1/*-
2 * Copyright (c) 2023 The FreeBSD Foundation
3 *
4 * This software was developed by Robert Clausecker <fuz@FreeBSD.org>
5 * under sponsorship from the FreeBSD Foundation.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 *    notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 *    notice, this list of conditions and the following disclaimer in the
14 *    documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ''AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE
27 */
28
29#include <sys/types.h>
30
31#include <machine/atomic.h>
32#include <machine/cpufunc.h>
33#include <machine/specialreg.h>
34
35#include <stddef.h>
36#include <string.h>
37
38#include "amd64_archlevel.h"
39#include "libc_private.h"
40
41#define ARCHLEVEL_ENV	"ARCHLEVEL"
42
43static volatile int amd64_archlevel = X86_64_UNDEFINED;
44
45static const struct archlevel {
46	char name[10];
47	/* CPUID feature bits that need to be present */
48	u_int feat_edx, feat_ecx, amd_ecx, ext_ebx;
49} levels[] = {
50	{
51		.name = "scalar",
52		.feat_edx = 0,
53		.feat_ecx = 0,
54		.amd_ecx = 0,
55		.ext_ebx = 0,
56	}, {
57#define FEAT_EDX_BASELINE (CPUID_FPU | CPUID_CX8 | CPUID_CMOV | CPUID_MMX | \
58	CPUID_FXSR | CPUID_SSE | CPUID_SSE2)
59		.name = "baseline",
60		.feat_edx = FEAT_EDX_BASELINE,
61		.feat_ecx = 0,
62		.amd_ecx = 0,
63		.ext_ebx = 0,
64	}, {
65#define FEAT_ECX_V2 (CPUID2_SSE3 | CPUID2_SSSE3 | CPUID2_CX16 | CPUID2_SSE41 | \
66	CPUID2_SSE42 | CPUID2_POPCNT)
67#define AMD_ECX_V2 AMDID2_LAHF
68		.name = "x86-64-v2",
69		.feat_edx = FEAT_EDX_BASELINE,
70		.feat_ecx = FEAT_ECX_V2,
71		.amd_ecx = AMD_ECX_V2,
72		.ext_ebx = 0,
73	}, {
74#define FEAT_ECX_V3 (FEAT_ECX_V2 | CPUID2_FMA | CPUID2_MOVBE | \
75	CPUID2_OSXSAVE | CPUID2_AVX | CPUID2_F16C)
76#define AMD_ECX_V3 (AMD_ECX_V2 | AMDID2_ABM)
77#define EXT_EBX_V3 (CPUID_STDEXT_BMI1 | CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2)
78		.name = "x86-64-v3",
79		.feat_edx = FEAT_EDX_BASELINE,
80		.feat_ecx = FEAT_ECX_V3,
81		.amd_ecx = AMD_ECX_V3,
82		.ext_ebx = EXT_EBX_V3,
83	}, {
84#define EXT_EBX_V4 (EXT_EBX_V3 | CPUID_STDEXT_AVX512F | \
85	CPUID_STDEXT_AVX512DQ | CPUID_STDEXT_AVX512CD | \
86	CPUID_STDEXT_AVX512BW | CPUID_STDEXT_AVX512VL)
87		.name = "x86-64-v4",
88		.feat_edx = FEAT_EDX_BASELINE,
89		.feat_ecx = FEAT_ECX_V3,
90		.amd_ecx = AMD_ECX_V3,
91		.ext_ebx = EXT_EBX_V4,
92	}
93};
94
95static int
96supported_archlevel(u_int feat_edx, u_int feat_ecx, u_int ext_ebx, u_int ext_ecx)
97{
98	int level;
99	u_int p[4], max_leaf;
100	u_int amd_ecx = 0;
101
102	(void)ext_ecx;
103
104	do_cpuid(0x80000000, p);
105	max_leaf = p[0];
106
107	if (max_leaf >= 0x80000001) {
108		do_cpuid(0x80000001, p);
109		amd_ecx = p[2];
110	}
111
112	for (level = X86_64_BASELINE; level <= X86_64_MAX; level++) {
113		const struct archlevel *lvl = &levels[level];
114
115		if ((lvl->feat_edx & feat_edx) != lvl->feat_edx ||
116		    (lvl->feat_ecx & feat_ecx) != lvl->feat_ecx ||
117		    (lvl->amd_ecx & amd_ecx) != lvl->amd_ecx ||
118		    (lvl->ext_ebx & ext_ebx) != lvl->ext_ebx)
119			return (level - 1);
120	}
121
122	return (X86_64_MAX);
123}
124
125static int
126match_archlevel(const char *str, int *force)
127{
128	int level, want_force = 0;
129
130	*force = 0;
131
132	if (str[0] == '!') {
133		str++;
134		want_force = 1;
135	}
136
137	for (level = 0; level <= X86_64_MAX; level++) {
138		size_t i;
139		const char *candidate = levels[level].name;
140
141		/* can't use strcmp here: would recurse during ifunc resolution */
142		for (i = 0; str[i] == candidate[i]; i++)
143			/* suffixes starting with : or + are ignored for future extensions */
144			if (str[i] == '\0' || str[i] == ':' || str[i] == '+') {
145				if (want_force)
146					*force = 1;
147
148				return (level);
149			}
150	}
151
152	return (X86_64_UNDEFINED);
153}
154
155/*
156 * We can't use getenv(), strcmp(), and a bunch of other functions here as
157 * they may in turn call SIMD-optimised string functions.
