3 * CPU-specific dispatch
5 * (c) 2015 Straylight/Edgeware
8 /*----- Licensing notice --------------------------------------------------*
10 * This file is part of Catacomb.
12 * Catacomb is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU Library General Public License as
14 * published by the Free Software Foundation; either version 2 of the
15 * License, or (at your option) any later version.
17 * Catacomb is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU Library General Public License for more details.
22 * You should have received a copy of the GNU Library General Public
23 * License along with Catacomb; if not, write to the Free
24 * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
28 /*----- Header files ------------------------------------------------------*/
38 #include <mLib/macros.h>
42 /*----- Intel x86/AMD64 feature probing -----------------------------------*/
44 #if CPUFAM_X86 || CPUFAM_AMD64
46 # define CPUID1D_SSE2 (1u << 26)
47 # define CPUID1D_FXSR (1u << 24)
48 # define CPUID1C_PCLMUL (1u << 1)
49 # define CPUID1C_SSSE3 (1u << 9)
50 # define CPUID1C_AESNI (1u << 25)
51 # define CPUID1C_AVX (1u << 28)
52 # define CPUID1C_RDRAND (1u << 30)
54 struct cpuid
{ unsigned a
, b
, c
, d
; };
55 extern int dispatch_x86ish_cpuid(struct cpuid
*, unsigned a
, unsigned c
);
56 extern int dispatch_x86ish_xmmregisters_p(void);
58 static void cpuid(struct cpuid
*cc
, unsigned a
, unsigned c
)
60 int rc
= dispatch_x86ish_cpuid(cc
, a
, c
);
62 dispatch_debug("CPUID instruction not available");
64 dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
65 a
, c
, cc
->a
, cc
->b
, cc
->c
, cc
->d
);
68 static unsigned cpuid_maxleaf(void)
69 { struct cpuid c
; cpuid(&c
, 0, 0); return (c
.a
); }
71 /* --- @cpuid_features_p@ --- *
73 * Arguments: @unsigned dbits@ = bits to check in EDX
74 * @unsigned cbits@ = bits to check in ECX
76 * Returns: Nonzero if all the requested bits are set in the CPUID result
80 static int cpuid_features_p(unsigned dbits
, unsigned cbits
)
83 if (cpuid_maxleaf() < 1) return (0);
85 return ((c
.d
& dbits
) == dbits
&& (c
.c
& cbits
) == cbits
);
88 /* --- @xmm_registers_available_p@ --- *
92 * Returns: Nonzero if the operating system has made the XMM registers
96 static int xmm_registers_available_p(void)
98 int f
= dispatch_x86ish_xmmregisters_p();
100 dispatch_debug("XMM registers %savailable", f ?
"" : "not ");
106 /*----- General feature probing using auxiliary vectors -------------------*/
108 /* Try to find the system's definitions for auxiliary vector entries. */
109 #ifdef HAVE_SYS_AUXV_H
110 # include <sys/auxv.h>
112 #ifdef HAVE_LINUX_AUXVEC_H
113 # include <linux/auxvec.h>
115 #ifdef HAVE_ASM_HWCAP_H
116 # include <asm/hwcap.h>
119 /* The type of entries in the auxiliary vector. I'm assuming that `unsigned
120 * long' matches each platform's word length; if this is false then we'll
121 * need some host-specific tweaking here.
123 union auxval
{ long i
; unsigned long u
; const void *p
; };
124 struct auxentry
{ unsigned long type
; union auxval value
; };
126 /* Register each CPU family's interest in the auxiliary vector. Make sure
127 * that the necessary entry types are defined. This is primarily ordered by
128 * entry type to minimize duplication.
130 #if defined(AT_HWCAP) && CPUFAM_ARMEL
132 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
135 #if defined(AT_HWCAP) && CPUFAM_ARM64
137 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
140 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
142 # define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
145 /* If we couldn't find any interesting entries then we can switch all of this
146 * machinery off. Also do that if we have no means for atomic updates.
148 #if WANT_ANY && CPU_DISPATCH_P
150 /* The main output of this section is a bitmask of detected features. The
151 * least significant bit will be set if we've tried to probe. Always access
152 * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
154 static unsigned hwcaps
= 0;
156 /* For each potentially interesting type which turned out not to exist or be
157 * wanted, define a dummy macro for the sake of the next step.
