crypto/x86_64cpuid.pl

   1 #!/usr/bin/env perl
   2
   3 $flavour = shift;
   4 $output  = shift;
   5 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
   6
   7 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
   8
   9 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  10 open STDOUT,"| $^X ${dir}perlasm/x86_64-xlate.pl $flavour $output";
  11
  12 if ($win64)     { $arg1="%rcx"; $arg2="%rdx"; }
  13 else            { $arg1="%rdi"; $arg2="%rsi"; }
  14 print<<___;
  15 .extern         OPENSSL_cpuid_setup
  16 .section        .init
  17         call    OPENSSL_cpuid_setup
  18
  19 .text
  20
  21 .globl  OPENSSL_atomic_add
  22 .type   OPENSSL_atomic_add,\@abi-omnipotent
  23 .align  16
  24 OPENSSL_atomic_add:
  25         movl    ($arg1),%eax
  26 .Lspin: leaq    ($arg2,%rax),%r8
  27         .byte   0xf0            # lock
  28         cmpxchgl        %r8d,($arg1)
  29         jne     .Lspin
  30         movl    %r8d,%eax
  31         .byte   0x48,0x98       # cltq/cdqe
  32         ret
  33 .size   OPENSSL_atomic_add,.-OPENSSL_atomic_add
  34
  35 .globl  OPENSSL_rdtsc
  36 .type   OPENSSL_rdtsc,\@abi-omnipotent
  37 .align  16
  38 OPENSSL_rdtsc:
  39         rdtsc
  40         shl     \$32,%rdx
  41         or      %rdx,%rax
  42         ret
  43 .size   OPENSSL_rdtsc,.-OPENSSL_rdtsc
  44
  45 .globl  OPENSSL_ia32_cpuid
  46 .type   OPENSSL_ia32_cpuid,\@abi-omnipotent
  47 .align  16
  48 OPENSSL_ia32_cpuid:
  49         mov     %rbx,%r8
  50
  51         xor     %eax,%eax
  52         cpuid
  53         xor     %eax,%eax
  54         cmp     \$0x756e6547,%ebx       # "Genu"
  55         setne   %al
  56         mov     %eax,%r9d
  57         cmp     \$0x49656e69,%edx       # "ineI"
  58         setne   %al
  59         or      %eax,%r9d
  60         cmp     \$0x6c65746e,%ecx       # "ntel"
  61         setne   %al
  62         or      %eax,%r9d
  63
  64         mov     \$1,%eax
  65         cpuid
  66         cmp     \$0,%r9d
  67         jne     .Lnotintel
  68         or      \$0x00100000,%edx       # use reserved 20th bit to engage RC4_CHAR
  69         and     \$15,%ah
  70         cmp     \$15,%ah                # examine Family ID
  71         je      .Lnotintel
  72         or      \$0x40000000,%edx       # use reserved bit to skip unrolled loop
  73 .Lnotintel:
  74         bt      \$28,%edx               # test hyper-threading bit
  75         jnc     .Ldone
  76         shr     \$16,%ebx
  77         cmp     \$1,%bl                 # see if cache is shared
  78         ja      .Ldone
  79         and     \$0xefffffff,%edx       # ~(1<<28)
  80 .Ldone:
  81         shl     \$32,%rcx
  82         mov     %edx,%eax
  83         mov     %r8,%rbx
  84         or      %rcx,%rax
  85         ret
  86 .size   OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
  87
  88 .globl  OPENSSL_cleanse
  89 .type   OPENSSL_cleanse,\@abi-omnipotent
  90 .align  16
  91 OPENSSL_cleanse:
  92         xor     %rax,%rax
  93         cmp     \$15,$arg2
  94         jae     .Lot
  95 .Little:
  96         mov     %al,($arg1)
  97         sub     \$1,$arg2
  98         lea     1($arg1),$arg1
  99         jnz     .Little
 100         ret
 101 .align  16
 102 .Lot:
 103         test    \$7,$arg1
 104         jz      .Laligned
 105         mov     %al,($arg1)
 106         lea     -1($arg2),$arg2
 107         lea     1($arg1),$arg1
 108         jmp     .Lot
 109 .Laligned:
 110         mov     %rax,($arg1)
 111         lea     -8($arg2),$arg2
 112         test    \$-8,$arg2
 113         lea     8($arg1),$arg1
 114         jnz     .Laligned
 115         cmp     \$0,$arg2
 116         jne     .Little
 117         ret
 118 .size   OPENSSL_cleanse,.-OPENSSL_cleanse
 119 ___
 120
 121 print<<___ if (!$win64);
 122 .globl  OPENSSL_wipe_cpu
 123 .type   OPENSSL_wipe_cpu,\@abi-omnipotent
 124 .align  16
 125 OPENSSL_wipe_cpu:
 126         pxor    %xmm0,%xmm0
 127         pxor    %xmm1,%xmm1
 128         pxor    %xmm2,%xmm2
 129         pxor    %xmm3,%xmm3
 130         pxor    %xmm4,%xmm4
 131         pxor    %xmm5,%xmm5
 132         pxor    %xmm6,%xmm6
 133         pxor    %xmm7,%xmm7
 134         pxor    %xmm8,%xmm8
 135         pxor    %xmm9,%xmm9
 136         pxor    %xmm10,%xmm10
 137         pxor    %xmm11,%xmm11
 138         pxor    %xmm12,%xmm12
 139         pxor    %xmm13,%xmm13
 140         pxor    %xmm14,%xmm14
 141         pxor    %xmm15,%xmm15
 142         xorq    %rcx,%rcx
 143         xorq    %rdx,%rdx
 144         xorq    %rsi,%rsi
 145         xorq    %rdi,%rdi
 146         xorq    %r8,%r8
 147         xorq    %r9,%r9
 148         xorq    %r10,%r10
 149         xorq    %r11,%r11
 150         leaq    8(%rsp),%rax
 151         ret
 152 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 153 ___
 154 print<<___ if ($win64);
 155 .globl  OPENSSL_wipe_cpu
 156 .type   OPENSSL_wipe_cpu,\@abi-omnipotent
 157 .align  16
 158 OPENSSL_wipe_cpu:
 159         pxor    %xmm0,%xmm0
 160         pxor    %xmm1,%xmm1
 161         pxor    %xmm2,%xmm2
 162         pxor    %xmm3,%xmm3
 163         pxor    %xmm4,%xmm4
 164         pxor    %xmm5,%xmm5
 165         xorq    %rcx,%rcx
 166         xorq    %rdx,%rdx
 167         xorq    %r8,%r8
 168         xorq    %r9,%r9
 169         xorq    %r10,%r10
 170         xorq    %r11,%r11
 171         leaq    8(%rsp),%rax
 172         ret
 173 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 174 ___
 175
 176 close STDOUT;   # flush