TEMP1, $ACC1, $ACC1
+       vpaddq          32*1-192($tp0), $ACC1, $ACC1
        vpmuludq        32*1-128($aap), $B1, $ACC2
-       vmovdqu         32*4-192($tp0,$tmp), $TEMP1
-       vpaddq          $TEMP2, $ACC2, $ACC2
+       vpaddq          32*2-192($tp0), $ACC2, $ACC2
        vpmuludq        32*2-128($aap), $B1, $ACC3
-       vmovdqu         32*5-192($tp0,$tmp), $TEMP2
-       vpaddq          $TEMP0, $ACC3, $ACC3
+       vpaddq          32*3-192($tp0), $ACC3, $ACC3
        vpmuludq        32*3-128($aap), $B1, $ACC4
-       vmovdqu         32*6-192($tp0,$tmp), $TEMP0
-       vpaddq          $TEMP1, $ACC4, $ACC4
+       vpaddq          32*4-192($tp0), $ACC4, $ACC4
        vpmuludq        32*4-128($aap), $B1, $ACC5
-       vmovdqu         32*7-192($tp0,$tmp), $TEMP1
-       vpaddq          $TEMP2, $ACC5, $ACC5
+       vpaddq          32*5-192($tp0), $ACC5, $ACC5
        vpmuludq        32*5-128($aap), $B1, $ACC6
-       vmovdqu         32*8-192($tp0,$tmp), $TEMP2
-       vpaddq          $TEMP0, $ACC6, $ACC6
+       vpaddq          32*6-192($tp0), $ACC6, $ACC6
        vpmuludq        32*6-128($aap), $B1, $ACC7
-       vpaddq          $TEMP1, $ACC7, $ACC7
+       vpaddq          32*7-192($tp0), $ACC7, $ACC7
        vpmuludq        32*7-128($aap), $B1, $ACC8
-        vpbroadcastq   32*2-128($ap,$tmp), $B1
-       vpaddq          $TEMP2, $ACC8, $ACC8
+        vpbroadcastq   32*2-128($tpa), $B1
+       vpaddq          32*8-192($tp0), $ACC8, $ACC8
 .Lsqr_entry_1024:
-       vmovdqu         $ACC0, 32*0(%rsp,$tmp)  # 32*0-192($tp0,$tmp)
-       vmovdqu         $ACC1, 32*1(%rsp,$tmp)  # 32*1-192($tp0,$tmp)
+       vmovdqu         $ACC0, 32*0-192($tp0)
+       vmovdqu         $ACC1, 32*1-192($tp0)
 
        vpmuludq        32*1-128($ap), $B2, $TEMP0
        vpaddq          $TEMP0, $ACC2, $ACC2
@@ -316,16 +308,15 @@ $code.=<<___;
        vpmuludq        32*4-128($aap), $B2, $TEMP1
        vpaddq          $TEMP1, $ACC6, $ACC6
        vpmuludq        32*5-128($aap), $B2, $TEMP2
-       vmovdqu         32*9-192($tp0,$tmp), $TEMP1
        vpaddq          $TEMP2, $ACC7, $ACC7
        vpmuludq        32*6-128($aap), $B2, $TEMP0
        vpaddq          $TEMP0, $ACC8, $ACC8
        vpmuludq        32*7-128($aap), $B2, $ACC0
-        vpbroadcastq   32*3-128($ap,$tmp), $B2
-       vpaddq          $TEMP1, $ACC0, $ACC0
+        vpbroadcastq   32*3-128($tpa), $B2
+       vpaddq          32*9-192($tp0), $ACC0, $ACC0
 
