]>
Commit | Line | Data |
---|---|---|
6aa36e8e | 1 | #! /usr/bin/env perl |
33388b44 | 2 | # Copyright 2005-2020 The OpenSSL Project Authors. All Rights Reserved. |
6aa36e8e | 3 | # |
677c7ab9 | 4 | # Licensed under the Apache License 2.0 (the "License"). You may not use |
6aa36e8e RS |
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 | ||
fec82dbc AP |
9 | # |
10 | # ==================================================================== | |
e3713c36 | 11 | # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL |
fec82dbc | 12 | # project. Rights for redistribution and usage in source and binary |
389c09fa | 13 | # forms are granted according to the License. |
fec82dbc AP |
14 | # ==================================================================== |
15 | # | |
16 | # whirlpool_block for x86_64. | |
17 | # | |
18 | # 2500 cycles per 64-byte input block on AMD64, which is *identical* | |
19 | # to 32-bit MMX version executed on same CPU. So why did I bother? | |
20 | # Well, it's faster than gcc 3.3.2 generated code by over 50%, and | |
21 | # over 80% faster than PathScale 1.4, an "ambitious" commercial | |
c6709b89 AP |
22 | # compiler. Furthermore it surpasses gcc 3.4.3 by 170% and Sun Studio |
23 | # 10 - by 360%[!]... What is it with x86_64 compilers? It's not the | |
24 | # first example when they fail to generate more optimal code, when | |
25 | # I believe they had *all* chances to... | |
fec82dbc AP |
26 | # |
27 | # Note that register and stack frame layout are virtually identical | |
28 | # to 32-bit MMX version, except that %r8-15 are used instead of | |
29 | # %mm0-8. You can even notice that K[i] and S[i] are loaded to | |
30 | # %eax:%ebx as pair of 32-bit values and not as single 64-bit one. | |
31 | # This is done in order to avoid 64-bit shift penalties on Intel | |
32 | # EM64T core. Speaking of which! I bet it's possible to improve | |
33 | # Opteron performance by compressing the table to 2KB and replacing | |
34 | # unaligned references with complementary rotations [which would | |
35 | # incidentally replace lea instructions], but it would definitely | |
36 | # just "kill" EM64T, because it has only 1 shifter/rotator [against | |
37 | # 3 on Opteron] and which is *unacceptably* slow with 64-bit | |
38 | # operand. | |
39 | ||
1aa89a7a RL |
40 | # $output is the last argument if it looks like a file (it has an extension) |
41 | # $flavour is the first argument if it doesn't look like a file | |
42 | $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef; | |
43 | $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef; | |
be01f79d AP |
44 | |
45 | $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); | |
abe7f8b4 AP |
46 | |
47 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; my $dir=$1; my $xlate; | |
48 | ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or | |
49 | ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or | |
50 | die "can't locate x86_64-xlate.pl"; | |
51 | ||
1aa89a7a RL |
52 | open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"" |
53 | or die "can't call $xlate: $!"; | |
46bf83f0 | 54 | *STDOUT=*OUT; |
fec82dbc AP |
55 | |
56 | sub L() { $code.=".byte ".join(',',@_)."\n"; } | |
57 | sub LL(){ $code.=".byte ".join(',',@_).",".join(',',@_)."\n"; } | |
58 | ||
59 | @mm=("%r8","%r9","%r10","%r11","%r12","%r13","%r14","%r15"); | |
60 | ||
61 | $func="whirlpool_block"; | |
62 | $table=".Ltable"; | |
63 | ||
64 | $code=<<___; | |
65 | .text | |
66 | ||
67 | .globl $func | |
