cpu-features.c revision 3ebb5017c5a6154992928ec2409bf3cc048e585b
1/* 2 * Copyright (C) 2010 The Android Open Source Project 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * * Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * * Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in 12 * the documentation and/or other materials provided with the 13 * distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 16 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 17 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 18 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 19 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 21 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 22 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 23 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 24 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 25 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29/* ChangeLog for this library: 30 * 31 * NDK r9?: Support for 64-bit CPUs (Intel, ARM & MIPS). 32 * 33 * NDK r8d: Add android_setCpu(). 34 * 35 * NDK r8c: Add new ARM CPU features: VFPv2, VFP_D32, VFP_FP16, 36 * VFP_FMA, NEON_FMA, IDIV_ARM, IDIV_THUMB2 and iWMMXt. 37 * 38 * Rewrite the code to parse /proc/self/auxv instead of 39 * the "Features" field in /proc/cpuinfo. 40 * 41 * Dynamically allocate the buffer that hold the content 42 * of /proc/cpuinfo to deal with newer hardware. 43 * 44 * NDK r7c: Fix CPU count computation. The old method only reported the 45 * number of _active_ CPUs when the library was initialized, 46 * which could be less than the real total. 47 * 48 * NDK r5: Handle buggy kernels which report a CPU Architecture number of 7 49 * for an ARMv6 CPU (see below). 50 * 51 * Handle kernels that only report 'neon', and not 'vfpv3' 52 * (VFPv3 is mandated by the ARM architecture is Neon is implemented) 53 * 54 * Handle kernels that only report 'vfpv3d16', and not 'vfpv3' 55 * 56 * Fix x86 compilation. Report ANDROID_CPU_FAMILY_X86 in 57 * android_getCpuFamily(). 58 * 59 * NDK r4: Initial release 60 */ 61 62#if defined(__le32__) || defined(__le64__) 63 64// When users enter this, we should only provide interface and 65// libportable will give the implementations. 66 67#else // !__le32__ && !__le64__ 68 69#include "cpu-features.h" 70 71#include <dlfcn.h> 72#include <errno.h> 73#include <fcntl.h> 74#include <pthread.h> 75#include <stdio.h> 76#include <stdlib.h> 77#include <sys/system_properties.h> 78 79static pthread_once_t g_once; 80static int g_inited; 81static AndroidCpuFamily g_cpuFamily; 82static uint64_t g_cpuFeatures; 83static int g_cpuCount; 84 85#ifdef __arm__ 86static uint32_t g_cpuIdArm; 87#endif 88 89static const int android_cpufeatures_debug = 0; 90 91#define D(...) \ 92 do { \ 93 if (android_cpufeatures_debug) { \ 94 printf(__VA_ARGS__); fflush(stdout); \ 95 } \ 96 } while (0) 97 98#ifdef __i386__ 99static __inline__ void x86_cpuid(int func, int values[4]) 100{ 101 int a, b, c, d; 102 /* We need to preserve ebx since we're compiling PIC code */ 103 /* this means we can't use "=b" for the second output register */ 104 __asm__ __volatile__ ( \ 105 "push %%ebx\n" 106 "cpuid\n" \ 107 "mov %%ebx, %1\n" 108 "pop %%ebx\n" 109 : "=a" (a), "=r" (b), "=c" (c), "=d" (d) \ 110 : "a" (func) \ 111 ); 112 values[0] = a; 113 values[1] = b; 114 values[2] = c; 115 values[3] = d; 116} 117#endif 118 119/* Get the size of a file by reading it until the end. This is needed 120 * because files under /proc do not always return a valid size when 121 * using fseek(0, SEEK_END) + ftell(). Nor can they be mmap()-ed. 122 */ 123static int 124get_file_size(const char* pathname) 125{ 126 int fd, ret, result = 0; 127 char buffer[256]; 128 129 fd = open(pathname, O_RDONLY); 130 if (fd < 0) { 131 D("Can't open %s: %s\n", pathname, strerror(errno)); 132 return -1; 133 } 134 135 for (;;) { 136 int ret = read(fd, buffer, sizeof buffer); 137 if (ret < 0) { 138 if (errno == EINTR) 139 continue; 140 D("Error while reading %s: %s\n", pathname, strerror(errno)); 141 break; 142 } 143 if (ret == 0) 144 break; 145 146 result += ret; 147 } 148 close(fd); 149 return result; 150} 151 152/* Read the content of /proc/cpuinfo into a user-provided buffer. 153 * Return the length of the data, or -1 on error. Does *not* 154 * zero-terminate the content. Will not read more 155 * than 'buffsize' bytes. 