sha512-x86_64.pl: >5% better performance.
authorAndy Polyakov <appro@openssl.org>
Mon, 28 May 2012 17:47:15 +0000 (17:47 +0000)
committerAndy Polyakov <appro@openssl.org>
Mon, 28 May 2012 17:47:15 +0000 (17:47 +0000)
crypto/sha/asm/sha512-x86_64.pl

index f611a2d..6a57b65 100755 (executable)
@@ -1,7 +1,7 @@
 #!/usr/bin/env perl
 #
 # ====================================================================
-# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
 # project. Rights for redistribution and usage in source and binary
 # forms are granted according to the OpenSSL license.
 # ====================================================================
 # contrary, 64-bit version, sha512_block, is ~30% *slower* than 32-bit
 # sha256_block:-( This is presumably because 64-bit shifts/rotates
 # apparently are not atomic instructions, but implemented in microcode.
+#
+# May 2012.
+#
+# Optimization including one of Pavel Semjanov's ideas, alternative
+# Maj, resulted in >=5% improvement on most CPUs, 20% on P4.
 
 $flavour = shift;
 $output  = shift;
@@ -59,7 +64,7 @@ if ($output =~ /512/) {
        $SZ=8;
        @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
                                        "%r8", "%r9", "%r10","%r11");
-       ($T1,$a0,$a1,$a2)=("%r12","%r13","%r14","%r15");
+       ($T1,$a0,$a1,$a2,$a3)=("%r12","%r13","%r14","%r15","%rdi");
        @Sigma0=(28,34,39);
        @Sigma1=(14,18,41);
        @sigma0=(1,  8, 7);
@@ -71,7 +76,7 @@ if ($output =~ /512/) {
        $SZ=4;
        @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
                                        "%r8d","%r9d","%r10d","%r11d");
-       ($T1,$a0,$a1,$a2)=("%r12d","%r13d","%r14d","%r15d");
+       ($T1,$a0,$a1,$a2,$a3)=("%r12d","%r13d","%r14d","%r15d","%edi");
        @Sigma0=( 2,13,22);
        @Sigma1=( 6,11,25);
        @sigma0=( 7,18, 3);
@@ -79,8 +84,7 @@ if ($output =~ /512/) {
        $rounds=64;
 }
 
-$ctx="%rdi";   # 1st arg
-$round="%rdi"; # zaps $ctx
+$ctx="%rdi";   # 1st arg, zapped by $a3
 $inp="%rsi";   # 2nd arg
 $Tbl="%rbp";
 
@@ -97,68 +101,71 @@ sub ROUND_00_15()
 $code.=<<___;
        ror     \$`$Sigma1[2]-$Sigma1[1]`,$a0
        mov     $f,$a2
-       mov     $T1,`$SZ*($i&0xf)`(%rsp)
 
        ror     \$`$Sigma0[2]-$Sigma0[1]`,$a1
        xor     $e,$a0
        xor     $g,$a2                  # f^g
 
        ror     \$`$Sigma1[1]-$Sigma1[0]`,$a0
-       add     $h,$T1                  # T1+=h
        xor     $a,$a1
-
-       add     ($Tbl,$round,$SZ),$T1   # T1+=K[round]
        and     $e,$a2                  # (f^g)&e
-       mov     $b,$h
+
+       mov     $T1,`$SZ*($i&0xf)`(%rsp)
+       add     $h,$T1                  # T1+=h
+       xor     $g,$a2                  # Ch(e,f,g)=((f^g)&e)^g
 
        ror     \$`$Sigma0[1]-$Sigma0[0]`,$a1
+       add     $a2,$T1                 # T1+=Ch(e,f,g)
        xor     $e,$a0
-       xor     $g,$a2                  # Ch(e,f,g)=((f^g)&e)^g
 
-       xor     $c,$h                   # b^c
+       add     ($Tbl),$T1              # T1+=K[round]
+       mov     $a,$a2
        xor     $a,$a1
-       add     $a2,$T1                 # T1+=Ch(e,f,g)
-       mov     $b,$a2
 
        ror     \$$Sigma1[0],$a0        # Sigma1(e)
-       and     $a,$h                   # h=(b^c)&a
-       and     $c,$a2                  # b&c
+       xor     $b,$a2                  # a^b, b^c in next round
+       mov     $b,$h
 
        ror     \$$Sigma0[0],$a1        # Sigma0(a)
+       and     $a2,$a3
        add     $a0,$T1                 # T1+=Sigma1(e)
-       add     $a2,$h                  # h+=b&c (completes +=Maj(a,b,c)
 
