]>
Commit | Line | Data |
---|---|---|
14e21f86 AP |
1 | #!/usr/bin/env perl |
2 | ||
3 | push(@INC,"perlasm"); | |
4 | require "x86asm.pl"; | |
5 | ||
6 | &asm_init($ARGV[0],"x86cpuid"); | |
7 | ||
c5022659 AP |
8 | for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); } |
9 | ||
14e21f86 AP |
10 | &function_begin("OPENSSL_ia32_cpuid"); |
11 | &xor ("edx","edx"); | |
12 | &pushf (); | |
13 | &pop ("eax"); | |
14 | &mov ("ecx","eax"); | |
15 | &xor ("eax",1<<21); | |
16 | &push ("eax"); | |
17 | &popf (); | |
18 | &pushf (); | |
19 | &pop ("eax"); | |
20 | &xor ("ecx","eax"); | |
21 | &bt ("ecx",21); | |
22 | &jnc (&label("nocpuid")); | |
162f677d AP |
23 | &xor ("eax","eax"); |
24 | &cpuid (); | |
25 | &xor ("eax","eax"); | |
26 | &cmp ("ebx",0x756e6547); # "Genu" | |
27 | &setne (&LB("eax")); | |
28 | &mov ("ebp","eax"); | |
29 | &cmp ("edx",0x49656e69); # "ineI" | |
30 | &setne (&LB("eax")); | |
31 | &or ("ebp","eax"); | |
32 | &cmp ("ecx",0x6c65746e); # "ntel" | |
33 | &setne (&LB("eax")); | |
34 | &or ("ebp","eax"); | |
14e21f86 AP |
35 | &mov ("eax",1); |
36 | &cpuid (); | |
162f677d AP |
37 | &bt ("edx",28); # test hyper-threading bit |
38 | &jnc (&label("nocpuid")); | |
39 | &cmp ("ebp",0); | |
40 | &jne (&label("notintel")); | |
41 | &or ("edx",1<<20); # use reserved bit to engage RC4_CHAR | |
42 | &set_label("notintel"); | |
43 | &shr ("ebx",16); | |
44 | &cmp (&LB("ebx"),1); # see if cache is shared(*) | |
45 | &ja (&label("nocpuid")); | |
46 | &and ("edx",~(1<<28)); # clear hyper-threading bit if not | |
14e21f86 AP |
47 | &set_label("nocpuid"); |
48 | &mov ("eax","edx"); | |
49 | &mov ("edx","ecx"); | |
50 | &function_end("OPENSSL_ia32_cpuid"); | |
162f677d AP |
51 | # (*) on Core2 this value is set to 2 denoting the fact that L2 |
52 | # cache is shared between cores. | |
14e21f86 | 53 | |
2b247cf8 | 54 | &external_label("OPENSSL_ia32cap_P"); |
14e21f86 | 55 | |
2b247cf8 | 56 | &function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD"); |
14e21f86 AP |
57 | &xor ("eax","eax"); |
58 | &xor ("edx","edx"); | |
2b247cf8 | 59 | &picmeup("ecx","OPENSSL_ia32cap_P"); |
14e21f86 AP |
60 | &bt (&DWP(0,"ecx"),4); |
61 | &jnc (&label("notsc")); | |
62 | &rdtsc (); | |
63 | &set_label("notsc"); | |
64 | &ret (); | |
65 | &function_end_B("OPENSSL_rdtsc"); | |
66 | ||
c85c5c40 AP |
67 | # This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host], |
68 | # but it's safe to call it on any [supported] 32-bit platform... | |
69 | # Just check for [non-]zero return value... | |
70 | &function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD"); | |
71 | &picmeup("ecx","OPENSSL_ia32cap_P"); | |
72 | &bt (&DWP(0,"ecx"),4); | |
73 | &jnc (&label("nohalt")); # no TSC | |
74 | ||
75 | &data_word(0x9058900e); # push %cs; pop %eax | |
76 | &and ("eax",3); | |
77 | &jnz (&label("nohalt")); # not enough privileges | |
78 | ||
79 | &pushf (); | |
80 | &pop ("eax") | |
81 | &bt ("eax",9); | |