156 */ 157static int 158read_file(const char* pathname, char* buffer, size_t buffsize) 159{ 160 int fd, count; 161 162 fd = open(pathname, O_RDONLY); 163 if (fd < 0) { 164 D("Could not open %s: %s\n", pathname, strerror(errno)); 165 return -1; 166 } 167 count = 0; 168 while (count < (int)buffsize) { 169 int ret = read(fd, buffer + count, buffsize - count); 170 if (ret < 0) { 171 if (errno == EINTR) 172 continue; 173 D("Error while reading from %s: %s\n", pathname, strerror(errno)); 174 if (count == 0) 175 count = -1; 176 break; 177 } 178 if (ret == 0) 179 break; 180 count += ret; 181 } 182 close(fd); 183 return count; 184} 185 186/* Extract the content of a the first occurence of a given field in 187 * the content of /proc/cpuinfo and return it as a heap-allocated 188 * string that must be freed by the caller. 189 * 190 * Return NULL if not found 191 */ 192static char* 193extract_cpuinfo_field(const char* buffer, int buflen, const char* field) 194{ 195 int fieldlen = strlen(field); 196 const char* bufend = buffer + buflen; 197 char* result = NULL; 198 int len, ignore; 199 const char *p, *q; 200 201 /* Look for first field occurence, and ensures it starts the line. */ 202 p = buffer; 203 for (;;) { 204 p = memmem(p, bufend-p, field, fieldlen); 205 if (p == NULL) 206 goto EXIT; 207 208 if (p == buffer || p[-1] == '\n') 209 break; 210 211 p += fieldlen; 212 } 213 214 /* Skip to the first column followed by a space */ 215 p += fieldlen; 216 p = memchr(p, ':', bufend-p); 217 if (p == NULL || p[1] != ' ') 218 goto EXIT; 219 220 /* Find the end of the line */ 221 p += 2; 222 q = memchr(p, '\n', bufend-p); 223 if (q == NULL) 224 q = bufend; 225 226 /* Copy the line into a heap-allocated buffer */ 227 len = q-p; 228 result = malloc(len+1); 229 if (result == NULL) 230 goto EXIT; 231 232 memcpy(result, p, len); 233 result[len] = '\0'; 234 235EXIT: 236 return result; 237} 238 239/* Checks that a space-separated list of items contains one given 'item'. 240 * Returns 1 if found, 0 otherwise. 241 */ 242static int 243has_list_item(const char* list, const char* item) 244{ 245 const char* p = list; 246 int itemlen = strlen(item); 247 248 if (list == NULL) 249 return 0; 250 251 while (*p) { 252 const char* q; 253 254 /* skip spaces */ 255 while (*p == ' ' || *p == '\t') 256 p++; 257 258 /* find end of current list item */ 259 q = p; 260 while (*q && *q != ' ' && *q != '\t') 261 q++; 262 263 if (itemlen == q-p && !memcmp(p, item, itemlen)) 264 return 1; 265 266 /* skip to next item */ 267 p = q; 268 } 269 return 0; 270} 271 272/* Parse a number starting from 'input', but not going further 273 * than 'limit'. Return the value into '*result'. 274 * 275 * NOTE: Does not skip over leading spaces, or deal with sign characters. 276 * NOTE: Ignores overflows. 277 * 278 * The function returns NULL in case of error (bad format), or the new 279 * position after the decimal number in case of success (which will always 280 * be <= 'limit'). 281 */ 282static const char* 283parse_number(const char* input, const char* limit, int base, int* result) 284{ 285 const char* p = input; 286 int val = 0; 287 while (p < limit) { 288 int d = (*p - '0'); 289 if ((unsigned)d >= 10U) { 290 d = (*p - 'a'); 291 if ((unsigned)d >= 6U) 292 d = (*p - 'A'); 293 if ((unsigned)d >= 6U) 294 break; 295 d += 10; 296 } 297 if (d >= base) 298 break; 299 val = val*base + d; 300 p++; 301 } 302 if (p == input) 303 return NULL; 304 305 *result = val; 306 return p; 307} 308 309static const char* 310parse_decimal(const char* input, const char* limit, int* result) 311{ 312 return parse_number(input, limit, 10, result); 313} 314 315static const char* 316parse_hexadecimal(const char* input, const char* limit, int* result) 317{ 318 return parse_number(input, limit, 16, result); 319} 320 321/* This small data type is used to represent a CPU list / mask, as read 322 * from sysfs on Linux. See http://www.kernel.org/doc/Documentation/cputopology.txt 323 * 324 * For now, we don't expect more than 32 cores on mobile devices, so keep 325 * everything simple. 326 */ 327typedef struct { 328 uint32_t mask; 329} CpuList; 330 331static __inline__ void 332cpulist_init(CpuList* list) { 333 list->mask = 0; 334} 335 336static __inline__ void 337cpulist_and(CpuList* list1, CpuList* list2) { 338 list1->mask &= list2->mask; 339} 340 341static __inline__ void 342cpulist_set(CpuList* list, int index) { 343 if ((unsigned)index < 32) { 344 list->mask |= (uint32_t)(1U << index); 345 } 346} 347 348static __inline__ int 349cpulist_count(CpuList* list) { 350 return __builtin_popcount(list->mask); 351} 352 353/* Parse a textual list of cpus and store the result inside a CpuList object. 354 * Input format is the following: 355 * - comma-separated list of items (no spaces) 356 * - each item is either a single decimal number (cpu index), or a range made 357 * of two numbers separated by a single dash (-). Ranges are inclusive. 