]>
Commit | Line | Data |
---|---|---|
e0a65194 | 1 | #! /usr/bin/env perl |
33388b44 | 2 | # Copyright 2005-2020 The OpenSSL Project Authors. All Rights Reserved. |
e0a65194 | 3 | # |
0e9725bc | 4 | # Licensed under the Apache License 2.0 (the "License"). You may not use |
e0a65194 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 | ||
14e21f86 | 9 | |
1aa89a7a RL |
10 | # $output is the last argument if it looks like a file (it has an extension) |
11 | # $flavour is the first argument if it doesn't look like a file | |
12 | $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef; | |
13 | $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef; | |
14e21f86 | 14 | |
aa8f38e4 | 15 | $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); |
376729e1 | 16 | |
aa8f38e4 | 17 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; |
301799b8 AP |
18 | ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or |
19 | ( $xlate="${dir}perlasm/x86_64-xlate.pl" and -f $xlate) or | |
20 | die "can't locate x86_64-xlate.pl"; | |
21 | ||
1aa89a7a RL |
22 | open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"" |
23 | or die "can't call $xlate: $!"; | |
46bf83f0 | 24 | *STDOUT=*OUT; |
376729e1 | 25 | |
5fabb88a AP |
26 | ($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order |
27 | ("%rdi","%rsi","%rdx","%rcx"); # Unix order | |
28 | ||
aa8f38e4 AP |
29 | print<<___; |
30 | .extern OPENSSL_cpuid_setup | |
ddc20d4d | 31 | .hidden OPENSSL_cpuid_setup |
aa8f38e4 AP |
32 | .section .init |
33 | call OPENSSL_cpuid_setup | |
932cc129 | 34 | |
ddc20d4d | 35 | .hidden OPENSSL_ia32cap_P |
c5cd28bd | 36 | .comm OPENSSL_ia32cap_P,16,4 |
ddc20d4d | 37 | |
14e21f86 | 38 | .text |
376729e1 AP |
39 | |
40 | .globl OPENSSL_atomic_add | |
aa8f38e4 | 41 | .type OPENSSL_atomic_add,\@abi-omnipotent |
376729e1 AP |
42 | .align 16 |
43 | OPENSSL_atomic_add: | |
8913378a | 44 | .cfi_startproc |
98ad3fe8 | 45 | endbranch |
aa8f38e4 AP |
46 | movl ($arg1),%eax |
47 | .Lspin: leaq ($arg2,%rax),%r8 | |
48 | .byte 0xf0 # lock | |
49 | cmpxchgl %r8d,($arg1) | |
376729e1 | 50 | jne .Lspin |
e442c362 | 51 | movl %r8d,%eax |
aa8f38e4 | 52 | .byte 0x48,0x98 # cltq/cdqe |
376729e1 | 53 | ret |
8913378a | 54 | .cfi_endproc |
376729e1 AP |
55 | .size OPENSSL_atomic_add,.-OPENSSL_atomic_add |
56 | ||
932cc129 AP |
57 | .globl OPENSSL_rdtsc |
58 | .type OPENSSL_rdtsc,\@abi-omnipotent | |
59 | .align 16 | |
60 | OPENSSL_rdtsc: | |
8913378a | 61 | .cfi_startproc |
98ad3fe8 | 62 | endbranch |
932cc129 AP |
63 | rdtsc |
64 | shl \$32,%rdx | |
65 | or %rdx,%rax | |
66 | ret | |
8913378a | 67 | .cfi_endproc |
932cc129 AP |
68 | .size OPENSSL_rdtsc,.-OPENSSL_rdtsc |
69 | ||
376729e1 | 70 | .globl OPENSSL_ia32_cpuid |
c5cd28bd | 71 | .type OPENSSL_ia32_cpuid,\@function,1 |
376729e1 AP |
72 | .align 16 |
73 | OPENSSL_ia32_cpuid: | |
5e32cfb2 | 74 | .cfi_startproc |
