2
0

wp-x86_64.pl 21 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619
  1. #! /usr/bin/env perl
  2. # Copyright 2005-2020 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. # ====================================================================
  10. # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  11. # project. Rights for redistribution and usage in source and binary
  12. # forms are granted according to the License.
  13. # ====================================================================
  14. #
  15. # whirlpool_block for x86_64.
  16. #
  17. # 2500 cycles per 64-byte input block on AMD64, which is *identical*
  18. # to 32-bit MMX version executed on same CPU. So why did I bother?
  19. # Well, it's faster than gcc 3.3.2 generated code by over 50%, and
  20. # over 80% faster than PathScale 1.4, an "ambitious" commercial
  21. # compiler. Furthermore it surpasses gcc 3.4.3 by 170% and Sun Studio
  22. # 10 - by 360%[!]... What is it with x86_64 compilers? It's not the
  23. # first example when they fail to generate more optimal code, when
  24. # I believe they had *all* chances to...
  25. #
  26. # Note that register and stack frame layout are virtually identical
  27. # to 32-bit MMX version, except that %r8-15 are used instead of
  28. # %mm0-8. You can even notice that K[i] and S[i] are loaded to
  29. # %eax:%ebx as pair of 32-bit values and not as single 64-bit one.
  30. # This is done in order to avoid 64-bit shift penalties on Intel
  31. # EM64T core. Speaking of which! I bet it's possible to improve
  32. # Opteron performance by compressing the table to 2KB and replacing
  33. # unaligned references with complementary rotations [which would
  34. # incidentally replace lea instructions], but it would definitely
  35. # just "kill" EM64T, because it has only 1 shifter/rotator [against
  36. # 3 on Opteron] and which is *unacceptably* slow with 64-bit
  37. # operand.
  38. # $output is the last argument if it looks like a file (it has an extension)
  39. # $flavour is the first argument if it doesn't look like a file
  40. $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
  41. $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
  42. $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  43. $0 =~ m/(.*[\/\\])[^\/\\]+$/; my $dir=$1; my $xlate;
  44. ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  45. ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  46. die "can't locate x86_64-xlate.pl";
  47. open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
  48. or die "can't call $xlate: $!";
  49. *STDOUT=*OUT;
  50. sub L() { $code.=".byte ".join(',',@_)."\n"; }
  51. sub LL(){ $code.=".byte ".join(',',@_).",".join(',',@_)."\n"; }
  52. @mm=("%r8","%r9","%r10","%r11","%r12","%r13","%r14","%r15");
  53. $func="whirlpool_block";
  54. $table=".Ltable";
  55. $code=<<___;
  56. .text
  57. .globl $func
  58. .type $func,\@function,3
  59. .align 16
  60. $func:
  61. .cfi_startproc
  62. mov %rsp,%rax
  63. .cfi_def_cfa_register %rax
  64. push %rbx
  65. .cfi_push %rbx
  66. push %rbp
