1 | /* |
2 | Copyright (c) 2005-2021 Intel Corporation |
3 | |
4 | Licensed under the Apache License, Version 2.0 (the "License"); |
5 | you may not use this file except in compliance with the License. |
6 | You may obtain a copy of the License at |
7 | |
8 | http://www.apache.org/licenses/LICENSE-2.0 |
9 | |
10 | Unless required by applicable law or agreed to in writing, software |
11 | distributed under the License is distributed on an "AS IS" BASIS, |
12 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | See the License for the specific language governing permissions and |
14 | limitations under the License. |
15 | */ |
16 | |
17 | #ifndef __TBB_detail__machine_H |
18 | #define __TBB_detail__machine_H |
19 | |
20 | #include "_config.h" |
21 | #include "_assert.h" |
22 | |
23 | #include <atomic> |
24 | #include <climits> |
25 | #include <cstdint> |
26 | #include <cstddef> |
27 | |
28 | #ifdef _WIN32 |
29 | #include <intrin.h> |
30 | #ifdef __TBBMALLOC_BUILD |
31 | #define WIN32_LEAN_AND_MEAN |
32 | #define NOMINMAX |
33 | #include <windows.h> // SwitchToThread() |
34 | #endif |
35 | #ifdef _MSC_VER |
36 | #if __TBB_x86_64 || __TBB_x86_32 |
37 | #pragma intrinsic(__rdtsc) |
38 | #endif |
39 | #endif |
40 | #endif |
41 | #if __TBB_x86_64 || __TBB_x86_32 |
42 | #include <immintrin.h> // _mm_pause |
43 | #endif |
44 | #if (_WIN32) |
45 | #include <float.h> // _control87 |
46 | #endif |
47 | |
48 | #if __TBB_GLIBCXX_THIS_THREAD_YIELD_BROKEN |
49 | #include <sched.h> // sched_yield |
50 | #else |
51 | #include <thread> // std::this_thread::yield() |
52 | #endif |
53 | |
54 | namespace tbb { |
55 | namespace detail { |
56 | inline namespace d0 { |
57 | |
58 | //-------------------------------------------------------------------------------------------------- |
59 | // Yield implementation |
60 | //-------------------------------------------------------------------------------------------------- |
61 | |
62 | #if __TBB_GLIBCXX_THIS_THREAD_YIELD_BROKEN |
63 | static inline void yield() { |
64 | int err = sched_yield(); |
65 | __TBB_ASSERT_EX(err == 0, "sched_yield has failed" ); |
66 | } |
67 | #elif __TBBMALLOC_BUILD && _WIN32 |
68 | // Use Windows API for yield in tbbmalloc to avoid dependency on C++ runtime with some implementations. |
69 | static inline void yield() { |
70 | SwitchToThread(); |
71 | } |
72 | #else |
73 | using std::this_thread::yield; |
74 | #endif |
75 | |
76 | //-------------------------------------------------------------------------------------------------- |
77 | // atomic_fence implementation |
78 | //-------------------------------------------------------------------------------------------------- |
79 | |
80 | #if _MSC_VER && (__TBB_x86_64 || __TBB_x86_32) |
81 | #pragma intrinsic(_mm_mfence) |
82 | #endif |
83 | |
84 | static inline void atomic_fence(std::memory_order order) { |
85 | #if _MSC_VER && (__TBB_x86_64 || __TBB_x86_32) |
86 | if (order == std::memory_order_seq_cst || |
87 | order == std::memory_order_acq_rel || |
