]>
Commit | Line | Data |
---|---|---|
c06b0f3d AP |
1 | #include <stdio.h> |
2 | #include <stdlib.h> | |
3 | #include <string.h> | |
c32fcca6 AP |
4 | #include <setjmp.h> |
5 | #include <signal.h> | |
c06b0f3d | 6 | #include <sys/time.h> |
68c06bf6 | 7 | #include <unistd.h> |
c06b0f3d AP |
8 | #include <openssl/bn.h> |
9 | ||
1fda639a | 10 | #include "sparc_arch.h" |
c32fcca6 | 11 | |
1fda639a | 12 | #if defined(__GNUC__) && defined(__linux) |
0f113f3e | 13 | __attribute__ ((visibility("hidden"))) |
1fda639a | 14 | #endif |
0f113f3e MC |
15 | unsigned int OPENSSL_sparcv9cap_P[2] = { SPARCV9_TICK_PRIVILEGED, 0 }; |
16 | ||
17 | int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
18 | const BN_ULONG *np, const BN_ULONG *n0, int num) | |
19 | { | |
20 | int bn_mul_mont_vis3(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
21 | const BN_ULONG *np, const BN_ULONG *n0, int num); | |
22 | int bn_mul_mont_fpu(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
23 | const BN_ULONG *np, const BN_ULONG *n0, int num); | |
24 | int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
25 | const BN_ULONG *np, const BN_ULONG *n0, int num); | |
26 | ||
27 | if (!(num & 1) && num >= 6) { | |
28 | if ((num & 15) == 0 && num <= 64 && | |
29 | (OPENSSL_sparcv9cap_P[1] & (CFR_MONTMUL | CFR_MONTSQR)) == | |
30 | (CFR_MONTMUL | CFR_MONTSQR)) { | |
31 | typedef int (*bn_mul_mont_f) (BN_ULONG *rp, const BN_ULONG *ap, | |
32 | const BN_ULONG *bp, | |
33 | const BN_ULONG *np, | |
34 | const BN_ULONG *n0); | |
35 | int bn_mul_mont_t4_8(BN_ULONG *rp, const BN_ULONG *ap, | |
36 | const BN_ULONG *bp, const BN_ULONG *np, | |
37 | const BN_ULONG *n0); | |
38 | int bn_mul_mont_t4_16(BN_ULONG *rp, const BN_ULONG *ap, | |
39 | const BN_ULONG *bp, const BN_ULONG *np, | |
40 | const BN_ULONG *n0); | |
41 | int bn_mul_mont_t4_24(BN_ULONG *rp, const BN_ULONG *ap, | |
42 | const BN_ULONG *bp, const BN_ULONG *np, | |
43 | const BN_ULONG *n0); | |
44 | int bn_mul_mont_t4_32(BN_ULONG *rp, const BN_ULONG *ap, | |
45 | const BN_ULONG *bp, const BN_ULONG *np, | |
46 | const BN_ULONG *n0); | |
47 | static const bn_mul_mont_f funcs[4] = { | |
48 | bn_mul_mont_t4_8, bn_mul_mont_t4_16, | |
49 | bn_mul_mont_t4_24, bn_mul_mont_t4_32 | |
50 | }; | |
51 | bn_mul_mont_f worker = funcs[num / 16 - 1]; | |
52 | ||
53 | if ((*worker) (rp, ap, bp, np, n0)) | |
54 | return 1; | |
55 | /* retry once and fall back */ | |
56 | if ((*worker) (rp, ap, bp, np, n0)) | |
57 | return 1; | |
58 | return bn_mul_mont_vis3(rp, ap, bp, np, n0, num); | |
59 | } | |
60 | if ((OPENSSL_sparcv9cap_P[0] & SPARCV9_VIS3)) | |
61 | return bn_mul_mont_vis3(rp, ap, bp, np, n0, num); | |
62 | else if (num >= 8 && | |
33ea23dc AP |
63 | /* |
64 | * bn_mul_mont_fpu doesn't use FMADD, we just use the | |
65 | * flag to detect when FPU path is preferable in cases | |
66 | * when current heuristics is unreliable. [it works | |