  67. .cfi_push %rbp
  68. push %r12
  69. .cfi_push %r12
  70. push %r13
  71. .cfi_push %r13
  72. push %r14
  73. .cfi_push %r14
  74. push %r15
  75. .cfi_push %r15
  76. sub \$128+40,%rsp
  77. and \$-64,%rsp
  78. lea 128(%rsp),%r10
  79. mov %rdi,0(%r10) # save parameter block
  80. mov %rsi,8(%r10)
  81. mov %rdx,16(%r10)
  82. mov %rax,32(%r10) # saved stack pointer
  83. .cfi_cfa_expression %rsp+`128+32`,deref,+8
  84. .Lprologue:
  85. mov %r10,%rbx
  86. lea $table(%rip),%rbp
  87. xor %rcx,%rcx
  88. xor %rdx,%rdx
  89. ___
  90. for($i=0;$i<8;$i++) { $code.="mov $i*8(%rdi),@mm[$i]\n"; } # L=H
  91. $code.=".Louterloop:\n";
  92. for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; } # K=L
  93. for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; } # L^=inp
  94. for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; } # S=L
  95. $code.=<<___;
  96. xor %rsi,%rsi
  97. mov %rsi,24(%rbx) # zero round counter
  98. jmp .Lround
  99. .align 16
  100. .Lround:
  101. mov 4096(%rbp,%rsi,8),@mm[0] # rc[r]
  102. mov 0(%rsp),%eax
  103. mov 4(%rsp),%ebx
  104. movz %al,%ecx
  105. movz %ah,%edx
  106. ___
  107. for($i=0;$i<8;$i++) {
  108. my $func = ($i==0)? "mov" : "xor";
  109. $code.=<<___;
  110. shr \$16,%eax
  111. lea (%rcx,%rcx),%rsi
  112. movz %al,%ecx
  113. lea (%rdx,%rdx),%rdi
  114. movz %ah,%edx
  115. xor 0(%rbp,%rsi,8),@mm[0]
  116. $func 7(%rbp,%rdi,8),@mm[1]
  117. mov $i*8+8(%rsp),%eax # ($i+1)*8
  118. lea (%rcx,%rcx),%rsi
  119. movz %bl,%ecx
  120. lea (%rdx,%rdx),%rdi
  121. movz %bh,%edx
  122. $func 6(%rbp,%rsi,8),@mm[2]
  123. $func 5(%rbp,%rdi,8),@mm[3]
  124. shr \$16,%ebx
  125. lea (%rcx,%rcx),%rsi
  126. movz %bl,%ecx
  127. lea (%rdx,%rdx),%rdi
  128. movz %bh,%edx
  129. $func 4(%rbp,%rsi,8),@mm[4]
  130. $func 3(%rbp,%rdi,8),@mm[5]
  131. mov $i*8+8+4(%rsp),%ebx # ($i+1)*8+4
  132. lea (%rcx,%rcx),%rsi
  133. movz %al,%ecx
  134. lea (%rdx,%rdx),%rdi
  135. movz %ah,%edx
  136. $func 2(%rbp,%rsi,8),@mm[6]
  137. $func 1(%rbp,%rdi,8),@mm[7]
  138. ___
  139. push(@mm,shift(@mm));
  140. }
  141. for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; } # K=L
  142. for($i=0;$i<8;$i++) {
  143. $code.=<<___;
  144. shr \$16,%eax
  145. lea (%rcx,%rcx),%rsi
  146. movz %al,%ecx
  147. lea (%rdx,%rdx),%rdi
  148. movz %ah,%edx
  149. xor 0(%rbp,%rsi,8),@mm[0]
  150. xor 7(%rbp,%rdi,8),@mm[1]
  151. `"mov 64+$i*8+8(%rsp),%eax" if($i<7);` # 64+($i+1)*8
  152. lea (%rcx,%rcx),%rsi
  153. movz %bl,%ecx
  154. lea (%rdx,%rdx),%rdi
  155. movz %bh,%edx
  156. xor 6(%rbp,%rsi,8),@mm[2]
  157. xor 5(%rbp,%rdi,8),@mm[3]
  158. shr \$16,%ebx
  159. lea (%rcx,%rcx),%rsi
  160. movz %bl,%ecx
  161. lea (%rdx,%rdx),%rdi
  162. movz %bh,%edx
  163. xor 4(%rbp,%rsi,8),@mm[4]
  164. xor 3(%rbp,%rdi,8),@mm[5]
  165. `"mov 64+$i*8+8+4(%rsp),%ebx" if($i<7);` # 64+($i+1)*8+4
  166. lea (%rcx,%rcx),%rsi
  167. movz %al,%ecx
  168. lea (%rdx,%rdx),%rdi
  169. movz %ah,%edx
  170. xor 2(%rbp,%rsi,8),@mm[6]
  171. xor 1(%rbp,%rdi,8),@mm[7]
  172. ___
  173. push(@mm,shift(@mm));
  174. }
  175. $code.=<<___;
