xref: /openssl/crypto/x86cpuid.pl (revision 33388b44)
1#! /usr/bin/env perl
2# Copyright 2004-2020 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$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
10push(@INC, "${dir}perlasm", "perlasm");
11require "x86asm.pl";
12
13$output = pop and open STDOUT,">$output";
14
15&asm_init($ARGV[0]);
16
17for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
18
19&function_begin("OPENSSL_ia32_cpuid");
20	&xor	("edx","edx");
21	&pushf	();
22	&pop	("eax");
23	&mov	("ecx","eax");
24	&xor	("eax",1<<21);
25	&push	("eax");
26	&popf	();
27	&pushf	();
28	&pop	("eax");
29	&xor	("ecx","eax");
30	&xor	("eax","eax");
31	&mov	("esi",&wparam(0));
32	&mov	(&DWP(8,"esi"),"eax");	# clear extended feature flags
33	&bt	("ecx",21);
34	&jnc	(&label("nocpuid"));
35	&cpuid	();
36	&mov	("edi","eax");		# max value for standard query level
37
38	&xor	("eax","eax");
39	&cmp	("ebx",0x756e6547);	# "Genu"
40	&setne	(&LB("eax"));
41	&mov	("ebp","eax");
42	&cmp	("edx",0x49656e69);	# "ineI"
43	&setne	(&LB("eax"));
44	&or	("ebp","eax");
45	&cmp	("ecx",0x6c65746e);	# "ntel"
46	&setne	(&LB("eax"));
47	&or	("ebp","eax");		# 0 indicates Intel CPU
48	&jz	(&label("intel"));
49
50	&cmp	("ebx",0x68747541);	# "Auth"
51	&setne	(&LB("eax"));
52	&mov	("esi","eax");
53	&cmp	("edx",0x69746E65);	# "enti"
54	&setne	(&LB("eax"));
55	&or	("esi","eax");
56	&cmp	("ecx",0x444D4163);	# "cAMD"
57	&setne	(&LB("eax"));
58	&or	("esi","eax");		# 0 indicates AMD CPU
59	&jnz	(&label("intel"));
60
61	# AMD specific
62	&mov	("eax",0x80000000);
63	&cpuid	();
64	&cmp	("eax",0x80000001);
65	&jb	(&label("intel"));
66	&mov	("esi","eax");
67	&mov	("eax",0x80000001);
68	&cpuid	();
69	&or	("ebp","ecx");
70	&and	("ebp",1<<11|1);	# isolate XOP bit
71	&cmp	("esi",0x80000008);
72	&jb	(&label("intel"));
73
74	&mov	("eax",0x80000008);
75	&cpuid	();
76	&movz	("esi",&LB("ecx"));	# number of cores - 1
77	&inc	("esi");		# number of cores
78
79	&mov	("eax",1);
80	&xor	("ecx","ecx");
81	&cpuid	();
82	&bt	("edx",28);
83	&jnc	(&label("generic"));
84	&shr	("ebx",16);
85	&and	("ebx",0xff);
86	&cmp	("ebx","esi");
87	&ja	(&label("generic"));
88	&and	("edx",0xefffffff);	# clear hyper-threading bit
89	&jmp	(&label("generic"));
90
91&set_label("intel");
92	&cmp	("edi",4);
93	&mov	("esi",-1);
94	&jb	(&label("nocacheinfo"));
95
96	&mov	("eax",4);
97	&mov	("ecx",0);		# query L1D
98	&cpuid	();
99	&mov	("esi","eax");
100	&shr	("esi",14);
101	&and	("esi",0xfff);		# number of cores -1 per L1D
102
103&set_label("nocacheinfo");
104	&mov	("eax",1);
105	&xor	("ecx","ecx");
106	&cpuid	();
107	&and	("edx",0xbfefffff);	# force reserved bits #20, #30 to 0
108	&cmp	("ebp",0);
109	&jne	(&label("notintel"));
110	&or	("edx",1<<30);		# set reserved bit#30 on Intel CPUs
111	&and	(&HB("eax"),15);	# family ID
112	&cmp	(&HB("eax"),15);	# P4?