98ad3fe8 | 75 | endbranch |
b9064221 | 76 | mov %rbx,%r8 # save %rbx |
5e32cfb2 | 77 | .cfi_register %rbx,%r8 |
9babf392 AP |
78 | |
79 | xor %eax,%eax | |
d6ee8f3d | 80 | mov %rax,8(%rdi) # clear extended feature flags |
9babf392 | 81 | cpuid |
761393bb AP |
82 | mov %eax,%r11d # max value for standard query level |
83 | ||
9babf392 AP |
84 | xor %eax,%eax |
85 | cmp \$0x756e6547,%ebx # "Genu" | |
86 | setne %al | |
87 | mov %eax,%r9d | |
88 | cmp \$0x49656e69,%edx # "ineI" | |
89 | setne %al | |
90 | or %eax,%r9d | |
91 | cmp \$0x6c65746e,%ecx # "ntel" | |
92 | setne %al | |
5cd91b50 | 93 | or %eax,%r9d # 0 indicates Intel CPU |
5cd91b50 AP |
94 | jz .Lintel |
95 | ||
96 | cmp \$0x68747541,%ebx # "Auth" | |
97 | setne %al | |
98 | mov %eax,%r10d | |
99 | cmp \$0x69746E65,%edx # "enti" | |
100 | setne %al | |
101 | or %eax,%r10d | |
102 | cmp \$0x444D4163,%ecx # "cAMD" | |
103 | setne %al | |
104 | or %eax,%r10d # 0 indicates AMD CPU | |
105 | jnz .Lintel | |
106 | ||
761393bb | 107 | # AMD specific |
5cd91b50 AP |
108 | mov \$0x80000000,%eax |
109 | cpuid | |
b9064221 AP |
110 | cmp \$0x80000001,%eax |
111 | jb .Lintel | |
112 | mov %eax,%r10d | |
113 | mov \$0x80000001,%eax | |
114 | cpuid | |
115 | or %ecx,%r9d | |
116 | and \$0x00000801,%r9d # isolate AMD XOP bit, 1<<11 | |
117 | ||
118 | cmp \$0x80000008,%r10d | |
5cd91b50 AP |
119 | jb .Lintel |
120 | ||
121 | mov \$0x80000008,%eax | |
122 | cpuid | |
123 | movzb %cl,%r10 # number of cores - 1 | |
124 | inc %r10 # number of cores | |
9babf392 | 125 | |
761393bb AP |
126 | mov \$1,%eax |
127 | cpuid | |
128 | bt \$28,%edx # test hyper-threading bit | |
b9064221 | 129 | jnc .Lgeneric |
761393bb AP |
130 | shr \$16,%ebx # number of logical processors |
131 | cmp %r10b,%bl | |
b9064221 | 132 | ja .Lgeneric |
761393bb | 133 | and \$0xefffffff,%edx # ~(1<<28) |
b9064221 | 134 | jmp .Lgeneric |
761393bb | 135 | |
5cd91b50 | 136 | .Lintel: |
761393bb AP |
137 | cmp \$4,%r11d |
138 | mov \$-1,%r10d | |
139 | jb .Lnocacheinfo | |
140 | ||
141 | mov \$4,%eax | |
142 | mov \$0,%ecx # query L1D | |
143 | cpuid | |
144 | mov %eax,%r10d | |
145 | shr \$14,%r10d | |
146 | and \$0xfff,%r10d # number of cores -1 per L1D | |
147 | ||
148 | .Lnocacheinfo: | |
932cc129 | 149 | mov \$1,%eax |
376729e1 | 150 | cpuid |
79337628 | 151 | movd %eax,%xmm0 # put aside processor id |
4bb90087 | 152 | and \$0xbfefffff,%edx # force reserved bits to 0 |
932cc129 | 153 | cmp \$0,%r9d |
9babf392 | 154 | jne .Lnotintel |
4bb90087 | 155 | or \$0x40000000,%edx # set reserved bit#30 on Intel CPUs |
9babf392 AP |
156 | and \$15,%ah |
157 | cmp \$15,%ah # examine Family ID | |
64d92d74 | 158 | jne .LnotP4 |
4bb90087 | 159 | or \$0x00100000,%edx # set reserved bit#20 to engage RC4_CHAR |
64d92d74 AP |
160 | .LnotP4: |
161 | cmp \$6,%ah | |
162 | jne .Lnotintel | |
d84df594 | 163 | and \$0x0fff0ff0,%eax |