358 * 359 * Examples: 0 360 * 2,4-127,128-143 361 * 0-1 362 */ 363static void 364cpulist_parse(CpuList* list, const char* line, int line_len) 365{ 366 const char* p = line; 367 const char* end = p + line_len; 368 const char* q; 369 370 /* NOTE: the input line coming from sysfs typically contains a 371 * trailing newline, so take care of it in the code below 372 */ 373 while (p < end && *p != '\n') 374 { 375 int val, start_value, end_value; 376 377 /* Find the end of current item, and put it into 'q' */ 378 q = memchr(p, ',', end-p); 379 if (q == NULL) { 380 q = end; 381 } 382 383 /* Get first value */ 384 p = parse_decimal(p, q, &start_value); 385 if (p == NULL) 386 goto BAD_FORMAT; 387 388 end_value = start_value; 389 390 /* If we're not at the end of the item, expect a dash and 391 * and integer; extract end value. 392 */ 393 if (p < q && *p == '-') { 394 p = parse_decimal(p+1, q, &end_value); 395 if (p == NULL) 396 goto BAD_FORMAT; 397 } 398 399 /* Set bits CPU list bits */ 400 for (val = start_value; val <= end_value; val++) { 401 cpulist_set(list, val); 402 } 403 404 /* Jump to next item */ 405 p = q; 406 if (p < end) 407 p++; 408 } 409 410BAD_FORMAT: 411 ; 412} 413 414/* Read a CPU list from one sysfs file */ 415static void 416cpulist_read_from(CpuList* list, const char* filename) 417{ 418 char file[64]; 419 int filelen; 420 421 cpulist_init(list); 422 423 filelen = read_file(filename, file, sizeof file); 424 if (filelen < 0) { 425 D("Could not read %s: %s\n", filename, strerror(errno)); 426 return; 427 } 428 429 cpulist_parse(list, file, filelen); 430} 431#if defined(__aarch64__) 432// see <uapi/asm/hwcap.h> kernel header 433#define HWCAP_FP (1 << 0) 434#define HWCAP_ASIMD (1 << 1) 435#define HWCAP_AES (1 << 3) 436#define HWCAP_PMULL (1 << 4) 437#define HWCAP_SHA1 (1 << 5) 438#define HWCAP_SHA2 (1 << 6) 439#define HWCAP_CRC32 (1 << 7) 440#endif 441 442#if defined(__arm__) 443 444// See <asm/hwcap.h> kernel header. 445#define HWCAP_VFP (1 << 6) 446#define HWCAP_IWMMXT (1 << 9) 447#define HWCAP_NEON (1 << 12) 448#define HWCAP_VFPv3 (1 << 13) 449#define HWCAP_VFPv3D16 (1 << 14) 450#define HWCAP_VFPv4 (1 << 16) 451#define HWCAP_IDIVA (1 << 17) 452#define HWCAP_IDIVT (1 << 18) 453 454// see <uapi/asm/hwcap.h> kernel header 455#define HWCAP2_AES (1 << 0) 456#define HWCAP2_PMULL (1 << 1) 457#define HWCAP2_SHA1 (1 << 2) 458#define HWCAP2_SHA2 (1 << 3) 459#define HWCAP2_CRC32 (1 << 4) 460 461// This is the list of 32-bit ARMv7 optional features that are _always_ 462// supported by ARMv8 CPUs, as mandated by the ARM Architecture Reference 463// Manual. 464#define HWCAP_SET_FOR_ARMV8 \ 465 ( HWCAP_VFP | \ 466 HWCAP_NEON | \ 467 HWCAP_VFPv3 | \ 468 HWCAP_VFPv4 | \ 469 HWCAP_IDIVA | \ 470 HWCAP_IDIVT ) 471#endif 472 473#if defined(__arm__) || defined(__aarch64__) 474 475#define AT_HWCAP 16 476#define AT_HWCAP2 26 477 478// Probe the system's C library for a 'getauxval' function and call it if 479// it exits, or return 0 for failure. This function is available since API 480// level 20. 481// 482// This code does *NOT* check for '__ANDROID_API__ >= 20' to support the 483// edge case where some NDK developers use headers for a platform that is 484// newer than the one really targetted by their application. 485// This is typically done to use newer native APIs only when running on more 486// recent Android versions, and requires careful symbol management. 487// 488// Note that getauxval() can't really be re-implemented here, because 489// its implementation does not parse /proc/self/auxv. Instead it depends 490// on values that are passed by the kernel at process-init time to the 491// C runtime initialization layer. 492static uint32_t 493get_elf_hwcap_from_getauxval(int hwcap_type) { 494 typedef unsigned long getauxval_func_t(unsigned long); 495 496 dlerror(); 497 void* libc_handle = dlopen("libc.so", RTLD_NOW); 498 if (!libc_handle) { 499 D("Could not dlopen() C library: %s\n", dlerror()); 500 return 0; 501 } 502 503 uint32_t ret = 0; 504 getauxval_func_t* func = (getauxval_func_t*) 505 dlsym(libc_handle, "getauxval"); 506 if (!func) { 507 D("Could not find getauxval() in C library\n"); 508 } else { 509 // Note: getauxval() returns 0 on failure. Doesn't touch errno. 510 ret = (uint32_t)(*func)(hwcap_type); 511 } 512 dlclose(libc_handle); 513 return ret; 514} 515#endif 516 517#if defined(__arm__) 518// Parse /proc/self/auxv to extract the ELF HW capabilities bitmap for the 519// current CPU. Note that this file is not accessible from regular 520// application processes on some Android platform releases. 521// On success, return new ELF hwcaps, or 0 on failure. 