88 | order == std::memory_order_acquire || |
89 | order == std::memory_order_release ) |
90 | { |
91 | _mm_mfence(); |
92 | return; |
93 | } |
94 | #endif /*_MSC_VER && (__TBB_x86_64 || __TBB_x86_32)*/ |
95 | std::atomic_thread_fence(m: order); |
96 | } |
97 | |
98 | //-------------------------------------------------------------------------------------------------- |
99 | // Pause implementation |
100 | //-------------------------------------------------------------------------------------------------- |
101 | |
102 | static inline void machine_pause(int32_t delay) { |
103 | #if __TBB_x86_64 || __TBB_x86_32 |
104 | while (delay-- > 0) { _mm_pause(); } |
105 | #elif __ARM_ARCH_7A__ || __aarch64__ |
106 | while (delay-- > 0) { __asm__ __volatile__("yield" ::: "memory" ); } |
107 | #else /* Generic */ |
108 | (void)delay; // suppress without including _template_helpers.h |
109 | yield(); |
110 | #endif |
111 | } |
112 | |
113 | //////////////////////////////////////////////////////////////////////////////////////////////////// |
114 | // tbb::detail::log2() implementation |
115 | //////////////////////////////////////////////////////////////////////////////////////////////////// |
116 | // TODO: Use log2p1() function that will be available in C++20 standard |
117 | |
118 | #if defined(__GNUC__) || defined(__clang__) |
119 | namespace gnu_builtins { |
120 | inline uintptr_t clz(unsigned int x) { return __builtin_clz(x); } |
121 | inline uintptr_t clz(unsigned long int x) { return __builtin_clzl(x); } |
122 | inline uintptr_t clz(unsigned long long int x) { return __builtin_clzll(x); } |
123 | } |
124 | #elif defined(_MSC_VER) |
125 | #pragma intrinsic(__TBB_W(_BitScanReverse)) |
126 | namespace msvc_intrinsics { |
127 | static inline uintptr_t bit_scan_reverse(uintptr_t i) { |
128 | unsigned long j; |
129 | __TBB_W(_BitScanReverse)( &j, i ); |
130 | return j; |
131 | } |
132 | } |
133 | #endif |
134 | |
135 | template <typename T> |
136 | constexpr std::uintptr_t number_of_bits() { |
137 | return sizeof(T) * CHAR_BIT; |
138 | } |
139 | |
140 | // logarithm is the index of the most significant non-zero bit |
141 | static inline uintptr_t machine_log2(uintptr_t x) { |
142 | #if defined(__GNUC__) || defined(__clang__) |
143 | // If P is a power of 2 and x<P, then (P-1)-x == (P-1) XOR x |
144 | return (number_of_bits<decltype(x)>() - 1) ^ gnu_builtins::clz(x); |
145 | #elif defined(_MSC_VER) |
146 | return msvc_intrinsics::bit_scan_reverse(x); |
147 | #elif __i386__ || __i386 /*for Sun OS*/ || __MINGW32__ |
148 | uintptr_t j, i = x; |
149 | __asm__("bsr %1,%0" : "=r" (j) : "r" (i)); |
150 | return j; |
151 | #elif __powerpc__ || __POWERPC__ |
152 | #if __TBB_WORDSIZE==8 |
153 | __asm__ __volatile__ ("cntlzd %0,%0" : "+r" (x)); |
154 | return 63 - static_cast<intptr_t>(x); |
155 | #else |
156 | __asm__ __volatile__ ("cntlzw %0,%0" : "+r" (x)); |
157 | return 31 - static_cast<intptr_t>(x); |
158 | #endif /*__TBB_WORDSIZE*/ |
159 | #elif __sparc |