64d92d74 AP |
164 | cmp \$0x00050670,%eax # Knights Landing |
165 | je .Lknights | |
166 | cmp \$0x00080650,%eax # Knights Mill (according to sde) | |
167 | jne .Lnotintel | |
168 | .Lknights: | |
169 | and \$0xfbffffff,%ecx # clear XSAVE flag to mimic Silvermont | |
170 | ||
9babf392 | 171 | .Lnotintel: |
3df2eff4 | 172 | bt \$28,%edx # test hyper-threading bit |
b9064221 | 173 | jnc .Lgeneric |
761393bb AP |
174 | and \$0xefffffff,%edx # ~(1<<28) |
175 | cmp \$0,%r10d | |
b9064221 | 176 | je .Lgeneric |
761393bb AP |
177 | |
178 | or \$0x10000000,%edx # 1<<28 | |
9babf392 | 179 | shr \$16,%ebx |
761393bb | 180 | cmp \$1,%bl # see if cache is shared |
b9064221 | 181 | ja .Lgeneric |
932cc129 | 182 | and \$0xefffffff,%edx # ~(1<<28) |
b9064221 AP |
183 | .Lgeneric: |
184 | and \$0x00000800,%r9d # isolate AMD XOP flag | |
185 | and \$0xfffff7ff,%ecx | |
2bc3ad28 | 186 | or %ecx,%r9d # merge AMD XOP flag |
b9064221 | 187 | |
2bc3ad28 | 188 | mov %edx,%r10d # %r9d:%r10d is copy of %ecx:%edx |
1aed5e1a AP |
189 | |
190 | cmp \$7,%r11d | |
191 | jb .Lno_extended_info | |
192 | mov \$7,%eax | |
193 | xor %ecx,%ecx | |
194 | cpuid | |
64d92d74 AP |
195 | bt \$26,%r9d # check XSAVE bit, cleared on Knights |
196 | jc .Lnotknights | |
197 | and \$0xfff7ffff,%ebx # clear ADCX/ADOX flag | |
198 | .Lnotknights: | |
79337628 AP |
199 | movd %xmm0,%eax # restore processor id |
200 | and \$0x0fff0ff0,%eax | |
201 | cmp \$0x00050650,%eax # Skylake-X | |
202 | jne .Lnotskylakex | |
203 | and \$0xfffeffff,%ebx # ~(1<<16) | |
204 | # suppress AVX512F flag on Skylake-X | |
205 | .Lnotskylakex: | |
1aed5e1a | 206 | mov %ebx,8(%rdi) # save extended feature flags |
d6ee8f3d | 207 | mov %ecx,12(%rdi) |
1aed5e1a AP |
208 | .Lno_extended_info: |
209 | ||
2bc3ad28 | 210 | bt \$27,%r9d # check OSXSAVE bit |
b9064221 AP |
211 | jnc .Lclear_avx |
212 | xor %ecx,%ecx # XCR0 | |
213 | .byte 0x0f,0x01,0xd0 # xgetbv | |
66bee01c AP |
214 | and \$0xe6,%eax # isolate XMM, YMM and ZMM state support |
215 | cmp \$0xe6,%eax | |
216 | je .Ldone | |
88ac224c AP |
217 | andl \$0x3fdeffff,8(%rdi) # ~(1<<31|1<<30|1<<21|1<<16) |
218 | # clear AVX512F+BW+VL+FIMA, all of | |
219 | # them are EVEX-encoded, which requires | |
220 | # ZMM state support even if one uses | |
221 | # only XMM and YMM :-( | |
b9064221 AP |
222 | and \$6,%eax # isolate XMM and YMM state support |
223 | cmp \$6,%eax | |
224 | je .Ldone | |
225 | .Lclear_avx: | |
226 | mov \$0xefffe7ff,%eax # ~(1<<28|1<<12|1<<11) | |
2bc3ad28 | 227 | and %eax,%r9d # clear AVX, FMA and AMD XOP bits |
66bee01c | 228 | mov \$0x3fdeffdf,%eax # ~(1<<31|1<<30|1<<21|1<<16|1<<5) |
d67e7554 | 229 | and %eax,8(%rdi) # clear AVX2 and AVX512* bits |
b9064221 | 230 | .Ldone: |
2bc3ad28 AP |
231 | shl \$32,%r9 |
232 | mov %r10d,%eax | |
b9064221 | 233 | mov %r8,%rbx # restore %rbx |
5e32cfb2 | 234 | .cfi_restore %rbx |