-       vmovdqu         $ACC2, 32*2-192($tp0,$tmp)
-       vmovdqu         $ACC3, 32*3-192($tp0,$tmp)
+       vmovdqu         $ACC2, 32*2-192($tp0)
+       vmovdqu         $ACC3, 32*3-192($tp0)
 
        vpmuludq        32*2-128($ap), $B1, $TEMP2
        vpaddq          $TEMP2, $ACC4, $ACC4
@@ -336,16 +327,15 @@ $code.=<<___;
        vpmuludq        32*4-128($aap), $B1, $TEMP2
        vpaddq          $TEMP2, $ACC7, $ACC7
        vpmuludq        32*5-128($aap), $B1, $TEMP0
-       vmovdqu         32*10-448($tp1,$tmp), $TEMP2
        vpaddq          $TEMP0, $ACC8, $ACC8
        vpmuludq        32*6-128($aap), $B1, $TEMP1
        vpaddq          $TEMP1, $ACC0, $ACC0
        vpmuludq        32*7-128($aap), $B1, $ACC1
-        vpbroadcastq   32*4-128($ap,$tmp), $B1
-       vpaddq          $TEMP2, $ACC1, $ACC1
+        vpbroadcastq   32*4-128($tpa), $B1
+       vpaddq          32*10-448($tp1), $ACC1, $ACC1
 
-       vmovdqu         $ACC4, 32*4-192($tp0,$tmp)
-       vmovdqu         $ACC5, 32*5-192($tp0,$tmp)
+       vmovdqu         $ACC4, 32*4-192($tp0)
+       vmovdqu         $ACC5, 32*5-192($tp0)
 
        vpmuludq        32*3-128($ap), $B2, $TEMP0
        vpaddq          $TEMP0, $ACC6, $ACC6
@@ -354,74 +344,70 @@ $code.=<<___;
        vpmuludq        32*4-128($aap), $B2, $TEMP2
        vpaddq          $TEMP2, $ACC8, $ACC8
        vpmuludq        32*5-128($aap), $B2, $TEMP0
-       vmovdqu         32*11-448($tp1,$tmp), $TEMP2
        vpaddq          $TEMP0, $ACC0, $ACC0
        vpmuludq        32*6-128($aap), $B2, $TEMP1
        vpaddq          $TEMP1, $ACC1, $ACC1
        vpmuludq        32*7-128($aap), $B2, $ACC2
-        vpbroadcastq   32*5-128($ap,$tmp), $B2
-       vpaddq          $TEMP2, $ACC2, $ACC2    
+        vpbroadcastq   32*5-128($tpa), $B2
+       vpaddq          32*11-448($tp1), $ACC2, $ACC2   
 
-       vmovdqu         $ACC6, 32*6-192($tp0,$tmp)
-       vmovdqu         $ACC7, 32*7-192($tp0,$tmp)
+       vmovdqu         $ACC6, 32*6-192($tp0)
+       vmovdqu         $ACC7, 32*7-192($tp0)
 
        vpmuludq        32*4-128($ap), $B1, $TEMP0
        vpaddq          $TEMP0, $ACC8, $ACC8
        vpmuludq        32*4-128($aap), $B1, $TEMP1
        vpaddq          $TEMP1, $ACC0, $ACC0
        vpmuludq        32*5-128($aap), $B1, $TEMP2
-       vmovdqu         32*12-448($tp1,$tmp), $TEMP1
        vpaddq          $TEMP2, $ACC1, $ACC1
        vpmuludq        32*6-128($aap), $B1, $TEMP0
        vpaddq          $TEMP0, $ACC2, $ACC2
        vpmuludq        32*7-128($aap), $B1, $ACC3
-        vpbroadcastq   32*6-128($ap,$tmp), $B1
-       vpaddq          $TEMP1, $ACC3, $ACC3
+        vpbroadcastq   32*6-128($tpa), $B1
+       vpaddq          32*12-448($tp1), $ACC3, $ACC3
 
