5 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
7 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
9 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
10 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
11 ( $xlate="${dir}perlasm/x86_64-xlate.pl" and -f $xlate) or
12 die "can't locate x86_64-xlate.pl";
14 open STDOUT,"| \"$^X\" $xlate $flavour $output";
16 ($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order
17 ("%rdi","%rsi","%rdx","%rcx"); # Unix order
20 .extern OPENSSL_cpuid_setup
21 .hidden OPENSSL_cpuid_setup
23 call OPENSSL_cpuid_setup
25 .hidden OPENSSL_ia32cap_P
26 .comm OPENSSL_ia32cap_P,16,4
30 .globl OPENSSL_atomic_add
31 .type OPENSSL_atomic_add,\@abi-omnipotent
35 .Lspin: leaq ($arg2,%rax),%r8
40 .byte 0x48,0x98 # cltq/cdqe
42 .size OPENSSL_atomic_add,.-OPENSSL_atomic_add
45 .type OPENSSL_rdtsc,\@abi-omnipotent
52 .size OPENSSL_rdtsc,.-OPENSSL_rdtsc
54 .globl OPENSSL_ia32_cpuid
55 .type OPENSSL_ia32_cpuid,\@function,1
58 mov %rbx,%r8 # save %rbx
61 mov %eax,8(%rdi) # clear 3rd word
63 mov %eax,%r11d # max value for standard query level
66 cmp \$0x756e6547,%ebx # "Genu"
69 cmp \$0x49656e69,%edx # "ineI"
72 cmp \$0x6c65746e,%ecx # "ntel"
74 or %eax,%r9d # 0 indicates Intel CPU
77 cmp \$0x68747541,%ebx # "Auth"
80 cmp \$0x69746E65,%edx # "enti"
83 cmp \$0x444D4163,%ecx # "cAMD"
85 or %eax,%r10d # 0 indicates AMD CPU
97 and \$0x00000801,%r9d # isolate AMD XOP bit, 1<<11
99 cmp \$0x80000008,%r10d
102 mov \$0x80000008,%eax
104 movzb %cl,%r10 # number of cores - 1
105 inc %r10 # number of cores
109 bt \$28,%edx # test hyper-threading bit
111 shr \$16,%ebx # number of logical processors
114 and \$0xefffffff,%edx # ~(1<<28)
123 mov \$0,%ecx # query L1D
127 and \$0xfff,%r10d # number of cores -1 per L1D
140 and \$0xbfefffff,%edx # force reserved bits to 0
143 or \$0x40000000,%edx # set reserved bit#30 on Intel CPUs
145 cmp \$15,%ah # examine Family ID
147 or \$0x00100000,%edx # set reserved bit#20 to engage RC4_CHAR
149 bt \$28,%edx # test hyper-threading bit
151 and \$0xefffffff,%edx # ~(1<<28)
155 or \$0x10000000,%edx # 1<<28
157 cmp \$1,%bl # see if cache is shared
159 and \$0xefffffff,%edx # ~(1<<28)
161 and \$0x00000800,%r9d # isolate AMD XOP flag
162 and \$0xfffff7ff,%ecx
163 or %ecx,%r9d # merge AMD XOP flag
165 mov %edx,%r10d # %r9d:%r10d is copy of %ecx:%edx
166 bt \$27,%r9d # check OSXSAVE bit
169 .byte 0x0f,0x01,0xd0 # xgetbv
170 and \$6,%eax # isolate XMM and YMM state support
174 mov \$0xefffe7ff,%eax # ~(1<<28|1<<12|1<<11)
175 and %eax,%r9d # clear AVX, FMA and AMD XOP bits
176 andl \$0xffffffdf,8(%rdi) # cleax AVX2, ~(1<<5)
180 mov %r8,%rbx # restore %rbx
183 .size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
185 .globl OPENSSL_cleanse
186 .type OPENSSL_cleanse,\@abi-omnipotent
218 .size OPENSSL_cleanse,.-OPENSSL_cleanse
221 print<<___ if (!$win64);
222 .globl OPENSSL_wipe_cpu
223 .type OPENSSL_wipe_cpu,\@abi-omnipotent
252 .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
254 print<<___ if ($win64);
255 .globl OPENSSL_wipe_cpu
256 .type OPENSSL_wipe_cpu,\@abi-omnipotent
273 .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
281 my $redzone=win64?8:-8;
284 .globl OPENSSL_instrument_bus
285 .type OPENSSL_instrument_bus,\@abi-omnipotent
287 OPENSSL_instrument_bus:
288 mov $arg1,$out # tribute to Win64
292 rdtsc # collect 1st tick
293 mov %eax,$lasttick # lasttick = tick
294 mov \$0,$lastdiff # lastdiff = 0
314 .size OPENSSL_instrument_bus,.-OPENSSL_instrument_bus
316 .globl OPENSSL_instrument_bus2
317 .type OPENSSL_instrument_bus2,\@abi-omnipotent
319 OPENSSL_instrument_bus2:
320 mov $arg1,$out # tribute to Win64
323 mov $cnt,$redzone(%rsp)
325 rdtsc # collect 1st tick
326 mov %eax,$lasttick # lasttick = tick
327 mov \$0,$lastdiff # lastdiff = 0
333 rdtsc # collect 1st diff
335 sub $lasttick,%eax # diff
336 mov %edx,$lasttick # lasttick = tick
337 mov %eax,$lastdiff # lastdiff = diff
341 add %eax,($out) # accumulate diff
348 sub $lasttick,%eax # diff
349 mov %edx,$lasttick # lasttick = tick
351 mov %eax,$lastdiff # lastdiff = diff
354 sub %rdx,$cnt # conditional --$cnt
355 lea ($out,%rdx,4),$out # conditional ++$out
359 mov $redzone(%rsp),%rax
362 .size OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2
367 .globl OPENSSL_ia32_rdrand
368 .type OPENSSL_ia32_rdrand,\@abi-omnipotent
380 .size OPENSSL_ia32_rdrand,.-OPENSSL_ia32_rdrand
383 close STDOUT; # flush