xref: /aosp_15_r20/external/flac/src/libFLAC/cpu.c (revision 600f14f40d737144c998e2ec7a483122d3776fbc)
1*600f14f4SXin Li /* libFLAC - Free Lossless Audio Codec library
2*600f14f4SXin Li  * Copyright (C) 2001-2009  Josh Coalson
3*600f14f4SXin Li  * Copyright (C) 2011-2023  Xiph.Org Foundation
4*600f14f4SXin Li  *
5*600f14f4SXin Li  * Redistribution and use in source and binary forms, with or without
6*600f14f4SXin Li  * modification, are permitted provided that the following conditions
7*600f14f4SXin Li  * are met:
8*600f14f4SXin Li  *
9*600f14f4SXin Li  * - Redistributions of source code must retain the above copyright
10*600f14f4SXin Li  * notice, this list of conditions and the following disclaimer.
11*600f14f4SXin Li  *
12*600f14f4SXin Li  * - Redistributions in binary form must reproduce the above copyright
13*600f14f4SXin Li  * notice, this list of conditions and the following disclaimer in the
14*600f14f4SXin Li  * documentation and/or other materials provided with the distribution.
15*600f14f4SXin Li  *
16*600f14f4SXin Li  * - Neither the name of the Xiph.org Foundation nor the names of its
17*600f14f4SXin Li  * contributors may be used to endorse or promote products derived from
18*600f14f4SXin Li  * this software without specific prior written permission.
19*600f14f4SXin Li  *
20*600f14f4SXin Li  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21*600f14f4SXin Li  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22*600f14f4SXin Li  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23*600f14f4SXin Li  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
24*600f14f4SXin Li  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
25*600f14f4SXin Li  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
26*600f14f4SXin Li  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
27*600f14f4SXin Li  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
28*600f14f4SXin Li  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
29*600f14f4SXin Li  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
30*600f14f4SXin Li  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31*600f14f4SXin Li  */
32*600f14f4SXin Li 
33*600f14f4SXin Li #ifdef HAVE_CONFIG_H
34*600f14f4SXin Li #  include <config.h>
35*600f14f4SXin Li #endif
36*600f14f4SXin Li 
37*600f14f4SXin Li #include "private/cpu.h"
38*600f14f4SXin Li #include "share/compat.h"
39*600f14f4SXin Li #include <stdlib.h>
40*600f14f4SXin Li #include <string.h>
41*600f14f4SXin Li 
42*600f14f4SXin Li #if defined _MSC_VER
43*600f14f4SXin Li #include <intrin.h> /* for __cpuid() and _xgetbv() */
44*600f14f4SXin Li #elif defined __GNUC__ && defined HAVE_CPUID_H
45*600f14f4SXin Li #include <cpuid.h> /* for __get_cpuid() and __get_cpuid_max() */
46*600f14f4SXin Li #endif
47*600f14f4SXin Li 
48*600f14f4SXin Li #ifndef NDEBUG
49*600f14f4SXin Li #include <stdio.h>
50*600f14f4SXin Li #define dfprintf fprintf
51*600f14f4SXin Li #else
52*600f14f4SXin Li /* This is bad practice, it should be a static void empty function */
53*600f14f4SXin Li #define dfprintf(file, format, ...)
54*600f14f4SXin Li #endif
55*600f14f4SXin Li 
56*600f14f4SXin Li #if defined(HAVE_SYS_AUXV_H)
57*600f14f4SXin Li #include <sys/auxv.h>
58*600f14f4SXin Li #endif
59*600f14f4SXin Li 
60*600f14f4SXin Li #if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN && !defined FLAC__NO_ASM
61*600f14f4SXin Li 
62*600f14f4SXin Li /* these are flags in EDX of CPUID AX=00000001 */
