crypto/x86_64cpuid.pl

   1 #!/usr/bin/env perl
   2
   3 $output=shift;
   4 $masm=1 if ($output =~ /\.asm/);
   5 open STDOUT,">$output" || die "can't open $output: $!";
   6
   7 print<<___ if(defined($masm));
   8 _TEXT   SEGMENT
   9 PUBLIC  OPENSSL_rdtsc
  10
  11 PUBLIC  OPENSSL_atomic_add
  12 ALIGN   16
  13 OPENSSL_atomic_add      PROC
  14         mov     eax,DWORD PTR[rcx]
  15 \$Lspin:        lea     r8,DWORD PTR[rdx+rax]
  16 lock    cmpxchg DWORD PTR[rcx],r8d
  17         jne     \$Lspin
  18         mov     eax,r8d
  19         cdqe
  20         ret
  21 OPENSSL_atomic_add      ENDP
  22
  23 PUBLIC  OPENSSL_wipe_cpu
  24 ALIGN   16
  25 OPENSSL_wipe_cpu        PROC
  26         pxor    xmm0,xmm0
  27         pxor    xmm1,xmm1
  28         pxor    xmm2,xmm2
  29         pxor    xmm3,xmm3
  30         pxor    xmm4,xmm4
  31         pxor    xmm5,xmm5
  32         xor     rcx,rcx
  33         xor     rdx,rdx
  34         xor     r8,r8
  35         xor     r9,r9
  36         xor     r10,r10
  37         xor     r11,r11
  38         lea     rax,QWORD PTR[rsp+8]
  39         ret
  40 OPENSSL_wipe_cpu        ENDP
  41 _TEXT   ENDS
  42
  43 CRT\$XIU        SEGMENT
  44 EXTRN   OPENSSL_cpuid_setup:PROC
  45 DQ      OPENSSL_cpuid_setup
  46 CRT\$XIU        ENDS
  47
  48 ___
  49 print<<___ if(!defined($masm));
  50 .text
  51
  52 .globl  OPENSSL_atomic_add
  53 .type   OPENSSL_atomic_add,\@function
  54 .align  16
  55 OPENSSL_atomic_add:
  56         movl    (%rdi),%eax
  57 .Lspin: leaq    (%rsi,%rax),%r8
  58 lock;   cmpxchgl        %r8d,(%rdi)
  59         jne     .Lspin
  60         movl    %r8d,%eax
  61         .byte   0x48,0x98
  62         ret
  63 .size   OPENSSL_atomic_add,.-OPENSSL_atomic_add
  64
  65 .globl  OPENSSL_wipe_cpu
  66 .type   OPENSSL_wipe_cpu,\@function
  67 .align  16
  68 OPENSSL_wipe_cpu:
  69         pxor    %xmm0,%xmm0
  70         pxor    %xmm1,%xmm1
  71         pxor    %xmm2,%xmm2
  72         pxor    %xmm3,%xmm3
  73         pxor    %xmm4,%xmm4
  74         pxor    %xmm5,%xmm5
  75         pxor    %xmm6,%xmm6
  76         pxor    %xmm7,%xmm7
  77         pxor    %xmm8,%xmm8
  78         pxor    %xmm9,%xmm9
  79         pxor    %xmm10,%xmm10
  80         pxor    %xmm11,%xmm11
  81         pxor    %xmm12,%xmm12
  82         pxor    %xmm13,%xmm13
  83         pxor    %xmm14,%xmm14
  84         pxor    %xmm15,%xmm15
  85         xorq    %rcx,%rcx
  86         xorq    %rdx,%rdx
  87         xorq    %rsi,%rsi
  88         xorq    %rdi,%rdi
  89         xorq    %r8,%r8
  90         xorq    %r9,%r9
  91         xorq    %r10,%r10
  92         xorq    %r11,%r11
  93         leaq    8(%rsp),%rax
  94         ret
  95 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
  96
  97 .section        .init
  98         call    OPENSSL_cpuid_setup
  99
 100 ___
 101
 102 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 103 $dir = "." if $dir eq "";
 104 open STDOUT,"| $^X $dir/perlasm/x86_64-xlate.pl $output";
 105
 106 print<<___;
 107 .text
 108
 109 .globl  OPENSSL_rdtsc
 110 .type   OPENSSL_rdtsc,\@abi-omnipotent
 111 .align  16
 112 OPENSSL_rdtsc:
 113         rdtsc
 114         shl     \$32,%rdx
 115         or      %rdx,%rax
 116         ret
 117 .size   OPENSSL_rdtsc,.-OPENSSL_rdtsc
 118
 119 .globl  OPENSSL_ia32_cpuid
 120 .type   OPENSSL_ia32_cpuid,\@abi-omnipotent
 121 .align  16
 122 OPENSSL_ia32_cpuid:
 123         mov     %rbx,%r8
 124
 125         xor     %eax,%eax
 126         cpuid
 127         xor     %eax,%eax
 128         cmp     \$0x756e6547,%ebx       # "Genu"
 129         setne   %al
 130         mov     %eax,%r9d
 131         cmp     \$0x49656e69,%edx       # "ineI"
 132         setne   %al
 133         or      %eax,%r9d
 134         cmp     \$0x6c65746e,%ecx       # "ntel"
 135         setne   %al
 136         or      %eax,%r9d
 137
 138         mov     \$1,%eax
 139         cpuid
 140         cmp     \$0,%r9d
 141         jne     .Lnotintel
 142         or      \$0x00100000,%edx       # use reserved 20th bit to engage RC4_CHAR
 143         and     \$15,%ah
 144         cmp     \$15,%ah                # examine Family ID
 145         je      .Lnotintel
 146         or      \$0x40000000,%edx       # use reserved bit to skip unrolled loop
 147 .Lnotintel:
 148         bt      \$28,%edx               # test hyper-threading bit
 149         jnc     .Ldone
 150         shr     \$16,%ebx
 151         cmp     \$1,%bl                 # see if cache is shared
 152         ja      .Ldone
 153         and     \$0xefffffff,%edx       # ~(1<<28)
 154 .Ldone:
 155         shl     \$32,%rcx
 156         mov     %edx,%eax
 157         mov     %r8,%rbx
 158         or      %rcx,%rax
 159         ret
 160 .size   OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
 161
 162 .globl  OPENSSL_cleanse
 163 .type   OPENSSL_cleanse,\@function,2
 164 .align  16
 165 OPENSSL_cleanse:
 166         xor     %rax,%rax
 167         cmp     \$15,%rsi
 168         jae     .Lot
 169 .Little:
 170         mov     %al,(%rdi)
 171         sub     \$1,%rsi
 172         lea     1(%rdi),%rdi
 173         jnz     .Little
 174         ret
 175 .align  16
 176 .Lot:
 177         test    \$7,%rdi
 178         jz      .Laligned
 179         mov     %al,(%rdi)
 180         lea     -1(%rsi),%rsi
 181         lea     1(%rdi),%rdi
 182         jmp     .Lot
 183 .Laligned:
 184         mov     %rax,(%rdi)
 185         lea     -8(%rsi),%rsi
 186         test    \$-8,%rsi
 187         lea     8(%rdi),%rdi
 188         jnz     .Laligned
 189         cmp     \$0,%rsi
 190         jne     .Little
 191         ret
 192 .size   OPENSSL_cleanse,.-OPENSSL_cleanse
 193 ___
 194 close STDOUT;   # flush