blob: 2058640377b24a9d96223aabf62301c6790fe1e3 [file] [log] [blame]
Andy Polyakovc06b0f32005-12-15 22:50:36 +00001#include <stdio.h>
2#include <stdlib.h>
3#include <string.h>
Andy Polyakovc32fcca2010-07-01 07:34:56 +00004#include <setjmp.h>
5#include <signal.h>
Andy Polyakovc06b0f32005-12-15 22:50:36 +00006#include <sys/time.h>
Andy Polyakov68c06bf2012-11-17 10:34:11 +00007#include <unistd.h>
Andy Polyakovc06b0f32005-12-15 22:50:36 +00008#include <openssl/bn.h>
9
Andy Polyakov1fda6392012-09-23 20:29:03 +000010#include "sparc_arch.h"
Andy Polyakovc32fcca2010-07-01 07:34:56 +000011
Andy Polyakov1fda6392012-09-23 20:29:03 +000012#if defined(__GNUC__) && defined(__linux)
Matt Caswell0f113f32015-01-22 03:40:55 +000013__attribute__ ((visibility("hidden")))
Andy Polyakov1fda6392012-09-23 20:29:03 +000014#endif
Matt Caswell0f113f32015-01-22 03:40:55 +000015unsigned int OPENSSL_sparcv9cap_P[2] = { SPARCV9_TICK_PRIVILEGED, 0 };
Andy Polyakovc06b0f32005-12-15 22:50:36 +000016
Matt Caswell0f113f32015-01-22 03:40:55 +000017int 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);
Andy Polyakovc06b0f32005-12-15 22:50:36 +000026
Matt Caswell0f113f32015-01-22 03:40:55 +000027 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];
Andy Polyakov68c06bf2012-11-17 10:34:11 +000052
Matt Caswell0f113f32015-01-22 03:40:55 +000053 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 &&
63 (OPENSSL_sparcv9cap_P[0] &
64 (SPARCV9_PREFER_FPU | SPARCV9_VIS1)) ==
65 (SPARCV9_PREFER_FPU | SPARCV9_VIS1))
66 return bn_mul_mont_fpu(rp, ap, bp, np, n0, num);
67 }
68 return bn_mul_mont_int(rp, ap, bp, np, n0, num);
69}
Andy Polyakovc06b0f32005-12-15 22:50:36 +000070
Matt Caswell0f113f32015-01-22 03:40:55 +000071unsigned long _sparcv9_rdtick(void);
72void _sparcv9_vis1_probe(void);
73unsigned long _sparcv9_vis1_instrument(void);
74void _sparcv9_vis2_probe(void);
75void _sparcv9_fmadd_probe(void);
76unsigned long _sparcv9_rdcfr(void);
77void _sparcv9_vis3_probe(void);
78unsigned long _sparcv9_random(void);
79size_t _sparcv9_vis1_instrument_bus(unsigned int *, size_t);
80size_t _sparcv9_vis1_instrument_bus2(unsigned int *, size_t, size_t);
Andy Polyakovc32fcca2010-07-01 07:34:56 +000081
Andy Polyakovc06b0f32005-12-15 22:50:36 +000082unsigned long OPENSSL_rdtsc(void)
Matt Caswell0f113f32015-01-22 03:40:55 +000083{
84 if (OPENSSL_sparcv9cap_P[0] & SPARCV9_TICK_PRIVILEGED)
Andy Polyakovc06b0f32005-12-15 22:50:36 +000085#if defined(__sun) && defined(__SVR4)
Matt Caswell0f113f32015-01-22 03:40:55 +000086 return gethrtime();
Andy Polyakovc06b0f32005-12-15 22:50:36 +000087#else
Matt Caswell0f113f32015-01-22 03:40:55 +000088 return 0;
Andy Polyakovc06b0f32005-12-15 22:50:36 +000089#endif
Matt Caswell0f113f32015-01-22 03:40:55 +000090 else
91 return _sparcv9_rdtick();
92}
Andy Polyakovc06b0f32005-12-15 22:50:36 +000093
Matt Caswell0f113f32015-01-22 03:40:55 +000094size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt)
95{
96 if ((OPENSSL_sparcv9cap_P[0] & (SPARCV9_TICK_PRIVILEGED | SPARCV9_BLK)) ==
97 SPARCV9_BLK)
98 return _sparcv9_vis1_instrument_bus(out, cnt);
99 else
100 return 0;
101}
Andy Polyakov5fabb882011-04-17 12:46:00 +0000102
Matt Caswell0f113f32015-01-22 03:40:55 +0000103size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max)
104{
105 if ((OPENSSL_sparcv9cap_P[0] & (SPARCV9_TICK_PRIVILEGED | SPARCV9_BLK)) ==