68 | .type $func,\@function,3 | |
69 | .align 16 | |
70 | $func: | |
fa62bc46 | 71 | .cfi_startproc |
384e6de4 | 72 | mov %rsp,%rax |
fa62bc46 | 73 | .cfi_def_cfa_register %rax |
fec82dbc | 74 | push %rbx |
fa62bc46 | 75 | .cfi_push %rbx |
fec82dbc | 76 | push %rbp |
fa62bc46 | 77 | .cfi_push %rbp |
fec82dbc | 78 | push %r12 |
fa62bc46 | 79 | .cfi_push %r12 |
fec82dbc | 80 | push %r13 |
fa62bc46 | 81 | .cfi_push %r13 |
fec82dbc | 82 | push %r14 |
fa62bc46 | 83 | .cfi_push %r14 |
fec82dbc | 84 | push %r15 |
fa62bc46 | 85 | .cfi_push %r15 |
fec82dbc | 86 | |
fec82dbc AP |
87 | sub \$128+40,%rsp |
88 | and \$-64,%rsp | |
89 | ||
be01f79d AP |
90 | lea 128(%rsp),%r10 |
91 | mov %rdi,0(%r10) # save parameter block | |
92 | mov %rsi,8(%r10) | |
93 | mov %rdx,16(%r10) | |
384e6de4 | 94 | mov %rax,32(%r10) # saved stack pointer |
fa62bc46 | 95 | .cfi_cfa_expression %rsp+`128+32`,deref,+8 |
be01f79d | 96 | .Lprologue: |
fec82dbc | 97 | |
be01f79d | 98 | mov %r10,%rbx |
aa8f38e4 | 99 | lea $table(%rip),%rbp |
fec82dbc AP |
100 | |
101 | xor %rcx,%rcx | |
102 | xor %rdx,%rdx | |
103 | ___ | |
104 | for($i=0;$i<8;$i++) { $code.="mov $i*8(%rdi),@mm[$i]\n"; } # L=H | |
105 | $code.=".Louterloop:\n"; | |
106 | for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; } # K=L | |
107 | for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; } # L^=inp | |
108 | for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; } # S=L | |
109 | $code.=<<___; | |
110 | xor %rsi,%rsi | |
111 | mov %rsi,24(%rbx) # zero round counter | |
701d593f | 112 | jmp .Lround |
fec82dbc AP |
113 | .align 16 |
114 | .Lround: | |
115 | mov 4096(%rbp,%rsi,8),@mm[0] # rc[r] | |
116 | mov 0(%rsp),%eax | |
117 | mov 4(%rsp),%ebx | |
701d593f AP |
118 | movz %al,%ecx |
119 | movz %ah,%edx | |
fec82dbc AP |
120 | ___ |
121 | for($i=0;$i<8;$i++) { | |
122 | my $func = ($i==0)? "mov" : "xor"; | |
123 | $code.=<<___; | |
701d593f | 124 | shr \$16,%eax |
fec82dbc | 125 | lea (%rcx,%rcx),%rsi |
701d593f | 126 | movz %al,%ecx |
fec82dbc | 127 | lea (%rdx,%rdx),%rdi |
701d593f | 128 | movz %ah,%edx |
fec82dbc AP |
129 | xor 0(%rbp,%rsi,8),@mm[0] |
130 | $func 7(%rbp,%rdi,8),@mm[1] | |
fec82dbc AP |
131 | mov $i*8+8(%rsp),%eax # ($i+1)*8 |
132 | lea (%rcx,%rcx),%rsi | |
701d593f | 133 | movz %bl,%ecx |
fec82dbc | 134 | lea (%rdx,%rdx),%rdi |
701d593f | 135 | movz %bh,%edx |
fec82dbc AP |
136 | $func 6(%rbp,%rsi,8),@mm[2] |
137 | $func 5(%rbp,%rdi,8),@mm[3] | |
701d593f | 138 | shr \$16,%ebx |
fec82dbc | 139 | lea (%rcx,%rcx),%rsi |
701d593f | 140 | movz %bl,%ecx |
fec82dbc | 141 | lea (%rdx,%rdx),%rdi |
701d593f | 142 | movz %bh,%edx |
fec82dbc AP |
143 | $func 4(%rbp,%rsi,8),@mm[4] |
144 | $func 3(%rbp,%rdi,8),@mm[5] | |
fec82dbc AP |
145 | mov $i*8+8+4(%rsp),%ebx # ($i+1)*8+4 |
146 | lea (%rcx,%rcx),%rsi | |
701d593f | 147 | movz %al,%ecx |
fec82dbc | 148 | lea (%rdx,%rdx),%rdi |
701d593f | 149 | movz %ah,%edx |
fec82dbc AP |
150 | $func 2(%rbp,%rsi,8),@mm[6] |
151 | $func 1(%rbp,%rdi,8),@mm[7] | |
152 | ___ | |
153 | push(@mm,shift(@mm)); | |
154 | } | |
155 | for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; } # K=L | |
156 | for($i=0;$i<8;$i++) { | |
157 | $code.=<<___; | |
701d593f | 158 | shr \$16,%eax |
fec82dbc | 159 | lea (%rcx,%rcx),%rsi |
701d593f | 160 | movz %al,%ecx |
fec82dbc | 161 | lea (%rdx,%rdx),%rdi |
701d593f | 162 | movz %ah,%edx |
fec82dbc AP |
163 | xor 0(%rbp,%rsi,8),@mm[0] |
164 | xor 7(%rbp,%rdi,8),@mm[1] | |
fec82dbc AP |
165 | `"mov 64+$i*8+8(%rsp),%eax" if($i<7);` # 64+($i+1)*8 |