67 | * out because FMADD-capable processors where FPU | |
68 | * code path is undesirable are also VIS3-capable and | |
69 | * VIS3 code path takes precedence.] | |
70 | */ | |
71 | ( (OPENSSL_sparcv9cap_P[0] & SPARCV9_FMADD) || | |
72 | (OPENSSL_sparcv9cap_P[0] & | |
73 | (SPARCV9_PREFER_FPU | SPARCV9_VIS1)) == | |
74 | (SPARCV9_PREFER_FPU | SPARCV9_VIS1) )) | |
0f113f3e MC |
75 | return bn_mul_mont_fpu(rp, ap, bp, np, n0, num); |
76 | } | |
77 | return bn_mul_mont_int(rp, ap, bp, np, n0, num); | |
78 | } | |
79 | ||
80 | unsigned long _sparcv9_rdtick(void); | |
81 | void _sparcv9_vis1_probe(void); | |
82 | unsigned long _sparcv9_vis1_instrument(void); | |
83 | void _sparcv9_vis2_probe(void); | |
84 | void _sparcv9_fmadd_probe(void); | |
85 | unsigned long _sparcv9_rdcfr(void); | |
86 | void _sparcv9_vis3_probe(void); | |
87 | unsigned long _sparcv9_random(void); | |
88 | size_t _sparcv9_vis1_instrument_bus(unsigned int *, size_t); | |
89 | size_t _sparcv9_vis1_instrument_bus2(unsigned int *, size_t, size_t); | |
c32fcca6 | 90 | |
c06b0f3d | 91 | unsigned long OPENSSL_rdtsc(void) |
0f113f3e MC |
92 | { |
93 | if (OPENSSL_sparcv9cap_P[0] & SPARCV9_TICK_PRIVILEGED) | |
c06b0f3d | 94 | #if defined(__sun) && defined(__SVR4) |
0f113f3e | 95 | return gethrtime(); |
c06b0f3d | 96 | #else |
0f113f3e | 97 | return 0; |
c06b0f3d | 98 | #endif |
0f113f3e MC |
99 | else |
100 | return _sparcv9_rdtick(); | |
101 | } | |
102 | ||
103 | size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt) | |
104 | { | |
105 | if ((OPENSSL_sparcv9cap_P[0] & (SPARCV9_TICK_PRIVILEGED | SPARCV9_BLK)) == | |
106 | SPARCV9_BLK) | |
107 | return _sparcv9_vis1_instrument_bus(out, cnt); | |
108 | else | |
109 | return 0; | |
110 | } | |
111 | ||
112 | size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max) | |
113 | { | |
114 | if ((OPENSSL_sparcv9cap_P[0] & (SPARCV9_TICK_PRIVILEGED | SPARCV9_BLK)) == | |
115 | SPARCV9_BLK) | |
116 | return _sparcv9_vis1_instrument_bus2(out, cnt, max); | |
117 | else | |
118 | return 0; | |
119 | } | |
5fabb88a | 120 | |
c32fcca6 | 121 | static sigjmp_buf common_jmp; |
0f113f3e MC |
122 | static void common_handler(int sig) |
123 | { | |
124 | siglongjmp(common_jmp, sig); | |
125 | } | |
c32fcca6 | 126 | |
2238e0e4 AP |
127 | #if defined(__sun) && defined(__SVR4) |
128 | # if defined(__GNUC__) && __GNUC__>=2 | |
129 | extern unsigned int getisax(unsigned int vec[], unsigned int sz) __attribute__ ((weak)); | |
130 | # elif defined(__SUNPRO_C) | |
131 | #pragma weak getisax | |
132 | extern unsigned int getisax(unsigned int vec[], unsigned int sz); | |
133 | # else | |
134 | static unsigned int (*getisax) (unsigned int vec[], unsigned int sz) = NULL; | |
135 | # endif | |
136 | #endif | |
137 | ||
5d7324e4 | 138 | void OPENSSL_cpuid_setup(void) |
0f113f3e MC |
139 | { |
140 | char *e; | |
141 | struct sigaction common_act, ill_oact, bus_oact; | |