113	&jne	(&label("notintel"));
114	&or	("edx",1<<20);		# set reserved bit#20 to engage RC4_CHAR
115&set_label("notintel");
116	&bt	("edx",28);		# test hyper-threading bit
117	&jnc	(&label("generic"));
118	&and	("edx",0xefffffff);
119	&cmp	("esi",0);
120	&je	(&label("generic"));
121
122	&or	("edx",0x10000000);
123	&shr	("ebx",16);
124	&cmp	(&LB("ebx"),1);
125	&ja	(&label("generic"));
126	&and	("edx",0xefffffff);	# clear hyper-threading bit if not
127
128&set_label("generic");
129	&and	("ebp",1<<11);		# isolate AMD XOP flag
130	&and	("ecx",0xfffff7ff);	# force 11th bit to 0
131	&mov	("esi","edx");		# %ebp:%esi is copy of %ecx:%edx
132	&or	("ebp","ecx");		# merge AMD XOP flag
133
134	&cmp	("edi",7);
135	&mov	("edi",&wparam(0));
136	&jb	(&label("no_extended_info"));
137	&mov	("eax",7);
138	&xor	("ecx","ecx");
139	&cpuid	();
140	&mov	(&DWP(8,"edi"),"ebx");	# save extended feature flag
141&set_label("no_extended_info");
142
143	&bt	("ebp",27);		# check OSXSAVE bit
144	&jnc	(&label("clear_avx"));
145	&xor	("ecx","ecx");
146	&data_byte(0x0f,0x01,0xd0);	# xgetbv
147	&and	("eax",6);
148	&cmp	("eax",6);
149	&je	(&label("done"));
150	&cmp	("eax",2);
151	&je	(&label("clear_avx"));
152&set_label("clear_xmm");
153	&and	("ebp",0xfdfffffd);	# clear AESNI and PCLMULQDQ bits
154	&and	("esi",0xfeffffff);	# clear FXSR
155&set_label("clear_avx");
156	&and	("ebp",0xefffe7ff);	# clear AVX, FMA and AMD XOP bits
157	&and	(&DWP(8,"edi"),0xffffffdf);	# clear AVX2
158&set_label("done");
159	&mov	("eax","esi");
160	&mov	("edx","ebp");
161&set_label("nocpuid");
162&function_end("OPENSSL_ia32_cpuid");
163
164&external_label("OPENSSL_ia32cap_P");
165
166&function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
167	&xor	("eax","eax");
168	&xor	("edx","edx");
169	&picmeup("ecx","OPENSSL_ia32cap_P");
170	&bt	(&DWP(0,"ecx"),4);
171	&jnc	(&label("notsc"));
172	&rdtsc	();
173&set_label("notsc");
174	&ret	();
175&function_end_B("OPENSSL_rdtsc");
176
177# This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
178# but it's safe to call it on any [supported] 32-bit platform...
179# Just check for [non-]zero return value...
180&function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
181	&picmeup("ecx","OPENSSL_ia32cap_P");
182	&bt	(&DWP(0,"ecx"),4);
183	&jnc	(&label("nohalt"));	# no TSC
184
185	&data_word(0x9058900e);		# push %cs; pop %eax
186	&and	("eax",3);
187	&jnz	(&label("nohalt"));	# not enough privileges
188
189	&pushf	();
190	&pop	("eax");
191	&bt	("eax",9);
192	&jnc	(&label("nohalt"));	# interrupts are disabled
193
194	&rdtsc	();
195	&push	("edx");
196	&push	("eax");
197	&halt	();
198	&rdtsc	();
199
200	&sub	("eax",&DWP(0,"esp"));
201	&sbb	("edx",&DWP(4,"esp"));
202	&add	("esp",8);
203	&ret	();
204
205&set_label("nohalt");
206	&xor	("eax","eax");
207	&xor	("edx","edx");
208	&ret	();
209&function_end_B("OPENSSL_instrument_halt");
210
211# Essentially there is only one use for this function. Under DJGPP:
212#
213#	#include <go32.h>
214#	...