2bc3ad28 | 235 | or %r9,%rax |
376729e1 | 236 | ret |
5e32cfb2 | 237 | .cfi_endproc |
376729e1 | 238 | .size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid |
b2dba9bf AP |
239 | |
240 | .globl OPENSSL_cleanse | |
aa8f38e4 | 241 | .type OPENSSL_cleanse,\@abi-omnipotent |
b2dba9bf AP |
242 | .align 16 |
243 | OPENSSL_cleanse: | |
8913378a | 244 | .cfi_startproc |
98ad3fe8 | 245 | endbranch |
b2dba9bf | 246 | xor %rax,%rax |
aa8f38e4 | 247 | cmp \$15,$arg2 |
b2dba9bf | 248 | jae .Lot |
7676eebf AP |
249 | cmp \$0,$arg2 |
250 | je .Lret | |
b2dba9bf | 251 | .Little: |
aa8f38e4 AP |
252 | mov %al,($arg1) |
253 | sub \$1,$arg2 | |
254 | lea 1($arg1),$arg1 | |
b2dba9bf | 255 | jnz .Little |
1fd79f66 AP |
256 | .Lret: |
257 | ret | |
b2dba9bf AP |
258 | .align 16 |
259 | .Lot: | |
aa8f38e4 | 260 | test \$7,$arg1 |
b2dba9bf | 261 | jz .Laligned |
aa8f38e4 AP |
262 | mov %al,($arg1) |
263 | lea -1($arg2),$arg2 | |
264 | lea 1($arg1),$arg1 | |
b2dba9bf AP |
265 | jmp .Lot |
266 | .Laligned: | |
aa8f38e4 AP |
267 | mov %rax,($arg1) |
268 | lea -8($arg2),$arg2 | |
269 | test \$-8,$arg2 | |
270 | lea 8($arg1),$arg1 | |
b2dba9bf | 271 | jnz .Laligned |
aa8f38e4 | 272 | cmp \$0,$arg2 |
b2dba9bf AP |
273 | jne .Little |
274 | ret | |
8913378a | 275 | .cfi_endproc |
b2dba9bf | 276 | .size OPENSSL_cleanse,.-OPENSSL_cleanse |
e33826f0 AP |
277 | |
278 | .globl CRYPTO_memcmp | |
279 | .type CRYPTO_memcmp,\@abi-omnipotent | |
280 | .align 16 | |
281 | CRYPTO_memcmp: | |
8913378a | 282 | .cfi_startproc |
98ad3fe8 | 283 | endbranch |
e33826f0 AP |
284 | xor %rax,%rax |
285 | xor %r10,%r10 | |
286 | cmp \$0,$arg3 | |
287 | je .Lno_data | |
9a708bf9 AP |
288 | cmp \$16,$arg3 |
289 | jne .Loop_cmp | |
290 | mov ($arg1),%r10 | |
291 | mov 8($arg1),%r11 | |
292 | mov \$1,$arg3 | |
293 | xor ($arg2),%r10 | |
294 | xor 8($arg2),%r11 | |
295 | or %r11,%r10 | |
296 | cmovnz $arg3,%rax | |
297 | ret | |
298 | ||
299 | .align 16 | |
e33826f0 AP |
300 | .Loop_cmp: |
301 | mov ($arg1),%r10b | |
302 | lea 1($arg1),$arg1 | |
303 | xor ($arg2),%r10b | |
304 | lea 1($arg2),$arg2 | |
305 | or %r10b,%al | |
306 | dec $arg3 | |
307 | jnz .Loop_cmp | |
308 | neg %rax | |
309 | shr \$63,%rax | |
310 | .Lno_data: | |
311 | ret | |
8913378a | 312 | .cfi_endproc |
e33826f0 | 313 | .size CRYPTO_memcmp,.-CRYPTO_memcmp |
14e21f86 | 314 | ___ |
aa8f38e4 AP |
315 | |
316 | print<<___ if (!$win64); | |
317 | .globl OPENSSL_wipe_cpu | |
318 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | |
319 | .align 16 | |
320 | OPENSSL_wipe_cpu: | |
8913378a | 321 | .cfi_startproc |
98ad3fe8 | 322 | endbranch |
aa8f38e4 AP |
323 | pxor %xmm0,%xmm0 |
324 | pxor %xmm1,%xmm1 | |
325 | pxor %xmm2,%xmm2 | |
326 | pxor %xmm3,%xmm3 | |
327 | pxor %xmm4,%xmm4 | |
328 | pxor %xmm5,%xmm5 | |
329 | pxor %xmm6,%xmm6 | |
330 | pxor %xmm7,%xmm7 | |
331 | pxor %xmm8,%xmm8 | |
