]>
Commit | Line | Data |
---|---|---|
14e21f86 AP |
1 | #!/usr/bin/env perl |
2 | ||
aa8f38e4 AP |
3 | $flavour = shift; |
4 | $output = shift; | |
5 | if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } | |
14e21f86 | 6 | |
aa8f38e4 | 7 | $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); |
376729e1 | 8 | |
aa8f38e4 | 9 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; |
301799b8 AP |
10 | ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or |
11 | ( $xlate="${dir}perlasm/x86_64-xlate.pl" and -f $xlate) or | |
12 | die "can't locate x86_64-xlate.pl"; | |
13 | ||
46bf83f0 AP |
14 | open OUT,"| \"$^X\" $xlate $flavour $output"; |
15 | *STDOUT=*OUT; | |
376729e1 | 16 | |
5fabb88a AP |
17 | ($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order |
18 | ("%rdi","%rsi","%rdx","%rcx"); # Unix order | |
19 | ||
aa8f38e4 AP |
20 | print<<___; |
21 | .extern OPENSSL_cpuid_setup | |
ddc20d4d | 22 | .hidden OPENSSL_cpuid_setup |
aa8f38e4 AP |
23 | .section .init |
24 | call OPENSSL_cpuid_setup | |
932cc129 | 25 | |
ddc20d4d | 26 | .hidden OPENSSL_ia32cap_P |
c5cd28bd | 27 | .comm OPENSSL_ia32cap_P,16,4 |
ddc20d4d | 28 | |
14e21f86 | 29 | .text |
376729e1 AP |
30 | |
31 | .globl OPENSSL_atomic_add | |
aa8f38e4 | 32 | .type OPENSSL_atomic_add,\@abi-omnipotent |
376729e1 AP |
33 | .align 16 |
34 | OPENSSL_atomic_add: | |
aa8f38e4 AP |
35 | movl ($arg1),%eax |
36 | .Lspin: leaq ($arg2,%rax),%r8 | |
37 | .byte 0xf0 # lock | |
38 | cmpxchgl %r8d,($arg1) | |
376729e1 | 39 | jne .Lspin |
e442c362 | 40 | movl %r8d,%eax |
aa8f38e4 | 41 | .byte 0x48,0x98 # cltq/cdqe |
376729e1 AP |
42 | ret |
43 | .size OPENSSL_atomic_add,.-OPENSSL_atomic_add | |
44 | ||
932cc129 AP |
45 | .globl OPENSSL_rdtsc |
46 | .type OPENSSL_rdtsc,\@abi-omnipotent | |
47 | .align 16 | |
48 | OPENSSL_rdtsc: | |
49 | rdtsc | |
50 | shl \$32,%rdx | |
51 | or %rdx,%rax | |
52 | ret | |
53 | .size OPENSSL_rdtsc,.-OPENSSL_rdtsc | |
54 | ||
376729e1 | 55 | .globl OPENSSL_ia32_cpuid |
c5cd28bd | 56 | .type OPENSSL_ia32_cpuid,\@function,1 |
376729e1 AP |
57 | .align 16 |
58 | OPENSSL_ia32_cpuid: | |
b9064221 | 59 | mov %rbx,%r8 # save %rbx |
9babf392 AP |
60 | |
61 | xor %eax,%eax | |
c5cd28bd | 62 | mov %eax,8(%rdi) # clear 3rd word |
9babf392 | 63 | cpuid |
761393bb AP |
64 | mov %eax,%r11d # max value for standard query level |
65 | ||
9babf392 AP |
66 | xor %eax,%eax |
67 | cmp \$0x756e6547,%ebx # "Genu" | |
68 | setne %al | |
69 | mov %eax,%r9d | |
70 | cmp \$0x49656e69,%edx # "ineI" | |
71 | setne %al | |
