projects / openssl.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
Skylake performance results.
[openssl.git] / crypto / aes / asm / aesni-sha1-x86_64.pl
diff --git a/crypto/aes/asm/aesni-sha1-x86_64.pl b/crypto/aes/asm/aesni-sha1-x86_64.pl
index ff0b068229cb4e6f32d939f754345b5a1b269b38..952b4cfdd668297ead9176cbc5ee68f3b3380fee 100644 (file)
--- a/crypto/aes/asm/aesni-sha1-x86_64.pl
+++ b/crypto/aes/asm/aesni-sha1-x86_64.pl
@@ -25,6 +25,7 @@
 # Sandy Bridge 5.05[+5.0(6.1)] 10.06(11.15)    5.98(7.05)  +68%(+58%)
 # Ivy Bridge   5.05[+4.6]      9.65            5.54        +74%
 # Haswell      4.43[+3.6(4.2)] 8.00(8.58)      4.55(5.21)  +75%(+65%)
 # Sandy Bridge 5.05[+5.0(6.1)] 10.06(11.15)    5.98(7.05)  +68%(+58%)
 # Ivy Bridge   5.05[+4.6]      9.65            5.54        +74%
 # Haswell      4.43[+3.6(4.2)] 8.00(8.58)      4.55(5.21)  +75%(+65%)
+# Skylake      2.63[+3.5(4.1)] 6.17(6.69)      4.23(4.44)  +46%(+51%)
 # Bulldozer    5.77[+6.0]      11.72           6.37        +84%
 #
 #              AES-192-CBC
 # Bulldozer    5.77[+6.0]      11.72           6.37        +84%
 #
 #              AES-192-CBC
@@ -39,6 +40,7 @@
 # Sandy Bridge 7.05            12.06(13.15)    7.12(7.72)  +69%(+70%)
 # Ivy Bridge   7.05            11.65           7.12        +64%
 # Haswell      6.19            9.76(10.34)     6.21(6.25)  +57%(+65%)
 # Sandy Bridge 7.05            12.06(13.15)    7.12(7.72)  +69%(+70%)
 # Ivy Bridge   7.05            11.65           7.12        +64%
 # Haswell      6.19            9.76(10.34)     6.21(6.25)  +57%(+65%)
+# Skylake      3.62            7.16(7.68)      4.56(4.76)  +57%(+61$)
 # Bulldozer    8.00            13.95           8.25        +69%
 #
 # (*)  There are two code paths: SSSE3 and AVX. See sha1-568.pl for
 # Bulldozer    8.00            13.95           8.25        +69%
 #
 # (*)  There are two code paths: SSSE3 and AVX. See sha1-568.pl for
@@ -94,6 +96,9 @@ $avx=1 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
 $avx=1 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
           `ml64 2>&1` =~ /Version ([0-9]+)\./ &&
           $1>=10);
 $avx=1 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
           `ml64 2>&1` =~ /Version ([0-9]+)\./ &&
           $1>=10);
+$avx=1 if (!$avx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/ && $2>=3.0);
+
+$shaext=1;     ### set to zero if compiling for 1.0.1
 
 $stitched_decrypt=0;
 
 
 $stitched_decrypt=0;
 
@@ -119,6 +124,8 @@ aesni_cbc_sha1_enc:
        # caller should check for SSSE3 and AES-NI bits
        mov     OPENSSL_ia32cap_P+0(%rip),%r10d
        mov     OPENSSL_ia32cap_P+4(%rip),%r11
        # caller should check for SSSE3 and AES-NI bits
        mov     OPENSSL_ia32cap_P+0(%rip),%r10d
        mov     OPENSSL_ia32cap_P+4(%rip),%r11
+___
+$code.=<<___ if ($shaext);
        bt      \$61,%r11               # check SHA bit
        jc      aesni_cbc_sha1_enc_shaext
 ___
        bt      \$61,%r11               # check SHA bit
        jc      aesni_cbc_sha1_enc_shaext
 ___
@@ -722,7 +729,7 @@ ___
                                                if ($stitched_decrypt) {{{
 # reset
 ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
                                                if ($stitched_decrypt) {{{
 # reset
 ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
-$j=$jj=$r=$sn=$rx=0;
+$j=$jj=$r=$rx=0;
 $Xi=4;
 
 # reassign for Atom Silvermont (see above)
 $Xi=4;
 
 # reassign for Atom Silvermont (see above)
@@ -990,7 +997,7 @@ $code.=<<___;
 .size  aesni256_cbc_sha1_dec_ssse3,.-aesni256_cbc_sha1_dec_ssse3
 ___
                                                }}}
 .size  aesni256_cbc_sha1_dec_ssse3,.-aesni256_cbc_sha1_dec_ssse3
 ___
                                                }}}
-$j=$jj=$r=$sn=$rx=0;
+$j=$jj=$r=$rx=0;
 
 if ($avx) {
 my ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
 
 if ($avx) {
 my ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
@@ -1436,7 +1443,7 @@ ___
 # reset
 ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
 