166 | lea (%rcx,%rcx),%rsi | |
701d593f | 167 | movz %bl,%ecx |
fec82dbc | 168 | lea (%rdx,%rdx),%rdi |
701d593f | 169 | movz %bh,%edx |
fec82dbc AP |
170 | xor 6(%rbp,%rsi,8),@mm[2] |
171 | xor 5(%rbp,%rdi,8),@mm[3] | |
701d593f | 172 | shr \$16,%ebx |
fec82dbc | 173 | lea (%rcx,%rcx),%rsi |
701d593f | 174 | movz %bl,%ecx |
fec82dbc | 175 | lea (%rdx,%rdx),%rdi |
701d593f | 176 | movz %bh,%edx |
fec82dbc AP |
177 | xor 4(%rbp,%rsi,8),@mm[4] |
178 | xor 3(%rbp,%rdi,8),@mm[5] | |
fec82dbc AP |
179 | `"mov 64+$i*8+8+4(%rsp),%ebx" if($i<7);` # 64+($i+1)*8+4 |
180 | lea (%rcx,%rcx),%rsi | |
701d593f | 181 | movz %al,%ecx |
fec82dbc | 182 | lea (%rdx,%rdx),%rdi |
701d593f | 183 | movz %ah,%edx |
fec82dbc AP |
184 | xor 2(%rbp,%rsi,8),@mm[6] |
185 | xor 1(%rbp,%rdi,8),@mm[7] | |
186 | ___ | |
187 | push(@mm,shift(@mm)); | |
188 | } | |
189 | $code.=<<___; | |
190 | lea 128(%rsp),%rbx | |
191 | mov 24(%rbx),%rsi # pull round counter | |
192 | add \$1,%rsi | |
193 | cmp \$10,%rsi | |
194 | je .Lroundsdone | |
195 | ||
196 | mov %rsi,24(%rbx) # update round counter | |
197 | ___ | |
198 | for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; } # S=L | |
199 | $code.=<<___; | |
200 | jmp .Lround | |
201 | .align 16 | |
202 | .Lroundsdone: | |
203 | mov 0(%rbx),%rdi # reload argument block | |
204 | mov 8(%rbx),%rsi | |
205 | mov 16(%rbx),%rax | |
206 | ___ | |
207 | for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; } # L^=inp | |
208 | for($i=0;$i<8;$i++) { $code.="xor $i*8(%rdi),@mm[$i]\n"; } # L^=H | |
209 | for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rdi)\n"; } # H=L | |
210 | $code.=<<___; | |
211 | lea 64(%rsi),%rsi # inp+=64 | |
212 | sub \$1,%rax # num-- | |
213 | jz .Lalldone | |
214 | mov %rsi,8(%rbx) # update parameter block | |
215 | mov %rax,16(%rbx) | |
216 | jmp .Louterloop | |
217 | .Lalldone: | |
be01f79d | 218 | mov 32(%rbx),%rsi # restore saved pointer |
fa62bc46 | 219 | .cfi_def_cfa %rsi,8 |
384e6de4 | 220 | mov -48(%rsi),%r15 |
fa62bc46 | 221 | .cfi_restore %r15 |
384e6de4 | 222 | mov -40(%rsi),%r14 |
fa62bc46 | 223 | .cfi_restore %r14 |
384e6de4 | 224 | mov -32(%rsi),%r13 |
fa62bc46 | 225 | .cfi_restore %r13 |
384e6de4 | 226 | mov -24(%rsi),%r12 |
fa62bc46 | 227 | .cfi_restore %r12 |
384e6de4 | 228 | mov -16(%rsi),%rbp |
fa62bc46 | 229 | .cfi_restore %rbp |
384e6de4 | 230 | mov -8(%rsi),%rbx |
fa62bc46 | 231 | .cfi_restore %rbx |
384e6de4 | 232 | lea (%rsi),%rsp |
fa62bc46 | 233 | .cfi_def_cfa_register %rsp |
be01f79d | 234 | .Lepilogue: |
fec82dbc | 235 | ret |
fa62bc46 | 236 | .cfi_endproc |
fec82dbc AP |
237 | .size $func,.-$func |
238 | ||
239 | .align 64 | |
240 | .type $table,\@object | |
241 | $table: | |
242 | ___ | |
243 | &LL(0x18,0x18,0x60,0x18,0xc0,0x78,0x30,0xd8); | |
244 | &LL(0x23,0x23,0x8c,0x23,0x05,0xaf,0x46,0x26); | |
245 | &LL(0xc6,0xc6,0x3f,0xc6,0x7e,0xf9,0x91,0xb8); | |
246 | &LL(0xe8,0xe8,0x87,0xe8,0x13,0x6f,0xcd,0xfb); | |
247 | &LL(0x87,0x87,0x26,0x87,0x4c,0xa1,0x13,0xcb); | |
248 | &LL(0xb8,0xb8,0xda,0xb8,0xa9,0x62,0x6d,0x11); | |
249 | &LL(0x01,0x01,0x04,0x01,0x08,0x05,0x02,0x09); | |
250 | &LL(0x4f,0x4f,0x21,0x4f,0x42,0x6e,0x9e,0x0d); | |
251 | &LL(0x36,0x36,0xd8,0x36,0xad,0xee,0x6c,0x9b); | |
252 | &LL(0xa6,0xa6,0xa2,0xa6,0x59,0x04,0x51,0xff); | |
253 | &LL(0xd2,0xd2,0x6f,0xd2,0xde,0xbd,0xb9,0x0c); | |
254 | &LL(0xf5,0xf5,0xf3,0xf5,0xfb,0x06,0xf7,0x0e); | |
255 | &LL(0x79,0x79,0xf9,0x79,0xef,0x80,0xf2,0x96); | |
256 | &LL(0x6f,0x6f,0xa1,0x6f,0x5f,0xce,0xde,0x30); | |