  176. lea 128(%rsp),%rbx
  177. mov 24(%rbx),%rsi # pull round counter
  178. add \$1,%rsi
  179. cmp \$10,%rsi
  180. je .Lroundsdone
  181. mov %rsi,24(%rbx) # update round counter
  182. ___
  183. for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; } # S=L
  184. $code.=<<___;
  185. jmp .Lround
  186. .align 16
  187. .Lroundsdone:
  188. mov 0(%rbx),%rdi # reload argument block
  189. mov 8(%rbx),%rsi
  190. mov 16(%rbx),%rax
  191. ___
  192. for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; } # L^=inp
  193. for($i=0;$i<8;$i++) { $code.="xor $i*8(%rdi),@mm[$i]\n"; } # L^=H
  194. for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rdi)\n"; } # H=L
  195. $code.=<<___;
  196. lea 64(%rsi),%rsi # inp+=64
  197. sub \$1,%rax # num--
  198. jz .Lalldone
  199. mov %rsi,8(%rbx) # update parameter block
  200. mov %rax,16(%rbx)
  201. jmp .Louterloop
  202. .Lalldone:
  203. mov 32(%rbx),%rsi # restore saved pointer
  204. .cfi_def_cfa %rsi,8
  205. mov -48(%rsi),%r15
  206. .cfi_restore %r15
  207. mov -40(%rsi),%r14
  208. .cfi_restore %r14
  209. mov -32(%rsi),%r13
  210. .cfi_restore %r13
  211. mov -24(%rsi),%r12
  212. .cfi_restore %r12
  213. mov -16(%rsi),%rbp
  214. .cfi_restore %rbp
  215. mov -8(%rsi),%rbx
  216. .cfi_restore %rbx
  217. lea (%rsi),%rsp
  218. .cfi_def_cfa_register %rsp
  219. .Lepilogue:
  220. ret
  221. .cfi_endproc
  222. .size $func,.-$func
  223. .align 64
  224. .type $table,\@object
  225. $table:
  226. ___
  227. &LL(0x18,0x18,0x60,0x18,0xc0,0x78,0x30,0xd8);
  228. &LL(0x23,0x23,0x8c,0x23,0x05,0xaf,0x46,0x26);
  229. &LL(0xc6,0xc6,0x3f,0xc6,0x7e,0xf9,0x91,0xb8);
  230. &LL(0xe8,0xe8,0x87,0xe8,0x13,0x6f,0xcd,0xfb);
  231. &LL(0x87,0x87,0x26,0x87,0x4c,0xa1,0x13,0xcb);
  232. &LL(0xb8,0xb8,0xda,0xb8,0xa9,0x62,0x6d,0x11);
  233. &LL(0x01,0x01,0x04,0x01,0x08,0x05,0x02,0x09);
  234. &LL(0x4f,0x4f,0x21,0x4f,0x42,0x6e,0x9e,0x0d);
  235. &LL(0x36,0x36,0xd8,0x36,0xad,0xee,0x6c,0x9b);
  236. &LL(0xa6,0xa6,0xa2,0xa6,0x59,0x04,0x51,0xff);
  237. &LL(0xd2,0xd2,0x6f,0xd2,0xde,0xbd,0xb9,0x0c);
  238. &LL(0xf5,0xf5,0xf3,0xf5,0xfb,0x06,0xf7,0x0e);
  239. &LL(0x79,0x79,0xf9,0x79,0xef,0x80,0xf2,0x96);
  240. &LL(0x6f,0x6f,0xa1,0x6f,0x5f,0xce,0xde,0x30);
  241. &LL(0x91,0x91,0x7e,0x91,0xfc,0xef,0x3f,0x6d);
  242. &LL(0x52,0x52,0x55,0x52,0xaa,0x07,0xa4,0xf8);
  243. &LL(0x60,0x60,0x9d,0x60,0x27,0xfd,0xc0,0x47);
  244. &LL(0xbc,0xbc,0xca,0xbc,0x89,0x76,0x65,0x35);
  245. &LL(0x9b,0x9b,0x56,0x9b,0xac,0xcd,0x2b,0x37);
  246. &LL(0x8e,0x8e,0x02,0x8e,0x04,0x8c,0x01,0x8a);
  247. &LL(0xa3,0xa3,0xb6,0xa3,0x71,0x15,0x5b,0xd2);
  248. &LL(0x0c,0x0c,0x30,0x0c,0x60,0x3c,0x18,0x6c);
  249. &LL(0x7b,0x7b,0xf1,0x7b,0xff,0x8a,0xf6,0x84);
  250. &LL(0x35,0x35,0xd4,0x35,0xb5,0xe1,0x6a,0x80);
  251. &LL(0x1d,0x1d,0x74,0x1d,0xe8,0x69,0x3a,0xf5);
  252. &LL(0xe0,0xe0,0xa7,0xe0,0x53,0x47,0xdd,0xb3);
  253. &LL(0xd7,0xd7,0x7b,0xd7,0xf6,0xac,0xb3,0x21);
  254. &LL(0xc2,0xc2,0x2f,0xc2,0x5e,0xed,0x99,0x9c);
  255. &LL(0x2e,0x2e,0xb8,0x2e,0x6d,0x96,0x5c,0x43);
  256. &LL(0x4b,0x4b,0x31,0x4b,0x62,0x7a,0x96,0x29);
