[src/trunk]: src/crypto/external/bsd/openssl/dist/crypto When loading a globa...

[joerg <joerg%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/261b724eb83d
branches:  trunk
changeset: 808659:261b724eb83d
user:      joerg <joerg%NetBSD.org@localhost>
date:      Thu May 28 19:13:32 2015 +0000

description:
When loading a global variable from the GOT, it must be dereferenced to
obtain the actual value. Adjust various tests to extract the correct bits.

diffstat:

 crypto/external/bsd/openssl/dist/crypto/aes/asm/aes-x86_64.pl        |   4 +-
 crypto/external/bsd/openssl/dist/crypto/aes/asm/aesni-sha1-x86_64.pl |  10 ++----
 crypto/external/bsd/openssl/dist/crypto/bn/asm/x86_64-gf2m.pl        |   2 +-
 crypto/external/bsd/openssl/dist/crypto/rc4/asm/rc4-x86_64.pl        |   8 ++--
 crypto/external/bsd/openssl/dist/crypto/sha/asm/sha1-x86_64.pl       |  15 ++++------
 5 files changed, 16 insertions(+), 23 deletions(-)

diffs (111 lines):

diff -r e7b2b7ef3246 -r 261b724eb83d crypto/external/bsd/openssl/dist/crypto/aes/asm/aes-x86_64.pl
--- a/crypto/external/bsd/openssl/dist/crypto/aes/asm/aes-x86_64.pl     Thu May 28 17:31:38 2015 +0000
+++ b/crypto/external/bsd/openssl/dist/crypto/aes/asm/aes-x86_64.pl     Thu May 28 19:13:32 2015 +0000
@@ -1679,12 +1679,12 @@
        lea     .LAES_Td(%rip),$sbox
 .Lcbc_picked_te:
 
-       mov     OPENSSL_ia32cap_P(%rip),%r10d
+       mov     OPENSSL_ia32cap_P(%rip),%r10
        cmp     \$$speed_limit,%rdx
        jb      .Lcbc_slow_prologue
        test    \$15,%rdx
        jnz     .Lcbc_slow_prologue
-       bt      \$28,%r10d
+       bt      \$28,(%r10)
        jc      .Lcbc_slow_prologue
 
        # allocate aligned stack frame...
diff -r e7b2b7ef3246 -r 261b724eb83d crypto/external/bsd/openssl/dist/crypto/aes/asm/aesni-sha1-x86_64.pl
--- a/crypto/external/bsd/openssl/dist/crypto/aes/asm/aesni-sha1-x86_64.pl      Thu May 28 17:31:38 2015 +0000
+++ b/crypto/external/bsd/openssl/dist/crypto/aes/asm/aesni-sha1-x86_64.pl      Thu May 28 19:13:32 2015 +0000
@@ -89,15 +89,11 @@
 .align 16
 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),%r11d
+       mov     OPENSSL_ia32cap_P(%rip),%r10
 ___
 $code.=<<___ if ($avx);
-       and     \$`1<<28`,%r11d         # mask AVX bit
-       and     \$`1<<30`,%r10d         # mask "Intel CPU" bit
-       or      %r11d,%r10d
-       cmp     \$`1<<28|1<<30`,%r10d
-       je      aesni_cbc_sha1_enc_avx
+       bt      \$28, 4(%r10)
+       jc      aesni_cbc_sha1_enc_avx
 ___
 $code.=<<___;
        jmp     aesni_cbc_sha1_enc_ssse3
diff -r e7b2b7ef3246 -r 261b724eb83d crypto/external/bsd/openssl/dist/crypto/bn/asm/x86_64-gf2m.pl
--- a/crypto/external/bsd/openssl/dist/crypto/bn/asm/x86_64-gf2m.pl     Thu May 28 17:31:38 2015 +0000
+++ b/crypto/external/bsd/openssl/dist/crypto/bn/asm/x86_64-gf2m.pl     Thu May 28 19:13:32 2015 +0000
@@ -168,7 +168,7 @@
 .align 16
 bn_GF2m_mul_2x2:
        mov     OPENSSL_ia32cap_P(%rip),%rax
-       bt      \$33,%rax
+       bt      \$1,4(%rax)
        jnc     .Lvanilla_mul_2x2
 
        movq            $a1,%xmm0
diff -r e7b2b7ef3246 -r 261b724eb83d crypto/external/bsd/openssl/dist/crypto/rc4/asm/rc4-x86_64.pl
--- a/crypto/external/bsd/openssl/dist/crypto/rc4/asm/rc4-x86_64.pl     Thu May 28 17:31:38 2015 +0000
+++ b/crypto/external/bsd/openssl/dist/crypto/rc4/asm/rc4-x86_64.pl     Thu May 28 19:13:32 2015 +0000
@@ -158,7 +158,7 @@
        mov     -4($dat),$YY#b
        cmpl    \$-1,256($dat)
        je      .LRC4_CHAR
-       mov     OPENSSL_ia32cap_P(%rip),%r8d
+       mov     OPENSSL_ia32cap_P(%rip),%r8
        xor     $TX[1],$TX[1]
        inc     $XX[0]#b
        sub     $XX[0],$TX[1]
@@ -166,7 +166,7 @@
        movl    ($dat,$XX[0],4),$TX[0]#d
        test    \$-16,$len
        jz      .Lloop1
-       bt      \$30,%r8d       # Intel CPU?
+       bt      \$30,(%r8)      # Intel CPU?
        jc      .Lintel
        and     \$7,$TX[1]
        lea     1($XX[0]),$XX[1]
@@ -443,8 +443,8 @@
        xor     %r10,%r10
        xor     %r11,%r11
 
-       mov     OPENSSL_ia32cap_P(%rip),$idx#d
-       bt      \$20,$idx#d     # RC4_CHAR?
+       mov     OPENSSL_ia32cap_P(%rip),$idx
+       bt      \$20,($idx)     # RC4_CHAR?
        jc      .Lc1stloop
        jmp     .Lw1stloop
 
diff -r e7b2b7ef3246 -r 261b724eb83d crypto/external/bsd/openssl/dist/crypto/sha/asm/sha1-x86_64.pl
--- a/crypto/external/bsd/openssl/dist/crypto/sha/asm/sha1-x86_64.pl    Thu May 28 17:31:38 2015 +0000
+++ b/crypto/external/bsd/openssl/dist/crypto/sha/asm/sha1-x86_64.pl    Thu May 28 19:13:32 2015 +0000
@@ -221,17 +221,14 @@
 .type  sha1_block_data_order,\@function,3
 .align 16
 sha1_block_data_order:
-       mov     OPENSSL_ia32cap_P+0(%rip),%r9d
-       mov     OPENSSL_ia32cap_P+4(%rip),%r8d
-       test    \$`1<<9`,%r8d           # check SSSE3 bit
-       jz      .Lialu
+       mov     OPENSSL_ia32cap_P+0(%rip),%r8
+       mov     4(%r8),%r8d
+       bt      \$9,%r8d
+       jnc     .Lialu
 ___
 $code.=<<___ if ($avx);
-       and     \$`1<<28`,%r8d          # mask AVX bit
-       and     \$`1<<30`,%r9d          # mask "Intel CPU" bit
-       or      %r9d,%r8d
-       cmp     \$`1<<28|1<<30`,%r8d
-       je      _avx_shortcut
+       bt      \$28,%r8d
+       jc      _avx_shortcut
 ___
 $code.=<<___;
        jmp     _ssse3_shortcut