[x86] Add a feature bit: Fast_Unaligned_Copy
[glibc.git] / sysdeps / x86 / cpu-features.h
blobbfe1f4c68d5aedcc3ca37b71548422e997340305
1 /* This file is part of the GNU C Library.
2 Copyright (C) 2008-2016 Free Software Foundation, Inc.
4 The GNU C Library is free software; you can redistribute it and/or
5 modify it under the terms of the GNU Lesser General Public
6 License as published by the Free Software Foundation; either
7 version 2.1 of the License, or (at your option) any later version.
9 The GNU C Library is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 Lesser General Public License for more details.
14 You should have received a copy of the GNU Lesser General Public
15 License along with the GNU C Library; if not, see
16 <http://www.gnu.org/licenses/>. */
18 #ifndef cpu_features_h
19 #define cpu_features_h
21 #define bit_arch_Fast_Rep_String (1 << 0)
22 #define bit_arch_Fast_Copy_Backward (1 << 1)
23 #define bit_arch_Slow_BSF (1 << 2)
24 #define bit_arch_Fast_Unaligned_Load (1 << 4)
25 #define bit_arch_Prefer_PMINUB_for_stringop (1 << 5)
26 #define bit_arch_AVX_Usable (1 << 6)
27 #define bit_arch_FMA_Usable (1 << 7)
28 #define bit_arch_FMA4_Usable (1 << 8)
29 #define bit_arch_Slow_SSE4_2 (1 << 9)
30 #define bit_arch_AVX2_Usable (1 << 10)
31 #define bit_arch_AVX_Fast_Unaligned_Load (1 << 11)
32 #define bit_arch_AVX512F_Usable (1 << 12)
33 #define bit_arch_AVX512DQ_Usable (1 << 13)
34 #define bit_arch_I586 (1 << 14)
35 #define bit_arch_I686 (1 << 15)
36 #define bit_arch_Prefer_MAP_32BIT_EXEC (1 << 16)
37 #define bit_arch_Prefer_No_VZEROUPPER (1 << 17)
38 #define bit_arch_Fast_Unaligned_Copy (1 << 18)
40 /* CPUID Feature flags. */
42 /* COMMON_CPUID_INDEX_1. */
43 #define bit_cpu_CX8 (1 << 8)
44 #define bit_cpu_CMOV (1 << 15)
45 #define bit_cpu_SSE2 (1 << 26)
46 #define bit_cpu_SSSE3 (1 << 9)
47 #define bit_cpu_SSE4_1 (1 << 19)
48 #define bit_cpu_SSE4_2 (1 << 20)
49 #define bit_cpu_OSXSAVE (1 << 27)
50 #define bit_cpu_AVX (1 << 28)
51 #define bit_cpu_POPCOUNT (1 << 23)
52 #define bit_cpu_FMA (1 << 12)
53 #define bit_cpu_FMA4 (1 << 16)
55 /* COMMON_CPUID_INDEX_7. */
56 #define bit_cpu_RTM (1 << 11)
57 #define bit_cpu_AVX2 (1 << 5)
58 #define bit_cpu_AVX512F (1 << 16)
59 #define bit_cpu_AVX512DQ (1 << 17)
61 /* XCR0 Feature flags. */
62 #define bit_XMM_state (1 << 1)
63 #define bit_YMM_state (2 << 1)
64 #define bit_Opmask_state (1 << 5)
65 #define bit_ZMM0_15_state (1 << 6)
66 #define bit_ZMM16_31_state (1 << 7)
68 /* The integer bit array index for the first set of internal feature bits. */
69 #define FEATURE_INDEX_1 0
71 /* The current maximum size of the feature integer bit array. */
72 #define FEATURE_INDEX_MAX 1
74 #ifdef __ASSEMBLER__
76 # include <ifunc-defines.h>
77 # include <rtld-global-offsets.h>
79 # define index_cpu_CX8 COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_EDX_OFFSET
80 # define index_cpu_CMOV COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_EDX_OFFSET
