]>
Commit | Line | Data |
---|---|---|
e0a65194 | 1 | #! /usr/bin/env perl |
082193ef | 2 | # Copyright 2005-2018 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 |
aa8f38e4 AP |
45 | movl ($arg1),%eax |
46 | .Lspin: leaq ($arg2,%rax),%r8 | |
47 | .byte 0xf0 # lock | |
48 | cmpxchgl %r8d,($arg1) | |
376729e1 | 49 | jne .Lspin |
e442c362 | 50 | movl %r8d,%eax |
aa8f38e4 | 51 | .byte 0x48,0x98 # cltq/cdqe |
376729e1 | 52 | ret |
8913378a | 53 | .cfi_endproc |
376729e1 AP |
54 | .size OPENSSL_atomic_add,.-OPENSSL_atomic_add |
55 | ||
932cc129 AP |
56 | .globl OPENSSL_rdtsc |
57 | .type OPENSSL_rdtsc,\@abi-omnipotent | |
58 | .align 16 | |
59 | OPENSSL_rdtsc: | |
8913378a | 60 | .cfi_startproc |
932cc129 AP |
61 | rdtsc |
62 | shl \$32,%rdx | |
63 | or %rdx,%rax | |
64 | ret | |
8913378a | 65 | .cfi_endproc |
932cc129 AP |
66 | .size OPENSSL_rdtsc,.-OPENSSL_rdtsc |
67 | ||
376729e1 | 68 | .globl OPENSSL_ia32_cpuid |
c5cd28bd | 69 | .type OPENSSL_ia32_cpuid,\@function,1 |
376729e1 AP |
70 | .align 16 |
71 | OPENSSL_ia32_cpuid: | |
5e32cfb2 | 72 | .cfi_startproc |
b9064221 | 73 | mov %rbx,%r8 # save %rbx |
5e32cfb2 | 74 | .cfi_register %rbx,%r8 |
9babf392 AP |
75 | |
76 | xor %eax,%eax | |
d6ee8f3d | 77 | mov %rax,8(%rdi) # clear extended feature flags |
9babf392 | 78 | cpuid |
761393bb AP |
79 | mov %eax,%r11d # max value for standard query level |
80 | ||
9babf392 AP |
81 | xor %eax,%eax |
82 | cmp \$0x756e6547,%ebx # "Genu" | |
83 | setne %al | |
84 | mov %eax,%r9d | |
85 | cmp \$0x49656e69,%edx # "ineI" | |
86 | setne %al | |
87 | or %eax,%r9d | |
88 | cmp \$0x6c65746e,%ecx # "ntel" | |
89 | setne %al | |
5cd91b50 | 90 | or %eax,%r9d # 0 indicates Intel CPU |
5cd91b50 AP |
91 | jz .Lintel |
92 | ||
93 | cmp \$0x68747541,%ebx # "Auth" | |
94 | setne %al | |
95 | mov %eax,%r10d | |
96 | cmp \$0x69746E65,%edx # "enti" | |
97 | setne %al | |
98 | or %eax,%r10d | |
99 | cmp \$0x444D4163,%ecx # "cAMD" | |
100 | setne %al | |
101 | or %eax,%r10d # 0 indicates AMD CPU | |
102 | jnz .Lintel | |
103 | ||
761393bb | 104 | # AMD specific |
5cd91b50 AP |
105 | mov \$0x80000000,%eax |
106 | cpuid | |
b9064221 AP |
107 | cmp \$0x80000001,%eax |
108 | jb .Lintel | |
109 | mov %eax,%r10d | |
110 | mov \$0x80000001,%eax | |
111 | cpuid | |
112 | or %ecx,%r9d | |
113 | and \$0x00000801,%r9d # isolate AMD XOP bit, 1<<11 | |
114 | ||
115 | cmp \$0x80000008,%r10d | |
5cd91b50 AP |
116 | jb .Lintel |
117 | ||
118 | mov \$0x80000008,%eax | |
119 | cpuid | |
120 | movzb %cl,%r10 # number of cores - 1 | |