72 | or %eax,%r9d | |
73 | cmp \$0x6c65746e,%ecx # "ntel" | |
74 | setne %al | |
5cd91b50 | 75 | or %eax,%r9d # 0 indicates Intel CPU |
5cd91b50 AP |
76 | jz .Lintel |
77 | ||
78 | cmp \$0x68747541,%ebx # "Auth" | |
79 | setne %al | |
80 | mov %eax,%r10d | |
81 | cmp \$0x69746E65,%edx # "enti" | |
82 | setne %al | |
83 | or %eax,%r10d | |
84 | cmp \$0x444D4163,%ecx # "cAMD" | |
85 | setne %al | |
86 | or %eax,%r10d # 0 indicates AMD CPU | |
87 | jnz .Lintel | |
88 | ||
761393bb | 89 | # AMD specific |
5cd91b50 AP |
90 | mov \$0x80000000,%eax |
91 | cpuid | |
b9064221 AP |
92 | cmp \$0x80000001,%eax |
93 | jb .Lintel | |
94 | mov %eax,%r10d | |
95 | mov \$0x80000001,%eax | |
96 | cpuid | |
97 | or %ecx,%r9d | |
98 | and \$0x00000801,%r9d # isolate AMD XOP bit, 1<<11 | |
99 | ||
100 | cmp \$0x80000008,%r10d | |
5cd91b50 AP |
101 | jb .Lintel |
102 | ||
103 | mov \$0x80000008,%eax | |
104 | cpuid | |
105 | movzb %cl,%r10 # number of cores - 1 | |
106 | inc %r10 # number of cores | |
9babf392 | 107 | |
761393bb AP |
108 | mov \$1,%eax |
109 | cpuid | |
110 | bt \$28,%edx # test hyper-threading bit | |
b9064221 | 111 | jnc .Lgeneric |
761393bb AP |
112 | shr \$16,%ebx # number of logical processors |
113 | cmp %r10b,%bl | |
b9064221 | 114 | ja .Lgeneric |
761393bb | 115 | and \$0xefffffff,%edx # ~(1<<28) |
b9064221 | 116 | jmp .Lgeneric |
761393bb | 117 | |
5cd91b50 | 118 | .Lintel: |
761393bb AP |
119 | cmp \$4,%r11d |
120 | mov \$-1,%r10d | |
121 | jb .Lnocacheinfo | |
122 | ||
123 | mov \$4,%eax | |
124 | mov \$0,%ecx # query L1D | |
125 | cpuid | |
126 | mov %eax,%r10d | |
127 | shr \$14,%r10d | |
128 | and \$0xfff,%r10d # number of cores -1 per L1D | |
129 | ||
c5cd28bd AP |
130 | cmp \$7,%r11d |
131 | jb .Lnocacheinfo | |
132 | ||
133 | mov \$7,%eax | |
134 | xor %ecx,%ecx | |
135 | cpuid | |
136 | mov %ebx,8(%rdi) | |
137 | ||
761393bb | 138 | .Lnocacheinfo: |
932cc129 | 139 | mov \$1,%eax |
376729e1 | 140 | cpuid |
4bb90087 | 141 | and \$0xbfefffff,%edx # force reserved bits to 0 |
932cc129 | 142 | cmp \$0,%r9d |
9babf392 | 143 | jne .Lnotintel |
4bb90087 | 144 | or \$0x40000000,%edx # set reserved bit#30 on Intel CPUs |
9babf392 AP |
145 | and \$15,%ah |
146 | cmp \$15,%ah # examine Family ID | |
4bb90087 AP |
147 | jne .Lnotintel |
148 | or \$0x00100000,%edx # set reserved bit#20 to engage RC4_CHAR | |
9babf392 | 149 | .Lnotintel: |
3df2eff4 | 150 | bt \$28,%edx # test hyper-threading bit |
b9064221 | 151 | jnc .Lgeneric |
761393bb AP |
152 | and \$0xefffffff,%edx # ~(1<<28) |
