1 #!/usr/bin/env perl
2
3 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
4 push(@INC, "${dir}", "perlasm");
5 require "x86asm.pl";
6
7 &asm_init($ARGV[0],"x86cpuid");
8
9 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
10
11 &function_begin("OPENSSL_ia32_cpuid");
12 &xor ("edx","edx");
13 &pushf ();
14 &pop ("eax");
15 &mov ("ecx","eax");
16 &xor ("eax",1<<21);
17 &push ("eax");
18 &popf ();
19 &pushf ();
20 &pop ("eax");
21 &xor ("ecx","eax");
22 &xor ("eax","eax");
23 &bt ("ecx",21);
24 &jnc (&label("nocpuid"));
25 &cpuid ();
26 &mov ("edi","eax"); # max value for standard query level
27
28 &xor ("eax","eax");
29 &cmp ("ebx",0x756e6547); # "Genu"
30 &setne (&LB("eax"));
31 &mov ("ebp","eax");
32 &cmp ("edx",0x49656e69); # "ineI"
33 &setne (&LB("eax"));
34 &or ("ebp","eax");
35 &cmp ("ecx",0x6c65746e); # "ntel"
36 &setne (&LB("eax"));
37 &or ("ebp","eax"); # 0 indicates Intel CPU
38 &jz (&label("intel"));
39
40 &cmp ("ebx",0x68747541); # "Auth"
41 &setne (&LB("eax"));
42 &mov ("esi","eax");
43 &cmp ("edx",0x69746E65); # "enti"
44 &setne (&LB("eax"));
45 &or ("esi","eax");
46 &cmp ("ecx",0x444D4163); # "cAMD"
47 &setne (&LB("eax"));
48 &or ("esi","eax"); # 0 indicates AMD CPU
49 &jnz (&label("intel"));
50
51 # AMD specific
52 &mov ("eax",0x80000000);
53 &cpuid ();
54 &cmp ("eax",0x80000001);
55 &jb (&label("intel"));
56 &mov ("esi","eax");
57 &mov ("eax",0x80000001);
58 &cpuid ();
59 &or ("ebp","ecx");
60 &and ("ebp",1<<11|1); # isolate XOP bit
61 &cmp ("esi",0x80000008);
62 &jb (&label("intel"));
63
64 &mov ("eax",0x80000008);
65 &cpuid ();
66 &movz ("esi",&LB("ecx")); # number of cores - 1
67 &inc ("esi"); # number of cores
68
69 &mov ("eax",1);
70 &xor ("ecx","ecx");
71 &cpuid ();
72 &bt ("edx",28);
73 &jnc (&label("generic"));
74 &shr ("ebx",16);
75 &and ("ebx",0xff);
76 &cmp ("ebx","esi");
77 &ja (&label("generic"));
78 &and ("edx",0xefffffff); # clear hyper-threading bit
79 &jmp (&label("generic"));
80
81 &set_label("intel");
82 &cmp ("edi",4);
83 &mov ("edi",-1);
84 &jb (&label("nocacheinfo"));
85
86 &mov ("eax",4);
87 &mov ("ecx",0); # query L1D
88 &cpuid ();
89 &mov ("edi","eax");
90 &shr ("edi",14);
91 &and ("edi",0xfff); # number of cores -1 per L1D
92
93 &set_label("nocacheinfo");
94 &mov ("eax",1);
95 &xor ("ecx","ecx");
96 &cpuid ();
97 &and ("edx",0xbfefffff); # force reserved bits #20, #30 to 0
98 &cmp ("ebp",0);
99 &jne (&label("notintel"));
100 &or ("edx",1<<30); # set reserved bit#30 on Intel CPUs
101 &and (&HB("eax"),15); # familiy ID
102 &cmp (&HB("eax"),15); # P4?
