engines/asm/e_padlock-x86_64.pl

   1 #!/usr/bin/env perl
   2
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
   5 # project. The module is, however, dual licensed under OpenSSL and
   6 # CRYPTOGAMS licenses depending on where you obtain it. For further
   7 # details see http://www.openssl.org/~appro/cryptogams/.
   8 # ====================================================================
   9
  10 # September 2011
  11 #
  12 # Assembler helpers for Padlock engine.
  13
  14 $flavour = shift;
  15 $output  = shift;
  16 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  17
  18 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  19
  20 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  21 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  22 ( $xlate="${dir}../../crypto/perlasm/x86_64-xlate.pl" and -f $xlate) or
  23 die "can't locate x86_64-xlate.pl";
  24
  25 open STDOUT,"| $^X $xlate $flavour $output";
  26
  27 $code=".text\n";
  28
  29 $PADLOCK_CHUNK=512;     # Must be a power of 2 larger than 16
  30
  31 $ctx="%rdx";
  32 $out="%rdi";
  33 $inp="%rsi";
  34 $len="%rcx";
  35 $chunk="%rbx";
  36
  37 ($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order
  38                                  ("%rdi","%rsi","%rdx","%rcx"); # Unix order
  39
  40 $code.=<<___;
  41 .globl  padlock_capability
  42 .type   padlock_capability,\@abi-omnipotent
  43 .align  16
  44 padlock_capability:
  45         mov     %rbx,%r8
  46         xor     %eax,%eax
  47         cpuid
  48         xor     %eax,%eax
  49         cmp     \$`"0x".unpack("H*",'tneC')`,%ebx
  50         jne     .Lnoluck
  51         cmp     \$`"0x".unpack("H*",'Hrua')`,%edx
  52         jne     .Lnoluck
  53         cmp     \$`"0x".unpack("H*",'slua')`,%ecx
  54         jne     .Lnoluck
  55         mov     \$0xC0000000,%eax
  56         cpuid
  57         mov     %eax,%edx
  58         xor     %eax,%eax
  59         cmp     \$0xC0000001,%edx
  60         jb      .Lnoluck
  61         mov     \$0xC0000001,%eax
  62         cpuid
  63         mov     %edx,%eax
  64         and     \$0xffffffef,%eax
  65         or      \$0x10,%eax             # set Nano bit#4
  66 .Lnoluck:
  67         mov     %r8,%rbx
  68         ret
  69 .size   padlock_capability,.-padlock_capability
  70
  71 .globl  padlock_key_bswap
  72 .type   padlock_key_bswap,\@abi-omnipotent,0
  73 .align  16
  74 padlock_key_bswap:
  75         mov     240($arg1),%edx
  76 .Lbswap_loop:
  77         mov     ($arg1),%eax
  78         bswap   %eax
  79         mov     %eax,($arg1)
  80         lea     4($arg1),$arg1
  81         sub     \$1,%edx
  82         jnz     .Lbswap_loop
  83         ret
  84 .size   padlock_key_bswap,.-padlock_key_bswap
  85
  86 .globl  padlock_verify_context
  87 .type   padlock_verify_context,\@abi-omnipotent
  88 .align  16
  89 padlock_verify_context:
  90         mov     $arg1,$ctx
  91         pushf
  92         lea     .Lpadlock_saved_context(%rip),%rax
  93         call    _padlock_verify_ctx
  94         lea     8(%rsp),%rsp
  95         ret
  96 .size   padlock_verify_context,.-padlock_verify_context
  97
  98 .type   _padlock_verify_ctx,\@abi-omnipotent
  99 .align  16
 100 _padlock_verify_ctx:
 101         mov     8(%rsp),%r8
 102         bt      \$30,%r8
 103         jnc     .Lverified
 104         cmp     (%rax),$ctx
 105         je      .Lverified
 106         pushf
 107         popf
 108 .Lverified:
 109         mov     $ctx,(%rax)
 110         ret
 111 .size   _padlock_verify_ctx,.-_padlock_verify_ctx
 112
 113 .globl  padlock_reload_key
 114 .type   padlock_reload_key,\@abi-omnipotent
 115 .align  16
 116 padlock_reload_key:
 117         pushf