160 | uint64_t count; |
161 | // one hot encode |
162 | x |= (x >> 1); |
163 | x |= (x >> 2); |
164 | x |= (x >> 4); |
165 | x |= (x >> 8); |
166 | x |= (x >> 16); |
167 | x |= (x >> 32); |
168 | // count 1's |
169 | __asm__ ("popc %1, %0" : "=r" (count) : "r" (x) ); |
170 | return count - 1; |
171 | #else |
172 | intptr_t result = 0; |
173 | |
174 | if( sizeof(x) > 4 && (uintptr_t tmp = x >> 32) ) { x = tmp; result += 32; } |
175 | if( uintptr_t tmp = x >> 16 ) { x = tmp; result += 16; } |
176 | if( uintptr_t tmp = x >> 8 ) { x = tmp; result += 8; } |
177 | if( uintptr_t tmp = x >> 4 ) { x = tmp; result += 4; } |
178 | if( uintptr_t tmp = x >> 2 ) { x = tmp; result += 2; } |
179 | |
180 | return (x & 2) ? result + 1 : result; |
181 | #endif |
182 | } |
183 | |
184 | //////////////////////////////////////////////////////////////////////////////////////////////////// |
185 | // tbb::detail::reverse_bits() implementation |
186 | //////////////////////////////////////////////////////////////////////////////////////////////////// |
187 | #if TBB_USE_CLANG_BITREVERSE_BUILTINS |
188 | namespace llvm_builtins { |
189 | inline uint8_t builtin_bitreverse(uint8_t x) { return __builtin_bitreverse8 (x); } |
190 | inline uint16_t builtin_bitreverse(uint16_t x) { return __builtin_bitreverse16(x); } |
191 | inline uint32_t builtin_bitreverse(uint32_t x) { return __builtin_bitreverse32(x); } |
192 | inline uint64_t builtin_bitreverse(uint64_t x) { return __builtin_bitreverse64(x); } |
193 | } |
194 | #else // generic |
195 | template<typename T> |
196 | struct reverse { |
197 | static const T byte_table[256]; |
198 | }; |
199 | |
200 | template<typename T> |
201 | const T reverse<T>::byte_table[256] = { |
202 | 0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0, 0x10, 0x90, 0x50, 0xD0, 0x30, 0xB0, 0x70, 0xF0, |
203 | 0x08, 0x88, 0x48, 0xC8, 0x28, 0xA8, 0x68, 0xE8, 0x18, 0x98, 0x58, 0xD8, 0x38, 0xB8, 0x78, 0xF8, |
204 | 0x04, 0x84, 0x44, 0xC4, 0x24, 0xA4, 0x64, 0xE4, 0x14, 0x94, 0x54, 0xD4, 0x34, 0xB4, 0x74, 0xF4, |
205 | 0x0C, 0x8C, 0x4C, 0xCC, 0x2C, 0xAC, 0x6C, 0xEC, 0x1C, 0x9C, 0x5C, 0xDC, 0x3C, 0xBC, 0x7C, 0xFC, |
206 | 0x02, 0x82, 0x42, 0xC2, 0x22, 0xA2, 0x62, 0xE2, 0x12, 0x92, 0x52, 0xD2, 0x32, 0xB2, 0x72, 0xF2, |
207 | 0x0A, 0x8A, 0x4A, 0xCA, 0x2A, 0xAA, 0x6A, 0xEA, 0x1A, 0x9A, 0x5A, 0xDA, 0x3A, 0xBA, 0x7A, 0xFA, |
208 | 0x06, 0x86, 0x46, 0xC6, 0x26, 0xA6, 0x66, 0xE6, 0x16, 0x96, 0x56, 0xD6, 0x36, 0xB6, 0x76, 0xF6, |
209 | 0x0E, 0x8E, 0x4E, 0xCE, 0x2E, 0xAE, 0x6E, 0xEE, 0x1E, 0x9E, 0x5E, 0xDE, 0x3E, 0xBE, 0x7E, 0xFE, |
210 | 0x01, 0x81, 0x41, 0xC1, 0x21, 0xA1, 0x61, 0xE1, 0x11, 0x91, 0x51, 0xD1, 0x31, 0xB1, 0x71, 0xF1, |
211 | 0x09, 0x89, 0x49, 0xC9, 0x29, 0xA9, 0x69, 0xE9, 0x19, 0x99, 0x59, 0xD9, 0x39, 0xB9, 0x79, 0xF9, |
212 | 0x05, 0x85, 0x45, 0xC5, 0x25, 0xA5, 0x65, 0xE5, 0x15, 0x95, 0x55, 0xD5, 0x35, 0xB5, 0x75, 0xF5, |
213 | 0x0D, 0x8D, 0x4D, 0xCD, 0x2D, 0xAD, 0x6D, 0xED, 0x1D, 0x9D, 0x5D, 0xDD, 0x3D, 0xBD, 0x7D, 0xFD, |