82 | &jnc (&label("nohalt")); # interrupts are disabled | |
83 | ||
84 | &rdtsc (); | |
85 | &push ("edx"); | |
86 | &push ("eax"); | |
87 | &halt (); | |
88 | &rdtsc (); | |
89 | ||
90 | &sub ("eax",&DWP(0,"esp")); | |
91 | &sbb ("edx",&DWP(4,"esp")); | |
92 | &add ("esp",8); | |
93 | &ret (); | |
94 | ||
95 | &set_label("nohalt"); | |
96 | &xor ("eax","eax"); | |
97 | &xor ("edx","edx"); | |
98 | &ret (); | |
99 | &function_end_B("OPENSSL_instrument_halt"); | |
100 | ||
cee73df3 AP |
101 | # Essentially there is only one use for this function. Under DJGPP: |
102 | # | |
103 | # #include <go32.h> | |
104 | # ... | |
105 | # i=OPENSSL_far_spin(_dos_ds,0x46c); | |
106 | # ... | |
107 | # to obtain the number of spins till closest timer interrupt. | |
108 | ||
109 | &function_begin_B("OPENSSL_far_spin"); | |
110 | &pushf (); | |
111 | &pop ("eax") | |
112 | &bt ("eax",9); | |
113 | &jnc (&label("nospin")); # interrupts are disabled | |
114 | ||
115 | &mov ("eax",&DWP(4,"esp")); | |
116 | &mov ("ecx",&DWP(8,"esp")); | |
117 | &data_word (0x90d88e1e); # push %ds, mov %eax,%ds | |
118 | &xor ("eax","eax"); | |
119 | &mov ("edx",&DWP(0,"ecx")); | |
120 | &jmp (&label("spin")); | |
121 | ||
122 | &align (16); | |
123 | &set_label("spin"); | |
124 | &inc ("eax"); | |
125 | &cmp ("edx",&DWP(0,"ecx")); | |
126 | &je (&label("spin")); | |
127 | ||
128 | &data_word (0x1f909090); # pop %ds | |
129 | &ret (); | |
130 | ||
131 | &set_label("nospin"); | |
132 | &xor ("eax","eax"); | |
133 | &xor ("edx","edx"); | |
134 | &ret (); | |
135 | &function_end_B("OPENSSL_far_spin"); | |
136 | ||
137 | &function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD"); | |
138 | &xor ("eax","eax"); | |
139 | &xor ("edx","edx"); | |
140 | &picmeup("ecx","OPENSSL_ia32cap_P"); | |
141 | &mov ("ecx",&DWP(0,"ecx")); | |
142 | &bt (&DWP(0,"ecx"),1); | |
143 | &jnc (&label("no_x87")); | |
c5022659 AP |
144 | if ($sse2) { |
145 | &bt (&DWP(0,"ecx"),26); | |
146 | &jnc (&label("no_sse2")); | |
147 | &pxor ("xmm0","xmm0"); | |
148 | &pxor ("xmm1","xmm1"); | |
149 | &pxor ("xmm2","xmm2"); | |
150 | &pxor ("xmm3","xmm3"); | |
151 | &pxor ("xmm4","xmm4"); | |
152 | &pxor ("xmm5","xmm5"); | |
153 | &pxor ("xmm6","xmm6"); | |
154 | &pxor ("xmm7","xmm7"); | |
155 | &set_label("no_sse2"); | |
156 | } | |
c49a0aa0 AP |
157 | # just a bunch of fldz to zap the fp/mm bank followed by finit... |
158 | &data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b); | |
cee73df3 AP |
159 | &set_label("no_x87"); |
160 | &lea ("eax",&DWP(4,"esp")); | |
161 | &ret (); | |
162 | &function_end_B("OPENSSL_wipe_cpu"); | |
163 | ||
164 | &function_begin_B("OPENSSL_atomic_add"); | |
165 | &mov ("edx",&DWP(4,"esp")); # fetch the pointer, 1st arg | |
166 | &mov ("ecx",&DWP(8,"esp")); # fetch the increment, 2nd arg | |
167 | &push ("ebx"); | |
168 | &nop (); | |
169 | &mov ("eax",&DWP(0,"edx")); | |
170 | &set_label("spin"); | |
171 | &lea ("ebx",&DWP(0,"eax","ecx")); | |