215#	i=OPENSSL_far_spin(_dos_ds,0x46c);
216#	...
217# to obtain the number of spins till closest timer interrupt.
218
219&function_begin_B("OPENSSL_far_spin");
220	&pushf	();
221	&pop	("eax");
222	&bt	("eax",9);
223	&jnc	(&label("nospin"));	# interrupts are disabled
224
225	&mov	("eax",&DWP(4,"esp"));
226	&mov	("ecx",&DWP(8,"esp"));
227	&data_word (0x90d88e1e);	# push %ds, mov %eax,%ds
228	&xor	("eax","eax");
229	&mov	("edx",&DWP(0,"ecx"));
230	&jmp	(&label("spin"));
231
232	&align	(16);
233&set_label("spin");
234	&inc	("eax");
235	&cmp	("edx",&DWP(0,"ecx"));
236	&je	(&label("spin"));
237
238	&data_word (0x1f909090);	# pop	%ds
239	&ret	();
240
241&set_label("nospin");
242	&xor	("eax","eax");
243	&xor	("edx","edx");
244	&ret	();
245&function_end_B("OPENSSL_far_spin");
246
247&function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
248	&xor	("eax","eax");
249	&xor	("edx","edx");
250	&picmeup("ecx","OPENSSL_ia32cap_P");
251	&mov	("ecx",&DWP(0,"ecx"));
252	&bt	(&DWP(0,"ecx"),1);
253	&jnc	(&label("no_x87"));
254	if ($sse2) {
255		&and	("ecx",1<<26|1<<24);	# check SSE2 and FXSR bits
256		&cmp	("ecx",1<<26|1<<24);
257		&jne	(&label("no_sse2"));
258		&pxor	("xmm0","xmm0");
259		&pxor	("xmm1","xmm1");
260		&pxor	("xmm2","xmm2");
261		&pxor	("xmm3","xmm3");
262		&pxor	("xmm4","xmm4");
263		&pxor	("xmm5","xmm5");
264		&pxor	("xmm6","xmm6");
265		&pxor	("xmm7","xmm7");
266	&set_label("no_sse2");
267	}
268	# just a bunch of fldz to zap the fp/mm bank followed by finit...
269	&data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
270&set_label("no_x87");
271	&lea	("eax",&DWP(4,"esp"));
272	&ret	();
273&function_end_B("OPENSSL_wipe_cpu");
274
275&function_begin_B("OPENSSL_atomic_add");
276	&mov	("edx",&DWP(4,"esp"));	# fetch the pointer, 1st arg
277	&mov	("ecx",&DWP(8,"esp"));	# fetch the increment, 2nd arg
278	&push	("ebx");
279	&nop	();
280	&mov	("eax",&DWP(0,"edx"));
281&set_label("spin");
282	&lea	("ebx",&DWP(0,"eax","ecx"));
283	&nop	();
284	&data_word(0x1ab10ff0);	# lock;	cmpxchg	%ebx,(%edx)	# %eax is involved and is always reloaded
285	&jne	(&label("spin"));
286	&mov	("eax","ebx");	# OpenSSL expects the new value
287	&pop	("ebx");
288	&ret	();
289&function_end_B("OPENSSL_atomic_add");
290
291&function_begin_B("OPENSSL_cleanse");
292	&mov	("edx",&wparam(0));
293	&mov	("ecx",&wparam(1));
294	&xor	("eax","eax");
295	&cmp	("ecx",7);
296	&jae	(&label("lot"));
297	&cmp	("ecx",0);
298	&je	(&label("ret"));
299&set_label("little");
300	&mov	(&BP(0,"edx"),"al");
301	&sub	("ecx",1);
302	&lea	("edx",&DWP(1,"edx"));
303	&jnz	(&label("little"));
304&set_label("ret");
305	&ret	();
306