257 | &LL(0x91,0x91,0x7e,0x91,0xfc,0xef,0x3f,0x6d); | |
258 | &LL(0x52,0x52,0x55,0x52,0xaa,0x07,0xa4,0xf8); | |
259 | &LL(0x60,0x60,0x9d,0x60,0x27,0xfd,0xc0,0x47); | |
260 | &LL(0xbc,0xbc,0xca,0xbc,0x89,0x76,0x65,0x35); | |
261 | &LL(0x9b,0x9b,0x56,0x9b,0xac,0xcd,0x2b,0x37); | |
262 | &LL(0x8e,0x8e,0x02,0x8e,0x04,0x8c,0x01,0x8a); | |
263 | &LL(0xa3,0xa3,0xb6,0xa3,0x71,0x15,0x5b,0xd2); | |
264 | &LL(0x0c,0x0c,0x30,0x0c,0x60,0x3c,0x18,0x6c); | |
265 | &LL(0x7b,0x7b,0xf1,0x7b,0xff,0x8a,0xf6,0x84); | |
266 | &LL(0x35,0x35,0xd4,0x35,0xb5,0xe1,0x6a,0x80); | |
267 | &LL(0x1d,0x1d,0x74,0x1d,0xe8,0x69,0x3a,0xf5); | |
268 | &LL(0xe0,0xe0,0xa7,0xe0,0x53,0x47,0xdd,0xb3); | |
269 | &LL(0xd7,0xd7,0x7b,0xd7,0xf6,0xac,0xb3,0x21); | |
270 | &LL(0xc2,0xc2,0x2f,0xc2,0x5e,0xed,0x99,0x9c); | |
271 | &LL(0x2e,0x2e,0xb8,0x2e,0x6d,0x96,0x5c,0x43); | |
272 | &LL(0x4b,0x4b,0x31,0x4b,0x62,0x7a,0x96,0x29); | |
273 | &LL(0xfe,0xfe,0xdf,0xfe,0xa3,0x21,0xe1,0x5d); | |
274 | &LL(0x57,0x57,0x41,0x57,0x82,0x16,0xae,0xd5); | |
275 | &LL(0x15,0x15,0x54,0x15,0xa8,0x41,0x2a,0xbd); | |
276 | &LL(0x77,0x77,0xc1,0x77,0x9f,0xb6,0xee,0xe8); | |
277 | &LL(0x37,0x37,0xdc,0x37,0xa5,0xeb,0x6e,0x92); | |
278 | &LL(0xe5,0xe5,0xb3,0xe5,0x7b,0x56,0xd7,0x9e); | |
279 | &LL(0x9f,0x9f,0x46,0x9f,0x8c,0xd9,0x23,0x13); | |
280 | &LL(0xf0,0xf0,0xe7,0xf0,0xd3,0x17,0xfd,0x23); | |
281 | &LL(0x4a,0x4a,0x35,0x4a,0x6a,0x7f,0x94,0x20); | |
282 | &LL(0xda,0xda,0x4f,0xda,0x9e,0x95,0xa9,0x44); | |
283 | &LL(0x58,0x58,0x7d,0x58,0xfa,0x25,0xb0,0xa2); | |
284 | &LL(0xc9,0xc9,0x03,0xc9,0x06,0xca,0x8f,0xcf); | |
285 | &LL(0x29,0x29,0xa4,0x29,0x55,0x8d,0x52,0x7c); | |
286 | &LL(0x0a,0x0a,0x28,0x0a,0x50,0x22,0x14,0x5a); | |
287 | &LL(0xb1,0xb1,0xfe,0xb1,0xe1,0x4f,0x7f,0x50); | |
288 | &LL(0xa0,0xa0,0xba,0xa0,0x69,0x1a,0x5d,0xc9); | |
289 | &LL(0x6b,0x6b,0xb1,0x6b,0x7f,0xda,0xd6,0x14); | |
290 | &LL(0x85,0x85,0x2e,0x85,0x5c,0xab,0x17,0xd9); | |
291 | &LL(0xbd,0xbd,0xce,0xbd,0x81,0x73,0x67,0x3c); | |
292 | &LL(0x5d,0x5d,0x69,0x5d,0xd2,0x34,0xba,0x8f); | |
293 | &LL(0x10,0x10,0x40,0x10,0x80,0x50,0x20,0x90); | |
294 | &LL(0xf4,0xf4,0xf7,0xf4,0xf3,0x03,0xf5,0x07); | |
295 | &LL(0xcb,0xcb,0x0b,0xcb,0x16,0xc0,0x8b,0xdd); | |
296 | &LL(0x3e,0x3e,0xf8,0x3e,0xed,0xc6,0x7c,0xd3); | |
297 | &LL(0x05,0x05,0x14,0x05,0x28,0x11,0x0a,0x2d); | |
298 | &LL(0x67,0x67,0x81,0x67,0x1f,0xe6,0xce,0x78); | |
299 | &LL(0xe4,0xe4,0xb7,0xe4,0x73,0x53,0xd5,0x97); | |
300 | &LL(0x27,0x27,0x9c,0x27,0x25,0xbb,0x4e,0x02); | |
301 | &LL(0x41,0x41,0x19,0x41,0x32,0x58,0x82,0x73); | |
302 | &LL(0x8b,0x8b,0x16,0x8b,0x2c,0x9d,0x0b,0xa7); | |
303 | &LL(0xa7,0xa7,0xa6,0xa7,0x51,0x01,0x53,0xf6); | |
304 | &LL(0x7d,0x7d,0xe9,0x7d,0xcf,0x94,0xfa,0xb2); | |
305 | &LL(0x95,0x95,0x6e,0x95,0xdc,0xfb,0x37,0x49); | |
306 | &LL(0xd8,0xd8,0x47,0xd8,0x8e,0x9f,0xad,0x56); | |
307 | &LL(0xfb,0xfb,0xcb,0xfb,0x8b,0x30,0xeb,0x70); | |
308 | &LL(0xee,0xee,0x9f,0xee,0x23,0x71,0xc1,0xcd); | |
309 | &LL(0x7c,0x7c,0xed,0x7c,0xc7,0x91,0xf8,0xbb); | |
310 | &LL(0x66,0x66,0x85,0x66,0x17,0xe3,0xcc,0x71); | |
311 | &LL(0xdd,0xdd,0x53,0xdd,0xa6,0x8e,0xa7,0x7b); | |
312 | &LL(0x17,0x17,0x5c,0x17,0xb8,0x4b,0x2e,0xaf); | |
313 | &LL(0x47,0x47,0x01,0x47,0x02,0x46,0x8e,0x45); | |
314 | &LL(0x9e,0x9e,0x42,0x9e,0x84,0xdc,0x21,0x1a); | |
315 | &LL(0xca,0xca,0x0f,0xca,0x1e,0xc5,0x89,0xd4); | |
316 | &LL(0x2d,0x2d,0xb4,0x2d,0x75,0x99,0x5a,0x58); | |
317 | &LL(0xbf,0xbf,0xc6,0xbf,0x91,0x79,0x63,0x2e); | |
318 | &LL(0x07,0x07,0x1c,0x07,0x38,0x1b,0x0e,0x3f); | |
319 | &LL(0xad,0xad,0x8e,0xad,0x01,0x23,0x47,0xac); | |