522static uint32_t 523get_elf_hwcap_from_proc_self_auxv(void) { 524 const char filepath[] = "/proc/self/auxv"; 525 int fd = TEMP_FAILURE_RETRY(open(filepath, O_RDONLY)); 526 if (fd < 0) { 527 D("Could not open %s: %s\n", filepath, strerror(errno)); 528 return 0; 529 } 530 531 struct { uint32_t tag; uint32_t value; } entry; 532 533 uint32_t result = 0; 534 for (;;) { 535 int ret = TEMP_FAILURE_RETRY(read(fd, (char*)&entry, sizeof entry)); 536 if (ret < 0) { 537 D("Error while reading %s: %s\n", filepath, strerror(errno)); 538 break; 539 } 540 // Detect end of list. 541 if (ret == 0 || (entry.tag == 0 && entry.value == 0)) 542 break; 543 if (entry.tag == AT_HWCAP) { 544 result = entry.value; 545 break; 546 } 547 } 548 close(fd); 549 return result; 550} 551 552/* Compute the ELF HWCAP flags from the content of /proc/cpuinfo. 553 * This works by parsing the 'Features' line, which lists which optional 554 * features the device's CPU supports, on top of its reference 555 * architecture. 556 */ 557static uint32_t 558get_elf_hwcap_from_proc_cpuinfo(const char* cpuinfo, int cpuinfo_len) { 559 uint32_t hwcaps = 0; 560 long architecture = 0; 561 char* cpuArch = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "CPU architecture"); 562 if (cpuArch) { 563 architecture = strtol(cpuArch, NULL, 10); 564 free(cpuArch); 565 566 if (architecture >= 8L) { 567 // This is a 32-bit ARM binary running on a 64-bit ARM64 kernel. 568 // The 'Features' line only lists the optional features that the 569 // device's CPU supports, compared to its reference architecture 570 // which are of no use for this process. 571 D("Faking 32-bit ARM HWCaps on ARMv%ld CPU\n", architecture); 572 return HWCAP_SET_FOR_ARMV8; 573 } 574 } 575 576 char* cpuFeatures = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "Features"); 577 if (cpuFeatures != NULL) { 578 D("Found cpuFeatures = '%s'\n", cpuFeatures); 579 580 if (has_list_item(cpuFeatures, "vfp")) 581 hwcaps |= HWCAP_VFP; 582 if (has_list_item(cpuFeatures, "vfpv3")) 583 hwcaps |= HWCAP_VFPv3; 584 if (has_list_item(cpuFeatures, "vfpv3d16")) 585 hwcaps |= HWCAP_VFPv3D16; 586 if (has_list_item(cpuFeatures, "vfpv4")) 587 hwcaps |= HWCAP_VFPv4; 588 if (has_list_item(cpuFeatures, "neon")) 589 hwcaps |= HWCAP_NEON; 590 if (has_list_item(cpuFeatures, "idiva")) 591 hwcaps |= HWCAP_IDIVA; 592 if (has_list_item(cpuFeatures, "idivt")) 593 hwcaps |= HWCAP_IDIVT; 594 if (has_list_item(cpuFeatures, "idiv")) 595 hwcaps |= HWCAP_IDIVA | HWCAP_IDIVT; 596 if (has_list_item(cpuFeatures, "iwmmxt")) 597 hwcaps |= HWCAP_IWMMXT; 598 599 free(cpuFeatures); 600 } 601 return hwcaps; 602} 603#endif /* __arm__ */ 604 605/* Return the number of cpus present on a given device. 606 * 607 * To handle all weird kernel configurations, we need to compute the 608 * intersection of the 'present' and 'possible' CPU lists and count 609 * the result. 610 */ 611static int 612get_cpu_count(void) 613{ 614 CpuList cpus_present[1]; 615 CpuList cpus_possible[1]; 616 617 cpulist_read_from(cpus_present, "/sys/devices/system/cpu/present"); 618 cpulist_read_from(cpus_possible, "/sys/devices/system/cpu/possible"); 619 620 /* Compute the intersection of both sets to get the actual number of 621 * CPU cores that can be used on this device by the kernel. 622 */ 623 cpulist_and(cpus_present, cpus_possible); 624 625 return cpulist_count(cpus_present); 626} 627 628static void 629android_cpuInitFamily(void) 630{ 631#if defined(__arm__) 632 g_cpuFamily = ANDROID_CPU_FAMILY_ARM; 633#elif defined(__i386__) 634 g_cpuFamily = ANDROID_CPU_FAMILY_X86; 635#elif defined(__mips64) 636/* Needs to be before __mips__ since the compiler defines both */ 637 g_cpuFamily = ANDROID_CPU_FAMILY_MIPS64; 638#elif defined(__mips__) 639 g_cpuFamily = ANDROID_CPU_FAMILY_MIPS; 640#elif defined(__aarch64__) 641 g_cpuFamily = ANDROID_CPU_FAMILY_ARM64; 642#elif defined(__x86_64__) 643 g_cpuFamily = ANDROID_CPU_FAMILY_X86_64; 644#else 645 g_cpuFamily = ANDROID_CPU_FAMILY_UNKNOWN; 646#endif 647} 648 649static void 650android_cpuInit(void) 651{ 652 char* cpuinfo = NULL; 653 int cpuinfo_len; 654 655 android_cpuInitFamily(); 656 657 g_cpuFeatures = 0; 658 g_cpuCount = 1; 659 g_inited = 1; 660 661 cpuinfo_len = get_file_size("/proc/cpuinfo"); 662 if (cpuinfo_len < 0) { 663 D("cpuinfo_len cannot be computed!"); 664 return; 665 } 666 cpuinfo = malloc(cpuinfo_len); 667 if (cpuinfo == NULL) { 668 D("cpuinfo buffer could not be allocated"); 669 return; 670 } 671 cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, cpuinfo_len); 672 D("cpuinfo_len is (%d):\n%.