63*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_CMOV    = 0x00008000;
64*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_MMX     = 0x00800000;
65*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_SSE     = 0x02000000;
66*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_SSE2    = 0x04000000;
67*600f14f4SXin Li 
68*600f14f4SXin Li /* these are flags in ECX of CPUID AX=00000001 */
69*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_SSE3    = 0x00000001;
70*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_SSSE3   = 0x00000200;
71*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_SSE41   = 0x00080000;
72*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_SSE42   = 0x00100000;
73*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_OSXSAVE = 0x08000000;
74*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_AVX     = 0x10000000;
75*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_FMA     = 0x00001000;
76*600f14f4SXin Li 
77*600f14f4SXin Li /* these are flags in EBX of CPUID AX=00000007 */
78*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_AVX2    = 0x00000020;
79*600f14f4SXin Li static const uint32_t FLAC__CPUINFO_X86_CPUID_BMI2    = 0x00000100;
80*600f14f4SXin Li 
81*600f14f4SXin Li static uint32_t
cpu_xgetbv_x86(void)82*600f14f4SXin Li cpu_xgetbv_x86(void)
83*600f14f4SXin Li {
84*600f14f4SXin Li #if (defined _MSC_VER || defined __INTEL_COMPILER) && FLAC__AVX_SUPPORTED
85*600f14f4SXin Li 	return (uint32_t)_xgetbv(0);
86*600f14f4SXin Li #elif defined __GNUC__
87*600f14f4SXin Li 	uint32_t lo, hi;
88*600f14f4SXin Li 	__asm__ volatile (".byte 0x0f, 0x01, 0xd0" : "=a"(lo), "=d"(hi) : "c" (0));
89*600f14f4SXin Li 	return lo;
90*600f14f4SXin Li #else
91*600f14f4SXin Li 	return 0;
92*600f14f4SXin Li #endif
93*600f14f4SXin Li }
94*600f14f4SXin Li 
95*600f14f4SXin Li static uint32_t
cpu_have_cpuid(void)96*600f14f4SXin Li cpu_have_cpuid(void)
97*600f14f4SXin Li {
98*600f14f4SXin Li #if defined FLAC__CPU_X86_64 || defined __i686__ || defined __SSE__ || (defined _M_IX86_FP && _M_IX86_FP > 0)
99*600f14f4SXin Li 	/* target CPU does have CPUID instruction */
100*600f14f4SXin Li 	return 1;
101*600f14f4SXin Li #elif defined __GNUC__ && defined HAVE_CPUID_H
102*600f14f4SXin Li 	if (__get_cpuid_max(0, 0) != 0)
103*600f14f4SXin Li 		return 1;
104*600f14f4SXin Li 	else
105*600f14f4SXin Li 		return 0;
106*600f14f4SXin Li #elif defined _MSC_VER
107*600f14f4SXin Li 	FLAC__uint32 flags1, flags2;
108*600f14f4SXin Li 	__asm {
109*600f14f4SXin Li 		pushfd
110*600f14f4SXin Li 		pushfd
111*600f14f4SXin Li 		pop		eax
112*600f14f4SXin Li 		mov		flags1, eax
113*600f14f4SXin Li 		xor		eax, 0x200000
114*600f14f4SXin Li 		push	eax
115*600f14f4SXin Li 		popfd
116*600f14f4SXin Li 		pushfd
117*600f14f4SXin Li 		pop		eax
118*600f14f4SXin Li 		mov		flags2, eax
119*600f14f4SXin Li 		popfd
120*600f14f4SXin Li 	}
121*600f14f4SXin Li 	if (((flags1^flags2) & 0x200000) != 0)
122*600f14f4SXin Li 		return 1;
123*600f14f4SXin Li 	else
124*600f14f4SXin Li 		return 0;
125*600f14f4SXin Li #else
126*600f14f4SXin Li 	return 0;
127*600f14f4SXin Li #endif
128*600f14f4SXin Li }
129*600f14f4SXin Li 
130*600f14f4SXin Li static void
cpuinfo_x86(FLAC__uint32 level,FLAC__uint32 * eax,FLAC__uint32 * ebx,FLAC__uint32 * ecx,FLAC__uint32 * edx)131*600f14f4SXin Li cpuinfo_x86(FLAC__uint32 level, FLAC__uint32 *eax, FLAC__uint32 *ebx, FLAC__uint32 *ecx, FLAC__uint32 *edx)
132*600f14f4SXin Li {