106 SPARCV9_BLK)
107 return _sparcv9_vis1_instrument_bus2(out, cnt, max);
108 else
109 return 0;
110}
Andy Polyakov5fabb882011-04-17 12:46:00 +0000111
Andy Polyakovc32fcca2010-07-01 07:34:56 +0000112static sigjmp_buf common_jmp;
Matt Caswell0f113f32015-01-22 03:40:55 +0000113static void common_handler(int sig)
114{
115 siglongjmp(common_jmp, sig);
116}
Andy Polyakovc32fcca2010-07-01 07:34:56 +0000117
Andy Polyakov2238e0e2015-12-01 12:21:08 +0100118#if defined(__sun) && defined(__SVR4)
119# if defined(__GNUC__) && __GNUC__>=2
120extern unsigned int getisax(unsigned int vec[], unsigned int sz) __attribute__ ((weak));
121# elif defined(__SUNPRO_C)
122#pragma weak getisax
123extern unsigned int getisax(unsigned int vec[], unsigned int sz);
124# else
125static unsigned int (*getisax) (unsigned int vec[], unsigned int sz) = NULL;
126# endif
127#endif
128
Andy Polyakov5d7324e2005-12-18 19:13:03 +0000129void OPENSSL_cpuid_setup(void)
Matt Caswell0f113f32015-01-22 03:40:55 +0000130{
131 char *e;
132 struct sigaction common_act, ill_oact, bus_oact;
133 sigset_t all_masked, oset;
134 static int trigger = 0;
Andy Polyakov4b2603e2010-09-05 19:41:41 +0000135
Matt Caswell0f113f32015-01-22 03:40:55 +0000136 if (trigger)
137 return;
138 trigger = 1;
Andy Polyakovc06b0f32005-12-15 22:50:36 +0000139
Matt Caswell0f113f32015-01-22 03:40:55 +0000140 if ((e = getenv("OPENSSL_sparcv9cap"))) {
141 OPENSSL_sparcv9cap_P[0] = strtoul(e, NULL, 0);
142 if ((e = strchr(e, ':')))
143 OPENSSL_sparcv9cap_P[1] = strtoul(e + 1, NULL, 0);
144 return;
145 }
Andy Polyakovc32fcca2010-07-01 07:34:56 +0000146
Andy Polyakov2238e0e2015-12-01 12:21:08 +0100147#if defined(__sun) && defined(__SVR4)
148 if (getisax != NULL) {
149 unsigned int vec[1];
150
151 if (getisax (vec,1)) {
152 if (vec[0]&0x0020) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1;
153 if (vec[0]&0x0040) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2;
154 if (vec[0]&0x0080) OPENSSL_sparcv9cap_P[0] |= SPARCV9_BLK;
155 if (vec[0]&0x0100) OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD;
156 if (vec[0]&0x0400) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3;
157
158 /* reconstruct %cfr copy */
159 OPENSSL_sparcv9cap_P[1] = (vec[0]>>17)&0x3ff;
160 OPENSSL_sparcv9cap_P[1] |= (OPENSSL_sparcv9cap_P[1]&CFR_MONTMUL)<<1;
161 if (vec[0]&0x20000000) OPENSSL_sparcv9cap_P[1] |= CFR_CRC32C;
162
163 /* Some heuristics */
164 /* all known VIS2-capable CPUs have unprivileged tick counter */
165 if (OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS2)
166 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED;
167
168 OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU;
169
170 /* detect UltraSPARC-Tx, see sparccpud.S for details... */
171 if ((OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS1) &&
172 _sparcv9_vis1_instrument() >= 12)
173 OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1 | SPARCV9_PREFER_FPU);
174 }
175
176 if (sizeof(size_t) == 8)
177 OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK;
178
179 return;
180 }
181#endif
182
Matt Caswell0f113f32015-01-22 03:40:55 +0000183 /* Initial value, fits UltraSPARC-I&II... */
184 OPENSSL_sparcv9cap_P[0] = SPARCV9_PREFER_FPU | SPARCV9_TICK_PRIVILEGED;
Andy Polyakovc32fcca2010-07-01 07:34:56 +0000185
Matt Caswell0f113f32015-01-22 03:40:55 +0000186 sigfillset(&all_masked);