307&set_label("lot",16);
308	&test	("edx",3);
309	&jz	(&label("aligned"));
310	&mov	(&BP(0,"edx"),"al");
311	&lea	("ecx",&DWP(-1,"ecx"));
312	&lea	("edx",&DWP(1,"edx"));
313	&jmp	(&label("lot"));
314&set_label("aligned");
315	&mov	(&DWP(0,"edx"),"eax");
316	&lea	("ecx",&DWP(-4,"ecx"));
317	&test	("ecx",-4);
318	&lea	("edx",&DWP(4,"edx"));
319	&jnz	(&label("aligned"));
320	&cmp	("ecx",0);
321	&jne	(&label("little"));
322	&ret	();
323&function_end_B("OPENSSL_cleanse");
324
325&function_begin_B("CRYPTO_memcmp");
326	&push	("esi");
327	&push	("edi");
328	&mov	("esi",&wparam(0));
329	&mov	("edi",&wparam(1));
330	&mov	("ecx",&wparam(2));
331	&xor	("eax","eax");
332	&xor	("edx","edx");
333	&cmp	("ecx",0);
334	&je	(&label("no_data"));
335&set_label("loop");
336	&mov	("dl",&BP(0,"esi"));
337	&lea	("esi",&DWP(1,"esi"));
338	&xor	("dl",&BP(0,"edi"));
339	&lea	("edi",&DWP(1,"edi"));
340	&or	("al","dl");
341	&dec	("ecx");
342	&jnz	(&label("loop"));
343	&neg	("eax");
344	&shr	("eax",31);
345&set_label("no_data");
346	&pop	("edi");
347	&pop	("esi");
348	&ret	();
349&function_end_B("CRYPTO_memcmp");
350{
351my $lasttick = "esi";
352my $lastdiff = "ebx";
353my $out = "edi";
354my $cnt = "ecx";
355my $max = "ebp";
356
357&function_begin("OPENSSL_instrument_bus");
358    &mov	("eax",0);
359    if ($sse2) {
360	&picmeup("edx","OPENSSL_ia32cap_P");
361	&bt	(&DWP(0,"edx"),4);
362	&jnc	(&label("nogo"));	# no TSC
363	&bt	(&DWP(0,"edx"),19);
364	&jnc	(&label("nogo"));	# no CLFLUSH
365
366	&mov	($out,&wparam(0));	# load arguments
367	&mov	($cnt,&wparam(1));
368
369	# collect 1st tick
370	&rdtsc	();
371	&mov	($lasttick,"eax");	# lasttick = tick
372	&mov	($lastdiff,0);		# lastdiff = 0
373	&clflush(&DWP(0,$out));
374	&data_byte(0xf0);		# lock
375	&add	(&DWP(0,$out),$lastdiff);
376	&jmp	(&label("loop"));
377
378&set_label("loop",16);
379	&rdtsc	();
380	&mov	("edx","eax");		# put aside tick (yes, I neglect edx)
381	&sub	("eax",$lasttick);	# diff
382	&mov	($lasttick,"edx");	# lasttick = tick
383	&mov	($lastdiff,"eax");	# lastdiff = diff
384	&clflush(&DWP(0,$out));
385	&data_byte(0xf0);		# lock
386	&add	(&DWP(0,$out),"eax");	# accumulate diff
387	&lea	($out,&DWP(4,$out));	# ++$out
388	&sub	($cnt,1);		# --$cnt
389	&jnz	(&label("loop"));
390
391	&mov	("eax",&wparam(1));
392&set_label("nogo");
393    }
394&function_end("OPENSSL_instrument_bus");
395
396&function_begin("OPENSSL_instrument_bus2");
397    &mov	("eax",0);
398    if ($sse2) {
399	&picmeup("edx","OPENSSL_ia32cap_P");
400	&bt	(&DWP(0,"edx"),4);
401	&jnc	(&label("nogo"));	# no TSC
402	&bt	(&DWP(0,"edx"),19);
403	&jnc	(&label("nogo"));	# no CLFLUSH
404