121 | inc %r10 # number of cores | |
9babf392 | 122 | |
761393bb AP |
123 | mov \$1,%eax |
124 | cpuid | |
125 | bt \$28,%edx # test hyper-threading bit | |
b9064221 | 126 | jnc .Lgeneric |
761393bb AP |
127 | shr \$16,%ebx # number of logical processors |
128 | cmp %r10b,%bl | |
b9064221 | 129 | ja .Lgeneric |
761393bb | 130 | and \$0xefffffff,%edx # ~(1<<28) |
b9064221 | 131 | jmp .Lgeneric |
761393bb | 132 | |
5cd91b50 | 133 | .Lintel: |
761393bb AP |
134 | cmp \$4,%r11d |
135 | mov \$-1,%r10d | |
136 | jb .Lnocacheinfo | |
137 | ||
138 | mov \$4,%eax | |
139 | mov \$0,%ecx # query L1D | |
140 | cpuid | |
141 | mov %eax,%r10d | |
142 | shr \$14,%r10d | |
143 | and \$0xfff,%r10d # number of cores -1 per L1D | |
144 | ||
145 | .Lnocacheinfo: | |
932cc129 | 146 | mov \$1,%eax |
376729e1 | 147 | cpuid |
79337628 | 148 | movd %eax,%xmm0 # put aside processor id |
4bb90087 | 149 | and \$0xbfefffff,%edx # force reserved bits to 0 |
932cc129 | 150 | cmp \$0,%r9d |
9babf392 | 151 | jne .Lnotintel |
4bb90087 | 152 | or \$0x40000000,%edx # set reserved bit#30 on Intel CPUs |
9babf392 AP |
153 | and \$15,%ah |
154 | cmp \$15,%ah # examine Family ID | |
64d92d74 | 155 | jne .LnotP4 |
4bb90087 | 156 | or \$0x00100000,%edx # set reserved bit#20 to engage RC4_CHAR |
64d92d74 AP |
157 | .LnotP4: |
158 | cmp \$6,%ah | |
159 | jne .Lnotintel | |
d84df594 | 160 | and \$0x0fff0ff0,%eax |
64d92d74 AP |
161 | cmp \$0x00050670,%eax # Knights Landing |
162 | je .Lknights | |
163 | cmp \$0x00080650,%eax # Knights Mill (according to sde) | |
164 | jne .Lnotintel | |
165 | .Lknights: | |
166 | and \$0xfbffffff,%ecx # clear XSAVE flag to mimic Silvermont | |
167 | ||
9babf392 | 168 | .Lnotintel: |
3df2eff4 | 169 | bt \$28,%edx # test hyper-threading bit |
b9064221 | 170 | jnc .Lgeneric |
761393bb AP |
171 | and \$0xefffffff,%edx # ~(1<<28) |
172 | cmp \$0,%r10d | |
b9064221 | 173 | je .Lgeneric |
761393bb AP |
174 | |
175 | or \$0x10000000,%edx # 1<<28 | |
9babf392 | 176 | shr \$16,%ebx |
761393bb | 177 | cmp \$1,%bl # see if cache is shared |
b9064221 | 178 | ja .Lgeneric |
932cc129 | 179 | and \$0xefffffff,%edx # ~(1<<28) |
b9064221 AP |
180 | .Lgeneric: |
181 | and \$0x00000800,%r9d # isolate AMD XOP flag | |
182 | and \$0xfffff7ff,%ecx | |
2bc3ad28 | 183 | or %ecx,%r9d # merge AMD XOP flag |
b9064221 | 184 | |
2bc3ad28 | 185 | mov %edx,%r10d # %r9d:%r10d is copy of %ecx:%edx |
1aed5e1a AP |
186 | |
187 | cmp \$7,%r11d | |
188 | jb .Lno_extended_info | |
189 | mov \$7,%eax | |
190 | xor %ecx,%ecx | |
191 | cpuid | |
