crypto/x86_64cpuid.pl

   1 #!/usr/bin/env perl
   2
   3 $output=shift;
   4 $win64a=1 if ($output =~ /win64a\.[s|asm]/);
   5 open STDOUT,">$output" || die "can't open $output: $!";
   6
   7 print<<___ if(defined($win64a));
   8 _TEXT   SEGMENT
   9 PUBLIC  OPENSSL_rdtsc
  10 ALIGN   16
  11 OPENSSL_rdtsc   PROC
  12         rdtsc
  13         shl     rdx,32
  14         or      rax,rdx
  15         ret
  16 OPENSSL_rdtsc   ENDP
  17
  18 PUBLIC  OPENSSL_atomic_add
  19 ALIGN   16
  20 OPENSSL_atomic_add      PROC
  21         mov     eax,DWORD PTR[rcx]
  22 \$Lspin:        lea     r8,DWORD PTR[rdx+rax]
  23 lock    cmpxchg DWORD PTR[rcx],r8d
  24         jne     \$Lspin
  25         mov     eax,r8d
  26         cdqe
  27         ret
  28 OPENSSL_atomic_add      ENDP
  29
  30 PUBLIC  OPENSSL_wipe_cpu
  31 ALIGN   16
  32 OPENSSL_wipe_cpu        PROC
  33         pxor    xmm0,xmm0
  34         pxor    xmm1,xmm1
  35         pxor    xmm2,xmm2
  36         pxor    xmm3,xmm3
  37         pxor    xmm4,xmm4
  38         pxor    xmm5,xmm5
  39         xor     rcx,rcx
  40         xor     rdx,rdx
  41         xor     r8,r8
  42         xor     r9,r9
  43         xor     r10,r10
  44         xor     r11,r11
  45         lea     rax,QWORD PTR[rsp+8]
  46         ret
  47 OPENSSL_wipe_cpu        ENDP
  48
  49 OPENSSL_ia32_cpuid      PROC
  50         mov     r8,rbx
  51
  52         xor     eax,eax
  53         cpuid
  54         xor     eax,eax
  55         cmp     ebx,0756e6547h
  56         setne   al
  57         mov     r9d,eax
  58         cmp     edx,049656e69h
  59         setne   al
  60         or      r9d,eax
  61         cmp     ecx,06c65746eh
  62         setne   al
  63         or      r9d,eax
  64
  65         mov     eax,1
  66         cpuid
  67         bt      edx,28
  68         jnc     \$Ldone
  69         cmp     r9,0
  70         jne     \$Lnotintel
  71         or      edx,000100000h
  72         and     ah,15
  73         cmp     ah,15
  74         je      \$Lnotintel
  75         or      edx,040000000h
  76 \$Lnotintel:
  77         shr     ebx,16
  78         cmp     bl,1
  79         ja      \$Ldone
  80         and     edx,0efffffffh
  81 \$Ldone:
  82         shl     rcx,32
  83         mov     eax,edx
  84         mov     rbx,r8
  85         or      rax,rcx
  86         ret
  87 OPENSSL_ia32_cpuid      ENDP
  88 _TEXT   ENDS
  89
  90 CRT\$XIU        SEGMENT
  91 EXTRN   OPENSSL_cpuid_setup:PROC
  92 DQ      OPENSSL_cpuid_setup
  93 CRT\$XIU        ENDS
  94 END
  95 ___
  96 print<<___ if(!defined($win64a));
  97 .text
  98 .globl  OPENSSL_rdtsc
  99 .align  16
 100 OPENSSL_rdtsc:
 101         rdtsc
 102         shlq    \$32,%rdx
 103         orq     %rdx,%rax
 104         ret
 105 .size   OPENSSL_rdtsc,.-OPENSSL_rdtsc
 106
 107 .globl  OPENSSL_atomic_add
 108 .type   OPENSSL_atomic_add,\@function
 109 .align  16
 110 OPENSSL_atomic_add:
 111         movl    (%rdi),%eax
 112 .Lspin: leaq    (%rsi,%rax),%r8
 113 lock;   cmpxchgl        %r8d,(%rdi)
 114         jne     .Lspin
 115         movl    %r8d,%eax
 116         .byte   0x48,0x98
 117         ret
 118 .size   OPENSSL_atomic_add,.-OPENSSL_atomic_add
 119
 120 .globl  OPENSSL_wipe_cpu
 121 .type   OPENSSL_wipe_cpu,\@function
 122 .align  16
 123 OPENSSL_wipe_cpu:
 124         pxor    %xmm0,%xmm0
 125         pxor    %xmm1,%xmm1
 126         pxor    %xmm2,%xmm2
 127         pxor    %xmm3,%xmm3
 128         pxor    %xmm4,%xmm4
 129         pxor    %xmm5,%xmm5
 130         pxor    %xmm6,%xmm6
 131         pxor    %xmm7,%xmm7
 132         pxor    %xmm8,%xmm8
 133         pxor    %xmm9,%xmm9
 134         pxor    %xmm10,%xmm10
 135         pxor    %xmm11,%xmm11
 136         pxor    %xmm12,%xmm12
 137         pxor    %xmm13,%xmm13
 138         pxor    %xmm14,%xmm14
 139         pxor    %xmm15,%xmm15
 140         xorq    %rcx,%rcx
 141         xorq    %rdx,%rdx
 142         xorq    %rsi,%rsi
 143         xorq    %rdi,%rdi
 144         xorq    %r8,%r8
 145         xorq    %r9,%r9
 146         xorq    %r10,%r10
 147         xorq    %r11,%r11
 148         leaq    8(%rsp),%rax
 149         ret
 150 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 151
 152 .globl  OPENSSL_ia32_cpuid
 153 .align  16
 154 OPENSSL_ia32_cpuid:
 155         movq    %rbx,%r8
 156
 157         xor     %eax,%eax
 158         cpuid
 159         xor     %eax,%eax
 160         cmp     \$0x756e6547,%ebx       # "Genu"
 161         setne   %al
 162         mov     %eax,%r9d
 163         cmp     \$0x49656e69,%edx       # "ineI"
 164         setne   %al
 165         or      %eax,%r9d
 166         cmp     \$0x6c65746e,%ecx       # "ntel"
 167         setne   %al
 168         or      %eax,%r9d
 169
 170         movl    \$1,%eax
 171         cpuid
 172         bt      \$28,%edx               # test hyper-threading bit
 173         jnc     .Ldone
 174         cmp     \$0,%r9
 175         jne     .Lnotintel
 176         or      \$1<<20,%edx            # use reserved bit to engage RC4_CHAR
 177         and     \$15,%ah
 178         cmp     \$15,%ah                # examine Family ID
 179         je      .Lnotintel
 180         or      \$1<<30,%edx            # use reserved bit to skip unrolled loop
 181 .Lnotintel:
 182         shr     \$16,%ebx
 183         cmp     \$1,%bl                 # see if cache is shared
 184         ja      .Ldone
 185         and     \$~(1<<28),%edx
 186 .Ldone:
 187         shlq    \$32,%rcx
 188         movl    %edx,%eax
 189         movq    %r8,%rbx
 190         orq     %rcx,%rax
 191         ret
 192 .size   OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
 193
 194 .section        .init
 195         call    OPENSSL_cpuid_setup
 196 ___