172 | &nop (); | |
173 | &data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx) # %eax is envolved and is always reloaded | |
174 | &jne (&label("spin")); | |
175 | &mov ("eax","ebx"); # OpenSSL expects the new value | |
176 | &pop ("ebx"); | |
177 | &ret (); | |
178 | &function_end_B("OPENSSL_atomic_add"); | |
179 | ||
061bebc0 AP |
180 | # This function can become handy under Win32 in situations when |
181 | # we don't know which calling convention, __stdcall or __cdecl(*), | |
182 | # indirect callee is using. In C it can be deployed as | |
183 | # | |
184 | #ifdef OPENSSL_CPUID_OBJ | |
185 | # type OPENSSL_indirect_call(void *f,...); | |
186 | # ... | |
187 | # OPENSSL_indirect_call(func,[up to $max arguments]); | |
188 | #endif | |
189 | # | |
190 | # (*) it's designed to work even for __fastcall if number of | |
191 | # arguments is 1 or 2! | |
192 | &function_begin_B("OPENSSL_indirect_call"); | |
193 | { | |
194 | my $i,$max=7; # $max has to be chosen as 4*n-1 | |
195 | # in order to preserve eventual | |
196 | # stack alignment | |
197 | &push ("ebp"); | |
198 | &mov ("ebp","esp"); | |
199 | &sub ("esp",$max*4); | |
200 | &mov ("ecx",&DWP(12,"ebp")); | |
201 | &mov (&DWP(0,"esp"),"ecx"); | |
202 | &mov ("edx",&DWP(16,"ebp")); | |
203 | &mov (&DWP(4,"esp"),"edx"); | |
204 | for($i=2;$i<$max;$i++) | |
205 | { | |
206 | # Some copies will be redundant/bogus... | |
207 | &mov ("eax",&DWP(12+$i*4,"ebp")); | |
208 | &mov (&DWP(0+$i*4,"esp"),"eax"); | |
209 | } | |
3ebf898e | 210 | &call_ptr (&DWP(8,"ebp"));# make the call... |
061bebc0 AP |
211 | &mov ("esp","ebp"); # ... and just restore the stack pointer |
212 | # without paying attention to what we called, | |
213 | # (__cdecl *func) or (__stdcall *one). | |
214 | &pop ("ebp"); | |
215 | &ret (); | |
216 | } | |
217 | &function_end_B("OPENSSL_indirect_call"); | |
218 | ||
b2dba9bf AP |
219 | &function_begin_B("OPENSSL_cleanse"); |
220 | &mov ("edx",&wparam(0)); | |
221 | &mov ("ecx",&wparam(1)); | |
222 | &xor ("eax","eax"); | |
223 | &cmp ("ecx",7); | |
224 | &jae (&label("lot")); | |
225 | &set_label("little"); | |
226 | &mov (&BP(0,"edx"),"al"); | |
227 | &sub ("ecx",1); | |
228 | &lea ("edx",&DWP(1,"edx")); | |
229 | &jnz (&label("little")); | |
230 | &ret (); | |
231 | ||
232 | &set_label("lot",16); | |
233 | &test ("edx",3); | |
234 | &jz (&label("aligned")); | |
235 | &mov (&BP(0,"edx"),"al"); | |
236 | &lea ("ecx",&DWP(-1,"ecx")); | |
237 | &lea ("edx",&DWP(1,"edx")); | |
238 | &jmp (&label("lot")); | |
239 | &set_label("aligned"); | |
240 | &mov (&DWP(0,"edx"),"eax"); | |
241 | &lea ("ecx",&DWP(-4,"ecx")); | |
242 | &test ("ecx",-4); | |
243 | &lea ("edx",&DWP(4,"edx")); | |
244 | &jnz (&label("aligned")); | |
245 | &cmp ("ecx",0); | |
246 | &jne (&label("little")); | |
247 | &ret (); | |
248 | &function_end_B("OPENSSL_cleanse"); | |
249 | ||
2b247cf8 | 250 | &initseg("OPENSSL_cpuid_setup"); |
14e21f86 AP |
251 | |
252 | &asm_finish(); |