64d92d74 AP |
192 | bt \$26,%r9d # check XSAVE bit, cleared on Knights |
193 | jc .Lnotknights | |
194 | and \$0xfff7ffff,%ebx # clear ADCX/ADOX flag | |
195 | .Lnotknights: | |
79337628 AP |
196 | movd %xmm0,%eax # restore processor id |
197 | and \$0x0fff0ff0,%eax | |
198 | cmp \$0x00050650,%eax # Skylake-X | |
199 | jne .Lnotskylakex | |
200 | and \$0xfffeffff,%ebx # ~(1<<16) | |
201 | # suppress AVX512F flag on Skylake-X | |
202 | .Lnotskylakex: | |
1aed5e1a | 203 | mov %ebx,8(%rdi) # save extended feature flags |
d6ee8f3d | 204 | mov %ecx,12(%rdi) |
1aed5e1a AP |
205 | .Lno_extended_info: |
206 | ||
2bc3ad28 | 207 | bt \$27,%r9d # check OSXSAVE bit |
b9064221 AP |
208 | jnc .Lclear_avx |
209 | xor %ecx,%ecx # XCR0 | |
210 | .byte 0x0f,0x01,0xd0 # xgetbv | |
66bee01c AP |
211 | and \$0xe6,%eax # isolate XMM, YMM and ZMM state support |
212 | cmp \$0xe6,%eax | |
213 | je .Ldone | |
88ac224c AP |
214 | andl \$0x3fdeffff,8(%rdi) # ~(1<<31|1<<30|1<<21|1<<16) |
215 | # clear AVX512F+BW+VL+FIMA, all of | |
216 | # them are EVEX-encoded, which requires | |
217 | # ZMM state support even if one uses | |
218 | # only XMM and YMM :-( | |
b9064221 AP |
219 | and \$6,%eax # isolate XMM and YMM state support |
220 | cmp \$6,%eax | |
221 | je .Ldone | |
222 | .Lclear_avx: | |
223 | mov \$0xefffe7ff,%eax # ~(1<<28|1<<12|1<<11) | |
2bc3ad28 | 224 | and %eax,%r9d # clear AVX, FMA and AMD XOP bits |
66bee01c | 225 | mov \$0x3fdeffdf,%eax # ~(1<<31|1<<30|1<<21|1<<16|1<<5) |
d67e7554 | 226 | and %eax,8(%rdi) # clear AVX2 and AVX512* bits |
b9064221 | 227 | .Ldone: |
2bc3ad28 AP |
228 | shl \$32,%r9 |
229 | mov %r10d,%eax | |
b9064221 | 230 | mov %r8,%rbx # restore %rbx |
5e32cfb2 | 231 | .cfi_restore %rbx |
2bc3ad28 | 232 | or %r9,%rax |
376729e1 | 233 | ret |
5e32cfb2 | 234 | .cfi_endproc |
376729e1 | 235 | .size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid |
b2dba9bf AP |
236 | |
237 | .globl OPENSSL_cleanse | |
aa8f38e4 | 238 | .type OPENSSL_cleanse,\@abi-omnipotent |
b2dba9bf AP |
239 | .align 16 |
240 | OPENSSL_cleanse: | |
8913378a | 241 | .cfi_startproc |
b2dba9bf | 242 | xor %rax,%rax |
aa8f38e4 | 243 | cmp \$15,$arg2 |
b2dba9bf | 244 | jae .Lot |
7676eebf AP |
245 | cmp \$0,$arg2 |
246 | je .Lret | |
b2dba9bf | 247 | .Little: |
aa8f38e4 AP |
248 | mov %al,($arg1) |
249 | sub \$1,$arg2 | |
250 | lea 1($arg1),$arg1 | |
b2dba9bf | 251 | jnz .Little |
1fd79f66 AP |
252 | .Lret: |
253 | ret | |
b2dba9bf AP |
254 | .align 16 |
255 | .Lot: | |
aa8f38e4 | 256 | test \$7,$arg1 |