-       vmovdqu         $ACC8, 32*8-192($tp0,$tmp)
-       vmovdqu         $ACC0, 32*9-192($tp0,$tmp)
+       vmovdqu         $ACC8, 32*8-192($tp0)
+       vmovdqu         $ACC0, 32*9-192($tp0)
+       lea             8($tp0), $tp0
 
        vpmuludq        32*5-128($ap), $B2, $TEMP2
        vpaddq          $TEMP2, $ACC1, $ACC1
        vpmuludq        32*5-128($aap), $B2, $TEMP0
-       vmovdqu         32*13-448($tp1,$tmp), $TEMP2
        vpaddq          $TEMP0, $ACC2, $ACC2
        vpmuludq        32*6-128($aap), $B2, $TEMP1
        vpaddq          $TEMP1, $ACC3, $ACC3
        vpmuludq        32*7-128($aap), $B2, $ACC4
-        vpbroadcastq   32*7-128($ap,$tmp), $B2
-       vpaddq          $TEMP2, $ACC4, $ACC4
+        vpbroadcastq   32*7-128($tpa), $B2
+       vpaddq          32*13-448($tp1), $ACC4, $ACC4
 
-       vmovdqu         $ACC1, 32*10-448($tp1,$tmp)
-       vmovdqu         $ACC2, 32*11-448($tp1,$tmp)
+       vmovdqu         $ACC1, 32*10-448($tp1)
+       vmovdqu         $ACC2, 32*11-448($tp1)
 
        vpmuludq        32*6-128($ap), $B1, $TEMP0
-       vmovdqu         32*14-448($tp1,$tmp), $TEMP2
        vpaddq          $TEMP0, $ACC3, $ACC3
        vpmuludq        32*6-128($aap), $B1, $TEMP1
-        vpbroadcastq   32*8-128($ap,$tmp), $ACC0       # borrow $ACC0 for $B1
+        vpbroadcastq   32*8-128($tpa), $ACC0           # borrow $ACC0 for $B1
        vpaddq          $TEMP1, $ACC4, $ACC4
        vpmuludq        32*7-128($aap), $B1, $ACC5
-        vpbroadcastq   32*0+8-128($ap,$tmp), $B1       # for next iteration
-       vpaddq          $TEMP2, $ACC5, $ACC5
-       vmovdqu         32*15-448($tp1,$tmp), $TEMP1
+        vpbroadcastq   32*0+8-128($tpa), $B1           # for next iteration
+       vpaddq          32*14-448($tp1), $ACC5, $ACC5
 
-       vmovdqu         $ACC3, 32*12-448($tp1,$tmp)
-       vmovdqu         $ACC4, 32*13-448($tp1,$tmp)
+       vmovdqu         $ACC3, 32*12-448($tp1)
+       vmovdqu         $ACC4, 32*13-448($tp1)
+       lea             8($tpa), $tpa
 
        vpmuludq        32*7-128($ap), $B2, $TEMP0
-       vmovdqu         32*16-448($tp1,$tmp), $TEMP2
        vpaddq          $TEMP0, $ACC5, $ACC5
        vpmuludq        32*7-128($aap), $B2, $ACC6
-       vpaddq          $TEMP1, $ACC6, $ACC6
+       vpaddq          32*15-448($tp1), $ACC6, $ACC6
 
        vpmuludq        32*8-128($ap), $ACC0, $ACC7
-       vmovdqu         $ACC5, 32*14-448($tp1,$tmp)
-       vpaddq          $TEMP2, $ACC7, $ACC7
-       vmovdqu         $ACC6, 32*15-448($tp1,$tmp)
-       vmovdqu         $ACC7, 32*16-448($tp1,$tmp)
+       vmovdqu         $ACC5, 32*14-448($tp1)
+       vpaddq          32*16-448($tp1), $ACC7, $ACC7
+       vmovdqu         $ACC6, 32*15-448($tp1)
+       vmovdqu         $ACC7, 32*16-448($tp1)
+       lea             8($tp1), $tp1
 