*s\n", cpuinfo_len, 673 cpuinfo_len >= 0 ? cpuinfo_len : 0, cpuinfo); 674 675 if (cpuinfo_len < 0) /* should not happen */ { 676 free(cpuinfo); 677 return; 678 } 679 680 /* Count the CPU cores, the value may be 0 for single-core CPUs */ 681 g_cpuCount = get_cpu_count(); 682 if (g_cpuCount == 0) { 683 g_cpuCount = 1; 684 } 685 686 D("found cpuCount = %d\n", g_cpuCount); 687 688#ifdef __arm__ 689 { 690 /* Extract architecture from the "CPU Architecture" field. 691 * The list is well-known, unlike the the output of 692 * the 'Processor' field which can vary greatly. 693 * 694 * See the definition of the 'proc_arch' array in 695 * $KERNEL/arch/arm/kernel/setup.c and the 'c_show' function in 696 * same file. 697 */ 698 char* cpuArch = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "CPU architecture"); 699 700 if (cpuArch != NULL) { 701 char* end; 702 long archNumber; 703 int hasARMv7 = 0; 704 705 D("found cpuArch = '%s'\n", cpuArch); 706 707 /* read the initial decimal number, ignore the rest */ 708 archNumber = strtol(cpuArch, &end, 10); 709 710 /* Note that ARMv8 is upwards compatible with ARMv7. */ 711 if (end > cpuArch && archNumber >= 7) { 712 hasARMv7 = 1; 713 } 714 715 /* Unfortunately, it seems that certain ARMv6-based CPUs 716 * report an incorrect architecture number of 7! 717 * 718 * See http://code.google.com/p/android/issues/detail?id=10812 719 * 720 * We try to correct this by looking at the 'elf_format' 721 * field reported by the 'Processor' field, which is of the 722 * form of "(v7l)" for an ARMv7-based CPU, and "(v6l)" for 723 * an ARMv6-one. 724 */ 725 if (hasARMv7) { 726 char* cpuProc = extract_cpuinfo_field(cpuinfo, cpuinfo_len, 727 "Processor"); 728 if (cpuProc != NULL) { 729 D("found cpuProc = '%s'\n", cpuProc); 730 if (has_list_item(cpuProc, "(v6l)")) { 731 D("CPU processor and architecture mismatch!!\n"); 732 hasARMv7 = 0; 733 } 734 free(cpuProc); 735 } 736 } 737 738 if (hasARMv7) { 739 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_ARMv7; 740 } 741 742 /* The LDREX / STREX instructions are available from ARMv6 */ 743 if (archNumber >= 6) { 744 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_LDREX_STREX; 745 } 746 747 free(cpuArch); 748 } 749 750 /* Extract the list of CPU features from ELF hwcaps */ 751 uint32_t hwcaps = 0; 752 hwcaps = get_elf_hwcap_from_getauxval(AT_HWCAP); 753 if (!hwcaps) { 754 D("Parsing /proc/self/auxv to extract ELF hwcaps!\n"); 755 hwcaps = get_elf_hwcap_from_proc_self_auxv(); 756 } 757 if (!hwcaps) { 758 // Parsing /proc/self/auxv will fail from regular application 759 // processes on some Android platform versions, when this happens 760 // parse proc/cpuinfo instead. 761 D("Parsing /proc/cpuinfo to extract ELF hwcaps!\n"); 762 hwcaps = get_elf_hwcap_from_proc_cpuinfo(cpuinfo, cpuinfo_len); 763 } 764 765 if (hwcaps != 0) { 766 int has_vfp = (hwcaps & HWCAP_VFP); 767 int has_vfpv3 = (hwcaps & HWCAP_VFPv3); 768 int has_vfpv3d16 = (hwcaps & HWCAP_VFPv3D16); 769 int has_vfpv4 = (hwcaps & HWCAP_VFPv4); 770 int has_neon = (hwcaps & HWCAP_NEON); 771 int has_idiva = (hwcaps & HWCAP_IDIVA); 772 int has_idivt = (hwcaps & HWCAP_IDIVT); 773 int has_iwmmxt = (hwcaps & HWCAP_IWMMXT); 774 775 // The kernel does a poor job at ensuring consistency when 776 // describing CPU features. So lots of guessing is needed. 777 778 // 'vfpv4' implies VFPv3|VFP_FMA|FP16 779 if (has_vfpv4) 780 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3 | 781 ANDROID_CPU_ARM_FEATURE_VFP_FP16 | 782 ANDROID_CPU_ARM_FEATURE_VFP_FMA; 783 784 // 'vfpv3' or 'vfpv3d16' imply VFPv3. Note that unlike GCC, 785 // a value of 'vfpv3' doesn't necessarily mean that the D32 786 // feature is present, so be conservative. All CPUs in the 787 // field that support D32 also support NEON, so this should 788 // not be a problem in practice. 789 if (has_vfpv3 || has_vfpv3d16) 790 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3; 791 792 // 'vfp' is super ambiguous. Depending on the kernel, it can 793 // either mean VFPv2 or VFPv3. Make it depend on ARMv7. 794 if (has_vfp) { 795 if (g_cpuFeatures & ANDROID_CPU_ARM_FEATURE_ARMv7) 796 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3; 797 else 798 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv2; 799 } 800 801 // Neon implies VFPv3|D32, and if vfpv4 is detected, NEON_FMA 802 if (has_neon) { 803 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3 | 804 ANDROID_CPU_ARM_FEATURE_NEON | 805 ANDROID_CPU_ARM_FEATURE_VFP_D32; 806 if (has_vfpv4) 807 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_NEON_FMA; 808 } 809 810 // VFPv3 implies VFPv2 and ARMv7 811 if (g_cpuFeatures & ANDROID_CPU_ARM_FEATURE_VFPv3) 812 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv2 | 813 ANDROID_CPU_ARM_FEATURE_ARMv7; 814 815 if (has_idiva) 816 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_ARM; 817 if (has_idivt) 