b2dba9bf | 257 | jz .Laligned |
aa8f38e4 AP |
258 | mov %al,($arg1) |
259 | lea -1($arg2),$arg2 | |
260 | lea 1($arg1),$arg1 | |
b2dba9bf AP |
261 | jmp .Lot |
262 | .Laligned: | |
aa8f38e4 AP |
263 | mov %rax,($arg1) |
264 | lea -8($arg2),$arg2 | |
265 | test \$-8,$arg2 | |
266 | lea 8($arg1),$arg1 | |
b2dba9bf | 267 | jnz .Laligned |
aa8f38e4 | 268 | cmp \$0,$arg2 |
b2dba9bf AP |
269 | jne .Little |
270 | ret | |
8913378a | 271 | .cfi_endproc |
b2dba9bf | 272 | .size OPENSSL_cleanse,.-OPENSSL_cleanse |
e33826f0 AP |
273 | |
274 | .globl CRYPTO_memcmp | |
275 | .type CRYPTO_memcmp,\@abi-omnipotent | |
276 | .align 16 | |
277 | CRYPTO_memcmp: | |
8913378a | 278 | .cfi_startproc |
e33826f0 AP |
279 | xor %rax,%rax |
280 | xor %r10,%r10 | |
281 | cmp \$0,$arg3 | |
282 | je .Lno_data | |
9a708bf9 AP |
283 | cmp \$16,$arg3 |
284 | jne .Loop_cmp | |
285 | mov ($arg1),%r10 | |
286 | mov 8($arg1),%r11 | |
287 | mov \$1,$arg3 | |
288 | xor ($arg2),%r10 | |
289 | xor 8($arg2),%r11 | |
290 | or %r11,%r10 | |
291 | cmovnz $arg3,%rax | |
292 | ret | |
293 | ||
294 | .align 16 | |
e33826f0 AP |
295 | .Loop_cmp: |
296 | mov ($arg1),%r10b | |
297 | lea 1($arg1),$arg1 | |
298 | xor ($arg2),%r10b | |
299 | lea 1($arg2),$arg2 | |
300 | or %r10b,%al | |
301 | dec $arg3 | |
302 | jnz .Loop_cmp | |
303 | neg %rax | |
304 | shr \$63,%rax | |
305 | .Lno_data: | |
306 | ret | |
8913378a | 307 | .cfi_endproc |
e33826f0 | 308 | .size CRYPTO_memcmp,.-CRYPTO_memcmp |
14e21f86 | 309 | ___ |
aa8f38e4 AP |
310 | |
311 | print<<___ if (!$win64); | |
312 | .globl OPENSSL_wipe_cpu | |
313 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | |
314 | .align 16 | |
315 | OPENSSL_wipe_cpu: | |
8913378a | 316 | .cfi_startproc |
aa8f38e4 AP |
317 | pxor %xmm0,%xmm0 |
318 | pxor %xmm1,%xmm1 | |
319 | pxor %xmm2,%xmm2 | |
320 | pxor %xmm3,%xmm3 | |
321 | pxor %xmm4,%xmm4 | |
322 | pxor %xmm5,%xmm5 | |
323 | pxor %xmm6,%xmm6 | |
324 | pxor %xmm7,%xmm7 | |
325 | pxor %xmm8,%xmm8 | |
326 | pxor %xmm9,%xmm9 | |
327 | pxor %xmm10,%xmm10 | |
328 | pxor %xmm11,%xmm11 | |
329 | pxor %xmm12,%xmm12 | |
330 | pxor %xmm13,%xmm13 | |
331 | pxor %xmm14,%xmm14 | |
332 | pxor %xmm15,%xmm15 | |
333 | xorq %rcx,%rcx | |
334 | xorq %rdx,%rdx | |
335 | xorq %rsi,%rsi | |
336 | xorq %rdi,%rdi | |
337 | xorq %r8,%r8 | |
338 | xorq %r9,%r9 | |
339 | xorq %r10,%r10 | |
340 | xorq %r11,%r11 | |
341 | leaq 8(%rsp),%rax | |
342 | ret | |
8913378a | 343 | .cfi_endproc |
aa8f38e4 AP |
344 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu |
345 | ___ | |