 118         popf
 119         ret
 120 .size   padlock_reload_key,.-padlock_reload_key
 121
 122 .globl  padlock_aes_block
 123 .type   padlock_aes_block,\@function,3
 124 .align  16
 125 padlock_aes_block:
 126         mov     %rbx,%r8
 127         mov     \$1,$len
 128         lea     32($ctx),%rbx           # key
 129         lea     16($ctx),$ctx           # control word
 130         .byte   0xf3,0x0f,0xa7,0xc8     # rep xcryptecb
 131         mov     %r8,%rbx
 132         ret
 133 .size   padlock_aes_block,.-padlock_aes_block
 134
 135 .globl  padlock_xstore
 136 .type   padlock_xstore,\@function,2
 137 .align  16
 138 padlock_xstore:
 139         mov     %esi,%edx
 140         .byte   0x0f,0xa7,0xc0          # xstore
 141         ret
 142 .size   padlock_xstore,.-padlock_xstore
 143
 144 .globl  padlock_sha1_oneshot
 145 .type   padlock_sha1_oneshot,\@function,3
 146 .align  16
 147 padlock_sha1_oneshot:
 148         xor     %rax,%rax
 149         mov     %rdx,%rcx
 150         .byte   0xf3,0x0f,0xa6,0xc8     # rep xsha1
 151         ret
 152 .size   padlock_sha1_oneshot,.-padlock_sha1_oneshot
 153
 154 .globl  padlock_sha1_blocks
 155 .type   padlock_sha1_blocks,\@function,3
 156 .align  16
 157 padlock_sha1_blocks:
 158         mov     \$-1,%rax
 159         mov     %rdx,%rcx
 160         .byte   0xf3,0x0f,0xa6,0xc8     # rep xsha1
 161         ret
 162 .size   padlock_sha1_blocks,.-padlock_sha1_blocks
 163
 164 .globl  padlock_sha256_oneshot
 165 .type   padlock_sha256_oneshot,\@function,3
 166 .align  16
 167 padlock_sha256_oneshot:
 168         xor     %rax,%rax
 169         mov     %rdx,%rcx
 170         .byte   0xf3,0x0f,0xa6,0xd0     # rep xsha256
 171         ret
 172 .size   padlock_sha256_oneshot,.-padlock_sha256_oneshot
 173
 174 .globl  padlock_sha256_blocks
 175 .type   padlock_sha256_blocks,\@function,3
 176 .align  16
 177 padlock_sha256_blocks:
 178         mov     \$-1,%rax
 179         mov     %rdx,%rcx
 180         .byte   0xf3,0x0f,0xa6,0xd0     # rep xsha256
 181         ret
 182 .size   padlock_sha256_blocks,.-padlock_sha256_blocks
 183
 184 .globl  padlock_sha512_blocks,\@function,3
 185 .align  16
 186 padlock_sha512_blocks:
 187         mov     %rdx,%rcx
 188         .byte   0xf3,0x0f,0xa6,0xe0     # rep xha512
 189         ret
 190 .size   padlock_sha512_blocks,.-padlock_sha512_blocks
 191 ___
 192
 193 sub generate_mode {
 194 my ($mode,$opcode) = @_;
 195 # int padlock_$mode_encrypt(void *out, const void *inp,
 196 #               struct padlock_cipher_data *ctx, size_t len);
 197 $code.=<<___;
 198 .globl  padlock_${mode}_encrypt
 199 .type   padlock_${mode}_encrypt,\@function,4
 200 .align  16
 201 padlock_${mode}_encrypt:
 202         push    %rbp
 203         push    %rbx
 204
 205         xor     %eax,%eax
 206         test    \$15,$ctx
 207         jnz     .L${mode}_abort
 208         test    \$15,$len
 209         jnz     .L${mode}_abort
 210         lea     .Lpadlock_saved_context(%rip),%rax
 211         pushf
 212         cld
 213         call    _padlock_verify_ctx
 214         lea     16($ctx),$ctx           # control word
 215         xor     %eax,%eax
 216         xor     %ebx,%ebx
 217         testl   \$`1<<5`,($ctx)         # align bit in control word
 218         jnz     .L${mode}_aligned
 219         test    \$0x0f,$out
 220         setz    %al                     # !out_misaligned
 221         test    \$0x0f,$inp
 222         setz    %bl                     # !inp_misaligned
 223         test    %ebx,%eax
 224         jnz     .L${mode}_aligned
 225         neg     %rax
 226         mov     \$$PADLOCK_CHUNK,$chunk
 227         not     %rax                    # out_misaligned?-1:0