81 # define index_cpu_SSE2 COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_EDX_OFFSET
82 # define index_cpu_SSSE3 COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_ECX_OFFSET
83 # define index_cpu_SSE4_1 COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_ECX_OFFSET
84 # define index_cpu_SSE4_2 COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_ECX_OFFSET
85 # define index_cpu_AVX COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_ECX_OFFSET
86 # define index_cpu_AVX2 COMMON_CPUID_INDEX_7*CPUID_SIZE+CPUID_EBX_OFFSET
88 # define index_arch_Fast_Rep_String FEATURE_INDEX_1*FEATURE_SIZE
89 # define index_arch_Fast_Copy_Backward FEATURE_INDEX_1*FEATURE_SIZE
90 # define index_arch_Slow_BSF FEATURE_INDEX_1*FEATURE_SIZE
91 # define index_arch_Fast_Unaligned_Load FEATURE_INDEX_1*FEATURE_SIZE
92 # define index_arch_Prefer_PMINUB_for_stringop FEATURE_INDEX_1*FEATURE_SIZE
93 # define index_arch_AVX_Usable FEATURE_INDEX_1*FEATURE_SIZE
94 # define index_arch_FMA_Usable FEATURE_INDEX_1*FEATURE_SIZE
95 # define index_arch_FMA4_Usable FEATURE_INDEX_1*FEATURE_SIZE
96 # define index_arch_Slow_SSE4_2 FEATURE_INDEX_1*FEATURE_SIZE
97 # define index_arch_AVX2_Usable FEATURE_INDEX_1*FEATURE_SIZE
98 # define index_arch_AVX_Fast_Unaligned_Load FEATURE_INDEX_1*FEATURE_SIZE
99 # define index_arch_AVX512F_Usable FEATURE_INDEX_1*FEATURE_SIZE
100 # define index_arch_AVX512DQ_Usable FEATURE_INDEX_1*FEATURE_SIZE
101 # define index_arch_I586 FEATURE_INDEX_1*FEATURE_SIZE
102 # define index_arch_I686 FEATURE_INDEX_1*FEATURE_SIZE
103 # define index_arch_Prefer_MAP_32BIT_EXEC FEATURE_INDEX_1*FEATURE_SIZE
104 # define index_arch_Prefer_No_VZEROUPPER FEATURE_INDEX_1*FEATURE_SIZE
105 # define index_arch_Fast_Unaligned_Copy FEATURE_INDEX_1*FEATURE_SIZE
108 # if defined (_LIBC) && !IS_IN (nonlib)
109 # ifdef __x86_64__
110 # ifdef SHARED
111 # if IS_IN (rtld)
112 # define LOAD_RTLD_GLOBAL_RO_RDX
113 # define HAS_FEATURE(offset, field, name) \
114 testl $(bit_##field##_##name), \
115 _rtld_local_ro+offset+(index_##field##_##name)(%rip)
116 # else
117 # define LOAD_RTLD_GLOBAL_RO_RDX \
118 mov _rtld_global_ro@GOTPCREL(%rip), %RDX_LP
119 # define HAS_FEATURE(offset, field, name) \
120 testl $(bit_##field##_##name), \
121 RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+offset+(index_##field##_##name)(%rdx)
122 # endif
123 # else /* SHARED */
124 # define LOAD_RTLD_GLOBAL_RO_RDX
125 # define HAS_FEATURE(offset, field, name) \
126 testl $(bit_##field##_##name), \
127 _dl_x86_cpu_features+offset+(index_##field##_##name)(%rip)
128 # endif /* !SHARED */
129 # else /* __x86_64__ */
130 # ifdef SHARED
131 # define LOAD_FUNC_GOT_EAX(func) \
132 leal func@GOTOFF(%edx), %eax
133 # if IS_IN (rtld)
134 # define LOAD_GOT_AND_RTLD_GLOBAL_RO \
135 LOAD_PIC_REG(dx)
136 # define HAS_FEATURE(offset, field, name) \
137 testl $(bit_##field##_##name), \
