crypto/c64xpluscpuid.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2012-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the Apache License 2.0 (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9 while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {}
  10 open STDOUT,">$output";
  11
  12 $code.=<<___;
  13         .text
  14
  15         .if     .ASSEMBLER_VERSION<7000000
  16         .asg    0,__TI_EABI__
  17         .endif
  18         .if     __TI_EABI__
  19         .asg    OPENSSL_rdtsc,_OPENSSL_rdtsc
  20         .asg    OPENSSL_cleanse,_OPENSSL_cleanse
  21         .asg    CRYPTO_memcmp,_CRYPTO_memcmp
  22         .asg    OPENSSL_atomic_add,_OPENSSL_atomic_add
  23         .asg    OPENSSL_wipe_cpu,_OPENSSL_wipe_cpu
  24         .asg    OPENSSL_instrument_bus,_OPENSSL_instrument_bus
  25         .asg    OPENSSL_instrument_bus2,_OPENSSL_instrument_bus2
  26         .endif
  27
  28         .asg    B3,RA
  29
  30         .global _OPENSSL_rdtsc
  31 _OPENSSL_rdtsc:
  32         .asmfunc
  33         B       RA
  34         MVC     TSCL,B0
  35         MVC     TSCH,B1
  36   [!B0] MVC     B0,TSCL         ; start TSC
  37         MV      B0,A4
  38         MV      B1,A5
  39         .endasmfunc
  40
  41         .global _OPENSSL_cleanse
  42 _OPENSSL_cleanse:
  43         .asmfunc
  44         ZERO    A3:A2
  45 ||      ZERO    B2
  46 ||      SHRU    B4,3,B0         ; is length >= 8
  47 ||      ADD     1,A4,B6
  48   [!B0] BNOP    RA
  49 ||      ZERO    A1
  50 ||      ZERO    B1
  51    [B0] MVC     B0,ILC
  52 ||[!B0] CMPLT   0,B4,A1
  53 ||[!B0] CMPLT   1,B4,B1
  54    [A1] STB     A2,*A4++[2]
  55 || [B1] STB     B2,*B6++[2]
  56 ||[!B0] CMPLT   2,B4,A1
  57 ||[!B0] CMPLT   3,B4,B1
  58    [A1] STB     A2,*A4++[2]
  59 || [B1] STB     B2,*B6++[2]
  60 ||[!B0] CMPLT   4,B4,A1
  61 ||[!B0] CMPLT   5,B4,B1
  62    [A1] STB     A2,*A4++[2]
  63 || [B1] STB     B2,*B6++[2]
  64 ||[!B0] CMPLT   6,B4,A1
  65    [A1] STB     A2,*A4++[2]
  66
  67         SPLOOP  1
  68         STNDW   A3:A2,*A4++
  69 ||      SUB     B4,8,B4
  70         SPKERNEL
  71
  72         MV      B4,B0           ; remaining bytes
  73 ||      ADD     1,A4,B6
  74 ||      BNOP    RA
  75    [B0] CMPLT   0,B0,A1
  76 || [B0] CMPLT   1,B0,B1
  77    [A1] STB     A2,*A4++[2]
  78 || [B1] STB     B2,*B6++[2]
  79 || [B0] CMPLT   2,B0,A1
  80 || [B0] CMPLT   3,B0,B1
  81    [A1] STB     A2,*A4++[2]
  82 || [B1] STB     B2,*B6++[2]
  83 || [B0] CMPLT   4,B0,A1
  84 || [B0] CMPLT   5,B0,B1
  85    [A1] STB     A2,*A4++[2]
  86 || [B1] STB     B2,*B6++[2]
  87 || [B0] CMPLT   6,B0,A1
  88    [A1] STB     A2,*A4++[2]
  89         .endasmfunc
  90
  91         .global _CRYPTO_memcmp
  92 _CRYPTO_memcmp:
  93         .asmfunc
  94         MV      A6,B0
  95   [!B0] BNOP    RA
  96 ||[!B0] ZERO    A4
  97    [B0] MVC     B0,ILC
  98 || [B0] ZERO    A0
  99         NOP     4
 100
 101         SPLOOP  1
 102         LDBU    *A4++,A1
 103 ||      LDBU    *B4++,B1
 104         NOP     4
 105         XOR.L   B1,A1,A2
 106         SPKERNEL 1,0
 107 ||      OR.S    A2,A0,A0
 108
 109         BNOP    RA,3
 110         ZERO.L  A4
 111   [A0]  MVK     1,A4
 112         .endasmfunc
 113
 114         .global _OPENSSL_atomic_add
 115 _OPENSSL_atomic_add:
 116         .asmfunc
 117         MV      A4,B0
 118 atomic_add?:
 119         LL      *B0,B5
 120         NOP     4
 121         ADD     B4,B5,B5
 122         SL      B5,*B0
 123         CMTL    *B0,B1
 124         NOP     4
 125   [!B1] B       atomic_add?