103 &jne (&label("notintel"));
104 &or ("edx",1<<20); # set reserved bit#20 to engage RC4_CHAR
105 &set_label("notintel");
106 &bt ("edx",28); # test hyper-threading bit
107 &jnc (&label("generic"));
108 &and ("edx",0xefffffff);
109 &cmp ("edi",0);
110 &je (&label("generic"));
111
112 &or ("edx",0x10000000);
113 &shr ("ebx",16);
114 &cmp (&LB("ebx"),1);
115 &ja (&label("generic"));
116 &and ("edx",0xefffffff); # clear hyper-threading bit if not
117
118 &set_label("generic");
119 &and ("ebp",1<<11); # isolate AMD XOP flag
120 &and ("ecx",0xfffff7ff); # force 11th bit to 0
121 &mov ("esi","edx");
122 &or ("ebp","ecx"); # merge AMD XOP flag
123
124 &bt ("ecx",27); # check OSXSAVE bit
125 &jnc (&label("clear_avx"));
126 &xor ("ecx","ecx");
127 &data_byte(0x0f,0x01,0xd0); # xgetbv
128 &and ("eax",6);
129 &cmp ("eax",6);
130 &je (&label("done"));
131 &cmp ("eax",2);
132 &je (&label("clear_avx"));
133 &set_label("clear_xmm");
134 &and ("ebp",0xfdfffffd); # clear AESNI and PCLMULQDQ bits
135 &and ("esi",0xfeffffff); # clear FXSR
136 &set_label("clear_avx");
137 &and ("ebp",0xefffe7ff); # clear AVX, FMA and AMD XOP bits
138 &set_label("done");
139 &mov ("eax","esi");
140 &mov ("edx","ebp");
141 &set_label("nocpuid");
142 &function_end("OPENSSL_ia32_cpuid");
143
144 &external_label("OPENSSL_ia32cap_P");
145
146 &function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
147 &xor ("eax","eax");
148 &xor ("edx","edx");
149 &picmeup("ecx","OPENSSL_ia32cap_P");
150 &bt (&DWP(0,"ecx"),4);
151 &jnc (&label("notsc"));
152 &rdtsc ();
153 &set_label("notsc");
154 &ret ();
155 &function_end_B("OPENSSL_rdtsc");
156
157 # This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
158 # but it's safe to call it on any [supported] 32-bit platform...
159 # Just check for [non-]zero return value...
160 &function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
161 &picmeup("ecx","OPENSSL_ia32cap_P");
162 &bt (&DWP(0,"ecx"),4);
163 &jnc (&label("nohalt")); # no TSC
164
165 &data_word(0x9058900e); # push %cs; pop %eax
166 &and ("eax",3);
167 &jnz (&label("nohalt")); # not enough privileges
168
169 &pushf ();
170 &pop ("eax");
171 &bt ("eax",9);
172 &jnc (&label("nohalt")); # interrupts are disabled
173
174 &rdtsc ();
175 &push ("edx");
176 &push ("eax");
177 &halt ();
178 &rdtsc ();
179
180 &sub ("eax",&DWP(0,"esp"));
181 &sbb ("edx",&DWP(4,"esp"));
182 &add ("esp",8);
183 &ret ();
184
185 &set_label("nohalt");
186 &xor ("eax","eax");
187 &xor ("edx","edx");
188 &ret ();
189 &function_end_B("OPENSSL_instrument_halt");
190
191 # Essentially there is only one use for this function. Under DJGPP:
192 #
193 # #include <go32.h>
194 # ...
195 # i=OPENSSL_far_spin(_dos_ds,0x46c);
196 # ...
197 # to obtain the number of spins till closest timer interrupt.
198
199 &function_begin_B("OPENSSL_far_spin");
200 &pushf ();
201 &pop ("eax")
202 &bt ("eax",9);
203 &jnc (&label("nospin")); # interrupts are disabled
204
205 &mov ("eax",&DWP(4,"esp"));
206 &mov ("ecx",&DWP(8,"esp"));
207 &data_word (0x90d88e1e); # push %ds, mov %eax,%ds
208 &xor ("eax","eax");
209 &mov ("edx",&DWP(0,"ecx"));
210 &jmp (&label("spin"));
211
212 &align (16);
213 &set_label("spin");
214 &inc ("eax");
215 &cmp ("edx",&DWP(0,"ecx"));
216 &je (&label("spin"));
217
218 &data_word (0x1f909090); # pop %ds
219 &ret ();
220
221 &set_label("nospin");
222 &xor ("eax","eax");
223 &xor ("edx","edx");
224 &ret ();
225 &function_end_B("OPENSSL_far_spin");
226
227 &function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
228 &xor ("eax","eax");
229 &xor ("edx","edx");
230 &picmeup("ecx","OPENSSL_ia32cap_P");
231 &mov ("ecx",&DWP(0,"ecx"));
232 &bt (&DWP(0,"ecx"),1);
233 &jnc (&label("no_x87"));
234 if ($sse2) {
235 &and ("ecx",1<<26|1<<24); # check SSE2 and FXSR bits
236 &cmp ("ecx",1<<26|1<<24);
237 &jne (&label("no_sse2"));
238 &pxor ("xmm0","xmm0");
239 &pxor ("xmm1","xmm1");
240 &pxor ("xmm2","xmm2");
241 &pxor ("xmm3","xmm3");
242 &pxor ("xmm4","xmm4");
243 &pxor ("xmm5","xmm5");
244 &pxor ("xmm6","xmm6");
245 &pxor ("xmm7","xmm7");
246 &set_label("no_sse2");
247 }
248 # just a bunch of fldz to zap the fp/mm bank followed by finit...