  257. &LL(0xfe,0xfe,0xdf,0xfe,0xa3,0x21,0xe1,0x5d);
  258. &LL(0x57,0x57,0x41,0x57,0x82,0x16,0xae,0xd5);
  259. &LL(0x15,0x15,0x54,0x15,0xa8,0x41,0x2a,0xbd);
  260. &LL(0x77,0x77,0xc1,0x77,0x9f,0xb6,0xee,0xe8);
  261. &LL(0x37,0x37,0xdc,0x37,0xa5,0xeb,0x6e,0x92);
  262. &LL(0xe5,0xe5,0xb3,0xe5,0x7b,0x56,0xd7,0x9e);
  263. &LL(0x9f,0x9f,0x46,0x9f,0x8c,0xd9,0x23,0x13);
  264. &LL(0xf0,0xf0,0xe7,0xf0,0xd3,0x17,0xfd,0x23);
  265. &LL(0x4a,0x4a,0x35,0x4a,0x6a,0x7f,0x94,0x20);
  266. &LL(0xda,0xda,0x4f,0xda,0x9e,0x95,0xa9,0x44);
  267. &LL(0x58,0x58,0x7d,0x58,0xfa,0x25,0xb0,0xa2);
  268. &LL(0xc9,0xc9,0x03,0xc9,0x06,0xca,0x8f,0xcf);
  269. &LL(0x29,0x29,0xa4,0x29,0x55,0x8d,0x52,0x7c);
  270. &LL(0x0a,0x0a,0x28,0x0a,0x50,0x22,0x14,0x5a);
  271. &LL(0xb1,0xb1,0xfe,0xb1,0xe1,0x4f,0x7f,0x50);
  272. &LL(0xa0,0xa0,0xba,0xa0,0x69,0x1a,0x5d,0xc9);
  273. &LL(0x6b,0x6b,0xb1,0x6b,0x7f,0xda,0xd6,0x14);
  274. &LL(0x85,0x85,0x2e,0x85,0x5c,0xab,0x17,0xd9);
  275. &LL(0xbd,0xbd,0xce,0xbd,0x81,0x73,0x67,0x3c);
  276. &LL(0x5d,0x5d,0x69,0x5d,0xd2,0x34,0xba,0x8f);
  277. &LL(0x10,0x10,0x40,0x10,0x80,0x50,0x20,0x90);
  278. &LL(0xf4,0xf4,0xf7,0xf4,0xf3,0x03,0xf5,0x07);
  279. &LL(0xcb,0xcb,0x0b,0xcb,0x16,0xc0,0x8b,0xdd);
  280. &LL(0x3e,0x3e,0xf8,0x3e,0xed,0xc6,0x7c,0xd3);
  281. &LL(0x05,0x05,0x14,0x05,0x28,0x11,0x0a,0x2d);
  282. &LL(0x67,0x67,0x81,0x67,0x1f,0xe6,0xce,0x78);
  283. &LL(0xe4,0xe4,0xb7,0xe4,0x73,0x53,0xd5,0x97);
  284. &LL(0x27,0x27,0x9c,0x27,0x25,0xbb,0x4e,0x02);
  285. &LL(0x41,0x41,0x19,0x41,0x32,0x58,0x82,0x73);
  286. &LL(0x8b,0x8b,0x16,0x8b,0x2c,0x9d,0x0b,0xa7);
  287. &LL(0xa7,0xa7,0xa6,0xa7,0x51,0x01,0x53,0xf6);
  288. &LL(0x7d,0x7d,0xe9,0x7d,0xcf,0x94,0xfa,0xb2);
  289. &LL(0x95,0x95,0x6e,0x95,0xdc,0xfb,0x37,0x49);
  290. &LL(0xd8,0xd8,0x47,0xd8,0x8e,0x9f,0xad,0x56);
  291. &LL(0xfb,0xfb,0xcb,0xfb,0x8b,0x30,0xeb,0x70);
  292. &LL(0xee,0xee,0x9f,0xee,0x23,0x71,0xc1,0xcd);
  293. &LL(0x7c,0x7c,0xed,0x7c,0xc7,0x91,0xf8,0xbb);
  294. &LL(0x66,0x66,0x85,0x66,0x17,0xe3,0xcc,0x71);
  295. &LL(0xdd,0xdd,0x53,0xdd,0xa6,0x8e,0xa7,0x7b);
  296. &LL(0x17,0x17,0x5c,0x17,0xb8,0x4b,0x2e,0xaf);
  297. &LL(0x47,0x47,0x01,0x47,0x02,0x46,0x8e,0x45);
  298. &LL(0x9e,0x9e,0x42,0x9e,0x84,0xdc,0x21,0x1a);
  299. &LL(0xca,0xca,0x0f,0xca,0x1e,0xc5,0x89,0xd4);
  300. &LL(0x2d,0x2d,0xb4,0x2d,0x75,0x99,0x5a,0x58);
  301. &LL(0xbf,0xbf,0xc6,0xbf,0x91,0x79,0x63,0x2e);
  302. &LL(0x07,0x07,0x1c,0x07,0x38,0x1b,0x0e,0x3f);
  303. &LL(0xad,0xad,0x8e,0xad,0x01,0x23,0x47,0xac);
  304. &LL(0x5a,0x5a,0x75,0x5a,0xea,0x2f,0xb4,0xb0);
  305. &LL(0x83,0x83,0x36,0x83,0x6c,0xb5,0x1b,0xef);
  306. &LL(0x33,0x33,0xcc,0x33,0x85,0xff,0x66,0xb6);
  307. &LL(0x63,0x63,0x91,0x63,0x3f,0xf2,0xc6,0x5c);
  308. &LL(0x02,0x02,0x08,0x02,0x10,0x0a,0x04,0x12);
  309. &LL(0xaa,0xaa,0x92,0xaa,0x39,0x38,0x49,0x93);
  310. &LL(0x71,0x71,0xd9,0x71,0xaf,0xa8,0xe2,0xde);
  311. &LL(0xc8,0xc8,0x07,0xc8,0x0e,0xcf,0x8d,0xc6);
  312. &LL(0x19,0x19,0x64,0x19,0xc8,0x7d,0x32,0xd1);
  313. &LL(0x49,0x49,0x39,0x49,0x72,0x70,0x92,0x3b);
  314. &LL(0xd9,0xd9,0x43,0xd9,0x86,0x9a,0xaf,0x5f);
  315. &LL(0xf2,0xf2,0xef,0xf2,0xc3,0x1d,0xf9,0x31);