320 | &LL(0x5a,0x5a,0x75,0x5a,0xea,0x2f,0xb4,0xb0); | |
321 | &LL(0x83,0x83,0x36,0x83,0x6c,0xb5,0x1b,0xef); | |
322 | &LL(0x33,0x33,0xcc,0x33,0x85,0xff,0x66,0xb6); | |
323 | &LL(0x63,0x63,0x91,0x63,0x3f,0xf2,0xc6,0x5c); | |
324 | &LL(0x02,0x02,0x08,0x02,0x10,0x0a,0x04,0x12); | |
325 | &LL(0xaa,0xaa,0x92,0xaa,0x39,0x38,0x49,0x93); | |
326 | &LL(0x71,0x71,0xd9,0x71,0xaf,0xa8,0xe2,0xde); | |
327 | &LL(0xc8,0xc8,0x07,0xc8,0x0e,0xcf,0x8d,0xc6); | |
328 | &LL(0x19,0x19,0x64,0x19,0xc8,0x7d,0x32,0xd1); | |
329 | &LL(0x49,0x49,0x39,0x49,0x72,0x70,0x92,0x3b); | |
330 | &LL(0xd9,0xd9,0x43,0xd9,0x86,0x9a,0xaf,0x5f); | |
331 | &LL(0xf2,0xf2,0xef,0xf2,0xc3,0x1d,0xf9,0x31); | |
332 | &LL(0xe3,0xe3,0xab,0xe3,0x4b,0x48,0xdb,0xa8); | |
333 | &LL(0x5b,0x5b,0x71,0x5b,0xe2,0x2a,0xb6,0xb9); | |
334 | &LL(0x88,0x88,0x1a,0x88,0x34,0x92,0x0d,0xbc); | |
335 | &LL(0x9a,0x9a,0x52,0x9a,0xa4,0xc8,0x29,0x3e); | |
336 | &LL(0x26,0x26,0x98,0x26,0x2d,0xbe,0x4c,0x0b); | |
337 | &LL(0x32,0x32,0xc8,0x32,0x8d,0xfa,0x64,0xbf); | |
338 | &LL(0xb0,0xb0,0xfa,0xb0,0xe9,0x4a,0x7d,0x59); | |
339 | &LL(0xe9,0xe9,0x83,0xe9,0x1b,0x6a,0xcf,0xf2); | |
340 | &LL(0x0f,0x0f,0x3c,0x0f,0x78,0x33,0x1e,0x77); | |
341 | &LL(0xd5,0xd5,0x73,0xd5,0xe6,0xa6,0xb7,0x33); | |
342 | &LL(0x80,0x80,0x3a,0x80,0x74,0xba,0x1d,0xf4); | |
343 | &LL(0xbe,0xbe,0xc2,0xbe,0x99,0x7c,0x61,0x27); | |
344 | &LL(0xcd,0xcd,0x13,0xcd,0x26,0xde,0x87,0xeb); | |
345 | &LL(0x34,0x34,0xd0,0x34,0xbd,0xe4,0x68,0x89); | |
346 | &LL(0x48,0x48,0x3d,0x48,0x7a,0x75,0x90,0x32); | |
347 | &LL(0xff,0xff,0xdb,0xff,0xab,0x24,0xe3,0x54); | |
348 | &LL(0x7a,0x7a,0xf5,0x7a,0xf7,0x8f,0xf4,0x8d); | |
349 | &LL(0x90,0x90,0x7a,0x90,0xf4,0xea,0x3d,0x64); | |
350 | &LL(0x5f,0x5f,0x61,0x5f,0xc2,0x3e,0xbe,0x9d); | |
351 | &LL(0x20,0x20,0x80,0x20,0x1d,0xa0,0x40,0x3d); | |
352 | &LL(0x68,0x68,0xbd,0x68,0x67,0xd5,0xd0,0x0f); | |
353 | &LL(0x1a,0x1a,0x68,0x1a,0xd0,0x72,0x34,0xca); | |
354 | &LL(0xae,0xae,0x82,0xae,0x19,0x2c,0x41,0xb7); | |
355 | &LL(0xb4,0xb4,0xea,0xb4,0xc9,0x5e,0x75,0x7d); | |
356 | &LL(0x54,0x54,0x4d,0x54,0x9a,0x19,0xa8,0xce); | |
357 | &LL(0x93,0x93,0x76,0x93,0xec,0xe5,0x3b,0x7f); | |
358 | &LL(0x22,0x22,0x88,0x22,0x0d,0xaa,0x44,0x2f); | |
359 | &LL(0x64,0x64,0x8d,0x64,0x07,0xe9,0xc8,0x63); | |
360 | &LL(0xf1,0xf1,0xe3,0xf1,0xdb,0x12,0xff,0x2a); | |
361 | &LL(0x73,0x73,0xd1,0x73,0xbf,0xa2,0xe6,0xcc); | |
362 | &LL(0x12,0x12,0x48,0x12,0x90,0x5a,0x24,0x82); | |
363 | &LL(0x40,0x40,0x1d,0x40,0x3a,0x5d,0x80,0x7a); | |
364 | &LL(0x08,0x08,0x20,0x08,0x40,0x28,0x10,0x48); | |
365 | &LL(0xc3,0xc3,0x2b,0xc3,0x56,0xe8,0x9b,0x95); | |
366 | &LL(0xec,0xec,0x97,0xec,0x33,0x7b,0xc5,0xdf); | |
367 | &LL(0xdb,0xdb,0x4b,0xdb,0x96,0x90,0xab,0x4d); | |
368 | &LL(0xa1,0xa1,0xbe,0xa1,0x61,0x1f,0x5f,0xc0); | |
369 | &LL(0x8d,0x8d,0x0e,0x8d,0x1c,0x83,0x07,0x91); | |
370 | &LL(0x3d,0x3d,0xf4,0x3d,0xf5,0xc9,0x7a,0xc8); | |
371 | &LL(0x97,0x97,0x66,0x97,0xcc,0xf1,0x33,0x5b); | |
372 | &LL(0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00); | |
373 | &LL(0xcf,0xcf,0x1b,0xcf,0x36,0xd4,0x83,0xf9); | |
374 | &LL(0x2b,0x2b,0xac,0x2b,0x45,0x87,0x56,0x6e); | |
375 | &LL(0x76,0x76,0xc5,0x76,0x97,0xb3,0xec,0xe1); | |
376 | &LL(0x82,0x82,0x32,0x82,0x64,0xb0,0x19,0xe6); | |
377 | &LL(0xd6,0xd6,0x7f,0xd6,0xfe,0xa9,0xb1,0x28); | |
378 | &LL(0x1b,0x1b,0x6c,0x1b,0xd8,0x77,0x36,0xc3); | |
379 | &LL(0xb5,0xb5,0xee,0xb5,0xc1,0x5b,0x77,0x74); | |
380 | &LL(0xaf,0xaf,0x86,0xaf,0x11,0x29,0x43,0xbe); | |
381 | &LL(0x6a,0x6a,0xb5,0x6a,0x77,0xdf,0xd4,0x1d); | |
382 | &LL(0x50,0x50,0x5d,0x50,0xba,0x0d,0xa0,0xea); | |