153 | cmp \$0,%r10d | |
b9064221 | 154 | je .Lgeneric |
761393bb AP |
155 | |
156 | or \$0x10000000,%edx # 1<<28 | |
9babf392 | 157 | shr \$16,%ebx |
761393bb | 158 | cmp \$1,%bl # see if cache is shared |
b9064221 | 159 | ja .Lgeneric |
932cc129 | 160 | and \$0xefffffff,%edx # ~(1<<28) |
b9064221 AP |
161 | .Lgeneric: |
162 | and \$0x00000800,%r9d # isolate AMD XOP flag | |
163 | and \$0xfffff7ff,%ecx | |
2bc3ad28 | 164 | or %ecx,%r9d # merge AMD XOP flag |
b9064221 | 165 | |
2bc3ad28 AP |
166 | mov %edx,%r10d # %r9d:%r10d is copy of %ecx:%edx |
167 | bt \$27,%r9d # check OSXSAVE bit | |
b9064221 AP |
168 | jnc .Lclear_avx |
169 | xor %ecx,%ecx # XCR0 | |
170 | .byte 0x0f,0x01,0xd0 # xgetbv | |
171 | and \$6,%eax # isolate XMM and YMM state support | |
172 | cmp \$6,%eax | |
173 | je .Ldone | |
174 | .Lclear_avx: | |
175 | mov \$0xefffe7ff,%eax # ~(1<<28|1<<12|1<<11) | |
2bc3ad28 | 176 | and %eax,%r9d # clear AVX, FMA and AMD XOP bits |
c5cd28bd | 177 | andl \$0xffffffdf,8(%rdi) # cleax AVX2, ~(1<<5) |
b9064221 | 178 | .Ldone: |
2bc3ad28 AP |
179 | shl \$32,%r9 |
180 | mov %r10d,%eax | |
b9064221 | 181 | mov %r8,%rbx # restore %rbx |
2bc3ad28 | 182 | or %r9,%rax |
376729e1 AP |
183 | ret |
184 | .size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid | |
b2dba9bf AP |
185 | |
186 | .globl OPENSSL_cleanse | |
aa8f38e4 | 187 | .type OPENSSL_cleanse,\@abi-omnipotent |
b2dba9bf AP |
188 | .align 16 |
189 | OPENSSL_cleanse: | |
190 | xor %rax,%rax | |
aa8f38e4 | 191 | cmp \$15,$arg2 |
b2dba9bf | 192 | jae .Lot |
7676eebf AP |
193 | cmp \$0,$arg2 |
194 | je .Lret | |
b2dba9bf | 195 | .Little: |
aa8f38e4 AP |
196 | mov %al,($arg1) |
197 | sub \$1,$arg2 | |
198 | lea 1($arg1),$arg1 | |
b2dba9bf | 199 | jnz .Little |
1fd79f66 AP |
200 | .Lret: |
201 | ret | |
b2dba9bf AP |
202 | .align 16 |
203 | .Lot: | |
aa8f38e4 | 204 | test \$7,$arg1 |
b2dba9bf | 205 | jz .Laligned |
aa8f38e4 AP |
206 | mov %al,($arg1) |
207 | lea -1($arg2),$arg2 | |
208 | lea 1($arg1),$arg1 | |
b2dba9bf AP |
209 | jmp .Lot |
210 | .Laligned: | |
aa8f38e4 AP |
211 | mov %rax,($arg1) |
212 | lea -8($arg2),$arg2 | |
213 | test \$-8,$arg2 | |
214 | lea 8($arg1),$arg1 | |
b2dba9bf | 215 | jnz .Laligned |
aa8f38e4 | 216 | cmp \$0,$arg2 |
b2dba9bf AP |
217 | jne .Little |
218 | ret | |
219 | .size OPENSSL_cleanse,.-OPENSSL_cleanse | |
14e21f86 | 220 | ___ |
aa8f38e4 AP |
221 | |
222 | print<<___ if (!$win64); | |
223 | .globl OPENSSL_wipe_cpu | |
224 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | |
