crypto/ppccpuid.pl

   1 #!/usr/bin/env perl
   2
   3 $flavour = shift;
   4
   5 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
   6 ( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
   7 ( $xlate="${dir}perlasm/ppc-xlate.pl" and -f $xlate) or
   8 die "can't locate ppc-xlate.pl";
   9
  10 open STDOUT,"| $^X $xlate $flavour ".shift || die "can't call $xlate: $!";
  11
  12 if ($flavour=~/64/) {
  13     $CMPLI="cmpldi";
  14     $SHRLI="srdi";
  15     $SIGNX="extsw";
  16 } else {
  17     $CMPLI="cmplwi";
  18     $SHRLI="srwi";
  19     $SIGNX="mr";
  20 }
  21
  22 $code=<<___;
  23 .machine        "any"
  24 .text
  25
  26 .globl  .OPENSSL_ppc64_probe
  27 .align  4
  28 .OPENSSL_ppc64_probe:
  29         fcfid   f1,f1
  30         extrdi  r0,r0,32,0
  31         blr
  32
  33 .globl  .OPENSSL_altivec_probe
  34 .align  4
  35 .OPENSSL_altivec_probe:
  36         .long   0x10000484      # vor   v0,v0,v0
  37         blr
  38
  39 .globl  .OPENSSL_wipe_cpu
  40 .align  4
  41 .OPENSSL_wipe_cpu:
  42         xor     r0,r0,r0
  43         fmr     f0,f31
  44         fmr     f1,f31
  45         fmr     f2,f31
  46         mr      r3,r1
  47         fmr     f3,f31
  48         xor     r4,r4,r4
  49         fmr     f4,f31
  50         xor     r5,r5,r5
  51         fmr     f5,f31
  52         xor     r6,r6,r6
  53         fmr     f6,f31
  54         xor     r7,r7,r7
  55         fmr     f7,f31
  56         xor     r8,r8,r8
  57         fmr     f8,f31
  58         xor     r9,r9,r9
  59         fmr     f9,f31
  60         xor     r10,r10,r10
  61         fmr     f10,f31
  62         xor     r11,r11,r11
  63         fmr     f11,f31
  64         xor     r12,r12,r12
  65         fmr     f12,f31
  66         fmr     f13,f31
  67         blr
  68
  69 .globl  .OPENSSL_atomic_add
  70 .align  4
  71 .OPENSSL_atomic_add:
  72 Ladd:   lwarx   r5,0,r3
  73         add     r0,r4,r5
  74         stwcx.  r0,0,r3
  75         bne-    Ladd
  76         $SIGNX  r3,r0
  77         blr
  78
  79 .globl  .OPENSSL_rdtsc
  80 .align  4
  81 .OPENSSL_rdtsc:
  82         mftb    r3
  83         mftbu   r4
  84         blr
  85
  86 .globl  .OPENSSL_cleanse
  87 .align  4
  88 .OPENSSL_cleanse:
  89         $CMPLI  r4,7
  90         li      r0,0
  91         bge     Lot
  92         $CMPLI  r4,0
  93         beqlr-
  94 Little: mtctr   r4
  95         stb     r0,0(r3)
  96         addi    r3,r3,1
  97         bdnz-   \$-8
  98         blr
  99 Lot:    andi.   r5,r3,3
 100         beq     Laligned
 101         stb     r0,0(r3)
 102         subi    r4,r4,1
 103         addi    r3,r3,1
 104         b       Lot
 105 Laligned:
 106         $SHRLI  r5,r4,2
 107         mtctr   r5
 108         stw     r0,0(r3)
 109         addi    r3,r3,4
 110         bdnz-   \$-8
 111         andi.   r4,r4,3
 112         bne     Little
 113         blr
 114 ___
 115 {
 116 my ($out,$cnt,$max)=("r3","r4","r5");
 117 my ($tick,$lasttick)=("r6","r7");
 118 my ($diff,$lastdiff)=("r8","r9");
 119
 120 $code.=<<___;
 121 .globl  .OPENSSL_instrument_bus
 122 .align  4
 123 .OPENSSL_instrument_bus:
 124         mtctr   $cnt
 125
 126         mftb    $lasttick               # collect 1st tick
 127         li      $diff,0
 128
 129         dcbf    0,$out                  # flush cache line
 130         lwarx   $tick,0,$out            # load and lock
 131         add     $tick,$tick,$diff
 132         stwcx.  $tick,0,$out
 133         stwx    $tick,0,$out
 134
 135 Loop:   mftb    $tick
 136         sub     $diff,$tick,$lasttick
 137         mr      $lasttick,$tick
 138         dcbf    0,$out                  # flush cache line
 139         lwarx   $tick,0,$out            # load and lock
 140         add     $tick,$tick,$diff
 141         stwcx.  $tick,0,$out
 142         stwx    $tick,0,$out
 143         addi    $out,$out,4             # ++$out
 144         bdnz    Loop
 145
 146         mr      r3,$cnt
 147         blr
 148
 149 .globl  .OPENSSL_instrument_bus2
 150 .align  4
 151 .OPENSSL_instrument_bus2:
 152         mr      r0,$cnt
 153         slwi    $cnt,$cnt,2
 154
 155         mftb    $lasttick               # collect 1st tick
 156         li      $diff,0
 157
 158         dcbf    0,$out                  # flush cache line
 159         lwarx   $tick,0,$out            # load and lock
 160         add     $tick,$tick,$diff
 161         stwcx.  $tick,0,$out
 162         stwx    $tick,0,$out
 163
 164         mftb    $tick                   # collect 1st diff
 165         sub     $diff,$tick,$lasttick
 166         mr      $lasttick,$tick
 167         mr      $lastdiff,$diff
 168 Loop2:
 169         dcbf    0,$out                  # flush cache line
 170         lwarx   $tick,0,$out            # load and lock
 171         add     $tick,$tick,$diff
 172         stwcx.  $tick,0,$out
 173         stwx    $tick,0,$out
 174
 175         addic.  $max,$max,-1
 176         beq     Ldone2
 177
 178         mftb    $tick
 179         sub     $diff,$tick,$lasttick
 180         mr      $lasttick,$tick
 181         cmplw   7,$diff,$lastdiff
 182         mr      $lastdiff,$diff
 183
 184         mfcr    $tick                   # pull cr
 185         not     $tick,$tick             # flip bits
 186         rlwinm  $tick,$tick,1,29,29     # isolate flipped eq bit and scale
 187
 188         sub.    $cnt,$cnt,$tick         # conditional --$cnt
 189         add     $out,$out,$tick         # conditional ++$out
 190         bne     Loop2
 191
 192 Ldone2:
 193         srwi    $cnt,$cnt,2
 194         sub     r3,r0,$cnt
 195         blr
 196 ___
 197 }
 198
 199 $code =~ s/\`([^\`]*)\`/eval $1/gem;
 200 print $code;
 201 close STDOUT;