  316. &LL(0xe3,0xe3,0xab,0xe3,0x4b,0x48,0xdb,0xa8);
  317. &LL(0x5b,0x5b,0x71,0x5b,0xe2,0x2a,0xb6,0xb9);
  318. &LL(0x88,0x88,0x1a,0x88,0x34,0x92,0x0d,0xbc);
  319. &LL(0x9a,0x9a,0x52,0x9a,0xa4,0xc8,0x29,0x3e);
  320. &LL(0x26,0x26,0x98,0x26,0x2d,0xbe,0x4c,0x0b);
  321. &LL(0x32,0x32,0xc8,0x32,0x8d,0xfa,0x64,0xbf);
  322. &LL(0xb0,0xb0,0xfa,0xb0,0xe9,0x4a,0x7d,0x59);
  323. &LL(0xe9,0xe9,0x83,0xe9,0x1b,0x6a,0xcf,0xf2);
  324. &LL(0x0f,0x0f,0x3c,0x0f,0x78,0x33,0x1e,0x77);
  325. &LL(0xd5,0xd5,0x73,0xd5,0xe6,0xa6,0xb7,0x33);
  326. &LL(0x80,0x80,0x3a,0x80,0x74,0xba,0x1d,0xf4);
  327. &LL(0xbe,0xbe,0xc2,0xbe,0x99,0x7c,0x61,0x27);
  328. &LL(0xcd,0xcd,0x13,0xcd,0x26,0xde,0x87,0xeb);
  329. &LL(0x34,0x34,0xd0,0x34,0xbd,0xe4,0x68,0x89);
  330. &LL(0x48,0x48,0x3d,0x48,0x7a,0x75,0x90,0x32);
  331. &LL(0xff,0xff,0xdb,0xff,0xab,0x24,0xe3,0x54);
  332. &LL(0x7a,0x7a,0xf5,0x7a,0xf7,0x8f,0xf4,0x8d);
  333. &LL(0x90,0x90,0x7a,0x90,0xf4,0xea,0x3d,0x64);
  334. &LL(0x5f,0x5f,0x61,0x5f,0xc2,0x3e,0xbe,0x9d);
  335. &LL(0x20,0x20,0x80,0x20,0x1d,0xa0,0x40,0x3d);
  336. &LL(0x68,0x68,0xbd,0x68,0x67,0xd5,0xd0,0x0f);
  337. &LL(0x1a,0x1a,0x68,0x1a,0xd0,0x72,0x34,0xca);
  338. &LL(0xae,0xae,0x82,0xae,0x19,0x2c,0x41,0xb7);
  339. &LL(0xb4,0xb4,0xea,0xb4,0xc9,0x5e,0x75,0x7d);
  340. &LL(0x54,0x54,0x4d,0x54,0x9a,0x19,0xa8,0xce);
  341. &LL(0x93,0x93,0x76,0x93,0xec,0xe5,0x3b,0x7f);
  342. &LL(0x22,0x22,0x88,0x22,0x0d,0xaa,0x44,0x2f);
  343. &LL(0x64,0x64,0x8d,0x64,0x07,0xe9,0xc8,0x63);
  344. &LL(0xf1,0xf1,0xe3,0xf1,0xdb,0x12,0xff,0x2a);
  345. &LL(0x73,0x73,0xd1,0x73,0xbf,0xa2,0xe6,0xcc);
  346. &LL(0x12,0x12,0x48,0x12,0x90,0x5a,0x24,0x82);
  347. &LL(0x40,0x40,0x1d,0x40,0x3a,0x5d,0x80,0x7a);
  348. &LL(0x08,0x08,0x20,0x08,0x40,0x28,0x10,0x48);
  349. &LL(0xc3,0xc3,0x2b,0xc3,0x56,0xe8,0x9b,0x95);
  350. &LL(0xec,0xec,0x97,0xec,0x33,0x7b,0xc5,0xdf);
  351. &LL(0xdb,0xdb,0x4b,0xdb,0x96,0x90,0xab,0x4d);
  352. &LL(0xa1,0xa1,0xbe,0xa1,0x61,0x1f,0x5f,0xc0);
  353. &LL(0x8d,0x8d,0x0e,0x8d,0x1c,0x83,0x07,0x91);
  354. &LL(0x3d,0x3d,0xf4,0x3d,0xf5,0xc9,0x7a,0xc8);
  355. &LL(0x97,0x97,0x66,0x97,0xcc,0xf1,0x33,0x5b);
  356. &LL(0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
  357. &LL(0xcf,0xcf,0x1b,0xcf,0x36,0xd4,0x83,0xf9);
  358. &LL(0x2b,0x2b,0xac,0x2b,0x45,0x87,0x56,0x6e);
  359. &LL(0x76,0x76,0xc5,0x76,0x97,0xb3,0xec,0xe1);
  360. &LL(0x82,0x82,0x32,0x82,0x64,0xb0,0x19,0xe6);
  361. &LL(0xd6,0xd6,0x7f,0xd6,0xfe,0xa9,0xb1,0x28);
  362. &LL(0x1b,0x1b,0x6c,0x1b,0xd8,0x77,0x36,0xc3);
  363. &LL(0xb5,0xb5,0xee,0xb5,0xc1,0x5b,0x77,0x74);
  364. &LL(0xaf,0xaf,0x86,0xaf,0x11,0x29,0x43,0xbe);
  365. &LL(0x6a,0x6a,0xb5,0x6a,0x77,0xdf,0xd4,0x1d);
  366. &LL(0x50,0x50,0x5d,0x50,0xba,0x0d,0xa0,0xea);
  367. &LL(0x45,0x45,0x09,0x45,0x12,0x4c,0x8a,0x57);
  368. &LL(0xf3,0xf3,0xeb,0xf3,0xcb,0x18,0xfb,0x38);
  369. &LL(0x30,0x30,0xc0,0x30,0x9d,0xf0,0x60,0xad);
  370. &LL(0xef,0xef,0x9b,0xef,0x2b,0x74,0xc3,0xc4);
  371. &LL(0x3f,0x3f,0xfc,0x3f,0xe5,0xc3,0x7e,0xda);
  372. &LL(0x55,0x55,0x49,0x55,0x92,0x1c,0xaa,0xc7);
  373. &LL(0xa2,0xa2,0xb2,0xa2,0x79,0x10,0x59,0xdb);