 126    [B1] BNOP    RA,4
 127         MV      B5,A4
 128         .endasmfunc
 129
 130         .global _OPENSSL_wipe_cpu
 131 _OPENSSL_wipe_cpu:
 132         .asmfunc
 133         ZERO    A0
 134 ||      ZERO    B0
 135 ||      ZERO    A1
 136 ||      ZERO    B1
 137         ZERO    A3:A2
 138 ||      MVD     B0,B2
 139 ||      ZERO    A4
 140 ||      ZERO    B4
 141 ||      ZERO    A5
 142 ||      ZERO    B5
 143 ||      BNOP    RA
 144         ZERO    A7:A6
 145 ||      ZERO    B7:B6
 146 ||      ZERO    A8
 147 ||      ZERO    B8
 148 ||      ZERO    A9
 149 ||      ZERO    B9
 150         ZERO    A17:A16
 151 ||      ZERO    B17:B16
 152 ||      ZERO    A18
 153 ||      ZERO    B18
 154 ||      ZERO    A19
 155 ||      ZERO    B19
 156         ZERO    A21:A20
 157 ||      ZERO    B21:B20
 158 ||      ZERO    A22
 159 ||      ZERO    B22
 160 ||      ZERO    A23
 161 ||      ZERO    B23
 162         ZERO    A25:A24
 163 ||      ZERO    B25:B24
 164 ||      ZERO    A26
 165 ||      ZERO    B26
 166 ||      ZERO    A27
 167 ||      ZERO    B27
 168         ZERO    A29:A28
 169 ||      ZERO    B29:B28
 170 ||      ZERO    A30
 171 ||      ZERO    B30
 172 ||      ZERO    A31
 173 ||      ZERO    B31
 174         .endasmfunc
 175
 176 CLFLUSH .macro  CONTROL,ADDR,LEN
 177         B       passthrough?
 178 ||      STW     ADDR,*CONTROL[0]
 179         STW     LEN,*CONTROL[1]
 180 spinlock?:
 181         LDW     *CONTROL[1],A0
 182         NOP     3
 183 passthrough?:
 184         NOP
 185   [A0]  BNOP    spinlock?,5
 186         .endm
 187
 188         .global _OPENSSL_instrument_bus
 189 _OPENSSL_instrument_bus:
 190         .asmfunc
 191         MV      B4,B0                   ; reassign sizeof(output)
 192 ||      MV      A4,B4                   ; reassign output
 193 ||      MVK     0x00004030,A3
 194         MV      B0,A4                   ; return value
 195 ||      MVK     1,A1
 196 ||      MVKH    0x01840000,A3           ; L1DWIBAR
 197         MVC     TSCL,B8                 ; collect 1st tick
 198 ||      MVK     0x00004010,A5
 199         MV      B8,B9                   ; lasttick = tick
 200 ||      MVK     0,B7                    ; lastdiff = 0
 201 ||      MVKH    0x01840000,A5           ; L2WIBAR
 202         CLFLUSH A3,B4,A1                ; write-back and invalidate L1D line
 203         CLFLUSH A5,B4,A1                ; write-back and invalidate L2 line
 204         LL      *B4,B5
 205         NOP     4
 206         ADD     B7,B5,B5
 207         SL      B5,*B4
 208         CMTL    *B4,B1
 209         NOP     4
 210         STW     B5,*B4
 211 bus_loop1?:
 212         MVC     TSCL,B8
 213 || [B0] SUB     B0,1,B0
 214         SUB     B8,B9,B7                ; lastdiff = tick - lasttick