158 *
159 * *force is set to 1 if the architecture level is valid and begins with a !
160 * and to 0 otherwise.
161 */
162static int
163env_archlevel(int *force)
164{
165	size_t i;
166
167	if (environ == NULL)
168		return (X86_64_UNDEFINED);
169
170	for (i = 0; environ[i] != NULL; i++) {
171		size_t j;
172
173		for (j = 0; environ[i][j] == ARCHLEVEL_ENV "="[j]; j++)
174			if (environ[i][j] == '=')
175				return (match_archlevel(&environ[i][j + 1], force));
176	}
177
178	*force = 0;
179
180	return (X86_64_UNDEFINED);
181
182}
183
184/*
185 * Determine the architecture level by checking the CPU capabilities
186 * and the environment:
187 *
188 * 1. If environment variable ARCHLEVEL starts with a ! and is followed
189 *    by a valid architecture level, that level is returned.
190 * 2. Else if ARCHLEVEL is set to a valid architecture level that is
191 *    supported by the CPU, that level is returned.
192 * 3. Else the highest architecture level supported by the CPU is
193 *    returned.
194 *
195 * Valid architecture levels are those defined in the levels array.
196 * The architecture level "scalar" indicates that SIMD enhancements
197 * shall not be used.
198 */
199static int
200archlevel(u_int feat_edx, u_int feat_ecx, u_int ext_ebx, u_int ext_ecx)
201{
202	int islevel, wantlevel, hwlevel, force;
203
204	islevel = atomic_load_int(&amd64_archlevel);
205	if (islevel != X86_64_UNDEFINED)
206		return (islevel);
207
208	wantlevel = env_archlevel(&force);
209	if (!force) {
210		hwlevel = supported_archlevel(feat_edx, feat_ecx, ext_ebx, ext_ecx);
211		if (wantlevel == X86_64_UNDEFINED || wantlevel > hwlevel)
212			wantlevel = hwlevel;
213	}
214
215	/*
216	 * Ensure amd64_archlevel is set only once and
217	 * all calls agree on what it was set to.
218	 */
219	if (atomic_cmpset_int(&amd64_archlevel, islevel, wantlevel))
220		return (wantlevel);
221	else
222		return (atomic_load_int(&amd64_archlevel));
223}
224
225/*
226 * Helper function for SIMD ifunc dispatch: select the highest level
227 * implementation up to the current architecture level.
228 */
229dlfunc_t
230__archlevel_resolve(u_int feat_edx, u_int feat_ecx, u_int ext_ebx,
231    u_int ext_ecx, int32_t funcs[static X86_64_MAX + 1])
232{
233	int level;
234
235	for (level = archlevel(feat_edx, feat_ecx, ext_ebx, ext_ecx); level >= 0; level--)
236		if (funcs[level] != 0)
237			return (dlfunc_t)((uintptr_t)funcs + (ptrdiff_t)funcs[level]);
238
239	/* no function is present -- what now? */
240	__builtin_trap();
241}
242