138 offset+(index_##field##_##name)+_rtld_local_ro@GOTOFF(%edx)
139 # else
140 # define LOAD_GOT_AND_RTLD_GLOBAL_RO \
141 LOAD_PIC_REG(dx); \
142 mov _rtld_global_ro@GOT(%edx), %ecx
143 # define HAS_FEATURE(offset, field, name) \
144 testl $(bit_##field##_##name), \
145 RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+offset+(index_##field##_##name)(%ecx)
146 # endif
147 # else /* SHARED */
148 # define LOAD_FUNC_GOT_EAX(func) \
149 leal func, %eax
150 # define LOAD_GOT_AND_RTLD_GLOBAL_RO
151 # define HAS_FEATURE(offset, field, name) \
152 testl $(bit_##field##_##name), \
153 _dl_x86_cpu_features+offset+(index_##field##_##name)
154 # endif /* !SHARED */
155 # endif /* !__x86_64__ */
156 # else /* _LIBC && !nonlib */
157 # error "Sorry, <cpu-features.h> is unimplemented for assembler"
158 # endif /* !_LIBC || nonlib */
160 /* HAS_* evaluates to true if we may use the feature at runtime. */
161 # define HAS_CPU_FEATURE(name) HAS_FEATURE (CPUID_OFFSET, cpu, name)
162 # define HAS_ARCH_FEATURE(name) HAS_FEATURE (FEATURE_OFFSET, arch, name)
164 #else /* __ASSEMBLER__ */
166 enum
168 COMMON_CPUID_INDEX_1 = 0,
169 COMMON_CPUID_INDEX_7,
170 COMMON_CPUID_INDEX_80000001, /* for AMD */
171 /* Keep the following line at the end. */
172 COMMON_CPUID_INDEX_MAX
175 struct cpu_features
177 enum cpu_features_kind
179 arch_kind_unknown = 0,
180 arch_kind_intel,
181 arch_kind_amd,
182 arch_kind_other
183 } kind;
184 int max_cpuid;
185 struct cpuid_registers
187 unsigned int eax;
188 unsigned int ebx;
189 unsigned int ecx;
190 unsigned int edx;
191 } cpuid[COMMON_CPUID_INDEX_MAX];
192 unsigned int family;
193 unsigned int model;
194 unsigned int feature[FEATURE_INDEX_MAX];
197 /* Used from outside of glibc to get access to the CPU features
198 structure. */
199 extern const struct cpu_features *__get_cpu_features (void)
200 __attribute__ ((const));
202 # if defined (_LIBC) && !IS_IN (nonlib)
203 /* Unused for x86. */
204 # define INIT_ARCH()
205 # define __get_cpu_features() (&GLRO(dl_x86_cpu_features))
206 # endif
209 /* Only used directly in cpu-features.c. */
210 # define CPU_FEATURES_CPU_P(ptr, name) \
211 ((ptr->cpuid[index_cpu_##name].reg_##name & (bit_cpu_##name)) != 0)
212 # define CPU_FEATURES_ARCH_P(ptr, name) \
213 ((ptr->feature[index_arch_##name] & (bit_arch_##name)) != 0)
215 /* HAS_* evaluates to true if we may use the feature at runtime. */
216 # define HAS_CPU_FEATURE(name) \
217 CPU_FEATURES_CPU_P (__get_cpu_features (), name)
218 # define HAS_ARCH_FEATURE(name) \
219 CPU_FEATURES_ARCH_P (__get_cpu_features (), name)
221 # define index_cpu_CX8 COMMON_CPUID_INDEX_1
222 # define index_cpu_CMOV COMMON_CPUID_INDEX_1
223 # define index_cpu_SSE2 COMMON_CPUID_INDEX_1
224 # define index_cpu_SSSE3 COMMON_CPUID_INDEX_1
225 # define index_cpu_SSE4_1 COMMON_CPUID_INDEX_1
226 # define index_cpu_SSE4_2 COMMON_CPUID_INDEX_1
227 # define index_cpu_AVX COMMON_CPUID_INDEX_1