 # reset
 ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
 
-$j=$jj=$r=$sn=$rx=0;
+$j=$jj=$r=$rx=0;
 $Xi=4;
 
 @aes256_dec = (
 $Xi=4;
 
 @aes256_dec = (
@@ -1657,7 +1664,7 @@ K_XX_XX:
 .asciz "AESNI-CBC+SHA1 stitch for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 .align 64
 ___
 .asciz "AESNI-CBC+SHA1 stitch for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 .align 64
 ___
-                                               {{{
+                                               if ($shaext) {{{
 ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
 
 $rounds="%r11d";
 ($in0,$out,$len,$key,$ivp,$ctx,$inp)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9","%r10");
 
 $rounds="%r11d";
@@ -1676,7 +1683,7 @@ aesni_cbc_sha1_enc_shaext:
        mov     `($win64?56:8)`(%rsp),$inp      # load 7th argument
 ___
 $code.=<<___ if ($win64);
        mov     `($win64?56:8)`(%rsp),$inp      # load 7th argument
 ___
 $code.=<<___ if ($win64);
-       lea     `-8-4*16`(%rsp),%rsp
+       lea     `-8-10*16`(%rsp),%rsp
        movaps  %xmm6,-8-10*16(%rax)
        movaps  %xmm7,-8-9*16(%rax)
        movaps  %xmm8,-8-8*16(%rax)
        movaps  %xmm6,-8-10*16(%rax)
        movaps  %xmm7,-8-9*16(%rax)
        movaps  %xmm8,-8-8*16(%rax)
@@ -1867,7 +1874,21 @@ ssse3_handler:
        lea     (%rsi,%r10),%r10        # epilogue label
        cmp     %r10,%rbx               # context->Rip>=epilogue label
        jae     .Lcommon_seh_tail
        lea     (%rsi,%r10),%r10        # epilogue label
        cmp     %r10,%rbx               # context->Rip>=epilogue label
        jae     .Lcommon_seh_tail
+___
+$code.=<<___ if ($shaext);
+       lea     aesni_cbc_sha1_enc_shaext(%rip),%r10
+       cmp     %r10,%rbx
+       jb      .Lseh_no_shaext
 
 
+       lea     (%rax),%rsi
+       lea     512($context),%rdi      # &context.Xmm6
+       mov     \$20,%ecx
+       .long   0xa548f3fc              # cld; rep movsq
+       lea     168(%rax),%rax          # adjust stack pointer
+       jmp     .Lcommon_seh_tail
+.Lseh_no_shaext:
+___
+$code.=<<___;
        lea     96(%rax),%rsi
        lea     512($context),%rdi      # &context.Xmm6
        mov     \$20,%ecx
        lea     96(%rax),%rsi
        lea     512($context),%rdi      # &context.Xmm6
        mov     \$20,%ecx
@@ -1939,6 +1960,11 @@ $code.=<<___ if ($avx);
        .rva    .LSEH_end_aesni_cbc_sha1_enc_avx
        .rva    .LSEH_info_aesni_cbc_sha1_enc_avx
 ___
        .rva    .LSEH_end_aesni_cbc_sha1_enc_avx
        .rva    .LSEH_info_aesni_cbc_sha1_enc_avx
 ___
+$code.=<<___ if ($shaext);
+       .rva    .LSEH_begin_aesni_cbc_sha1_enc_shaext
+       .rva    .LSEH_end_aesni_cbc_sha1_enc_shaext
+       .rva    .LSEH_info_aesni_cbc_sha1_enc_shaext
+___
 $code.=<<___;
 .section       .xdata
 .align 8
 $code.=<<___;
 .section       .xdata
 .align 8
@@ -1953,6 +1979,12 @@ $code.=<<___ if ($avx);
        .rva    ssse3_handler
        .rva    .Lprologue_avx,.Lepilogue_avx           # HandlerData[]
 ___
        .rva    ssse3_handler
        .rva    .Lprologue_avx,.Lepilogue_avx           # HandlerData[]
 ___
+$code.=<<___ if ($shaext);
+.LSEH_info_aesni_cbc_sha1_enc_shaext:
+       .byte   9,0,0,0
+       .rva    ssse3_handler
+       .rva    .Lprologue_shaext,.Lepilogue_shaext     # HandlerData[]
+___
 }
 
 ####################################################################
 }
 
 ####################################################################
Unnamed repository; edit this file 'description' to name the repository.