332 | pxor %xmm9,%xmm9 | |
333 | pxor %xmm10,%xmm10 | |
334 | pxor %xmm11,%xmm11 | |
335 | pxor %xmm12,%xmm12 | |
336 | pxor %xmm13,%xmm13 | |
337 | pxor %xmm14,%xmm14 | |
338 | pxor %xmm15,%xmm15 | |
339 | xorq %rcx,%rcx | |
340 | xorq %rdx,%rdx | |
341 | xorq %rsi,%rsi | |
342 | xorq %rdi,%rdi | |
343 | xorq %r8,%r8 | |
344 | xorq %r9,%r9 | |
345 | xorq %r10,%r10 | |
346 | xorq %r11,%r11 | |
347 | leaq 8(%rsp),%rax | |
348 | ret | |
8913378a | 349 | .cfi_endproc |
aa8f38e4 AP |
350 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu |
351 | ___ | |
352 | print<<___ if ($win64); | |
353 | .globl OPENSSL_wipe_cpu | |
354 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | |
355 | .align 16 | |
356 | OPENSSL_wipe_cpu: | |
357 | pxor %xmm0,%xmm0 | |
358 | pxor %xmm1,%xmm1 | |
359 | pxor %xmm2,%xmm2 | |
360 | pxor %xmm3,%xmm3 | |
361 | pxor %xmm4,%xmm4 | |
362 | pxor %xmm5,%xmm5 | |
363 | xorq %rcx,%rcx | |
364 | xorq %rdx,%rdx | |
365 | xorq %r8,%r8 | |
366 | xorq %r9,%r9 | |
367 | xorq %r10,%r10 | |
368 | xorq %r11,%r11 | |
369 | leaq 8(%rsp),%rax | |
370 | ret | |
371 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | |
372 | ___ | |
5fabb88a AP |
373 | { |
374 | my $out="%r10"; | |
375 | my $cnt="%rcx"; | |
376 | my $max="%r11"; | |
377 | my $lasttick="%r8d"; | |
378 | my $lastdiff="%r9d"; | |
379 | my $redzone=win64?8:-8; | |
380 | ||
381 | print<<___; | |
382 | .globl OPENSSL_instrument_bus | |
383 | .type OPENSSL_instrument_bus,\@abi-omnipotent | |
384 | .align 16 | |
385 | OPENSSL_instrument_bus: | |
8913378a | 386 | .cfi_startproc |
98ad3fe8 | 387 | endbranch |
5fabb88a AP |
388 | mov $arg1,$out # tribute to Win64 |
389 | mov $arg2,$cnt | |
390 | mov $arg2,$max | |
391 | ||
392 | rdtsc # collect 1st tick | |
393 | mov %eax,$lasttick # lasttick = tick | |
394 | mov \$0,$lastdiff # lastdiff = 0 | |
395 | clflush ($out) | |
b9064221 | 396 | .byte 0xf0 # lock |
5fabb88a AP |
397 | add $lastdiff,($out) |
398 | jmp .Loop | |
399 | .align 16 | |
400 | .Loop: rdtsc | |
401 | mov %eax,%edx | |
402 | sub $lasttick,%eax | |
403 | mov %edx,$lasttick | |
404 | mov %eax,$lastdiff | |
405 | clflush ($out) | |
b9064221 | 406 | .byte 0xf0 # lock |
5fabb88a AP |
407 | add %eax,($out) |
408 | lea 4($out),$out | |
409 | sub \$1,$cnt | |
410 | jnz .Loop | |
411 | ||
412 | mov $max,%rax | |
413 | ret | |
8913378a | 414 | .cfi_endproc |
5fabb88a AP |
415 | .size OPENSSL_instrument_bus,.-OPENSSL_instrument_bus |
416 | ||
417 | .globl OPENSSL_instrument_bus2 | |
418 | .type OPENSSL_instrument_bus2,\@abi-omnipotent | |
419 | .align 16 | |
420 | OPENSSL_instrument_bus2: | |
8913378a | 421 | .cfi_startproc |
98ad3fe8 | 422 | endbranch |
5fabb88a AP |
423 | mov $arg1,$out # tribute to Win64 |
424 | mov $arg2,$cnt | |
425 | mov $arg3,$max | |
426 | mov $cnt,$redzone(%rsp) | |
427 | ||
428 | rdtsc # collect 1st tick | |
429 | mov %eax,$lasttick # lasttick = tick | |