-       lea     8($tmp), $tmp
        dec     $i        
        jnz     .LOOP_SQR_1024
 ___
@@ -432,9 +418,10 @@ $TEMP3 = $Y1;
 $TEMP4 = $Y2;
 $code.=<<___;
        #we need to fix indexes 32-39 to avoid overflow
-       vmovdqu         32*8-192($tp0), $ACC8
-       vmovdqu         32*9-192($tp0), $ACC1
-       vmovdqu         32*10-448($tp1), $ACC2
+       vmovdqu         32*8(%rsp), $ACC8               # 32*8-192($tp0),
+       vmovdqu         32*9(%rsp), $ACC1               # 32*9-192($tp0)
+       vmovdqu         32*10(%rsp), $ACC2              # 32*10-192($tp0)
+       lea             192(%rsp), $tp0                 # 64+128=192
 
        vpsrlq          \$29, $ACC8, $TEMP1
        vpand           $AND_MASK, $ACC8, $ACC8
@@ -452,7 +439,7 @@ $code.=<<___;
        vpaddq          $TEMP1, $ACC1, $ACC1
        vpaddq          $TEMP2, $ACC2, $ACC2
        vmovdqu         $ACC1, 32*9-192($tp0)
-       vmovdqu         $ACC2, 32*10-448($tp1)
+       vmovdqu         $ACC2, 32*10-192($tp0)
 
        mov     (%rsp), %rax
        mov     8(%rsp), $r1
@@ -502,13 +489,15 @@ $code.=<<___;
         mov    %rax, %rdx
         imulq  -128($np), %rax
        vpaddq          $TEMP0, $ACC1, $ACC1
-       vpmuludq        32*2-128($np), $Y1, $TEMP1
         add    %rax, $r1
+       vpmuludq        32*2-128($np), $Y1, $TEMP1
         mov    %rdx, %rax
         imulq  8-128($np), %rax
        vpaddq          $TEMP1, $ACC2, $ACC2
        vpmuludq        32*3-128($np), $Y1, $TEMP2
+        .byte  0x67
         add    %rax, $r2
+        .byte  0x67
         mov    %rdx, %rax
         imulq  16-128($np), %rax
         shr    \$29, $r1
@@ -528,17 +517,17 @@ $code.=<<___;
        vpaddq          $TEMP0, $ACC7, $ACC7
        vpmuludq        32*8-128($np), $Y1, $TEMP1
         vmovd  %eax, $Y1
-        vmovdqu        32*1-8-128($np), $TEMP2
+        #vmovdqu       32*1-8-128($np), $TEMP2         # moved below
        vpaddq          $TEMP1, $ACC8, $ACC8
-        vmovdqu        32*2-8-128($np), $TEMP0
+        #vmovdqu       32*2-8-128($np), $TEMP0         # moved below
         vpbroadcastq   $Y1, $Y1
 