133*600f14f4SXin Li #if defined _MSC_VER
134*600f14f4SXin Li 	int cpuinfo[4];
135*600f14f4SXin Li 	int ext = level & 0x80000000;
136*600f14f4SXin Li 	__cpuid(cpuinfo, ext);
137*600f14f4SXin Li 	if ((uint32_t)cpuinfo[0] >= level) {
138*600f14f4SXin Li #if FLAC__AVX_SUPPORTED
139*600f14f4SXin Li 		__cpuidex(cpuinfo, level, 0); /* for AVX2 detection */
140*600f14f4SXin Li #else
141*600f14f4SXin Li 		__cpuid(cpuinfo, level); /* some old compilers don't support __cpuidex */
142*600f14f4SXin Li #endif
143*600f14f4SXin Li 		*eax = cpuinfo[0]; *ebx = cpuinfo[1]; *ecx = cpuinfo[2]; *edx = cpuinfo[3];
144*600f14f4SXin Li 		return;
145*600f14f4SXin Li 	}
146*600f14f4SXin Li #elif defined __GNUC__ && defined HAVE_CPUID_H
147*600f14f4SXin Li 	FLAC__uint32 ext = level & 0x80000000;
148*600f14f4SXin Li 	__cpuid(ext, *eax, *ebx, *ecx, *edx);
149*600f14f4SXin Li 	if (*eax >= level) {
150*600f14f4SXin Li 		__cpuid_count(level, 0, *eax, *ebx, *ecx, *edx);
151*600f14f4SXin Li 		return;
152*600f14f4SXin Li 	}
153*600f14f4SXin Li #endif
154*600f14f4SXin Li 	*eax = *ebx = *ecx = *edx = 0;
155*600f14f4SXin Li }
156*600f14f4SXin Li 
157*600f14f4SXin Li #endif
158*600f14f4SXin Li 
159*600f14f4SXin Li static void
x86_cpu_info(FLAC__CPUInfo * info)160*600f14f4SXin Li x86_cpu_info (FLAC__CPUInfo *info)
161*600f14f4SXin Li {
162*600f14f4SXin Li #if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN && !defined FLAC__NO_ASM
163*600f14f4SXin Li 	FLAC__bool x86_osxsave = false;
164*600f14f4SXin Li 	FLAC__bool os_avx = false;
165*600f14f4SXin Li 	FLAC__uint32 flags_eax, flags_ebx, flags_ecx, flags_edx;
166*600f14f4SXin Li 
167*600f14f4SXin Li 	info->use_asm = true; /* we assume a minimum of 80386 */
168*600f14f4SXin Li 	if (!cpu_have_cpuid())
169*600f14f4SXin Li 		return;
170*600f14f4SXin Li 
171*600f14f4SXin Li 	cpuinfo_x86(0, &flags_eax, &flags_ebx, &flags_ecx, &flags_edx);
172*600f14f4SXin Li 	info->x86.intel = (flags_ebx == 0x756E6547 && flags_edx == 0x49656E69 && flags_ecx == 0x6C65746E) ? true : false; /* GenuineIntel */
173*600f14f4SXin Li 	cpuinfo_x86(1, &flags_eax, &flags_ebx, &flags_ecx, &flags_edx);
174*600f14f4SXin Li 
175*600f14f4SXin Li 	info->x86.cmov  = (flags_edx & FLAC__CPUINFO_X86_CPUID_CMOV ) ? true : false;
176*600f14f4SXin Li 	info->x86.mmx   = (flags_edx & FLAC__CPUINFO_X86_CPUID_MMX  ) ? true : false;
177*600f14f4SXin Li 	info->x86.sse   = (flags_edx & FLAC__CPUINFO_X86_CPUID_SSE  ) ? true : false;
178*600f14f4SXin Li 	info->x86.sse2  = (flags_edx & FLAC__CPUINFO_X86_CPUID_SSE2 ) ? true : false;
179*600f14f4SXin Li 	info->x86.sse3  = (flags_ecx & FLAC__CPUINFO_X86_CPUID_SSE3 ) ? true : false;
180*600f14f4SXin Li 	info->x86.ssse3 = (flags_ecx & FLAC__CPUINFO_X86_CPUID_SSSE3) ? true : false;
181*600f14f4SXin Li 	info->x86.sse41 = (flags_ecx & FLAC__CPUINFO_X86_CPUID_SSE41) ? true : false;
182*600f14f4SXin Li 	info->x86.sse42 = (flags_ecx & FLAC__CPUINFO_X86_CPUID_SSE42) ? true : false;
183*600f14f4SXin Li 
184*600f14f4SXin Li 	if (FLAC__AVX_SUPPORTED) {
185*600f14f4SXin Li 		x86_osxsave     = (flags_ecx & FLAC__CPUINFO_X86_CPUID_OSXSAVE) ? true : false;
186*600f14f4SXin Li 		info->x86.avx   = (flags_ecx & FLAC__CPUINFO_X86_CPUID_AVX    ) ? true : false;
187*600f14f4SXin Li 		info->x86.fma   = (flags_ecx & FLAC__CPUINFO_X86_CPUID_FMA    ) ? true : false;
188*600f14f4SXin Li 		cpuinfo_x86(7, &flags_eax, &flags_ebx, &flags_ecx, &flags_edx);