225 | .align 16 | |
226 | OPENSSL_wipe_cpu: | |
227 | pxor %xmm0,%xmm0 | |
228 | pxor %xmm1,%xmm1 | |
229 | pxor %xmm2,%xmm2 | |
230 | pxor %xmm3,%xmm3 | |
231 | pxor %xmm4,%xmm4 | |
232 | pxor %xmm5,%xmm5 | |
233 | pxor %xmm6,%xmm6 | |
234 | pxor %xmm7,%xmm7 | |
235 | pxor %xmm8,%xmm8 | |
236 | pxor %xmm9,%xmm9 | |
237 | pxor %xmm10,%xmm10 | |
238 | pxor %xmm11,%xmm11 | |
239 | pxor %xmm12,%xmm12 | |
240 | pxor %xmm13,%xmm13 | |
241 | pxor %xmm14,%xmm14 | |
242 | pxor %xmm15,%xmm15 | |
243 | xorq %rcx,%rcx | |
244 | xorq %rdx,%rdx | |
245 | xorq %rsi,%rsi | |
246 | xorq %rdi,%rdi | |
247 | xorq %r8,%r8 | |
248 | xorq %r9,%r9 | |
249 | xorq %r10,%r10 | |
250 | xorq %r11,%r11 | |
251 | leaq 8(%rsp),%rax | |
252 | ret | |
253 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | |
254 | ___ | |
255 | print<<___ if ($win64); | |
256 | .globl OPENSSL_wipe_cpu | |
257 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | |
258 | .align 16 | |
259 | OPENSSL_wipe_cpu: | |
260 | pxor %xmm0,%xmm0 | |
261 | pxor %xmm1,%xmm1 | |
262 | pxor %xmm2,%xmm2 | |
263 | pxor %xmm3,%xmm3 | |
264 | pxor %xmm4,%xmm4 | |
265 | pxor %xmm5,%xmm5 | |
266 | xorq %rcx,%rcx | |
267 | xorq %rdx,%rdx | |
268 | xorq %r8,%r8 | |
269 | xorq %r9,%r9 | |
270 | xorq %r10,%r10 | |
271 | xorq %r11,%r11 | |
272 | leaq 8(%rsp),%rax | |
273 | ret | |
274 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | |
275 | ___ | |
5fabb88a AP |
276 | { |
277 | my $out="%r10"; | |
278 | my $cnt="%rcx"; | |
279 | my $max="%r11"; | |
280 | my $lasttick="%r8d"; | |
281 | my $lastdiff="%r9d"; | |
282 | my $redzone=win64?8:-8; | |
283 | ||
284 | print<<___; | |
285 | .globl OPENSSL_instrument_bus | |
286 | .type OPENSSL_instrument_bus,\@abi-omnipotent | |
287 | .align 16 | |
288 | OPENSSL_instrument_bus: | |
289 | mov $arg1,$out # tribute to Win64 | |
290 | mov $arg2,$cnt | |
291 | mov $arg2,$max | |
292 | ||
293 | rdtsc # collect 1st tick | |
294 | mov %eax,$lasttick # lasttick = tick | |
295 | mov \$0,$lastdiff # lastdiff = 0 | |
296 | clflush ($out) | |
b9064221 | 297 | .byte 0xf0 # lock |
5fabb88a AP |
298 | add $lastdiff,($out) |
299 | jmp .Loop | |
300 | .align 16 | |
301 | .Loop: rdtsc | |
302 | mov %eax,%edx | |
303 | sub $lasttick,%eax | |
304 | mov %edx,$lasttick | |
305 | mov %eax,$lastdiff | |
306 | clflush ($out) | |
b9064221 | 307 | .byte 0xf0 # lock |
5fabb88a AP |
308 | add %eax,($out) |
309 | lea 4($out),$out | |
310 | sub \$1,$cnt | |
311 | jnz .Loop | |
312 | ||
313 | mov $max,%rax | |
314 | ret | |
