2 # Copyright 2005-2016 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the OpenSSL license (the "License"). You may not use
5 # this file except in compliance with the License. You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
12 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
14 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
16 $0 =~ m/(.*[\/\\])[^\
/\\]+$/; $dir=$1;
17 ( $xlate="${dir}x86_64-xlate.pl" and -f
$xlate ) or
18 ( $xlate="${dir}perlasm/x86_64-xlate.pl" and -f
$xlate) or
19 die "can't locate x86_64-xlate.pl";
21 open OUT
,"| \"$^X\" $xlate $flavour $output";
24 ($arg1,$arg2,$arg3,$arg4)=$win64?
("%rcx","%rdx","%r8", "%r9") : # Win64 order
25 ("%rdi","%rsi","%rdx","%rcx"); # Unix order
28 .extern OPENSSL_cpuid_setup
29 .hidden OPENSSL_cpuid_setup
31 call OPENSSL_cpuid_setup
33 .hidden OPENSSL_ia32cap_P
34 .comm OPENSSL_ia32cap_P
,16,4
38 .globl OPENSSL_atomic_add
39 .type OPENSSL_atomic_add
,\
@abi-omnipotent
43 .Lspin
: leaq
($arg2,%rax),%r8
48 .byte
0x48,0x98 # cltq/cdqe
50 .size OPENSSL_atomic_add
,.-OPENSSL_atomic_add
53 .type OPENSSL_rdtsc
,\
@abi-omnipotent
60 .size OPENSSL_rdtsc
,.-OPENSSL_rdtsc
62 .globl OPENSSL_ia32_cpuid
63 .type OPENSSL_ia32_cpuid
,\
@function,1
66 mov
%rbx,%r8 # save %rbx
69 mov
%eax,8(%rdi) # clear 3rd word
71 mov
%eax,%r11d # max value for standard query level
74 cmp \
$0x756e6547,%ebx # "Genu"
77 cmp \
$0x49656e69,%edx # "ineI"
80 cmp \
$0x6c65746e,%ecx # "ntel"
82 or %eax,%r9d # 0 indicates Intel CPU
85 cmp \
$0x68747541,%ebx # "Auth"
88 cmp \
$0x69746E65,%edx # "enti"
91 cmp \
$0x444D4163,%ecx # "cAMD"
93 or %eax,%r10d # 0 indicates AMD CPU
102 mov \
$0x80000001,%eax
105 and \
$0x00000801,%r9d # isolate AMD XOP bit, 1<<11
107 cmp \
$0x80000008,%r10d
110 mov \
$0x80000008,%eax
112 movzb
%cl,%r10 # number of cores - 1
113 inc
%r10 # number of cores
117 bt \
$28,%edx # test hyper-threading bit
119 shr \
$16,%ebx # number of logical processors
122 and \
$0xefffffff,%edx # ~(1<<28)
131 mov \
$0,%ecx # query L1D
135 and \
$0xfff,%r10d # number of cores -1 per L1D
148 and \
$0xbfefffff,%edx # force reserved bits to 0
151 or \
$0x40000000,%edx # set reserved bit#30 on Intel CPUs
153 cmp \
$15,%ah # examine Family ID
155 or \
$0x00100000,%edx # set reserved bit#20 to engage RC4_CHAR
157 bt \
$28,%edx # test hyper-threading bit
159 and \
$0xefffffff,%edx # ~(1<<28)
163 or \
$0x10000000,%edx # 1<<28
165 cmp \
$1,%bl # see if cache is shared
167 and \
$0xefffffff,%edx # ~(1<<28)
169 and \
$0x00000800,%r9d # isolate AMD XOP flag
170 and \
$0xfffff7ff,%ecx
171 or %ecx,%r9d # merge AMD XOP flag
173 mov
%edx,%r10d # %r9d:%r10d is copy of %ecx:%edx
174 bt \
$27,%r9d # check OSXSAVE bit
177 .byte
0x0f,0x01,0xd0 # xgetbv
178 and \
$6,%eax # isolate XMM and YMM state support
182 mov \
$0xefffe7ff,%eax # ~(1<<28|1<<12|1<<11)
183 and %eax,%r9d # clear AVX, FMA and AMD XOP bits
184 andl \
$0xffffffdf,8(%rdi) # cleax AVX2, ~(1<<5)
188 mov
%r8,%rbx # restore %rbx
191 .size OPENSSL_ia32_cpuid
,.-OPENSSL_ia32_cpuid
193 .globl OPENSSL_cleanse
194 .type OPENSSL_cleanse
,\
@abi-omnipotent
226 .size OPENSSL_cleanse
,.-OPENSSL_cleanse
229 .type CRYPTO_memcmp
,\
@abi-omnipotent
248 .size CRYPTO_memcmp
,.-CRYPTO_memcmp
251 print<<___
if (!$win64);
252 .globl OPENSSL_wipe_cpu
253 .type OPENSSL_wipe_cpu
,\
@abi-omnipotent
282 .size OPENSSL_wipe_cpu
,.-OPENSSL_wipe_cpu
284 print<<___
if ($win64);
285 .globl OPENSSL_wipe_cpu
286 .type OPENSSL_wipe_cpu
,\
@abi-omnipotent
303 .size OPENSSL_wipe_cpu
,.-OPENSSL_wipe_cpu
311 my $redzone=win64?
8:-8;
314 .globl OPENSSL_instrument_bus
315 .type OPENSSL_instrument_bus
,\
@abi-omnipotent
317 OPENSSL_instrument_bus
:
318 mov
$arg1,$out # tribute to Win64
322 rdtsc
# collect 1st tick
323 mov
%eax,$lasttick # lasttick = tick
324 mov \
$0,$lastdiff # lastdiff = 0
344 .size OPENSSL_instrument_bus
,.-OPENSSL_instrument_bus
346 .globl OPENSSL_instrument_bus2
347 .type OPENSSL_instrument_bus2
,\
@abi-omnipotent
349 OPENSSL_instrument_bus2
:
350 mov
$arg1,$out # tribute to Win64
353 mov
$cnt,$redzone(%rsp)
355 rdtsc
# collect 1st tick
356 mov
%eax,$lasttick # lasttick = tick
357 mov \
$0,$lastdiff # lastdiff = 0
363 rdtsc
# collect 1st diff
365 sub $lasttick,%eax # diff
366 mov
%edx,$lasttick # lasttick = tick
367 mov
%eax,$lastdiff # lastdiff = diff
371 add
%eax,($out) # accumulate diff
378 sub $lasttick,%eax # diff
379 mov
%edx,$lasttick # lasttick = tick
381 mov
%eax,$lastdiff # lastdiff = diff
384 sub %rdx,$cnt # conditional --$cnt
385 lea
($out,%rdx,4),$out # conditional ++$out
389 mov
$redzone(%rsp),%rax
392 .size OPENSSL_instrument_bus2
,.-OPENSSL_instrument_bus2
397 .globl OPENSSL_ia32_rdrand
398 .type OPENSSL_ia32_rdrand
,\
@abi-omnipotent
410 .size OPENSSL_ia32_rdrand
,.-OPENSSL_ia32_rdrand
412 .globl OPENSSL_ia32_rdseed
413 .type OPENSSL_ia32_rdseed
,\
@abi-omnipotent
425 .size OPENSSL_ia32_rdseed
,.-OPENSSL_ia32_rdseed
428 close STDOUT
; # flush