  374. &LL(0xea,0xea,0x8f,0xea,0x03,0x65,0xc9,0xe9);
  375. &LL(0x65,0x65,0x89,0x65,0x0f,0xec,0xca,0x6a);
  376. &LL(0xba,0xba,0xd2,0xba,0xb9,0x68,0x69,0x03);
  377. &LL(0x2f,0x2f,0xbc,0x2f,0x65,0x93,0x5e,0x4a);
  378. &LL(0xc0,0xc0,0x27,0xc0,0x4e,0xe7,0x9d,0x8e);
  379. &LL(0xde,0xde,0x5f,0xde,0xbe,0x81,0xa1,0x60);
  380. &LL(0x1c,0x1c,0x70,0x1c,0xe0,0x6c,0x38,0xfc);
  381. &LL(0xfd,0xfd,0xd3,0xfd,0xbb,0x2e,0xe7,0x46);
  382. &LL(0x4d,0x4d,0x29,0x4d,0x52,0x64,0x9a,0x1f);
  383. &LL(0x92,0x92,0x72,0x92,0xe4,0xe0,0x39,0x76);
  384. &LL(0x75,0x75,0xc9,0x75,0x8f,0xbc,0xea,0xfa);
  385. &LL(0x06,0x06,0x18,0x06,0x30,0x1e,0x0c,0x36);
  386. &LL(0x8a,0x8a,0x12,0x8a,0x24,0x98,0x09,0xae);
  387. &LL(0xb2,0xb2,0xf2,0xb2,0xf9,0x40,0x79,0x4b);
  388. &LL(0xe6,0xe6,0xbf,0xe6,0x63,0x59,0xd1,0x85);
  389. &LL(0x0e,0x0e,0x38,0x0e,0x70,0x36,0x1c,0x7e);
  390. &LL(0x1f,0x1f,0x7c,0x1f,0xf8,0x63,0x3e,0xe7);
  391. &LL(0x62,0x62,0x95,0x62,0x37,0xf7,0xc4,0x55);
  392. &LL(0xd4,0xd4,0x77,0xd4,0xee,0xa3,0xb5,0x3a);
  393. &LL(0xa8,0xa8,0x9a,0xa8,0x29,0x32,0x4d,0x81);
  394. &LL(0x96,0x96,0x62,0x96,0xc4,0xf4,0x31,0x52);
  395. &LL(0xf9,0xf9,0xc3,0xf9,0x9b,0x3a,0xef,0x62);
  396. &LL(0xc5,0xc5,0x33,0xc5,0x66,0xf6,0x97,0xa3);
  397. &LL(0x25,0x25,0x94,0x25,0x35,0xb1,0x4a,0x10);
  398. &LL(0x59,0x59,0x79,0x59,0xf2,0x20,0xb2,0xab);
  399. &LL(0x84,0x84,0x2a,0x84,0x54,0xae,0x15,0xd0);
  400. &LL(0x72,0x72,0xd5,0x72,0xb7,0xa7,0xe4,0xc5);
  401. &LL(0x39,0x39,0xe4,0x39,0xd5,0xdd,0x72,0xec);
  402. &LL(0x4c,0x4c,0x2d,0x4c,0x5a,0x61,0x98,0x16);
  403. &LL(0x5e,0x5e,0x65,0x5e,0xca,0x3b,0xbc,0x94);
  404. &LL(0x78,0x78,0xfd,0x78,0xe7,0x85,0xf0,0x9f);
  405. &LL(0x38,0x38,0xe0,0x38,0xdd,0xd8,0x70,0xe5);
  406. &LL(0x8c,0x8c,0x0a,0x8c,0x14,0x86,0x05,0x98);
  407. &LL(0xd1,0xd1,0x63,0xd1,0xc6,0xb2,0xbf,0x17);
  408. &LL(0xa5,0xa5,0xae,0xa5,0x41,0x0b,0x57,0xe4);
  409. &LL(0xe2,0xe2,0xaf,0xe2,0x43,0x4d,0xd9,0xa1);
  410. &LL(0x61,0x61,0x99,0x61,0x2f,0xf8,0xc2,0x4e);
  411. &LL(0xb3,0xb3,0xf6,0xb3,0xf1,0x45,0x7b,0x42);
  412. &LL(0x21,0x21,0x84,0x21,0x15,0xa5,0x42,0x34);
  413. &LL(0x9c,0x9c,0x4a,0x9c,0x94,0xd6,0x25,0x08);
  414. &LL(0x1e,0x1e,0x78,0x1e,0xf0,0x66,0x3c,0xee);
  415. &LL(0x43,0x43,0x11,0x43,0x22,0x52,0x86,0x61);
  416. &LL(0xc7,0xc7,0x3b,0xc7,0x76,0xfc,0x93,0xb1);
  417. &LL(0xfc,0xfc,0xd7,0xfc,0xb3,0x2b,0xe5,0x4f);
  418. &LL(0x04,0x04,0x10,0x04,0x20,0x14,0x08,0x24);
  419. &LL(0x51,0x51,0x59,0x51,0xb2,0x08,0xa2,0xe3);
  420. &LL(0x99,0x99,0x5e,0x99,0xbc,0xc7,0x2f,0x25);
  421. &LL(0x6d,0x6d,0xa9,0x6d,0x4f,0xc4,0xda,0x22);
  422. &LL(0x0d,0x0d,0x34,0x0d,0x68,0x39,0x1a,0x65);
  423. &LL(0xfa,0xfa,0xcf,0xfa,0x83,0x35,0xe9,0x79);
  424. &LL(0xdf,0xdf,0x5b,0xdf,0xb6,0x84,0xa3,0x69);
  425. &LL(0x7e,0x7e,0xe5,0x7e,0xd7,0x9b,0xfc,0xa9);
  426. &LL(0x24,0x24,0x90,0x24,0x3d,0xb4,0x48,0x19);
  427. &LL(0x3b,0x3b,0xec,0x3b,0xc5,0xd7,0x76,0xfe);
  428. &LL(0xab,0xab,0x96,0xab,0x31,0x3d,0x4b,0x9a);
  429. &LL(0xce,0xce,0x1f,0xce,0x3e,0xd1,0x81,0xf0);
  430. &LL(0x11,0x11,0x44,0x11,0x88,0x55,0x22,0x99);
  431. &LL(0x8f,0x8f,0x06,0x8f,0x0c,0x89,0x03,0x83);
  432. &LL(0x4e,0x4e,0x25,0x4e,0x4a,0x6b,0x9c,0x04);