 215 ||      MV      B8,B9                   ; lasttick = tick
 216         CLFLUSH A3,B4,A1                ; write-back and invalidate L1D line
 217         CLFLUSH A5,B4,A1                ; write-back and invalidate L2 line
 218         LL      *B4,B5
 219         NOP     4
 220         ADD     B7,B5,B5
 221         SL      B5,*B4
 222         CMTL    *B4,B1
 223         STW     B5,*B4                  ; [!B1] is removed to flatten samples
 224 ||      ADDK    4,B4
 225 || [B0] BNOP    bus_loop1?,5
 226
 227         BNOP    RA,5
 228         .endasmfunc
 229
 230         .global _OPENSSL_instrument_bus2
 231 _OPENSSL_instrument_bus2:
 232         .asmfunc
 233         MV      A6,B0                   ; reassign max
 234 ||      MV      B4,A6                   ; reassign sizeof(output)
 235 ||      MVK     0x00004030,A3
 236         MV      A4,B4                   ; reassign output
 237 ||      MVK     0,A4                    ; return value
 238 ||      MVK     1,A1
 239 ||      MVKH    0x01840000,A3           ; L1DWIBAR
 240
 241         MVC     TSCL,B8                 ; collect 1st tick
 242 ||      MVK     0x00004010,A5
 243         MV      B8,B9                   ; lasttick = tick
 244 ||      MVK     0,B7                    ; lastdiff = 0
 245 ||      MVKH    0x01840000,A5           ; L2WIBAR
 246         CLFLUSH A3,B4,A1                ; write-back and invalidate L1D line
 247         CLFLUSH A5,B4,A1                ; write-back and invalidate L2 line
 248         LL      *B4,B5
 249         NOP     4
 250         ADD     B7,B5,B5
 251         SL      B5,*B4
 252         CMTL    *B4,B1
 253         NOP     4
 254         STW     B5,*B4
 255
 256         MVC     TSCL,B8                 ; collect 1st diff
 257         SUB     B8,B9,B7                ; lastdiff = tick - lasttick
 258 ||      MV      B8,B9                   ; lasttick = tick
 259 ||      SUB     B0,1,B0
 260 bus_loop2?:
 261         CLFLUSH A3,B4,A1                ; write-back and invalidate L1D line
 262         CLFLUSH A5,B4,A1                ; write-back and invalidate L2 line
 263         LL      *B4,B5
 264         NOP     4
 265         ADD     B7,B5,B5
 266         SL      B5,*B4
 267         CMTL    *B4,B1
 268         STW     B5,*B4                  ; [!B1] is removed to flatten samples
 269 ||[!B0] BNOP    bus_loop2_done?,2
 270 ||      SUB     B0,1,B0
 271         MVC     TSCL,B8
 272         SUB     B8,B9,B8
 273 ||      MV      B8,B9
 274         CMPEQ   B8,B7,B2
 275 ||      MV      B8,B7
 276   [!B2] ADDAW   B4,1,B4
 277 ||[!B2] ADDK    1,A4
 278         CMPEQ   A4,A6,A2
 279   [!A2] BNOP    bus_loop2?,5
 280
 281 bus_loop2_done?:
 282         BNOP    RA,5
 283         .endasmfunc
 284 ___
 285
 286 print $code;
 287 close STDOUT;