383 | &LL(0x45,0x45,0x09,0x45,0x12,0x4c,0x8a,0x57); | |
384 | &LL(0xf3,0xf3,0xeb,0xf3,0xcb,0x18,0xfb,0x38); | |
385 | &LL(0x30,0x30,0xc0,0x30,0x9d,0xf0,0x60,0xad); | |
386 | &LL(0xef,0xef,0x9b,0xef,0x2b,0x74,0xc3,0xc4); | |
387 | &LL(0x3f,0x3f,0xfc,0x3f,0xe5,0xc3,0x7e,0xda); | |
388 | &LL(0x55,0x55,0x49,0x55,0x92,0x1c,0xaa,0xc7); | |
389 | &LL(0xa2,0xa2,0xb2,0xa2,0x79,0x10,0x59,0xdb); | |
390 | &LL(0xea,0xea,0x8f,0xea,0x03,0x65,0xc9,0xe9); | |
391 | &LL(0x65,0x65,0x89,0x65,0x0f,0xec,0xca,0x6a); | |
392 | &LL(0xba,0xba,0xd2,0xba,0xb9,0x68,0x69,0x03); | |
393 | &LL(0x2f,0x2f,0xbc,0x2f,0x65,0x93,0x5e,0x4a); | |
394 | &LL(0xc0,0xc0,0x27,0xc0,0x4e,0xe7,0x9d,0x8e); | |
395 | &LL(0xde,0xde,0x5f,0xde,0xbe,0x81,0xa1,0x60); | |
396 | &LL(0x1c,0x1c,0x70,0x1c,0xe0,0x6c,0x38,0xfc); | |
397 | &LL(0xfd,0xfd,0xd3,0xfd,0xbb,0x2e,0xe7,0x46); | |
398 | &LL(0x4d,0x4d,0x29,0x4d,0x52,0x64,0x9a,0x1f); | |
399 | &LL(0x92,0x92,0x72,0x92,0xe4,0xe0,0x39,0x76); | |
400 | &LL(0x75,0x75,0xc9,0x75,0x8f,0xbc,0xea,0xfa); | |
401 | &LL(0x06,0x06,0x18,0x06,0x30,0x1e,0x0c,0x36); | |
402 | &LL(0x8a,0x8a,0x12,0x8a,0x24,0x98,0x09,0xae); | |
403 | &LL(0xb2,0xb2,0xf2,0xb2,0xf9,0x40,0x79,0x4b); | |
404 | &LL(0xe6,0xe6,0xbf,0xe6,0x63,0x59,0xd1,0x85); | |
405 | &LL(0x0e,0x0e,0x38,0x0e,0x70,0x36,0x1c,0x7e); | |
406 | &LL(0x1f,0x1f,0x7c,0x1f,0xf8,0x63,0x3e,0xe7); | |
407 | &LL(0x62,0x62,0x95,0x62,0x37,0xf7,0xc4,0x55); | |
408 | &LL(0xd4,0xd4,0x77,0xd4,0xee,0xa3,0xb5,0x3a); | |
409 | &LL(0xa8,0xa8,0x9a,0xa8,0x29,0x32,0x4d,0x81); | |
410 | &LL(0x96,0x96,0x62,0x96,0xc4,0xf4,0x31,0x52); | |
411 | &LL(0xf9,0xf9,0xc3,0xf9,0x9b,0x3a,0xef,0x62); | |
412 | &LL(0xc5,0xc5,0x33,0xc5,0x66,0xf6,0x97,0xa3); | |
413 | &LL(0x25,0x25,0x94,0x25,0x35,0xb1,0x4a,0x10); | |
414 | &LL(0x59,0x59,0x79,0x59,0xf2,0x20,0xb2,0xab); | |
415 | &LL(0x84,0x84,0x2a,0x84,0x54,0xae,0x15,0xd0); | |
416 | &LL(0x72,0x72,0xd5,0x72,0xb7,0xa7,0xe4,0xc5); | |
417 | &LL(0x39,0x39,0xe4,0x39,0xd5,0xdd,0x72,0xec); | |
418 | &LL(0x4c,0x4c,0x2d,0x4c,0x5a,0x61,0x98,0x16); | |
419 | &LL(0x5e,0x5e,0x65,0x5e,0xca,0x3b,0xbc,0x94); | |
420 | &LL(0x78,0x78,0xfd,0x78,0xe7,0x85,0xf0,0x9f); | |
421 | &LL(0x38,0x38,0xe0,0x38,0xdd,0xd8,0x70,0xe5); | |
422 | &LL(0x8c,0x8c,0x0a,0x8c,0x14,0x86,0x05,0x98); | |
423 | &LL(0xd1,0xd1,0x63,0xd1,0xc6,0xb2,0xbf,0x17); | |
424 | &LL(0xa5,0xa5,0xae,0xa5,0x41,0x0b,0x57,0xe4); | |
425 | &LL(0xe2,0xe2,0xaf,0xe2,0x43,0x4d,0xd9,0xa1); | |
426 | &LL(0x61,0x61,0x99,0x61,0x2f,0xf8,0xc2,0x4e); | |
427 | &LL(0xb3,0xb3,0xf6,0xb3,0xf1,0x45,0x7b,0x42); | |
428 | &LL(0x21,0x21,0x84,0x21,0x15,0xa5,0x42,0x34); | |
429 | &LL(0x9c,0x9c,0x4a,0x9c,0x94,0xd6,0x25,0x08); | |
430 | &LL(0x1e,0x1e,0x78,0x1e,0xf0,0x66,0x3c,0xee); | |
431 | &LL(0x43,0x43,0x11,0x43,0x22,0x52,0x86,0x61); | |
432 | &LL(0xc7,0xc7,0x3b,0xc7,0x76,0xfc,0x93,0xb1); | |
433 | &LL(0xfc,0xfc,0xd7,0xfc,0xb3,0x2b,0xe5,0x4f); | |
434 | &LL(0x04,0x04,0x10,0x04,0x20,0x14,0x08,0x24); | |
435 | &LL(0x51,0x51,0x59,0x51,0xb2,0x08,0xa2,0xe3); | |
436 | &LL(0x99,0x99,0x5e,0x99,0xbc,0xc7,0x2f,0x25); | |
437 | &LL(0x6d,0x6d,0xa9,0x6d,0x4f,0xc4,0xda,0x22); | |
438 | &LL(0x0d,0x0d,0x34,0x0d,0x68,0x39,0x1a,0x65); | |
439 | &LL(0xfa,0xfa,0xcf,0xfa,0x83,0x35,0xe9,0x79); | |
440 | &LL(0xdf,0xdf,0x5b,0xdf,0xb6,0x84,0xa3,0x69); | |
441 | &LL(0x7e,0x7e,0xe5,0x7e,0xd7,0x9b,0xfc,0xa9); | |
442 | &LL(0x24,0x24,0x90,0x24,0x3d,0xb4,0x48,0x19); | |
443 | &LL(0x3b,0x3b,0xec,0x3b,0xc5,0xd7,0x76,0xfe); | |
444 | &LL(0xab,0xab,0x96,0xab,0x31,0x3d,0x4b,0x9a); | |