-       vpmuludq        $Y2, $TEMP2, $TEMP2
+       vpmuludq        32*1-8-128($np), $Y2, $TEMP2    # see above
        vmovdqu         32*3-8-128($np), $TEMP1
         mov    %rax, %rdx
         imulq  -128($np), %rax
        vpaddq          $TEMP2, $ACC1, $ACC1
-       vpmuludq        $Y2, $TEMP0, $TEMP0
+       vpmuludq        32*2-8-128($np), $Y2, $TEMP0    # see above
        vmovdqu         32*4-8-128($np), $TEMP2
         add    %rax, $r2
         mov    %rdx, %rax
@@ -552,11 +541,12 @@ $code.=<<___;
        vpaddq          $TEMP1, $ACC3, $ACC3
        vpmuludq        $Y2, $TEMP2, $TEMP2
        vmovdqu         32*6-8-128($np), $TEMP1
+        .byte  0x67
         mov    %rax, $r3
         imull  $n0, %eax
        vpaddq          $TEMP2, $ACC4, $ACC4
        vpmuludq        $Y2, $TEMP0, $TEMP0
-       vmovdqu         32*7-8-128($np), $TEMP2
+       .byte   0xc4,0x41,0x7e,0x6f,0x9d,0x58,0x00,0x00,0x00    # vmovdqu               32*7-8-128($np), $TEMP2
         and    \$0x1fffffff, %eax
        vpaddq          $TEMP0, $ACC5, $ACC5
        vpmuludq        $Y2, $TEMP1, $TEMP1
@@ -584,11 +574,12 @@ $code.=<<___;
        vpaddq          $TEMP1, $ACC1, $ACC1
         vpmuludq       $Y2, $ACC0, $ACC0
        vpmuludq        $Y1, $TEMP2, $TEMP2
-       vmovdqu         32*4-16-128($np), $TEMP1
+       .byte   0xc4,0x41,0x7e,0x6f,0xb5,0xf0,0xff,0xff,0xff    # vmovdqu               32*4-16-128($np), $TEMP1
         vpaddq         $ACC1, $ACC0, $ACC0
        vpaddq          $TEMP2, $ACC2, $ACC2
        vpmuludq        $Y1, $TEMP0, $TEMP0
        vmovdqu         32*5-16-128($np), $TEMP2
+        .byte  0x67
         vmovq          $ACC0, %rax
         vmovdqu        $ACC0, (%rsp)           # transfer $r0-$r3
        vpaddq          $TEMP0, $ACC3, $ACC3
@@ -602,12 +593,12 @@ $code.=<<___;
        vmovdqu         32*8-16-128($np), $TEMP2
        vpaddq          $TEMP0, $ACC6, $ACC6
        vpmuludq        $Y1, $TEMP1, $TEMP1
-       vmovdqu         32*9-16-128($np), $TEMP0
         shr    \$29, $r3
+       vmovdqu         32*9-16-128($np), $TEMP0
+        add    $r3, %rax
        vpaddq          $TEMP1, $ACC7, $ACC7
        vpmuludq        $Y1, $TEMP2, $TEMP2
-        vmovdqu        32*2-24-128($np), $TEMP1
-        add    $r3, %rax
+        #vmovdqu       32*2-24-128($np), $TEMP1        # moved below
         mov    %rax, $r0
         imull  $n0, %eax
        vpaddq          $TEMP2, $ACC8, $ACC8
@@ -615,10 +606,11 @@ $code.=<<___;
         and    \$0x1fffffff, %eax
         vmovd  %eax, $Y1
         vmovdqu        32*3-24-128($np), $TEMP2
+       .byte   0x67
        vpaddq          $TEMP0, $ACC9, $ACC9
         vpbroadcastq   $Y1, $Y1
 
-       vpmuludq        $Y2, $TEMP1, $TEMP1
+       vpmuludq        32*2-24-128($np), $Y2, $TEMP1   # see above
        vmovdqu         32*4-24-128($np), $TEMP0
         mov    %rax, %rdx
         imulq  -128($np), %rax
@@ -629,6 +621,7 @@ $code.=<<___;
         add    %rax, $r0
         mov    %rdx, %rax
         imulq  8-128($np), %rax
+        .byte  0x67
         shr    \$29, $r0
         mov    16(%rsp), $r2
        vpaddq          $TEMP2, $ACC3, $ACC2
@@ -648,8 +641,8 @@ $code.=<<___;
        vmovdqu         32*8-24-128($np), $TEMP1
         mov    %rax, $r1
         imull  $n0, %eax
-       vpaddq          $TEMP2, $ACC6, $ACC5
        vpmuludq        $Y2, $TEMP0, $TEMP0
+       vpaddq          $TEMP2, $ACC6, $ACC5
        vmovdqu         32*9-24-128($np), $TEMP2
         and    \$0x1fffffff, %eax
        vpaddq          $TEMP0, $ACC7, $ACC6