214 | 0x03, 0x83, 0x43, 0xC3, 0x23, 0xA3, 0x63, 0xE3, 0x13, 0x93, 0x53, 0xD3, 0x33, 0xB3, 0x73, 0xF3, |
215 | 0x0B, 0x8B, 0x4B, 0xCB, 0x2B, 0xAB, 0x6B, 0xEB, 0x1B, 0x9B, 0x5B, 0xDB, 0x3B, 0xBB, 0x7B, 0xFB, |
216 | 0x07, 0x87, 0x47, 0xC7, 0x27, 0xA7, 0x67, 0xE7, 0x17, 0x97, 0x57, 0xD7, 0x37, 0xB7, 0x77, 0xF7, |
217 | 0x0F, 0x8F, 0x4F, 0xCF, 0x2F, 0xAF, 0x6F, 0xEF, 0x1F, 0x9F, 0x5F, 0xDF, 0x3F, 0xBF, 0x7F, 0xFF |
218 | }; |
219 | |
220 | inline unsigned char reverse_byte(unsigned char src) { |
221 | return reverse<unsigned char>::byte_table[src]; |
222 | } |
223 | #endif // TBB_USE_CLANG_BITREVERSE_BUILTINS |
224 | |
225 | template<typename T> |
226 | T machine_reverse_bits(T src) { |
227 | #if TBB_USE_CLANG_BITREVERSE_BUILTINS |
228 | return builtin_bitreverse(fixed_width_cast(src)); |
229 | #else /* Generic */ |
230 | T dst; |
231 | unsigned char *original = (unsigned char *) &src; |
232 | unsigned char *reversed = (unsigned char *) &dst; |
233 | |
234 | for ( int i = sizeof(T) - 1; i >= 0; i-- ) { |
235 | reversed[i] = reverse_byte( src: original[sizeof(T) - i - 1] ); |
236 | } |
237 | |
238 | return dst; |
239 | #endif // TBB_USE_CLANG_BITREVERSE_BUILTINS |
240 | } |
241 | |
242 | } // inline namespace d0 |
243 | |
244 | namespace d1 { |
245 | |
246 | #if (_WIN32) |
247 | // API to retrieve/update FPU control setting |
248 | #define __TBB_CPU_CTL_ENV_PRESENT 1 |
249 | struct cpu_ctl_env { |
250 | unsigned int x87cw{}; |
251 | #if (__TBB_x86_64) |
252 | // Changing the infinity mode or the floating-point precision is not supported on x64. |
253 | // The attempt causes an assertion. See |
254 | // https://docs.microsoft.com/en-us/cpp/c-runtime-library/reference/control87-controlfp-control87-2 |
255 | static constexpr unsigned int X87CW_CONTROL_MASK = _MCW_DN | _MCW_EM | _MCW_RC; |
256 | #else |
257 | static constexpr unsigned int X87CW_CONTROL_MASK = ~0U; |
258 | #endif |
259 | #if (__TBB_x86_32 || __TBB_x86_64) |
260 | unsigned int mxcsr{}; |
261 | static constexpr unsigned int MXCSR_CONTROL_MASK = ~0x3fu; /* all except last six status bits */ |
262 | #endif |
263 | |
264 | bool operator!=( const cpu_ctl_env& ctl ) const { |
265 | return |
266 | #if (__TBB_x86_32 || __TBB_x86_64) |
267 | mxcsr != ctl.mxcsr || |
268 | #endif |
269 | x87cw != ctl.x87cw; |
270 | } |
271 | void get_env() { |
272 | x87cw = _control87(0, 0); |
273 | #if (__TBB_x86_32 || __TBB_x86_64) |
274 | mxcsr = _mm_getcsr(); |
275 | #endif |
276 | } |
277 | void set_env() const { |
278 | _control87(x87cw, X87CW_CONTROL_MASK); |
279 | #if (__TBB_x86_32 || __TBB_x86_64) |
280 | _mm_setcsr(mxcsr & MXCSR_CONTROL_MASK); |
281 | #endif |
282 | } |
283 | }; |
284 | #elif (__TBB_x86_32 || __TBB_x86_64) |
285 | // API to retrieve/update FPU control setting |
286 | #define __TBB_CPU_CTL_ENV_PRESENT 1 |
287 | struct cpu_ctl_env { |
288 | int mxcsr{}; |
289 | short x87cw{}; |
290 | static const int MXCSR_CONTROL_MASK = ~0x3f; /* all except last six status bits */ |
291 | |
292 | bool operator!=(const cpu_ctl_env& ctl) const { |