445 | &LL(0xce,0xce,0x1f,0xce,0x3e,0xd1,0x81,0xf0); | |
446 | &LL(0x11,0x11,0x44,0x11,0x88,0x55,0x22,0x99); | |
447 | &LL(0x8f,0x8f,0x06,0x8f,0x0c,0x89,0x03,0x83); | |
448 | &LL(0x4e,0x4e,0x25,0x4e,0x4a,0x6b,0x9c,0x04); | |
449 | &LL(0xb7,0xb7,0xe6,0xb7,0xd1,0x51,0x73,0x66); | |
450 | &LL(0xeb,0xeb,0x8b,0xeb,0x0b,0x60,0xcb,0xe0); | |
451 | &LL(0x3c,0x3c,0xf0,0x3c,0xfd,0xcc,0x78,0xc1); | |
452 | &LL(0x81,0x81,0x3e,0x81,0x7c,0xbf,0x1f,0xfd); | |
453 | &LL(0x94,0x94,0x6a,0x94,0xd4,0xfe,0x35,0x40); | |
454 | &LL(0xf7,0xf7,0xfb,0xf7,0xeb,0x0c,0xf3,0x1c); | |
455 | &LL(0xb9,0xb9,0xde,0xb9,0xa1,0x67,0x6f,0x18); | |
456 | &LL(0x13,0x13,0x4c,0x13,0x98,0x5f,0x26,0x8b); | |
457 | &LL(0x2c,0x2c,0xb0,0x2c,0x7d,0x9c,0x58,0x51); | |
458 | &LL(0xd3,0xd3,0x6b,0xd3,0xd6,0xb8,0xbb,0x05); | |
459 | &LL(0xe7,0xe7,0xbb,0xe7,0x6b,0x5c,0xd3,0x8c); | |
460 | &LL(0x6e,0x6e,0xa5,0x6e,0x57,0xcb,0xdc,0x39); | |
461 | &LL(0xc4,0xc4,0x37,0xc4,0x6e,0xf3,0x95,0xaa); | |
462 | &LL(0x03,0x03,0x0c,0x03,0x18,0x0f,0x06,0x1b); | |
463 | &LL(0x56,0x56,0x45,0x56,0x8a,0x13,0xac,0xdc); | |
464 | &LL(0x44,0x44,0x0d,0x44,0x1a,0x49,0x88,0x5e); | |
465 | &LL(0x7f,0x7f,0xe1,0x7f,0xdf,0x9e,0xfe,0xa0); | |
466 | &LL(0xa9,0xa9,0x9e,0xa9,0x21,0x37,0x4f,0x88); | |
467 | &LL(0x2a,0x2a,0xa8,0x2a,0x4d,0x82,0x54,0x67); | |
468 | &LL(0xbb,0xbb,0xd6,0xbb,0xb1,0x6d,0x6b,0x0a); | |
469 | &LL(0xc1,0xc1,0x23,0xc1,0x46,0xe2,0x9f,0x87); | |
470 | &LL(0x53,0x53,0x51,0x53,0xa2,0x02,0xa6,0xf1); | |
471 | &LL(0xdc,0xdc,0x57,0xdc,0xae,0x8b,0xa5,0x72); | |
472 | &LL(0x0b,0x0b,0x2c,0x0b,0x58,0x27,0x16,0x53); | |
473 | &LL(0x9d,0x9d,0x4e,0x9d,0x9c,0xd3,0x27,0x01); | |
474 | &LL(0x6c,0x6c,0xad,0x6c,0x47,0xc1,0xd8,0x2b); | |
475 | &LL(0x31,0x31,0xc4,0x31,0x95,0xf5,0x62,0xa4); | |
476 | &LL(0x74,0x74,0xcd,0x74,0x87,0xb9,0xe8,0xf3); | |
477 | &LL(0xf6,0xf6,0xff,0xf6,0xe3,0x09,0xf1,0x15); | |
478 | &LL(0x46,0x46,0x05,0x46,0x0a,0x43,0x8c,0x4c); | |
479 | &LL(0xac,0xac,0x8a,0xac,0x09,0x26,0x45,0xa5); | |
480 | &LL(0x89,0x89,0x1e,0x89,0x3c,0x97,0x0f,0xb5); | |
481 | &LL(0x14,0x14,0x50,0x14,0xa0,0x44,0x28,0xb4); | |
482 | &LL(0xe1,0xe1,0xa3,0xe1,0x5b,0x42,0xdf,0xba); | |
483 | &LL(0x16,0x16,0x58,0x16,0xb0,0x4e,0x2c,0xa6); | |
484 | &LL(0x3a,0x3a,0xe8,0x3a,0xcd,0xd2,0x74,0xf7); | |
485 | &LL(0x69,0x69,0xb9,0x69,0x6f,0xd0,0xd2,0x06); | |
486 | &LL(0x09,0x09,0x24,0x09,0x48,0x2d,0x12,0x41); | |
487 | &LL(0x70,0x70,0xdd,0x70,0xa7,0xad,0xe0,0xd7); | |
488 | &LL(0xb6,0xb6,0xe2,0xb6,0xd9,0x54,0x71,0x6f); | |
489 | &LL(0xd0,0xd0,0x67,0xd0,0xce,0xb7,0xbd,0x1e); | |
490 | &LL(0xed,0xed,0x93,0xed,0x3b,0x7e,0xc7,0xd6); | |
491 | &LL(0xcc,0xcc,0x17,0xcc,0x2e,0xdb,0x85,0xe2); | |
492 | &LL(0x42,0x42,0x15,0x42,0x2a,0x57,0x84,0x68); | |
493 | &LL(0x98,0x98,0x5a,0x98,0xb4,0xc2,0x2d,0x2c); | |
494 | &LL(0xa4,0xa4,0xaa,0xa4,0x49,0x0e,0x55,0xed); | |
495 | &LL(0x28,0x28,0xa0,0x28,0x5d,0x88,0x50,0x75); | |
496 | &LL(0x5c,0x5c,0x6d,0x5c,0xda,0x31,0xb8,0x86); | |
497 | &LL(0xf8,0xf8,0xc7,0xf8,0x93,0x3f,0xed,0x6b); | |
498 | &LL(0x86,0x86,0x22,0x86,0x44,0xa4,0x11,0xc2); | |
499 | ||
500 | &L(0x18,0x23,0xc6,0xe8,0x87,0xb8,0x01,0x4f); # rc[ROUNDS] | |
501 | &L(0x36,0xa6,0xd2,0xf5,0x79,0x6f,0x91,0x52); | |
502 | &L(0x60,0xbc,0x9b,0x8e,0xa3,0x0c,0x7b,0x35); | |
503 | &L(0x1d,0xe0,0xd7,0xc2,0x2e,0x4b,0xfe,0x57); | |
504 | &L(0x15,0x77,0x37,0xe5,0x9f,0xf0,0x4a,0xda); | |
505 | &L(0x58,0xc9,0x29,0x0a,0xb1,0xa0,0x6b,0x85); | |
506 | &L(0xbd,0x5d,0x10,0xf4,0xcb,0x3e,0x05,0x67); | |
507 | &L(0xe4,0x27,0x41,0x8b,0xa7,0x7d,0x95,0xd8); | |
508 | &L(0xfb,0xee,0x7c,0x66,0xdd,0x17,0x47,0x9e); | |