  433. &LL(0xb7,0xb7,0xe6,0xb7,0xd1,0x51,0x73,0x66);
  434. &LL(0xeb,0xeb,0x8b,0xeb,0x0b,0x60,0xcb,0xe0);
  435. &LL(0x3c,0x3c,0xf0,0x3c,0xfd,0xcc,0x78,0xc1);
  436. &LL(0x81,0x81,0x3e,0x81,0x7c,0xbf,0x1f,0xfd);
  437. &LL(0x94,0x94,0x6a,0x94,0xd4,0xfe,0x35,0x40);
  438. &LL(0xf7,0xf7,0xfb,0xf7,0xeb,0x0c,0xf3,0x1c);
  439. &LL(0xb9,0xb9,0xde,0xb9,0xa1,0x67,0x6f,0x18);
  440. &LL(0x13,0x13,0x4c,0x13,0x98,0x5f,0x26,0x8b);
  441. &LL(0x2c,0x2c,0xb0,0x2c,0x7d,0x9c,0x58,0x51);
  442. &LL(0xd3,0xd3,0x6b,0xd3,0xd6,0xb8,0xbb,0x05);
  443. &LL(0xe7,0xe7,0xbb,0xe7,0x6b,0x5c,0xd3,0x8c);
  444. &LL(0x6e,0x6e,0xa5,0x6e,0x57,0xcb,0xdc,0x39);
  445. &LL(0xc4,0xc4,0x37,0xc4,0x6e,0xf3,0x95,0xaa);
  446. &LL(0x03,0x03,0x0c,0x03,0x18,0x0f,0x06,0x1b);
  447. &LL(0x56,0x56,0x45,0x56,0x8a,0x13,0xac,0xdc);
  448. &LL(0x44,0x44,0x0d,0x44,0x1a,0x49,0x88,0x5e);
  449. &LL(0x7f,0x7f,0xe1,0x7f,0xdf,0x9e,0xfe,0xa0);
  450. &LL(0xa9,0xa9,0x9e,0xa9,0x21,0x37,0x4f,0x88);
  451. &LL(0x2a,0x2a,0xa8,0x2a,0x4d,0x82,0x54,0x67);
  452. &LL(0xbb,0xbb,0xd6,0xbb,0xb1,0x6d,0x6b,0x0a);
  453. &LL(0xc1,0xc1,0x23,0xc1,0x46,0xe2,0x9f,0x87);
  454. &LL(0x53,0x53,0x51,0x53,0xa2,0x02,0xa6,0xf1);
  455. &LL(0xdc,0xdc,0x57,0xdc,0xae,0x8b,0xa5,0x72);
  456. &LL(0x0b,0x0b,0x2c,0x0b,0x58,0x27,0x16,0x53);
  457. &LL(0x9d,0x9d,0x4e,0x9d,0x9c,0xd3,0x27,0x01);
  458. &LL(0x6c,0x6c,0xad,0x6c,0x47,0xc1,0xd8,0x2b);
  459. &LL(0x31,0x31,0xc4,0x31,0x95,0xf5,0x62,0xa4);
  460. &LL(0x74,0x74,0xcd,0x74,0x87,0xb9,0xe8,0xf3);
  461. &LL(0xf6,0xf6,0xff,0xf6,0xe3,0x09,0xf1,0x15);
  462. &LL(0x46,0x46,0x05,0x46,0x0a,0x43,0x8c,0x4c);
  463. &LL(0xac,0xac,0x8a,0xac,0x09,0x26,0x45,0xa5);
  464. &LL(0x89,0x89,0x1e,0x89,0x3c,0x97,0x0f,0xb5);
  465. &LL(0x14,0x14,0x50,0x14,0xa0,0x44,0x28,0xb4);
  466. &LL(0xe1,0xe1,0xa3,0xe1,0x5b,0x42,0xdf,0xba);
  467. &LL(0x16,0x16,0x58,0x16,0xb0,0x4e,0x2c,0xa6);
  468. &LL(0x3a,0x3a,0xe8,0x3a,0xcd,0xd2,0x74,0xf7);
  469. &LL(0x69,0x69,0xb9,0x69,0x6f,0xd0,0xd2,0x06);
  470. &LL(0x09,0x09,0x24,0x09,0x48,0x2d,0x12,0x41);
  471. &LL(0x70,0x70,0xdd,0x70,0xa7,0xad,0xe0,0xd7);
  472. &LL(0xb6,0xb6,0xe2,0xb6,0xd9,0x54,0x71,0x6f);
  473. &LL(0xd0,0xd0,0x67,0xd0,0xce,0xb7,0xbd,0x1e);
  474. &LL(0xed,0xed,0x93,0xed,0x3b,0x7e,0xc7,0xd6);
  475. &LL(0xcc,0xcc,0x17,0xcc,0x2e,0xdb,0x85,0xe2);
  476. &LL(0x42,0x42,0x15,0x42,0x2a,0x57,0x84,0x68);
  477. &LL(0x98,0x98,0x5a,0x98,0xb4,0xc2,0x2d,0x2c);
  478. &LL(0xa4,0xa4,0xaa,0xa4,0x49,0x0e,0x55,0xed);
  479. &LL(0x28,0x28,0xa0,0x28,0x5d,0x88,0x50,0x75);
  480. &LL(0x5c,0x5c,0x6d,0x5c,0xda,0x31,0xb8,0x86);
  481. &LL(0xf8,0xf8,0xc7,0xf8,0x93,0x3f,0xed,0x6b);
  482. &LL(0x86,0x86,0x22,0x86,0x44,0xa4,0x11,0xc2);
  483. &L(0x18,0x23,0xc6,0xe8,0x87,0xb8,0x01,0x4f); # rc[ROUNDS]
  484. &L(0x36,0xa6,0xd2,0xf5,0x79,0x6f,0x91,0x52);
  485. &L(0x60,0xbc,0x9b,0x8e,0xa3,0x0c,0x7b,0x35);
  486. &L(0x1d,0xe0,0xd7,0xc2,0x2e,0x4b,0xfe,0x57);
  487. &L(0x15,0x77,0x37,0xe5,0x9f,0xf0,0x4a,0xda);
  488. &L(0x58,0xc9,0x29,0x0a,0xb1,0xa0,0x6b,0x85);
  489. &L(0xbd,0x5d,0x10,0xf4,0xcb,0x3e,0x05,0x67);
  490. &L(0xe4,0x27,0x41,0x8b,0xa7,0x7d,0x95,0xd8);