159 #ifndef WANT_AT_HWCAP
160 # define WANT_AT_HWCAP(_)
162 #ifndef WANT_AT_HWCAP2
163 # define WANT_AT_HWCAP2(_)
166 /* For each CPU family, define two lists.
168 * * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
169 * family tried to register interest in above. Each entry contains the
170 * interesting auxiliary vector entry type, the name of the union branch
171 * for its value, and the name of the slot in `struct auxprobe' in which
172 * to store the value.
174 * * `CAPMAP' is a list describing the output features which the CPU family
175 * intends to satisfy from the auxiliary vector. Each entry contains a
176 * feature name suffix, and the token name (for `check_env').
179 # define WANTAUX(_) \
183 _(ARM_VFP, "arm:vfp") \
184 _(ARM_NEON, "arm:neon") \
185 _(ARM_V4, "arm:v4") \
186 _(ARM_D32, "arm:d32") \
187 _(ARM_AES, "arm:aes") \
188 _(ARM_PMULL, "arm:pmull")
191 # define WANTAUX(_) \
194 _(ARM_AES, "arm:aes") \
195 _(ARM_PMULL, "arm:pmull")
198 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
201 #define HFI__ENUM(feat, tok) HFI_##feat,
208 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
214 /* Build a structure in which we can capture the interesting data from the
217 #define AUXUTYPE_i long
218 #define AUXUTYPE_u unsigned long
219 #define AUXUTYPE_p const void *
221 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
222 WANTAUX(AUXPROBE__SLOT
)
226 /* --- @probe_hwcaps@ --- *
232 * Use: Attempt to find the auxiliary vector (which is well hidden)
233 * and discover interesting features from it.
236 static void probe_hwcaps(void)
238 unsigned hw
= HF_PROBED
;
239 struct auxprobe probed
= { 0 };
241 /* Populate `probed' with the information we manage to retrieve from the
242 * auxiliary vector. Slots we couldn't find are left zero-valued.
244 #if defined(HAVE_GETAUXVAL)
245 /* Shiny new libc lets us request individual entry types. This is almost
248 # define CAP__GET(type, ubranch, slot) \
249 probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
252 /* Otherwise we're a bit stuck, really. Modern Linux kernels make a copy
253 * of the vector available in `/procc' so we could try that.
255 * The usual place is stuck on the end of the environment vector, but that
256 * may well have moved, and we have no way of telling whether it has or
257 * whether there was ever an auxiliary vector there at all; so don't do
262 unsigned char *p
= 0, *q
= 0;
263 const struct auxentry
*a
;
266 /* Open the file and read it into a memory chunk. */
267 if ((fp
= fopen("/proc/self/auxv", "rb")) == 0) goto clean
;
269 if ((p
= malloc(sz
)) == 0) goto clean
;
271 n
= fread(p
+ off
, 1, sz
- off
, fp
);
274 sz
*= 2; if ((q
= realloc(p
, sz
)) == 0) break;
278 /* Work through the vector (or as much of it as we found) and extract the
279 * types we're interested in.
281 for (a
= (const struct auxentry
*)p
,
282 n
= sz
/sizeof(struct auxentry
);
285 #define CAP__SWITCH(type, ubranch, slot) \
286 case type: probed.slot = a->value.ubranch; break;
288 case AT_NULL
: goto clean
;
298 /* Each CPU family now has to pick through what was found and stashed in
299 * `probed', and set the appropriate flag bits in `hw'.
302 if (probed
.hwcap
& HWCAP_VFPv3
) hw
|= HF_ARM_VFP
;
303 if (probed
.hwcap
& HWCAP_NEON
) hw
|= HF_ARM_NEON
;
304 if (probed
.hwcap
& HWCAP_VFPD32
) hw
|= HF_ARM_D32
;
305 if (probed
.hwcap
& HWCAP_VFPv4
) hw
|= HF_ARM_V4
;
307 if (probed
.hwcap2
& HWCAP2_AES
) hw
|= HF_ARM_AES
;
310 if (probed
.hwcap2
& HWCAP2_PMULL
) hw
|= HF_ARM_PMULL
;
314 if (probed
.hwcap
& HWCAP_AES
) hw
|= HF_ARM_AES
;
315 if (probed
.hwcap
& HWCAP_PMULL
) hw
|= HF_ARM_PMULL
;
318 /* Store the bitmask of features we probed for everyone to see. */
319 DISPATCH_STORE(hwcaps
, hw
);
321 /* Finally, make a report about the things we found. (Doing this earlier
322 * will pointlessly widen the window in which multiple threads will do the
323 * above auxiliary-vector probing.)