818 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2; 819 820 if (has_iwmmxt) 821 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_iWMMXt; 822 } 823 824 /* Extract the list of CPU features from ELF hwcaps2 */ 825 uint32_t hwcaps2 = 0; 826 hwcaps2 = get_elf_hwcap_from_getauxval(AT_HWCAP2); 827 if (hwcaps2 != 0) { 828 int has_aes = (hwcaps2 & HWCAP2_AES); 829 int has_pmull = (hwcaps2 & HWCAP2_PMULL); 830 int has_sha1 = (hwcaps2 & HWCAP2_SHA1); 831 int has_sha2 = (hwcaps2 & HWCAP2_SHA2); 832 int has_crc32 = (hwcaps2 & HWCAP2_CRC32); 833 834 if (has_aes) 835 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_AES; 836 if (has_pmull) 837 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_PMULL; 838 if (has_sha1) 839 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_SHA1; 840 if (has_sha2) 841 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_SHA2; 842 if (has_crc32) 843 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_CRC32; 844 } 845 /* Extract the cpuid value from various fields */ 846 // The CPUID value is broken up in several entries in /proc/cpuinfo. 847 // This table is used to rebuild it from the entries. 848 static const struct CpuIdEntry { 849 const char* field; 850 char format; 851 char bit_lshift; 852 char bit_length; 853 } cpu_id_entries[] = { 854 { "CPU implementer", 'x', 24, 8 }, 855 { "CPU variant", 'x', 20, 4 }, 856 { "CPU part", 'x', 4, 12 }, 857 { "CPU revision", 'd', 0, 4 }, 858 }; 859 size_t i; 860 D("Parsing /proc/cpuinfo to recover CPUID\n"); 861 for (i = 0; 862 i < sizeof(cpu_id_entries)/sizeof(cpu_id_entries[0]); 863 ++i) { 864 const struct CpuIdEntry* entry = &cpu_id_entries[i]; 865 char* value = extract_cpuinfo_field(cpuinfo, 866 cpuinfo_len, 867 entry->field); 868 if (value == NULL) 869 continue; 870 871 D("field=%s value='%s'\n", entry->field, value); 872 char* value_end = value + strlen(value); 873 int val = 0; 874 const char* start = value; 875 const char* p; 876 if (value[0] == '0' && (value[1] == 'x' || value[1] == 'X')) { 877 start += 2; 878 p = parse_hexadecimal(start, value_end, &val); 879 } else if (entry->format == 'x') 880 p = parse_hexadecimal(value, value_end, &val); 881 else 882 p = parse_decimal(value, value_end, &val); 883 884 if (p > (const char*)start) { 885 val &= ((1 << entry->bit_length)-1); 886 val <<= entry->bit_lshift; 887 g_cpuIdArm |= (uint32_t) val; 888 } 889 890 free(value); 891 } 892 893 // Handle kernel configuration bugs that prevent the correct 894 // reporting of CPU features. 895 static const struct CpuFix { 896 uint32_t cpuid; 897 uint64_t or_flags; 898 } cpu_fixes[] = { 899 /* The Nexus 4 (Qualcomm Krait) kernel configuration 900 * forgets to report IDIV support. */ 901 { 0x510006f2, ANDROID_CPU_ARM_FEATURE_IDIV_ARM | 902 ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2 }, 903 }; 904 size_t n; 905 for (n = 0; n < sizeof(cpu_fixes)/sizeof(cpu_fixes[0]); ++n) { 906 const struct CpuFix* entry = &cpu_fixes[n]; 907 908 if (g_cpuIdArm == entry->cpuid) 909 g_cpuFeatures |= entry->or_flags; 910 } 911 912 // Special case: The emulator-specific Android 4.2 kernel fails 913 // to report support for the 32-bit ARM IDIV instruction. 914 // Technically, this is a feature of the virtual CPU implemented 915 // by the emulator. Note that it could also support Thumb IDIV 916 // in the future, and this will have to be slightly updated. 917 char* hardware = extract_cpuinfo_field(cpuinfo, 918 cpuinfo_len, 919 "Hardware"); 920 if (hardware) { 921 if (!strcmp(hardware, "Goldfish") && 922 g_cpuIdArm == 0x4100c080 && 923 (g_cpuFamily & ANDROID_CPU_ARM_FEATURE_ARMv7) != 0) { 924 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_ARM; 925 } 926 free(hardware); 927 } 928 } 929#endif /* __arm__ */ 930#ifdef __aarch64__ 931 { 932 /* Extract the list of CPU features from ELF hwcaps */ 933 uint32_t hwcaps = 0; 934 hwcaps = get_elf_hwcap_from_getauxval(AT_HWCAP); 935 if (hwcaps != 0) { 936 int has_fp = (hwcaps & HWCAP_FP); 937 int has_asimd = (hwcaps & HWCAP_ASIMD); 938 int has_aes = (hwcaps & HWCAP_AES); 939 int has_pmull = (hwcaps & HWCAP_PMULL); 940 int has_sha1 = (hwcaps & HWCAP_SHA1); 941 int has_sha2 = (hwcaps & HWCAP_SHA2); 942 int has_crc32 = (hwcaps & HWCAP_CRC32); 943 944 if(has_fp == 0) { 945 D("ERROR: Floating-point unit missing, but is required by Android on AArch64 CPUs\n"); 946 } 947 if(has_asimd == 0) { 948 D("ERROR: ASIMD unit missing, but is required by Android on AArch64 CPUs\n"); 949 } 950 951 if (has_fp) 952 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_FP; 953 if (has_asimd) 954 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_ASIMD; 955 if (has_aes) 956 