346 | print<<___ if ($win64); | |
347 | .globl OPENSSL_wipe_cpu | |
348 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | |
349 | .align 16 | |
350 | OPENSSL_wipe_cpu: | |
351 | pxor %xmm0,%xmm0 | |
352 | pxor %xmm1,%xmm1 | |
353 | pxor %xmm2,%xmm2 | |
354 | pxor %xmm3,%xmm3 | |
355 | pxor %xmm4,%xmm4 | |
356 | pxor %xmm5,%xmm5 | |
357 | xorq %rcx,%rcx | |
358 | xorq %rdx,%rdx | |
359 | xorq %r8,%r8 | |
360 | xorq %r9,%r9 | |
361 | xorq %r10,%r10 | |
362 | xorq %r11,%r11 | |
363 | leaq 8(%rsp),%rax | |
364 | ret | |
365 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | |
366 | ___ | |
5fabb88a AP |
367 | { |
368 | my $out="%r10"; | |
369 | my $cnt="%rcx"; | |
370 | my $max="%r11"; | |
371 | my $lasttick="%r8d"; | |
372 | my $lastdiff="%r9d"; | |
373 | my $redzone=win64?8:-8; | |
374 | ||
375 | print<<___; | |
376 | .globl OPENSSL_instrument_bus | |
377 | .type OPENSSL_instrument_bus,\@abi-omnipotent | |
378 | .align 16 | |
379 | OPENSSL_instrument_bus: | |
8913378a | 380 | .cfi_startproc |
5fabb88a AP |
381 | mov $arg1,$out # tribute to Win64 |
382 | mov $arg2,$cnt | |
383 | mov $arg2,$max | |
384 | ||
385 | rdtsc # collect 1st tick | |
386 | mov %eax,$lasttick # lasttick = tick | |
387 | mov \$0,$lastdiff # lastdiff = 0 | |
388 | clflush ($out) | |
b9064221 | 389 | .byte 0xf0 # lock |
5fabb88a AP |
390 | add $lastdiff,($out) |
391 | jmp .Loop | |
392 | .align 16 | |
393 | .Loop: rdtsc | |
394 | mov %eax,%edx | |
395 | sub $lasttick,%eax | |
396 | mov %edx,$lasttick | |
397 | mov %eax,$lastdiff | |
398 | clflush ($out) | |
b9064221 | 399 | .byte 0xf0 # lock |
5fabb88a AP |
400 | add %eax,($out) |
401 | lea 4($out),$out | |
402 | sub \$1,$cnt | |
403 | jnz .Loop | |
404 | ||
405 | mov $max,%rax | |
406 | ret | |
8913378a | 407 | .cfi_endproc |
5fabb88a AP |
408 | .size OPENSSL_instrument_bus,.-OPENSSL_instrument_bus |
409 | ||
410 | .globl OPENSSL_instrument_bus2 | |
411 | .type OPENSSL_instrument_bus2,\@abi-omnipotent | |
412 | .align 16 | |
413 | OPENSSL_instrument_bus2: | |
8913378a | 414 | .cfi_startproc |
5fabb88a AP |
415 | mov $arg1,$out # tribute to Win64 |
416 | mov $arg2,$cnt | |
417 | mov $arg3,$max | |
418 | mov $cnt,$redzone(%rsp) | |
419 | ||
420 | rdtsc # collect 1st tick | |
421 | mov %eax,$lasttick # lasttick = tick | |
422 | mov \$0,$lastdiff # lastdiff = 0 | |
423 | ||
424 | clflush ($out) | |
b9064221 | 425 | .byte 0xf0 # lock |
5fabb88a AP |
426 | add $lastdiff,($out) |
427 | ||
428 | rdtsc # collect 1st diff | |
429 | mov %eax,%edx | |
430 | sub $lasttick,%eax # diff | |