315 | .size OPENSSL_instrument_bus,.-OPENSSL_instrument_bus | |
316 | ||
317 | .globl OPENSSL_instrument_bus2 | |
318 | .type OPENSSL_instrument_bus2,\@abi-omnipotent | |
319 | .align 16 | |
320 | OPENSSL_instrument_bus2: | |
321 | mov $arg1,$out # tribute to Win64 | |
322 | mov $arg2,$cnt | |
323 | mov $arg3,$max | |
324 | mov $cnt,$redzone(%rsp) | |
325 | ||
326 | rdtsc # collect 1st tick | |
327 | mov %eax,$lasttick # lasttick = tick | |
328 | mov \$0,$lastdiff # lastdiff = 0 | |
329 | ||
330 | clflush ($out) | |
b9064221 | 331 | .byte 0xf0 # lock |
5fabb88a AP |
332 | add $lastdiff,($out) |
333 | ||
334 | rdtsc # collect 1st diff | |
335 | mov %eax,%edx | |
336 | sub $lasttick,%eax # diff | |
337 | mov %edx,$lasttick # lasttick = tick | |
338 | mov %eax,$lastdiff # lastdiff = diff | |
339 | .Loop2: | |
340 | clflush ($out) | |
b9064221 | 341 | .byte 0xf0 # lock |
5fabb88a AP |
342 | add %eax,($out) # accumulate diff |
343 | ||
344 | sub \$1,$max | |
345 | jz .Ldone2 | |
346 | ||
347 | rdtsc | |
348 | mov %eax,%edx | |
349 | sub $lasttick,%eax # diff | |
350 | mov %edx,$lasttick # lasttick = tick | |
351 | cmp $lastdiff,%eax | |
352 | mov %eax,$lastdiff # lastdiff = diff | |
353 | mov \$0,%edx | |
354 | setne %dl | |
355 | sub %rdx,$cnt # conditional --$cnt | |
356 | lea ($out,%rdx,4),$out # conditional ++$out | |
357 | jnz .Loop2 | |
358 | ||
359 | .Ldone2: | |
360 | mov $redzone(%rsp),%rax | |
361 | sub $cnt,%rax | |
362 | ret | |
363 | .size OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2 | |
364 | ___ | |
365 | } | |
aa8f38e4 | 366 | |
301799b8 AP |
367 | print<<___; |
368 | .globl OPENSSL_ia32_rdrand | |
369 | .type OPENSSL_ia32_rdrand,\@abi-omnipotent | |
370 | .align 16 | |
371 | OPENSSL_ia32_rdrand: | |
372 | mov \$8,%ecx | |
373 | .Loop_rdrand: | |
374 | rdrand %rax | |
375 | jc .Lbreak_rdrand | |
376 | loop .Loop_rdrand | |
377 | .Lbreak_rdrand: | |
378 | cmp \$0,%rax | |
379 | cmove %rcx,%rax | |
380 | ret | |
4d01f276 | 381 | .size OPENSSL_ia32_rdrand,.-OPENSSL_ia32_rdrand |
f4d45640 AP |
382 | |
383 | .globl OPENSSL_ia32_rdseed | |
384 | .type OPENSSL_ia32_rdseed,\@abi-omnipotent | |
385 | .align 16 | |
386 | OPENSSL_ia32_rdseed: | |
387 | mov \$8,%ecx | |
388 | .Loop_rdseed: | |
389 | rdseed %rax | |
390 | jc .Lbreak_rdseed | |
391 | loop .Loop_rdseed | |
392 | .Lbreak_rdseed: | |
393 | cmp \$0,%rax | |
394 | cmove %rcx,%rax | |
395 | ret | |
396 | .size OPENSSL_ia32_rdseed,.-OPENSSL_ia32_rdseed | |
301799b8 AP |
397 | ___ |
398 | ||
5d863367 | 399 | close STDOUT; # flush |