249 &data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
250 &set_label("no_x87");
251 &lea ("eax",&DWP(4,"esp"));
252 &ret ();
253 &function_end_B("OPENSSL_wipe_cpu");
254
255 &function_begin_B("OPENSSL_atomic_add");
256 &mov ("edx",&DWP(4,"esp")); # fetch the pointer, 1st arg
257 &mov ("ecx",&DWP(8,"esp")); # fetch the increment, 2nd arg
258 &push ("ebx");
259 &nop ();
260 &mov ("eax",&DWP(0,"edx"));
261 &set_label("spin");
262 &lea ("ebx",&DWP(0,"eax","ecx"));
263 &nop ();
264 &data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx) # %eax is envolved and is always reloaded
265 &jne (&label("spin"));
266 &mov ("eax","ebx"); # OpenSSL expects the new value
267 &pop ("ebx");
268 &ret ();
269 &function_end_B("OPENSSL_atomic_add");
270
271 # This function can become handy under Win32 in situations when
272 # we don't know which calling convention, __stdcall or __cdecl(*),
273 # indirect callee is using. In C it can be deployed as
274 #
275 #ifdef OPENSSL_CPUID_OBJ
276 # type OPENSSL_indirect_call(void *f,...);
277 # ...
278 # OPENSSL_indirect_call(func,[up to $max arguments]);
279 #endif
280 #
281 # (*) it's designed to work even for __fastcall if number of
282 # arguments is 1 or 2!
283 &function_begin_B("OPENSSL_indirect_call");
284 {
285 my ($max,$i)=(7,); # $max has to be chosen as 4*n-1
286 # in order to preserve eventual
287 # stack alignment
288 &push ("ebp");
289 &mov ("ebp","esp");
290 &sub ("esp",$max*4);
291 &mov ("ecx",&DWP(12,"ebp"));
292 &mov (&DWP(0,"esp"),"ecx");
293 &mov ("edx",&DWP(16,"ebp"));
294 &mov (&DWP(4,"esp"),"edx");
295 for($i=2;$i<$max;$i++)
296 {
297 # Some copies will be redundant/bogus...
298 &mov ("eax",&DWP(12+$i*4,"ebp"));
299 &mov (&DWP(0+$i*4,"esp"),"eax");
300 }
301 &call_ptr (&DWP(8,"ebp"));# make the call...
302 &mov ("esp","ebp"); # ... and just restore the stack pointer
303 # without paying attention to what we called,
304 # (__cdecl *func) or (__stdcall *one).
305 &pop ("ebp");
306 &ret ();
307 }
308 &function_end_B("OPENSSL_indirect_call");
309
310 &function_begin_B("OPENSSL_cleanse");
311 &mov ("edx",&wparam(0));
312 &mov ("ecx",&wparam(1));
313 &xor ("eax","eax");
314 &cmp ("ecx",7);
315 &jae (&label("lot"));
316 &cmp ("ecx",0);
317 &je (&label("ret"));
318 &set_label("little");
319 &mov (&BP(0,"edx"),"al");
320 &sub ("ecx",1);
321 &lea ("edx",&DWP(1,"edx"));
322 &jnz (&label("little"));
323 &set_label("ret");
324 &ret ();
325
326 &set_label("lot",16);
327 &test ("edx",3);
328 &jz (&label("aligned"));
329 &mov (&BP(0,"edx"),"al");
330 &lea ("ecx",&DWP(-1,"ecx"));
331 &lea ("edx",&DWP(1,"edx"));
332 &jmp (&label("lot"));
333 &set_label("aligned");
334 &mov (&DWP(0,"edx"),"eax");
335 &lea ("ecx",&DWP(-4,"ecx"));
336 &test ("ecx",-4);
337 &lea ("edx",&DWP(4,"edx"));
338 &jnz (&label("aligned"));
339 &cmp ("ecx",0);
340 &jne (&label("little"));
341 &ret ();
342 &function_end_B("OPENSSL_cleanse");
343
344 &function_begin_B("OPENSSL_ia32_rdrand");
345 &mov ("ecx",8);
346 &set_label("loop");
347 &rdrand ("eax");
348 &jc (&label("break"));
349 &loop (&label("loop"));
350 &set_label("break");
351 &cmp ("eax",0);
352 &cmove ("eax","ecx");
353 &ret ();
354 &function_end_B("OPENSSL_ia32_rdrand");
355
356 &initseg("illumos_locking_setup");
357 &initseg("OPENSSL_cpuid_setup");
358
359 &hidden("illumos_locking_setup");
360 &hidden("OPENSSL_cpuid_setup");
361 &hidden("OPENSSL_ia32cap_P");
362
363 &asm_finish();