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_AES; 957 if (has_pmull) 958 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_PMULL; 959 if (has_sha1) 960 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_SHA1; 961 if (has_sha2) 962 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_SHA2; 963 if (has_crc32) 964 g_cpuFeatures |= ANDROID_CPU_ARM64_FEATURE_CRC32; 965 } 966 } 967#endif /* __aarch64__ */ 968 969#ifdef __i386__ 970 int regs[4]; 971 972/* According to http://en.wikipedia.org/wiki/CPUID */ 973#define VENDOR_INTEL_b 0x756e6547 974#define VENDOR_INTEL_c 0x6c65746e 975#define VENDOR_INTEL_d 0x49656e69 976 977 x86_cpuid(0, regs); 978 int vendorIsIntel = (regs[1] == VENDOR_INTEL_b && 979 regs[2] == VENDOR_INTEL_c && 980 regs[3] == VENDOR_INTEL_d); 981 982 x86_cpuid(1, regs); 983 if ((regs[2] & (1 << 9)) != 0) { 984 g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_SSSE3; 985 } 986 if ((regs[2] & (1 << 23)) != 0) { 987 g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_POPCNT; 988 } 989 if (vendorIsIntel && (regs[2] & (1 << 22)) != 0) { 990 g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_MOVBE; 991 } 992#endif 993 994 free(cpuinfo); 995} 996 997 998AndroidCpuFamily 999android_getCpuFamily(void) 1000{ 1001 pthread_once(&g_once, android_cpuInit); 1002 return g_cpuFamily; 1003} 1004 1005 1006uint64_t 1007android_getCpuFeatures(void) 1008{ 1009 pthread_once(&g_once, android_cpuInit); 1010 return g_cpuFeatures; 1011} 1012 1013 1014int 1015android_getCpuCount(void) 1016{ 1017 pthread_once(&g_once, android_cpuInit); 1018 return g_cpuCount; 1019} 1020 1021static void 1022android_cpuInitDummy(void) 1023{ 1024 g_inited = 1; 1025} 1026 1027int 1028android_setCpu(int cpu_count, uint64_t cpu_features) 1029{ 1030 /* Fail if the library was already initialized. */ 1031 if (g_inited) 1032 return 0; 1033 1034 android_cpuInitFamily(); 1035 g_cpuCount = (cpu_count <= 0 ? 1 : cpu_count); 1036 g_cpuFeatures = cpu_features; 1037 pthread_once(&g_once, android_cpuInitDummy); 1038 1039 return 1; 1040} 1041 1042#ifdef __arm__ 1043uint32_t 1044android_getCpuIdArm(void) 1045{ 1046 pthread_once(&g_once, android_cpuInit); 1047 return g_cpuIdArm; 1048} 1049 1050int 1051android_setCpuArm(int cpu_count, uint64_t cpu_features, uint32_t cpu_id) 1052{ 1053 if (!android_setCpu(cpu_count, cpu_features)) 1054 return 0; 1055 1056 g_cpuIdArm = cpu_id; 1057 return 1; 1058} 1059#endif /* __arm__ */ 1060 1061/* 1062 * Technical note: Making sense of ARM's FPU architecture versions. 1063 * 1064 * FPA was ARM's first attempt at an FPU architecture. There is no Android 1065 * device that actually uses it since this technology was already obsolete 1066 * when the project started. If you see references to FPA instructions 1067 * somewhere, you can be sure that this doesn't apply to Android at all. 1068 * 1069 * FPA was followed by "VFP", soon renamed "VFPv1" due to the emergence of 1070 * new versions / additions to it. ARM considers this obsolete right now, 1071 * and no known Android device implements it either. 1072 * 1073 * VFPv2 added a few instructions to VFPv1, and is an *optional* extension 1074 * supported by some ARMv5TE, ARMv6 and ARMv6T2 CPUs. Note that a device 1075 * supporting the 'armeabi' ABI doesn't necessarily support these. 1076 * 1077 * VFPv3-D16 adds a few instructions on top of VFPv2 and is typically used 1078 * on ARMv7-A CPUs which implement a FPU. Note that it is also mandated 1079 * by the Android 'armeabi-v7a' ABI. The -D16 suffix in its name means 1080 * that it provides 16 double-precision FPU registers (d0-d15) and 32 1081 * single-precision ones (s0-s31) which happen to be mapped to the same 1082 * register banks. 1083 * 1084 * VFPv3-D32 is the name of an extension to VFPv3-D16 that provides 16 1085 * additional double precision registers (d16-d31). Note that there are 1086 * still only 32 single precision registers. 1087 * 1088 * VFPv3xD is a *subset* of VFPv3-D16 that only provides single-precision 1089 * registers. It is only used on ARMv7-M (i.e. on micro-controllers) which 1090 * are not supported by Android. Note that it is not compatible with VFPv2. 1091 * 1092 * NOTE: The term 'VFPv3' usually designate either VFPv3-D16 or VFPv3-D32 1093 * depending on context. For example GCC uses it for VFPv3-D32, but 1094 * the Linux kernel code uses it for VFPv3-D16 (especially in 1095 * /proc/cpuinfo). Always try to use the full designation when 1096 * possible. 1097 * 1098 * NEON, a.k.a. "ARM Advanced SIMD" is an extension that provides 1099 * instructions to perform parallel computations on vectors of 8, 16, 1100 * 32, 64 and 128 bit quantities. NEON requires VFPv32-D32 since all 1101 * NEON registers are also mapped to the same register banks. 