  491. &L(0xfb,0xee,0x7c,0x66,0xdd,0x17,0x47,0x9e);
  492. &L(0xca,0x2d,0xbf,0x07,0xad,0x5a,0x83,0x33);
  493. # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
  494. # CONTEXT *context,DISPATCHER_CONTEXT *disp)
  495. if ($win64) {
  496. $rec="%rcx";
  497. $frame="%rdx";
  498. $context="%r8";
  499. $disp="%r9";
  500. $code.=<<___;
  501. .extern __imp_RtlVirtualUnwind
  502. .type se_handler,\@abi-omnipotent
  503. .align 16
  504. se_handler:
  505. push %rsi
  506. push %rdi
  507. push %rbx
  508. push %rbp
  509. push %r12
  510. push %r13
  511. push %r14
  512. push %r15
  513. pushfq
  514. sub \$64,%rsp
  515. mov 120($context),%rax # pull context->Rax
  516. mov 248($context),%rbx # pull context->Rip
  517. lea .Lprologue(%rip),%r10
  518. cmp %r10,%rbx # context->Rip<.Lprologue
  519. jb .Lin_prologue
  520. mov 152($context),%rax # pull context->Rsp
  521. lea .Lepilogue(%rip),%r10
  522. cmp %r10,%rbx # context->Rip>=.Lepilogue
  523. jae .Lin_prologue
  524. mov 128+32(%rax),%rax # pull saved stack pointer
  525. mov -8(%rax),%rbx
  526. mov -16(%rax),%rbp
  527. mov -24(%rax),%r12
  528. mov -32(%rax),%r13
  529. mov -40(%rax),%r14
  530. mov -48(%rax),%r15
  531. mov %rbx,144($context) # restore context->Rbx
  532. mov %rbp,160($context) # restore context->Rbp
  533. mov %r12,216($context) # restore context->R12
  534. mov %r13,224($context) # restore context->R13
  535. mov %r14,232($context) # restore context->R14
  536. mov %r15,240($context) # restore context->R15
  537. .Lin_prologue:
  538. mov 8(%rax),%rdi
  539. mov 16(%rax),%rsi
  540. mov %rax,152($context) # restore context->Rsp
  541. mov %rsi,168($context) # restore context->Rsi
  542. mov %rdi,176($context) # restore context->Rdi
  543. mov 40($disp),%rdi # disp->ContextRecord
  544. mov $context,%rsi # context
  545. mov \$154,%ecx # sizeof(CONTEXT)
  546. .long 0xa548f3fc # cld; rep movsq
  547. mov $disp,%rsi
  548. xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
  549. mov 8(%rsi),%rdx # arg2, disp->ImageBase
  550. mov 0(%rsi),%r8 # arg3, disp->ControlPc
  551. mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
  552. mov 40(%rsi),%r10 # disp->ContextRecord
  553. lea 56(%rsi),%r11 # &disp->HandlerData
  554. lea 24(%rsi),%r12 # &disp->EstablisherFrame
  555. mov %r10,32(%rsp) # arg5
  556. mov %r11,40(%rsp) # arg6
  557. mov %r12,48(%rsp) # arg7
  558. mov %rcx,56(%rsp) # arg8, (NULL)
  559. call *__imp_RtlVirtualUnwind(%rip)
  560. mov \$1,%eax # ExceptionContinueSearch
  561. add \$64,%rsp
  562. popfq
  563. pop %r15
  564. pop %r14
  565. pop %r13
  566. pop %r12
  567. pop %rbp
  568. pop %rbx
  569. pop %rdi
  570. pop %rsi
  571. ret
  572. .size se_handler,.-se_handler
  573. .section .pdata
  574. .align 4
  575. .rva .LSEH_begin_$func
  576. .rva .LSEH_end_$func
  577. .rva .LSEH_info_$func
  578. .section .xdata
  579. .align 8
  580. .LSEH_info_$func:
  581. .byte 9,0,0,0
  582. .rva se_handler
  583. ___
  584. }
  585. $code =~ s/\`([^\`]*)\`/eval $1/gem;
  586. print $code;
  587. close STDOUT or die "error closing STDOUT: $!";