142 | sigset_t all_masked, oset; | |
143 | static int trigger = 0; | |
144 | ||
145 | if (trigger) | |
146 | return; | |
147 | trigger = 1; | |
148 | ||
149 | if ((e = getenv("OPENSSL_sparcv9cap"))) { | |
150 | OPENSSL_sparcv9cap_P[0] = strtoul(e, NULL, 0); | |
151 | if ((e = strchr(e, ':'))) | |
152 | OPENSSL_sparcv9cap_P[1] = strtoul(e + 1, NULL, 0); | |
153 | return; | |
154 | } | |
155 | ||
2238e0e4 AP |
156 | #if defined(__sun) && defined(__SVR4) |
157 | if (getisax != NULL) { | |
158 | unsigned int vec[1]; | |
159 | ||
160 | if (getisax (vec,1)) { | |
299ccadc AP |
161 | if (vec[0]&0x00020) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1; |
162 | if (vec[0]&0x00040) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2; | |
163 | if (vec[0]&0x00080) OPENSSL_sparcv9cap_P[0] |= SPARCV9_BLK; | |
164 | if (vec[0]&0x00100) OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD; | |
165 | if (vec[0]&0x00400) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3; | |
166 | if (vec[0]&0x01000) OPENSSL_sparcv9cap_P[0] |= SPARCV9_FJHPCACE; | |
167 | if (vec[0]&0x02000) OPENSSL_sparcv9cap_P[0] |= SPARCV9_FJDESX; | |
168 | if (vec[0]&0x08000) OPENSSL_sparcv9cap_P[0] |= SPARCV9_IMA; | |
4400f6c6 | 169 | if (vec[0]&0x10000) OPENSSL_sparcv9cap_P[0] |= SPARCV9_FJAESX; |
299ccadc | 170 | if (vec[1]&0x00008) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS4; |
2238e0e4 AP |
171 | |
172 | /* reconstruct %cfr copy */ | |
173 | OPENSSL_sparcv9cap_P[1] = (vec[0]>>17)&0x3ff; | |
174 | OPENSSL_sparcv9cap_P[1] |= (OPENSSL_sparcv9cap_P[1]&CFR_MONTMUL)<<1; | |
175 | if (vec[0]&0x20000000) OPENSSL_sparcv9cap_P[1] |= CFR_CRC32C; | |
299ccadc AP |
176 | if (vec[1]&0x00000020) OPENSSL_sparcv9cap_P[1] |= CFR_XMPMUL; |
177 | if (vec[1]&0x00000040) | |
178 | OPENSSL_sparcv9cap_P[1] |= CFR_XMONTMUL|CFR_XMONTSQR; | |
2238e0e4 AP |
179 | |
180 | /* Some heuristics */ | |
181 | /* all known VIS2-capable CPUs have unprivileged tick counter */ | |
182 | if (OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS2) | |
183 | OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; | |
184 | ||
185 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU; | |
186 | ||
187 | /* detect UltraSPARC-Tx, see sparccpud.S for details... */ | |
188 | if ((OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS1) && | |
189 | _sparcv9_vis1_instrument() >= 12) | |
190 | OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1 | SPARCV9_PREFER_FPU); | |
191 | } | |
192 | ||
193 | if (sizeof(size_t) == 8) | |
194 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK; | |
195 | ||
196 | return; | |
197 | } | |
198 | #endif | |
199 | ||
0f113f3e MC |
200 | /* Initial value, fits UltraSPARC-I&II... */ |
201 | OPENSSL_sparcv9cap_P[0] = SPARCV9_PREFER_FPU | SPARCV9_TICK_PRIVILEGED; | |
202 | ||
203 | sigfillset(&all_masked); | |
204 | sigdelset(&all_masked, SIGILL); | |
205 | sigdelset(&all_masked, SIGTRAP); | |