430 | mov \$0,$lastdiff # lastdiff = 0 | |
431 | ||
432 | clflush ($out) | |
b9064221 | 433 | .byte 0xf0 # lock |
5fabb88a AP |
434 | add $lastdiff,($out) |
435 | ||
436 | rdtsc # collect 1st diff | |
437 | mov %eax,%edx | |
438 | sub $lasttick,%eax # diff | |
439 | mov %edx,$lasttick # lasttick = tick | |
440 | mov %eax,$lastdiff # lastdiff = diff | |
441 | .Loop2: | |
442 | clflush ($out) | |
b9064221 | 443 | .byte 0xf0 # lock |
5fabb88a AP |
444 | add %eax,($out) # accumulate diff |
445 | ||
446 | sub \$1,$max | |
447 | jz .Ldone2 | |
448 | ||
449 | rdtsc | |
450 | mov %eax,%edx | |
451 | sub $lasttick,%eax # diff | |
452 | mov %edx,$lasttick # lasttick = tick | |
453 | cmp $lastdiff,%eax | |
454 | mov %eax,$lastdiff # lastdiff = diff | |
455 | mov \$0,%edx | |
456 | setne %dl | |
457 | sub %rdx,$cnt # conditional --$cnt | |
458 | lea ($out,%rdx,4),$out # conditional ++$out | |
459 | jnz .Loop2 | |
460 | ||
461 | .Ldone2: | |
462 | mov $redzone(%rsp),%rax | |
463 | sub $cnt,%rax | |
464 | ret | |
8913378a | 465 | .cfi_endproc |
5fabb88a AP |
466 | .size OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2 |
467 | ___ | |
468 | } | |
aa8f38e4 | 469 | |
9c940446 AP |
470 | sub gen_random { |
471 | my $rdop = shift; | |
301799b8 | 472 | print<<___; |
9c940446 AP |
473 | .globl OPENSSL_ia32_${rdop}_bytes |
474 | .type OPENSSL_ia32_${rdop}_bytes,\@abi-omnipotent | |
f4d45640 | 475 | .align 16 |
9c940446 | 476 | OPENSSL_ia32_${rdop}_bytes: |
8913378a | 477 | .cfi_startproc |
98ad3fe8 | 478 | endbranch |
9c940446 AP |
479 | xor %rax, %rax # return value |
480 | cmp \$0,$arg2 | |
481 | je .Ldone_${rdop}_bytes | |
482 | ||
483 | mov \$8,%r11 | |
484 | .Loop_${rdop}_bytes: | |
485 | ${rdop} %r10 | |
486 | jc .Lbreak_${rdop}_bytes | |
487 | dec %r11 | |
488 | jnz .Loop_${rdop}_bytes | |
489 | jmp .Ldone_${rdop}_bytes | |
490 | ||
491 | .align 16 | |
492 | .Lbreak_${rdop}_bytes: | |
493 | cmp \$8,$arg2 | |
494 | jb .Ltail_${rdop}_bytes | |
495 | mov %r10,($arg1) | |
496 | lea 8($arg1),$arg1 | |
497 | add \$8,%rax | |
498 | sub \$8,$arg2 | |
499 | jz .Ldone_${rdop}_bytes | |
500 | mov \$8,%r11 | |
501 | jmp .Loop_${rdop}_bytes | |
502 | ||
503 | .align 16 | |
504 | .Ltail_${rdop}_bytes: | |
505 | mov %r10b,($arg1) | |
506 | lea 1($arg1),$arg1 | |
507 | inc %rax | |
082193ef | 508 | shr \$8,%r10 |
9c940446 AP |
509 | dec $arg2 |
510 | jnz .Ltail_${rdop}_bytes | |
511 | ||
512 | .Ldone_${rdop}_bytes: | |
082193ef | 513 | xor %r10,%r10 # Clear sensitive data from register |
f4d45640 | 514 | ret |
8913378a | 515 | .cfi_endproc |
9c940446 | 516 | .size OPENSSL_ia32_${rdop}_bytes,.-OPENSSL_ia32_${rdop}_bytes |
301799b8 | 517 | ___ |
9c940446 AP |
518 | } |
519 | gen_random("rdrand"); | |
520 | gen_random("rdseed"); | |
301799b8 | 521 | |
a21314db | 522 | close STDOUT or die "error closing STDOUT: $!"; # flush |