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"; $arg2="%rsi"; }
  14 print<<___;
  15 .extern         OPENSSL_cpuid_setup
  16 .section        .init
  17         call    OPENSSL_cpuid_setup
  18
  19 ___
  20
  21 if ($flavour =~ /elf/) {
  22   print ".section .note.GNU-stack,\"\",\@progbits\n"
  23 }
  24
  25 print<<___;
  26 .text
  27
  28 .globl  OPENSSL_atomic_add
  29 .type   OPENSSL_atomic_add,\@abi-omnipotent
  30 .align  16
  31 OPENSSL_atomic_add:
  32         movl    ($arg1),%eax
  33 .Lspin: leaq    ($arg2,%rax),%r8
  34         .byte   0xf0            # lock
  35         cmpxchgl        %r8d,($arg1)
  36         jne     .Lspin
  37         movl    %r8d,%eax
  38         .byte   0x48,0x98       # cltq/cdqe
  39         ret
  40 .size   OPENSSL_atomic_add,.-OPENSSL_atomic_add
  41
  42 .globl  OPENSSL_rdtsc
  43 .type   OPENSSL_rdtsc,\@abi-omnipotent
  44 .align  16
  45 OPENSSL_rdtsc:
  46         rdtsc
  47         shl     \$32,%rdx
  48         or      %rdx,%rax
  49         ret
  50 .size   OPENSSL_rdtsc,.-OPENSSL_rdtsc
  51
  52 .globl  OPENSSL_ia32_cpuid
  53 .type   OPENSSL_ia32_cpuid,\@abi-omnipotent
  54 .align  16
  55 OPENSSL_ia32_cpuid:
  56         mov     %rbx,%r8
  57
  58         xor     %eax,%eax
  59         cpuid
  60         mov     %eax,%r11d              # max value for standard query level
  61
  62         xor     %eax,%eax
  63         cmp     \$0x756e6547,%ebx       # "Genu"
  64         setne   %al
  65         mov     %eax,%r9d
  66         cmp     \$0x49656e69,%edx       # "ineI"
  67         setne   %al
  68         or      %eax,%r9d
  69         cmp     \$0x6c65746e,%ecx       # "ntel"
  70         setne   %al
  71         or      %eax,%r9d               # 0 indicates Intel CPU
  72         jz      .Lintel
  73
  74         cmp     \$0x68747541,%ebx       # "Auth"
  75         setne   %al
  76         mov     %eax,%r10d
  77         cmp     \$0x69746E65,%edx       # "enti"
  78         setne   %al
  79         or      %eax,%r10d
  80         cmp     \$0x444D4163,%ecx       # "cAMD"
  81         setne   %al
  82         or      %eax,%r10d              # 0 indicates AMD CPU
  83         jnz     .Lintel
  84
  85         # AMD specific
  86         mov     \$0x80000000,%eax
  87         cpuid
  88         cmp     \$0x80000008,%eax
  89         jb      .Lintel
  90
  91         mov     \$0x80000008,%eax
  92         cpuid
  93         movzb   %cl,%r10                # number of cores - 1
  94         inc     %r10                    # number of cores
  95
  96         mov     \$1,%eax
  97         cpuid
  98         bt      \$28,%edx               # test hyper-threading bit
  99         jnc     .Ldone
 100         shr     \$16,%ebx               # number of logical processors
 101         cmp     %r10b,%bl
 102         ja      .Ldone
 103         and     \$0xefffffff,%edx       # ~(1<<28)
 104         jmp     .Ldone
 105
 106 .Lintel:
 107         cmp     \$4,%r11d
 108         mov     \$-1,%r10d
 109         jb      .Lnocacheinfo
 110
 111         mov     \$4,%eax
 112         mov     \$0,%ecx                # query L1D
 113         cpuid
 114         mov     %eax,%r10d
 115         shr     \$14,%r10d
 116         and     \$0xfff,%r10d           # number of cores -1 per L1D
 117
 118 .Lnocacheinfo:
 119         mov     \$1,%eax
 120         cpuid
 121         cmp     \$0,%r9d
 122         jne     .Lnotintel
 123         or      \$0x00100000,%edx       # use reserved 20th bit to engage RC4_CHAR
 124         and     \$15,%ah