187 sigdelset(&all_masked, SIGILL);
188 sigdelset(&all_masked, SIGTRAP);
189# ifdef SIGEMT
190 sigdelset(&all_masked, SIGEMT);
191# endif
192 sigdelset(&all_masked, SIGFPE);
193 sigdelset(&all_masked, SIGBUS);
194 sigdelset(&all_masked, SIGSEGV);
195 sigprocmask(SIG_SETMASK, &all_masked, &oset);
Andy Polyakovc32fcca2010-07-01 07:34:56 +0000196
Matt Caswell0f113f32015-01-22 03:40:55 +0000197 memset(&common_act, 0, sizeof(common_act));
198 common_act.sa_handler = common_handler;
199 common_act.sa_mask = all_masked;
Andy Polyakov4b2603e2010-09-05 19:41:41 +0000200
Matt Caswell0f113f32015-01-22 03:40:55 +0000201 sigaction(SIGILL, &common_act, &ill_oact);
202 sigaction(SIGBUS, &common_act, &bus_oact); /* T1 fails 16-bit ldda [on
203 * Linux] */
Andy Polyakovc32fcca2010-07-01 07:34:56 +0000204
Matt Caswell0f113f32015-01-22 03:40:55 +0000205 if (sigsetjmp(common_jmp, 1) == 0) {
206 _sparcv9_rdtick();
207 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED;
208 }
Andy Polyakov4b2603e2010-09-05 19:41:41 +0000209
Matt Caswell0f113f32015-01-22 03:40:55 +0000210 if (sigsetjmp(common_jmp, 1) == 0) {
211 _sparcv9_vis1_probe();
212 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1 | SPARCV9_BLK;
213 /* detect UltraSPARC-Tx, see sparccpud.S for details... */
214 if (_sparcv9_vis1_instrument() >= 12)
215 OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1 | SPARCV9_PREFER_FPU);
216 else {
217 _sparcv9_vis2_probe();
218 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2;
219 }
220 }
Andy Polyakov1fda6392012-09-23 20:29:03 +0000221
Matt Caswell0f113f32015-01-22 03:40:55 +0000222 if (sigsetjmp(common_jmp, 1) == 0) {
223 _sparcv9_fmadd_probe();
224 OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD;
225 }
Andy Polyakov1fda6392012-09-23 20:29:03 +0000226
Matt Caswell0f113f32015-01-22 03:40:55 +0000227 /*
228 * VIS3 flag is tested independently from VIS1, unlike VIS2 that is,
229 * because VIS3 defines even integer instructions.
230 */
231 if (sigsetjmp(common_jmp, 1) == 0) {
232 _sparcv9_vis3_probe();
233 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3;
234 }
Andy Polyakov1fda6392012-09-23 20:29:03 +0000235
Matt Caswell0f113f32015-01-22 03:40:55 +0000236 /*
237 * In wait for better solution _sparcv9_rdcfr is masked by
238 * VIS3 flag, because it goes to uninterruptable endless
239 * loop on UltraSPARC II running Solaris. Things might be
240 * different on Linux...
241 */
242 if ((OPENSSL_sparcv9cap_P[0] & SPARCV9_VIS3) &&
243 sigsetjmp(common_jmp, 1) == 0) {
244 OPENSSL_sparcv9cap_P[1] = (unsigned int)_sparcv9_rdcfr();
245 }
Andy Polyakov4b2603e2010-09-05 19:41:41 +0000246
Matt Caswell0f113f32015-01-22 03:40:55 +0000247 sigaction(SIGBUS, &bus_oact, NULL);
248 sigaction(SIGILL, &ill_oact, NULL);
Andy Polyakov7c5889b2010-07-08 07:47:35 +0000249
Matt Caswell0f113f32015-01-22 03:40:55 +0000250 sigprocmask(SIG_SETMASK, &oset, NULL);
Andy Polyakov68c06bf2012-11-17 10:34:11 +0000251
Matt Caswell0f113f32015-01-22 03:40:55 +0000252 if (sizeof(size_t) == 8)
253 OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK;
254# ifdef __linux
255 else {
256 int ret = syscall(340);
Andy Polyakov68c06bf2012-11-17 10:34:11 +0000257
Matt Caswell0f113f32015-01-22 03:40:55 +0000258 if (ret >= 0 && ret & 1)
259 OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK;
260 }
261# endif
262}