228 # define index_cpu_AVX2 COMMON_CPUID_INDEX_7
229 # define index_cpu_AVX512F COMMON_CPUID_INDEX_7
230 # define index_cpu_AVX512DQ COMMON_CPUID_INDEX_7
231 # define index_cpu_RTM COMMON_CPUID_INDEX_7
232 # define index_cpu_FMA COMMON_CPUID_INDEX_1
233 # define index_cpu_FMA4 COMMON_CPUID_INDEX_80000001
234 # define index_cpu_POPCOUNT COMMON_CPUID_INDEX_1
235 # define index_cpu_OSXSAVE COMMON_CPUID_INDEX_1
237 # define reg_CX8 edx
238 # define reg_CMOV edx
239 # define reg_SSE2 edx
240 # define reg_SSSE3 ecx
241 # define reg_SSE4_1 ecx
242 # define reg_SSE4_2 ecx
243 # define reg_AVX ecx
244 # define reg_AVX2 ebx
245 # define reg_AVX512F ebx
246 # define reg_AVX512DQ ebx
247 # define reg_RTM ebx
248 # define reg_FMA ecx
249 # define reg_FMA4 ecx
250 # define reg_POPCOUNT ecx
251 # define reg_OSXSAVE ecx
253 # define index_arch_Fast_Rep_String FEATURE_INDEX_1
254 # define index_arch_Fast_Copy_Backward FEATURE_INDEX_1
255 # define index_arch_Slow_BSF FEATURE_INDEX_1
256 # define index_arch_Fast_Unaligned_Load FEATURE_INDEX_1
257 # define index_arch_Prefer_PMINUB_for_stringop FEATURE_INDEX_1
258 # define index_arch_AVX_Usable FEATURE_INDEX_1
259 # define index_arch_FMA_Usable FEATURE_INDEX_1
260 # define index_arch_FMA4_Usable FEATURE_INDEX_1
261 # define index_arch_Slow_SSE4_2 FEATURE_INDEX_1
262 # define index_arch_AVX2_Usable FEATURE_INDEX_1
263 # define index_arch_AVX_Fast_Unaligned_Load FEATURE_INDEX_1
264 # define index_arch_AVX512F_Usable FEATURE_INDEX_1
265 # define index_arch_AVX512DQ_Usable FEATURE_INDEX_1
266 # define index_arch_I586 FEATURE_INDEX_1
267 # define index_arch_I686 FEATURE_INDEX_1
268 # define index_arch_Prefer_MAP_32BIT_EXEC FEATURE_INDEX_1
269 # define index_arch_Prefer_No_VZEROUPPER FEATURE_INDEX_1
270 # define index_arch_Fast_Unaligned_Copy FEATURE_INDEX_1
272 #endif /* !__ASSEMBLER__ */
274 #ifdef __x86_64__
275 # define HAS_CPUID 1
276 #elif defined __i586__ || defined __pentium__
277 # define HAS_CPUID 1
278 # define HAS_I586 1
279 # define HAS_I686 HAS_ARCH_FEATURE (I686)
280 #elif (defined __i686__ || defined __pentiumpro__ \
281 || defined __pentium4__ || defined __nocona__ \
282 || defined __atom__ || defined __core2__ \
283 || defined __corei7__ || defined __corei7_avx__ \
284 || defined __core_avx2__ || defined __nehalem__ \
285 || defined __sandybridge__ || defined __haswell__ \
286 || defined __knl__ || defined __bonnell__ \
287 || defined __silvermont__ \
288 || defined __k6__ || defined __k8__ \
289 || defined __athlon__ || defined __amdfam10__ \
290 || defined __bdver1__ || defined __bdver2__ \
291 || defined __bdver3__ || defined __bdver4__ \
292 || defined __btver1__ || defined __btver2__)
293 # define HAS_CPUID 1
294 # define HAS_I586 1
295 # define HAS_I686 1
296 #else
297 # define HAS_CPUID 0
298 # define HAS_I586 HAS_ARCH_FEATURE (I586)
299 # define HAS_I686 HAS_ARCH_FEATURE (I686)
300 #endif
302 #endif /* cpu_features_h */