 125         cmp     \$15,%ah                # examine Family ID
 126         je      .Lnotintel
 127         or      \$0x40000000,%edx       # use reserved bit to skip unrolled loop
 128 .Lnotintel:
 129         bt      \$28,%edx               # test hyper-threading bit
 130         jnc     .Ldone
 131         and     \$0xefffffff,%edx       # ~(1<<28)
 132         cmp     \$0,%r10d
 133         je      .Ldone
 134
 135         or      \$0x10000000,%edx       # 1<<28
 136         shr     \$16,%ebx
 137         cmp     \$1,%bl                 # see if cache is shared
 138         ja      .Ldone
 139         and     \$0xefffffff,%edx       # ~(1<<28)
 140 .Ldone:
 141         shl     \$32,%rcx
 142         mov     %edx,%eax
 143         mov     %r8,%rbx
 144         or      %rcx,%rax
 145         ret
 146 .size   OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
 147
 148 .globl  OPENSSL_cleanse
 149 .type   OPENSSL_cleanse,\@abi-omnipotent
 150 .align  16
 151 OPENSSL_cleanse:
 152         xor     %rax,%rax
 153         cmp     \$15,$arg2
 154         jae     .Lot
 155         cmp     \$0,$arg2
 156         je      .Lret
 157 .Little:
 158         mov     %al,($arg1)
 159         sub     \$1,$arg2
 160         lea     1($arg1),$arg1
 161         jnz     .Little
 162 .Lret:
 163         ret
 164 .align  16
 165 .Lot:
 166         test    \$7,$arg1
 167         jz      .Laligned
 168         mov     %al,($arg1)
 169         lea     -1($arg2),$arg2
 170         lea     1($arg1),$arg1
 171         jmp     .Lot
 172 .Laligned:
 173         mov     %rax,($arg1)
 174         lea     -8($arg2),$arg2
 175         test    \$-8,$arg2
 176         lea     8($arg1),$arg1
 177         jnz     .Laligned
 178         cmp     \$0,$arg2
 179         jne     .Little
 180         ret
 181 .size   OPENSSL_cleanse,.-OPENSSL_cleanse
 182 ___
 183
 184 print<<___ if (!$win64);
 185 .globl  OPENSSL_wipe_cpu
 186 .type   OPENSSL_wipe_cpu,\@abi-omnipotent
 187 .align  16
 188 OPENSSL_wipe_cpu:
 189         pxor    %xmm0,%xmm0
 190         pxor    %xmm1,%xmm1
 191         pxor    %xmm2,%xmm2
 192         pxor    %xmm3,%xmm3
 193         pxor    %xmm4,%xmm4
 194         pxor    %xmm5,%xmm5
 195         pxor    %xmm6,%xmm6
 196         pxor    %xmm7,%xmm7
 197         pxor    %xmm8,%xmm8
 198         pxor    %xmm9,%xmm9
 199         pxor    %xmm10,%xmm10
 200         pxor    %xmm11,%xmm11
 201         pxor    %xmm12,%xmm12
 202         pxor    %xmm13,%xmm13
 203         pxor    %xmm14,%xmm14
 204         pxor    %xmm15,%xmm15
 205         xorq    %rcx,%rcx
 206         xorq    %rdx,%rdx
 207         xorq    %rsi,%rsi
 208         xorq    %rdi,%rdi
 209         xorq    %r8,%r8
 210         xorq    %r9,%r9
 211         xorq    %r10,%r10
 212         xorq    %r11,%r11
 213         leaq    8(%rsp),%rax
 214         ret
 215 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 216 ___
 217 print<<___ if ($win64);
 218 .globl  OPENSSL_wipe_cpu
 219 .type   OPENSSL_wipe_cpu,\@abi-omnipotent
 220 .align  16
 221 OPENSSL_wipe_cpu:
 222         pxor    %xmm0,%xmm0
 223         pxor    %xmm1,%xmm1
 224         pxor    %xmm2,%xmm2
 225         pxor    %xmm3,%xmm3
 226         pxor    %xmm4,%xmm4
 227         pxor    %xmm5,%xmm5
 228         xorq    %rcx,%rcx
 229         xorq    %rdx,%rdx
 230         xorq    %r8,%r8
 231         xorq    %r9,%r9
 232         xorq    %r10,%r10
 233         xorq    %r11,%r11
 234         leaq    8(%rsp),%rax
 235         ret
 236 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 237 ___
 238
 239 close STDOUT;   # flush