1102 * 1103 * VFPv4-D16, adds a few instructions on top of VFPv3-D16 in order to 1104 * perform fused multiply-accumulate on VFP registers, as well as 1105 * half-precision (16-bit) conversion operations. 1106 * 1107 * VFPv4-D32 is VFPv4-D16 with 32, instead of 16, FPU double precision 1108 * registers. 1109 * 1110 * VPFv4-NEON is VFPv4-D32 with NEON instructions. It also adds fused 1111 * multiply-accumulate instructions that work on the NEON registers. 1112 * 1113 * NOTE: Similarly, "VFPv4" might either reference VFPv4-D16 or VFPv4-D32 1114 * depending on context. 1115 * 1116 * The following information was determined by scanning the binutils-2.22 1117 * sources: 1118 * 1119 * Basic VFP instruction subsets: 1120 * 1121 * #define FPU_VFP_EXT_V1xD 0x08000000 // Base VFP instruction set. 1122 * #define FPU_VFP_EXT_V1 0x04000000 // Double-precision insns. 1123 * #define FPU_VFP_EXT_V2 0x02000000 // ARM10E VFPr1. 1124 * #define FPU_VFP_EXT_V3xD 0x01000000 // VFPv3 single-precision. 1125 * #define FPU_VFP_EXT_V3 0x00800000 // VFPv3 double-precision. 1126 * #define FPU_NEON_EXT_V1 0x00400000 // Neon (SIMD) insns. 1127 * #define FPU_VFP_EXT_D32 0x00200000 // Registers D16-D31. 1128 * #define FPU_VFP_EXT_FP16 0x00100000 // Half-precision extensions. 1129 * #define FPU_NEON_EXT_FMA 0x00080000 // Neon fused multiply-add 1130 * #define FPU_VFP_EXT_FMA 0x00040000 // VFP fused multiply-add 1131 * 1132 * FPU types (excluding NEON) 1133 * 1134 * FPU_VFP_V1xD (EXT_V1xD) 1135 * | 1136 * +--------------------------+ 1137 * | | 1138 * FPU_VFP_V1 (+EXT_V1) FPU_VFP_V3xD (+EXT_V2+EXT_V3xD) 1139 * | | 1140 * | | 1141 * FPU_VFP_V2 (+EXT_V2) FPU_VFP_V4_SP_D16 (+EXT_FP16+EXT_FMA) 1142 * | 1143 * FPU_VFP_V3D16 (+EXT_Vx3D+EXT_V3) 1144 * | 1145 * +--------------------------+ 1146 * | | 1147 * FPU_VFP_V3 (+EXT_D32) FPU_VFP_V4D16 (+EXT_FP16+EXT_FMA) 1148 * | | 1149 * | FPU_VFP_V4 (+EXT_D32) 1150 * | 1151 * FPU_VFP_HARD (+EXT_FMA+NEON_EXT_FMA) 1152 * 1153 * VFP architectures: 1154 * 1155 * ARCH_VFP_V1xD (EXT_V1xD) 1156 * | 1157 * +------------------+ 1158 * | | 1159 * | ARCH_VFP_V3xD (+EXT_V2+EXT_V3xD) 1160 * | | 1161 * | ARCH_VFP_V3xD_FP16 (+EXT_FP16) 1162 * | | 1163 * | ARCH_VFP_V4_SP_D16 (+EXT_FMA) 1164 * | 1165 * ARCH_VFP_V1 (+EXT_V1) 1166 * | 1167 * ARCH_VFP_V2 (+EXT_V2) 1168 * | 1169 * ARCH_VFP_V3D16 (+EXT_V3xD+EXT_V3) 1170 * | 1171 * +-------------------+ 1172 * | | 1173 * | ARCH_VFP_V3D16_FP16 (+EXT_FP16) 1174 * | 1175 * +-------------------+ 1176 * | | 1177 * | ARCH_VFP_V4_D16 (+EXT_FP16+EXT_FMA) 1178 * | | 1179 * | ARCH_VFP_V4 (+EXT_D32) 1180 * | | 1181 * | ARCH_NEON_VFP_V4 (+EXT_NEON+EXT_NEON_FMA) 1182 * | 1183 * ARCH_VFP_V3 (+EXT_D32) 1184 * | 1185 * +-------------------+ 1186 * | | 1187 * | ARCH_VFP_V3_FP16 (+EXT_FP16) 1188 * | 1189 * ARCH_VFP_V3_PLUS_NEON_V1 (+EXT_NEON) 1190 * | 1191 * ARCH_NEON_FP16 (+EXT_FP16) 1192 * 1193 * -fpu=<name> values and their correspondance with FPU architectures above: 1194 * 1195 * {"vfp", FPU_ARCH_VFP_V2}, 1196 * {"vfp9", FPU_ARCH_VFP_V2}, 1197 * {"vfp3", FPU_ARCH_VFP_V3}, // For backwards compatbility. 1198 * {"vfp10", FPU_ARCH_VFP_V2}, 1199 * {"vfp10-r0", FPU_ARCH_VFP_V1}, 1200 * {"vfpxd", FPU_ARCH_VFP_V1xD}, 1201 * {"vfpv2", FPU_ARCH_VFP_V2}, 1202 * {"vfpv3", FPU_ARCH_VFP_V3}, 1203 * {"vfpv3-fp16", FPU_ARCH_VFP_V3_FP16}, 1204 * {"vfpv3-d16", FPU_ARCH_VFP_V3D16}, 1205 * {"vfpv3-d16-fp16", FPU_ARCH_VFP_V3D16_FP16}, 1206 * {"vfpv3xd", FPU_ARCH_VFP_V3xD}, 1207 * {"vfpv3xd-fp16", FPU_ARCH_VFP_V3xD_FP16}, 1208 * {"neon", FPU_ARCH_VFP_V3_PLUS_NEON_V1}, 1209 * {"neon-fp16", FPU_ARCH_NEON_FP16}, 1210 * {"vfpv4", FPU_ARCH_VFP_V4}, 1211 * {"vfpv4-d16", FPU_ARCH_VFP_V4D16}, 1212 * {"fpv4-sp-d16", FPU_ARCH_VFP_V4_SP_D16}, 1213 * {"neon-vfpv4", FPU_ARCH_NEON_VFP_V4}, 1214 * 1215 * 1216 * Simplified diagram that only includes FPUs supported by Android: 1217 * Only ARCH_VFP_V3D16 is actually mandated by the armeabi-v7a ABI, 1218 * all others are optional and must be probed at runtime. 1219 * 1220 * ARCH_VFP_V3D16 (EXT_V1xD+EXT_V1+EXT_V2+EXT_V3xD+EXT_V3) 1221 * | 1222 * +-------------------+ 1223 * | | 1224 * | ARCH_VFP_V3D16_FP16 (+EXT_FP16) 1225 * | 1226 * +-------------------+ 1227 * | | 1228 * | ARCH_VFP_V4_D16 (+EXT_FP16+EXT_FMA) 1229 * | | 1230 * | ARCH_VFP_V4 (+EXT_D32) 1231 * | | 1232 * | ARCH_NEON_VFP_V4 (+EXT_NEON+EXT_NEON_FMA) 1233 * | 1234 * ARCH_VFP_V3 (+EXT_D32) 1235 * | 1236 * +-------------------+ 1237 * | | 1238 * | ARCH_VFP_V3_FP16 (+EXT_FP16) 1239 * | 1240 * ARCH_VFP_V3_PLUS_NEON_V1 (+EXT_NEON) 1241 * | 1242 * ARCH_NEON_FP16 (+EXT_FP16) 1243 * 1244 */ 1245 1246#endif // defined(__le32__) || defined(__le64__) 1247