 228         lea     (%rsp),%rbp
 229         cmp     $chunk,$len
 230         cmovc   $len,$chunk             # chunk=len>PADLOCK_CHUNK?PADLOCK_CHUNK:len
 231         and     $chunk,%rax             # out_misaligned?chunk:0
 232         mov     $len,$chunk
 233         neg     %rax
 234         and     \$$PADLOCK_CHUNK-1,$chunk       # chunk%=PADLOCK_CHUNK
 235         lea     (%rax,%rbp),%rsp
 236         jmp     .L${mode}_loop
 237 .align  16
 238 .L${mode}_loop:
 239         mov     $out,%r8                # save parameters
 240         mov     $inp,%r9
 241         mov     $len,%r10
 242         mov     $chunk,$len
 243         mov     $chunk,%r11
 244         test    \$0x0f,$out             # out_misaligned
 245         cmovnz  %rsp,$out
 246         test    \$0x0f,$inp             # inp_misaligned
 247         jz      .L${mode}_inp_aligned
 248         shr     \$3,$len
 249         .byte   0xf3,0x48,0xa5          # rep movsq
 250         sub     $chunk,$out
 251         mov     $chunk,$len
 252         mov     $out,$inp
 253 .L${mode}_inp_aligned:
 254         lea     -16($ctx),%rax          # ivp
 255         lea     16($ctx),%rbx           # key
 256         shr     \$4,$len
 257         .byte   0xf3,0x0f,0xa7,$opcode  # rep xcrypt*
 258 ___
 259 $code.=<<___                            if ($mode !~ /ecb|ctr/);
 260         movdqa  (%rax),%xmm0
 261         movdqa  %xmm0,-16($ctx)         # copy [or refresh] iv
 262 ___
 263 $code.=<<___;
 264         mov     %r8,$out                # restore paramters
 265         mov     %r11,$chunk
 266         test    \$0x0f,$out
 267         jz      .L${mode}_out_aligned
 268         mov     $chunk,$len
 269         shr     \$3,$len
 270         lea     (%rsp),$inp
 271         .byte   0xf3,0x48,0xa5          # rep movsq
 272         sub     $chunk,$out
 273 .L${mode}_out_aligned:
 274         mov     %r9,$inp
 275         mov     %r10,$len
 276         add     $chunk,$out
 277         add     $chunk,$inp
 278         sub     $chunk,$len
 279         mov     \$$PADLOCK_CHUNK,$chunk
 280         jnz     .L${mode}_loop
 281
 282         test    \$0x0f,$out
 283         jz      .L${mode}_done
 284
 285         mov     %rbp,$len
 286         mov     %rsp,$out
 287         sub     %rsp,$len
 288         xor     %rax,%rax
 289         shr     \$3,$len
 290         .byte   0xf3,0x48,0xab          # rep stosq
 291 .L${mode}_done:
 292         lea     (%rbp),%rsp
 293         jmp     .L${mode}_exit
 294
 295 .align  16
 296 .L${mode}_aligned:
 297         lea     -16($ctx),%rax          # ivp
 298         lea     16($ctx),%rbx           # key
 299         shr     \$4,$len                # len/=AES_BLOCK_SIZE
 300         .byte   0xf3,0x0f,0xa7,$opcode  # rep xcrypt*
 301 ___
 302 $code.=<<___                            if ($mode !~ /ecb|ctr/);
 303         movdqa  (%rax),%xmm0
 304         movdqa  %xmm0,-16($ctx)         # copy [or refresh] iv
 305 ___
 306 $code.=<<___;
 307 .L${mode}_exit:
 308         mov     \$1,%eax
 309         lea     8(%rsp),%rsp
 310 .L${mode}_abort:
 311         pop     %rbx
 312         pop     %rbp
 313         ret
 314 .size   padlock_${mode}_encrypt,.-padlock_${mode}_encrypt
 315 ___
 316 }
 317
 318 &generate_mode("ecb",0xc8);
 319 &generate_mode("cbc",0xd0);
 320 &generate_mode("cfb",0xe0);
 321 &generate_mode("ofb",0xe8);
 322 &generate_mode("ctr16",0xd8);
 323
 324 $code.=<<___;
 325 .asciz  "VIA Padlock x86_64 module, CRYPTOGAMS by <appro\@openssl.org>"
 326 .align  16
 327 .data
 328 .align  8
 329 .Lpadlock_saved_context:
 330         .quad   0
 331 ___
 332 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
 333
 334 print $code;
 335
 336 close STDOUT;