405	&mov	($out,&wparam(0));	# load arguments
406	&mov	($cnt,&wparam(1));
407	&mov	($max,&wparam(2));
408
409	&rdtsc	();			# collect 1st tick
410	&mov	($lasttick,"eax");	# lasttick = tick
411	&mov	($lastdiff,0);		# lastdiff = 0
412
413	&clflush(&DWP(0,$out));
414	&data_byte(0xf0);		# lock
415	&add	(&DWP(0,$out),$lastdiff);
416
417	&rdtsc	();			# collect 1st diff
418	&mov	("edx","eax");		# put aside tick (yes, I neglect edx)
419	&sub	("eax",$lasttick);	# diff
420	&mov	($lasttick,"edx");	# lasttick = tick
421	&mov	($lastdiff,"eax");	# lastdiff = diff
422	&jmp	(&label("loop2"));
423
424&set_label("loop2",16);
425	&clflush(&DWP(0,$out));
426	&data_byte(0xf0);		# lock
427	&add	(&DWP(0,$out),"eax");	# accumulate diff
428
429	&sub	($max,1);
430	&jz	(&label("done2"));
431
432	&rdtsc	();
433	&mov	("edx","eax");		# put aside tick (yes, I neglect edx)
434	&sub	("eax",$lasttick);	# diff
435	&mov	($lasttick,"edx");	# lasttick = tick
436	&cmp	("eax",$lastdiff);
437	&mov	($lastdiff,"eax");	# lastdiff = diff
438	&mov	("edx",0);
439	&setne	("dl");
440	&sub	($cnt,"edx");		# conditional --$cnt
441	&lea	($out,&DWP(0,$out,"edx",4));	# conditional ++$out
442	&jnz	(&label("loop2"));
443
444&set_label("done2");
445	&mov	("eax",&wparam(1));
446	&sub	("eax",$cnt);
447&set_label("nogo");
448    }
449&function_end("OPENSSL_instrument_bus2");
450}
451
452sub gen_random {
453my $rdop = shift;
454&function_begin_B("OPENSSL_ia32_${rdop}_bytes");
455	&push	("edi");
456	&push	("ebx");
457	&xor	("eax","eax");		# return value
458	&mov	("edi",&wparam(0));
459	&mov	("ebx",&wparam(1));
460
461	&cmp	("ebx",0);
462	&je	(&label("done"));
463
464	&mov	("ecx",8);
465&set_label("loop");
466	&${rdop}("edx");
467	&jc	(&label("break"));
468	&loop	(&label("loop"));
469	&jmp	(&label("done"));
470
471&set_label("break",16);
472	&cmp	("ebx",4);
473	&jb	(&label("tail"));
474	&mov	(&DWP(0,"edi"),"edx");
475	&lea	("edi",&DWP(4,"edi"));
476	&add	("eax",4);
477	&sub	("ebx",4);
478	&jz	(&label("done"));
479	&mov	("ecx",8);
480	&jmp	(&label("loop"));
481
482&set_label("tail",16);
483	&mov	(&BP(0,"edi"),"dl");
484	&lea	("edi",&DWP(1,"edi"));
485	&inc	("eax");
486	&shr	("edx",8);
487	&dec	("ebx");
488	&jnz	(&label("tail"));
489
490&set_label("done");
491	&xor	("edx","edx");		# Clear random value from registers
492	&pop	("ebx");
493	&pop	("edi");
494	&ret	();
495&function_end_B("OPENSSL_ia32_${rdop}_bytes");
496}
497&gen_random("rdrand");
498&gen_random("rdseed");
499
500&initseg("OPENSSL_cpuid_setup");
501
502&hidden("OPENSSL_cpuid_setup");
503&hidden("OPENSSL_ia32cap_P");
504
505&asm_finish();
506
507close STDOUT or die "error closing STDOUT: $!";
508