+       xor     $a3,$h                  # h=Maj(a,b,c)=Ch(a^b,c,b)
        add     $T1,$d                  # d+=T1
        add     $T1,$h                  # h+=T1
-       lea     1($round),$round        # round++
+___
+$code.=<<___ if ($i>=15);
+       mov     `$SZ*(($i+2)&0xf)`(%rsp),$a0
+___
+$code.=<<___;
+       lea     $SZ($Tbl),$Tbl          # round++
        add     $a1,$h                  # h+=Sigma0(a)
 
 ___
+       ($a2,$a3) = ($a3,$a2);
 }
 
 sub ROUND_16_XX()
 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 
 $code.=<<___;
-       mov     `$SZ*(($i+1)&0xf)`(%rsp),$a0
+       #mov    `$SZ*(($i+1)&0xf)`(%rsp),$a0
        mov     `$SZ*(($i+14)&0xf)`(%rsp),$a1
+
        mov     $a0,$T1
+       ror     \$`$sigma0[1]-$sigma0[0]`,$a0
        mov     $a1,$a2
+       ror     \$`$sigma1[1]-$sigma1[0]`,$a1
 
-       ror     \$`$sigma0[1]-$sigma0[0]`,$T1
-       xor     $a0,$T1
-       shr     \$$sigma0[2],$a0
+       xor     $T1,$a0
+       shr     \$$sigma0[2],$T1
+       ror     \$$sigma0[0],$a0
+       xor     $a2,$a1
+       shr     \$$sigma1[2],$a2
 
-       ror     \$$sigma0[0],$T1
-       xor     $T1,$a0                 # sigma0(X[(i+1)&0xf])
-       mov     `$SZ*(($i+9)&0xf)`(%rsp),$T1
-
-       ror     \$`$sigma1[1]-$sigma1[0]`,$a2
-       xor     $a1,$a2
-       shr     \$$sigma1[2],$a1
-
-       ror     \$$sigma1[0],$a2
-       add     $a0,$T1
+       xor     $a0,$T1                 # sigma0(X[(i+1)&0xf])
+       ror     \$$sigma1[0],$a1
+       add     `$SZ*(($i+9)&0xf)`(%rsp),$T1
        xor     $a2,$a1                 # sigma1(X[(i+14)&0xf])
 
        add     `$SZ*($i&0xf)`(%rsp),$T1
@@ -193,8 +200,6 @@ $func:
        mov     %r11,$_rsp              # save copy of %rsp
 .Lprologue:
 
-       lea     $TABLE(%rip),$Tbl
-
        mov     $SZ*0($ctx),$A
        mov     $SZ*1($ctx),$B
        mov     $SZ*2($ctx),$C
@@ -207,7 +212,9 @@ $func:
 
 .align 16
 .Lloop:
-       xor     $round,$round
+       mov     $B,$a3
+       lea     $TABLE(%rip),$Tbl
+       xor     $C,$a3                  # magic
 ___
        for($i=0;$i<16;$i++) {
                $code.="        mov     $SZ*$i($inp),$T1\n";
@@ -228,8 +235,8 @@ ___
        }
 
 $code.=<<___;
-       cmp     \$$rounds,$round
-       j     .Lrounds_16_xx
+       testl   \$-1,($Tbl)
+       jnz     .Lrounds_16_xx
 
        mov     $_ctx,$ctx
        lea     16*$SZ($inp),$inp
@@ -289,6 +296,8 @@ $TABLE:
        .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
        .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
        .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
+       .long   0
+       .asciz  "SHA256 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 ___
 } else {
 $code.=<<___;
@@ -335,6 +344,8 @@ $TABLE:
        .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
        .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
        .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
+       .long   0
+       .asciz  "SHA512 block transfort for x86_64, CRYPTOGAMS by <appro\@openssl.org>
 ___
 }