206 | # ifdef SIGEMT | |
207 | sigdelset(&all_masked, SIGEMT); | |
208 | # endif | |
209 | sigdelset(&all_masked, SIGFPE); | |
210 | sigdelset(&all_masked, SIGBUS); | |
211 | sigdelset(&all_masked, SIGSEGV); | |
212 | sigprocmask(SIG_SETMASK, &all_masked, &oset); | |
213 | ||
214 | memset(&common_act, 0, sizeof(common_act)); | |
215 | common_act.sa_handler = common_handler; | |
216 | common_act.sa_mask = all_masked; | |
217 | ||
218 | sigaction(SIGILL, &common_act, &ill_oact); | |
219 | sigaction(SIGBUS, &common_act, &bus_oact); /* T1 fails 16-bit ldda [on | |
220 | * Linux] */ | |
221 | ||
222 | if (sigsetjmp(common_jmp, 1) == 0) { | |
223 | _sparcv9_rdtick(); | |
224 | OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; | |
225 | } | |
226 | ||
227 | if (sigsetjmp(common_jmp, 1) == 0) { | |
228 | _sparcv9_vis1_probe(); | |
229 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1 | SPARCV9_BLK; | |
230 | /* detect UltraSPARC-Tx, see sparccpud.S for details... */ | |
231 | if (_sparcv9_vis1_instrument() >= 12) | |
232 | OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1 | SPARCV9_PREFER_FPU); | |
233 | else { | |
234 | _sparcv9_vis2_probe(); | |
235 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2; | |
236 | } | |
237 | } | |
238 | ||
239 | if (sigsetjmp(common_jmp, 1) == 0) { | |
240 | _sparcv9_fmadd_probe(); | |
241 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD; | |
242 | } | |
243 | ||
244 | /* | |
245 | * VIS3 flag is tested independently from VIS1, unlike VIS2 that is, | |
246 | * because VIS3 defines even integer instructions. | |
247 | */ | |
248 | if (sigsetjmp(common_jmp, 1) == 0) { | |
249 | _sparcv9_vis3_probe(); | |
250 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3; | |
251 | } | |
0f113f3e | 252 | |
4400f6c6 AP |
253 | if (sigsetjmp(common_jmp, 1) == 0) { |
254 | _sparcv9_fjaesx_probe(); | |
255 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_FJAESX; | |
256 | } | |
257 | ||
0f113f3e MC |
258 | /* |
259 | * In wait for better solution _sparcv9_rdcfr is masked by | |
260 | * VIS3 flag, because it goes to uninterruptable endless | |
261 | * loop on UltraSPARC II running Solaris. Things might be | |
262 | * different on Linux... | |
263 | */ | |
264 | if ((OPENSSL_sparcv9cap_P[0] & SPARCV9_VIS3) && | |
265 | sigsetjmp(common_jmp, 1) == 0) { | |
266 | OPENSSL_sparcv9cap_P[1] = (unsigned int)_sparcv9_rdcfr(); | |
267 | } | |
268 | ||
269 | sigaction(SIGBUS, &bus_oact, NULL); | |
270 | sigaction(SIGILL, &ill_oact, NULL); | |
271 | ||
272 | sigprocmask(SIG_SETMASK, &oset, NULL); | |
273 | ||
274 | if (sizeof(size_t) == 8) | |
275 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK; | |
276 | # ifdef __linux | |
277 | else { | |
278 | int ret = syscall(340); | |
279 | ||
280 | if (ret >= 0 && ret & 1) | |
281 | OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK; | |
282 | } | |
283 | # endif | |
284 | } |