189*600f14f4SXin Li 		info->x86.avx2  = (flags_ebx & FLAC__CPUINFO_X86_CPUID_AVX2   ) ? true : false;
190*600f14f4SXin Li 		info->x86.bmi2  = (flags_ebx & FLAC__CPUINFO_X86_CPUID_BMI2   ) ? true : false;
191*600f14f4SXin Li 	}
192*600f14f4SXin Li 
193*600f14f4SXin Li #if defined FLAC__CPU_IA32
194*600f14f4SXin Li 	dfprintf(stderr, "CPU info (IA-32):\n");
195*600f14f4SXin Li #else
196*600f14f4SXin Li 	dfprintf(stderr, "CPU info (x86-64):\n");
197*600f14f4SXin Li #endif
198*600f14f4SXin Li 	dfprintf(stderr, "  CMOV ....... %c\n", info->x86.cmov    ? 'Y' : 'n');
199*600f14f4SXin Li 	dfprintf(stderr, "  MMX ........ %c\n", info->x86.mmx     ? 'Y' : 'n');
200*600f14f4SXin Li 	dfprintf(stderr, "  SSE ........ %c\n", info->x86.sse     ? 'Y' : 'n');
201*600f14f4SXin Li 	dfprintf(stderr, "  SSE2 ....... %c\n", info->x86.sse2    ? 'Y' : 'n');
202*600f14f4SXin Li 	dfprintf(stderr, "  SSE3 ....... %c\n", info->x86.sse3    ? 'Y' : 'n');
203*600f14f4SXin Li 	dfprintf(stderr, "  SSSE3 ...... %c\n", info->x86.ssse3   ? 'Y' : 'n');
204*600f14f4SXin Li 	dfprintf(stderr, "  SSE41 ...... %c\n", info->x86.sse41   ? 'Y' : 'n');
205*600f14f4SXin Li 	dfprintf(stderr, "  SSE42 ...... %c\n", info->x86.sse42   ? 'Y' : 'n');
206*600f14f4SXin Li 
207*600f14f4SXin Li 	if (FLAC__AVX_SUPPORTED) {
208*600f14f4SXin Li 		dfprintf(stderr, "  AVX ........ %c\n", info->x86.avx     ? 'Y' : 'n');
209*600f14f4SXin Li 		dfprintf(stderr, "  FMA ........ %c\n", info->x86.fma     ? 'Y' : 'n');
210*600f14f4SXin Li 		dfprintf(stderr, "  AVX2 ....... %c\n", info->x86.avx2    ? 'Y' : 'n');
211*600f14f4SXin Li 		dfprintf(stderr, "  BMI2 ....... %c\n", info->x86.bmi2    ? 'Y' : 'n');
212*600f14f4SXin Li 	}
213*600f14f4SXin Li 
214*600f14f4SXin Li 	/*
215*600f14f4SXin Li 	 * now have to check for OS support of AVX instructions
216*600f14f4SXin Li 	 */
217*600f14f4SXin Li 	if (FLAC__AVX_SUPPORTED && info->x86.avx && x86_osxsave && (cpu_xgetbv_x86() & 0x6) == 0x6) {
218*600f14f4SXin Li 		os_avx = true;
219*600f14f4SXin Li 	}
220*600f14f4SXin Li 	if (os_avx) {
221*600f14f4SXin Li 		dfprintf(stderr, "  AVX OS sup . %c\n", info->x86.avx ? 'Y' : 'n');
222*600f14f4SXin Li 	}
223*600f14f4SXin Li 	if (!os_avx) {
224*600f14f4SXin Li 		/* no OS AVX support */
225*600f14f4SXin Li 		info->x86.avx     = false;
226*600f14f4SXin Li 		info->x86.avx2    = false;
227*600f14f4SXin Li 		info->x86.fma     = false;
228*600f14f4SXin Li 	}
229*600f14f4SXin Li #else
230*600f14f4SXin Li 	info->use_asm = false;
231*600f14f4SXin Li #endif
232*600f14f4SXin Li }
233*600f14f4SXin Li 
FLAC__cpu_info(FLAC__CPUInfo * info)234*600f14f4SXin Li void FLAC__cpu_info (FLAC__CPUInfo *info)
235*600f14f4SXin Li {
236*600f14f4SXin Li 	memset(info, 0, sizeof(*info));
237*600f14f4SXin Li 
238*600f14f4SXin Li #ifdef FLAC__CPU_IA32
239*600f14f4SXin Li 	info->type = FLAC__CPUINFO_TYPE_IA32;
240*600f14f4SXin Li #elif defined FLAC__CPU_X86_64
241*600f14f4SXin Li 	info->type = FLAC__CPUINFO_TYPE_X86_64;
242*600f14f4SXin Li #else
243*600f14f4SXin Li 	info->type = FLAC__CPUINFO_TYPE_UNKNOWN;
244*600f14f4SXin Li #endif
245*600f14f4SXin Li 
246*600f14f4SXin Li 	switch (info->type) {
247*600f14f4SXin Li 	case FLAC__CPUINFO_TYPE_IA32: /* fallthrough */
248*600f14f4SXin Li 	case FLAC__CPUINFO_TYPE_X86_64:
249*600f14f4SXin Li 		x86_cpu_info (info);
250*600f14f4SXin Li 		break;
251*600f14f4SXin Li 	default:
252*600f14f4SXin Li 		info->use_asm = false;
253*600f14f4SXin Li 		break;
254*600f14f4SXin Li 	}
255*600f14f4SXin Li }
256