293 | return mxcsr != ctl.mxcsr || x87cw != ctl.x87cw; |
294 | } |
295 | void get_env() { |
296 | __asm__ __volatile__( |
297 | "stmxcsr %0\n\t" |
298 | "fstcw %1" |
299 | : "=m" (mxcsr), "=m" (x87cw) |
300 | ); |
301 | mxcsr &= MXCSR_CONTROL_MASK; |
302 | } |
303 | void set_env() const { |
304 | __asm__ __volatile__( |
305 | "ldmxcsr %0\n\t" |
306 | "fldcw %1" |
307 | : : "m" (mxcsr), "m" (x87cw) |
308 | ); |
309 | } |
310 | }; |
311 | #endif |
312 | |
313 | } // namespace d1 |
314 | |
315 | } // namespace detail |
316 | } // namespace tbb |
317 | |
318 | #if !__TBB_CPU_CTL_ENV_PRESENT |
319 | #include <fenv.h> |
320 | |
321 | #include <cstring> |
322 | |
323 | namespace tbb { |
324 | namespace detail { |
325 | |
326 | namespace r1 { |
327 | void* __TBB_EXPORTED_FUNC cache_aligned_allocate(std::size_t size); |
328 | void __TBB_EXPORTED_FUNC cache_aligned_deallocate(void* p); |
329 | } // namespace r1 |
330 | |
331 | namespace d1 { |
332 | |
333 | class cpu_ctl_env { |
334 | fenv_t *my_fenv_ptr; |
335 | public: |
336 | cpu_ctl_env() : my_fenv_ptr(NULL) {} |
337 | ~cpu_ctl_env() { |
338 | if ( my_fenv_ptr ) |
339 | r1::cache_aligned_deallocate( (void*)my_fenv_ptr ); |
340 | } |
341 | // It is possible not to copy memory but just to copy pointers but the following issues should be addressed: |
342 | // 1. The arena lifetime and the context lifetime are independent; |
343 | // 2. The user is allowed to recapture different FPU settings to context so 'current FPU settings' inside |
344 | // dispatch loop may become invalid. |
345 | // But do we really want to improve the fenv implementation? It seems to be better to replace the fenv implementation |
346 | // with a platform specific implementation. |
347 | cpu_ctl_env( const cpu_ctl_env &src ) : my_fenv_ptr(NULL) { |
348 | *this = src; |
349 | } |
350 | cpu_ctl_env& operator=( const cpu_ctl_env &src ) { |
351 | __TBB_ASSERT( src.my_fenv_ptr, NULL ); |
352 | if ( !my_fenv_ptr ) |
353 | my_fenv_ptr = (fenv_t*)r1::cache_aligned_allocate(sizeof(fenv_t)); |
354 | *my_fenv_ptr = *src.my_fenv_ptr; |
355 | return *this; |
356 | } |
357 | bool operator!=( const cpu_ctl_env &ctl ) const { |
358 | __TBB_ASSERT( my_fenv_ptr, "cpu_ctl_env is not initialized." ); |
359 | __TBB_ASSERT( ctl.my_fenv_ptr, "cpu_ctl_env is not initialized." ); |
360 | return std::memcmp( (void*)my_fenv_ptr, (void*)ctl.my_fenv_ptr, sizeof(fenv_t) ); |
361 | } |
362 | void get_env () { |
363 | if ( !my_fenv_ptr ) |
364 | my_fenv_ptr = (fenv_t*)r1::cache_aligned_allocate(sizeof(fenv_t)); |
365 | fegetenv( my_fenv_ptr ); |
366 | } |
367 | const cpu_ctl_env& set_env () const { |
368 | __TBB_ASSERT( my_fenv_ptr, "cpu_ctl_env is not initialized." ); |
369 | fesetenv( my_fenv_ptr ); |
370 | return *this; |
371 | } |
372 | }; |
373 | |
374 | } // namespace d1 |
375 | } // namespace detail |
376 | } // namespace tbb |
377 | |
378 | #endif /* !__TBB_CPU_CTL_ENV_PRESENT */ |
379 | |
380 | #endif // __TBB_detail__machine_H |
381 | |