509 | &L(0xca,0x2d,0xbf,0x07,0xad,0x5a,0x83,0x33); | |
510 | ||
be01f79d AP |
511 | # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, |
512 | # CONTEXT *context,DISPATCHER_CONTEXT *disp) | |
513 | if ($win64) { | |
514 | $rec="%rcx"; | |
515 | $frame="%rdx"; | |
516 | $context="%r8"; | |
517 | $disp="%r9"; | |
518 | ||
519 | $code.=<<___; | |
520 | .extern __imp_RtlVirtualUnwind | |
521 | .type se_handler,\@abi-omnipotent | |
522 | .align 16 | |
523 | se_handler: | |
524 | push %rsi | |
525 | push %rdi | |
526 | push %rbx | |
527 | push %rbp | |
528 | push %r12 | |
529 | push %r13 | |
530 | push %r14 | |
531 | push %r15 | |
532 | pushfq | |
533 | sub \$64,%rsp | |
534 | ||
535 | mov 120($context),%rax # pull context->Rax | |
536 | mov 248($context),%rbx # pull context->Rip | |
537 | ||
538 | lea .Lprologue(%rip),%r10 | |
539 | cmp %r10,%rbx # context->Rip<.Lprologue | |
540 | jb .Lin_prologue | |
541 | ||
542 | mov 152($context),%rax # pull context->Rsp | |
543 | ||
544 | lea .Lepilogue(%rip),%r10 | |
545 | cmp %r10,%rbx # context->Rip>=.Lepilogue | |
546 | jae .Lin_prologue | |
547 | ||
548 | mov 128+32(%rax),%rax # pull saved stack pointer | |
be01f79d AP |
549 | |
550 | mov -8(%rax),%rbx | |
551 | mov -16(%rax),%rbp | |
552 | mov -24(%rax),%r12 | |
553 | mov -32(%rax),%r13 | |
554 | mov -40(%rax),%r14 | |
555 | mov -48(%rax),%r15 | |
556 | mov %rbx,144($context) # restore context->Rbx | |
557 | mov %rbp,160($context) # restore context->Rbp | |
558 | mov %r12,216($context) # restore context->R12 | |
559 | mov %r13,224($context) # restore context->R13 | |
560 | mov %r14,232($context) # restore context->R14 | |
561 | mov %r15,240($context) # restore context->R15 | |
562 | ||
563 | .Lin_prologue: | |
564 | mov 8(%rax),%rdi | |
565 | mov 16(%rax),%rsi | |
566 | mov %rax,152($context) # restore context->Rsp | |
567 | mov %rsi,168($context) # restore context->Rsi | |
568 | mov %rdi,176($context) # restore context->Rdi | |
569 | ||
570 | mov 40($disp),%rdi # disp->ContextRecord | |
571 | mov $context,%rsi # context | |
572 | mov \$154,%ecx # sizeof(CONTEXT) | |
573 | .long 0xa548f3fc # cld; rep movsq | |
574 | ||
575 | mov $disp,%rsi | |
576 | xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER | |
577 | mov 8(%rsi),%rdx # arg2, disp->ImageBase | |
578 | mov 0(%rsi),%r8 # arg3, disp->ControlPc | |
579 | mov 16(%rsi),%r9 # arg4, disp->FunctionEntry | |
580 | mov 40(%rsi),%r10 # disp->ContextRecord | |
581 | lea 56(%rsi),%r11 # &disp->HandlerData | |
582 | lea 24(%rsi),%r12 # &disp->EstablisherFrame | |
583 | mov %r10,32(%rsp) # arg5 | |
584 | mov %r11,40(%rsp) # arg6 | |
585 | mov %r12,48(%rsp) # arg7 | |
586 | mov %rcx,56(%rsp) # arg8, (NULL) | |
587 | call *__imp_RtlVirtualUnwind(%rip) | |
588 | ||
589 | mov \$1,%eax # ExceptionContinueSearch | |
590 | add \$64,%rsp | |
591 | popfq | |
592 | pop %r15 | |
593 | pop %r14 | |
594 | pop %r13 | |
595 | pop %r12 | |
596 | pop %rbp | |
597 | pop %rbx | |
598 | pop %rdi | |
599 | pop %rsi | |
600 | ret | |
601 | .size se_handler,.-se_handler | |
602 | ||
603 | .section .pdata | |
604 | .align 4 | |
605 | .rva .LSEH_begin_$func | |
606 | .rva .LSEH_end_$func | |
607 | .rva .LSEH_info_$func | |
608 | ||
609 | .section .xdata | |
610 | .align 8 | |
611 | .LSEH_info_$func: | |
612 | .byte 9,0,0,0 | |
613 | .rva se_handler | |
614 | ___ | |
615 | } | |
616 | ||
fec82dbc AP |
617 | $code =~ s/\`([^\`]*)\`/eval $1/gem; |
618 | print $code; | |
a21314db | 619 | close STDOUT or die "error closing STDOUT: $!"; |