431 | mov %edx,$lasttick # lasttick = tick | |
432 | mov %eax,$lastdiff # lastdiff = diff | |
433 | .Loop2: | |
434 | clflush ($out) | |
b9064221 | 435 | .byte 0xf0 # lock |
5fabb88a AP |
436 | add %eax,($out) # accumulate diff |
437 | ||
438 | sub \$1,$max | |
439 | jz .Ldone2 | |
440 | ||
441 | rdtsc | |
442 | mov %eax,%edx | |
443 | sub $lasttick,%eax # diff | |
444 | mov %edx,$lasttick # lasttick = tick | |
445 | cmp $lastdiff,%eax | |
446 | mov %eax,$lastdiff # lastdiff = diff | |
447 | mov \$0,%edx | |
448 | setne %dl | |
449 | sub %rdx,$cnt # conditional --$cnt | |
450 | lea ($out,%rdx,4),$out # conditional ++$out | |
451 | jnz .Loop2 | |
452 | ||
453 | .Ldone2: | |
454 | mov $redzone(%rsp),%rax | |
455 | sub $cnt,%rax | |
456 | ret | |
8913378a | 457 | .cfi_endproc |
5fabb88a AP |
458 | .size OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2 |
459 | ___ | |
460 | } | |
aa8f38e4 | 461 | |
9c940446 AP |
462 | sub gen_random { |
463 | my $rdop = shift; | |
301799b8 | 464 | print<<___; |
9c940446 AP |
465 | .globl OPENSSL_ia32_${rdop}_bytes |
466 | .type OPENSSL_ia32_${rdop}_bytes,\@abi-omnipotent | |
f4d45640 | 467 | .align 16 |
9c940446 | 468 | OPENSSL_ia32_${rdop}_bytes: |
8913378a | 469 | .cfi_startproc |
9c940446 AP |
470 | xor %rax, %rax # return value |
471 | cmp \$0,$arg2 | |
472 | je .Ldone_${rdop}_bytes | |
473 | ||
474 | mov \$8,%r11 | |
475 | .Loop_${rdop}_bytes: | |
476 | ${rdop} %r10 | |
477 | jc .Lbreak_${rdop}_bytes | |
478 | dec %r11 | |
479 | jnz .Loop_${rdop}_bytes | |
480 | jmp .Ldone_${rdop}_bytes | |
481 | ||
482 | .align 16 | |
483 | .Lbreak_${rdop}_bytes: | |
484 | cmp \$8,$arg2 | |
485 | jb .Ltail_${rdop}_bytes | |
486 | mov %r10,($arg1) | |
487 | lea 8($arg1),$arg1 | |
488 | add \$8,%rax | |
489 | sub \$8,$arg2 | |
490 | jz .Ldone_${rdop}_bytes | |
491 | mov \$8,%r11 | |
492 | jmp .Loop_${rdop}_bytes | |
493 | ||
494 | .align 16 | |
495 | .Ltail_${rdop}_bytes: | |
496 | mov %r10b,($arg1) | |
497 | lea 1($arg1),$arg1 | |
498 | inc %rax | |
082193ef | 499 | shr \$8,%r10 |
9c940446 AP |
500 | dec $arg2 |
501 | jnz .Ltail_${rdop}_bytes | |
502 | ||
503 | .Ldone_${rdop}_bytes: | |
082193ef | 504 | xor %r10,%r10 # Clear sensitive data from register |
f4d45640 | 505 | ret |
8913378a | 506 | .cfi_endproc |
9c940446 | 507 | .size OPENSSL_ia32_${rdop}_bytes,.-OPENSSL_ia32_${rdop}_bytes |
301799b8 | 508 | ___ |
9c940446 AP |
509 | } |
510 | gen_random("rdrand"); | |
511 | gen_random("rdseed"); | |
301799b8 | 512 | |
5d863367 | 513 | close STDOUT; # flush |