325 #define CAP__DEBUG(feat, tok) \
326 dispatch_debug("check auxv for feature `%s': %s", tok, \
327 hw & HF_##feat ? "available" : "absent");
332 /* --- @get_hwcaps@ --- *
336 * Returns: A mask of hardware capabilities and other features, as probed
337 * from the auxiliary vector.
340 static unsigned get_hwcaps(void)
344 DISPATCH_LOAD(hwcaps
, hw
);
345 if (!(hwcaps
& HF_PROBED
)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps
, hw
); }
351 /*----- External interface ------------------------------------------------*/
353 /* --- @dispatch_debug@ --- *
355 * Arguments: @const char *fmt@ = a format string
356 * @...@ = additional arguments
360 * Use: Writes a formatted message to standard output if dispatch
361 * debugging is enabled.
364 void dispatch_debug(const char *fmt
, ...)
367 const char *e
= getenv("CATACOMB_CPUDISPATCH_DEBUG");
369 if (e
&& *e
!= 'n' && *e
!= '0') {
371 fputs("Catacomb CPUDISPATCH: ", stderr
);
372 vfprintf(stderr
, fmt
, ap
);
378 /* --- @check_env@ --- *
380 * Arguments: @const char *ftok@ = feature token
382 * Returns: Zero if the feature is forced off; positive if it's forced
383 * on; negative if the user hasn't decided.
385 * Use: Checks the environment variable `CATACOMB_CPUFEAT' for the
386 * feature token @ftok@. The variable, if it exists, should be
387 * a space-separated sequence of `+tok' and `-tok' items. These
388 * tokens may end in `*', which matches any suffix.
391 static int IGNORABLE
check_env(const char *ftok
)
393 const char *p
, *q
, *pp
;
396 p
= getenv("CATACOMB_CPUFEAT");
400 while (isspace((unsigned char)*p
)) p
++;
401 if (!*p
) return (-1);
403 case '+': d
= +1; p
++; break;
404 case '-': d
= 0; p
++; break;
405 default: d
= -1; break;
407 for (q
= p
; *q
&& !isspace((unsigned char)*q
); q
++);
409 for (pp
= ftok
; p
< q
&& *pp
&& *p
== *pp
; p
++, pp
++);
410 if ((p
== q
&& !*pp
) || (*p
== '*' && p
+ 1 == q
)) return (d
);
417 /* --- @cpu_feature_p@ --- *
419 * Arguments: @unsigned feat@ = a @CPUFEAT_...@ code
421 * Returns: Nonzero if the feature is available.
427 feat_debug(const char *ftok
, const char *check
, int verdict
)
430 dispatch_debug("feature `%s': %s -> %s", ftok
, check
,
431 verdict ?
"available" : "absent");
436 int cpu_feature_p(int feat
)
440 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat: \
441 if ((f = feat_debug(ftok, "environment override", \
442 check_env(ftok))) >= 0) \
445 return (feat_debug(ftok, "runtime probe", cond));
448 #if CPUFAM_X86 || CPUFAM_AMD64
449 CASE_CPUFEAT(X86_SSE2
, "x86:sse2",
450 cpuid_features_p(CPUID1D_SSE2
, 0) &&
451 xmm_registers_available_p());
452 CASE_CPUFEAT(X86_AESNI
, "x86:aesni",
453 cpuid_features_p(CPUID1D_SSE2
, CPUID1C_AESNI
) &&
454 xmm_registers_available_p());
455 CASE_CPUFEAT(X86_RDRAND
, "x86:rdrand",
456 cpuid_features_p(0, CPUID1C_RDRAND
));
457 CASE_CPUFEAT(X86_AVX
, "x86:avx",
458 cpuid_features_p(0, CPUID1C_AVX
) &&
459 xmm_registers_available_p());
460 CASE_CPUFEAT(X86_SSSE3
, "x86:ssse3",
461 cpuid_features_p(0, CPUID1C_SSSE3
) &&
462 xmm_registers_available_p());
463 CASE_CPUFEAT(X86_PCLMUL
, "x86:pclmul",
464 cpuid_features_p(0, CPUID1C_PCLMUL
) &&
465 xmm_registers_available_p());
468 # define FEATP__CASE(feat, tok) \
469 CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
474 dispatch_debug("denying unknown feature %d", feat
);
480 /*----- That's all, folks -------------------------------------------------*/