1 /*
2 * @file architecture specific interfaces
3 * @remark Copyright 2008 Intel Corporation
4 * @remark Read the file COPYING
5 * @author Andi Kleen
6 */
7
8 #if (defined(__i386__) || defined(__x86_64__)) && !defined(ANDROID_HOST)
9
10 /* Assume we run on the same host as the profilee */
11
12 #define num_to_mask(x) ((1U << (x)) - 1)
13
cpuid_vendor(char * vnd)14 static inline int cpuid_vendor(char *vnd)
15 {
16 union {
17 struct {
18 unsigned b,d,c;
19 };
20 char v[12];
21 } v;
22 unsigned eax;
23 #ifdef __PIC__
24 __asm__ __volatile__(
25 "pushl %%ebx\n" /* must be preserved due to PIC code */
26 "cpuid\n"
27 "mov %%ebx, 0(%%edi)\n"
28 "mov %%ecx, 4(%%edi)\n"
29 "mov %%edx, 8(%%edi)\n"
30 "popl %%ebx\n"
31 : "=a" (eax)
32 : "a"(0), "D"(v.v)
33 : "%ecx", "%edx"
34 );
35 #else
36 asm("cpuid" : "=a" (eax), "=b" (v.b), "=c" (v.c), "=d" (v.d) : "0" (0));
37 #endif
38 return !strncmp(v.v, vnd, 12);
39 }
40
arch_cpuid_1(int code)41 static inline unsigned arch_cpuid_1(int code)
42 {
43 unsigned val;
44 #ifdef __PIC__
45 __asm__ __volatile__ (
46 "pushl %%ebx\n"
47 "cpuid\n"
48 "popl %%ebx\n"
49 : "=a" (val)
50 : "a" (code)
51 : "ecx", "edx"
52 );
53 #else
54 asm("cpuid" : "=a" (v.eax) : "a" (code) : "ecx","ebx","edx");
55 #endif
56 return val;
57 }
58
cpuid_signature()59 static inline unsigned int cpuid_signature()
60 {
61 return arch_cpuid_1(1);
62 }
63
cpu_model(unsigned int eax)64 static inline unsigned int cpu_model(unsigned int eax)
65 {
66 unsigned model = (eax & 0xf0) >> 4;
67 unsigned ext_model = (eax & 0xf0000) >> 12;
68 return ext_model + model;
69 }
70
cpu_family(unsigned int eax)71 static inline unsigned int cpu_family(unsigned int eax)
72 {
73 unsigned family = (eax & 0xf00) >> 8;
74 unsigned ext_family = (eax & 0xff00000) >> 20;
75 return ext_family + family;
76 }
77
cpu_stepping(unsigned int eax)78 static inline unsigned int cpu_stepping(unsigned int eax)
79 {
80 return (eax & 0xf);
81 }
82
83
84 /* Work around Nehalem spec update AAJ79: CPUID incorrectly indicates
85 unhalted reference cycle architectural event is supported. We assume
86 steppings after C0 report correct data in CPUID. */
workaround_nehalem_aaj79(unsigned * ebx)87 static inline void workaround_nehalem_aaj79(unsigned *ebx)
88 {
89 unsigned eax;
90
91 if (!cpuid_vendor("GenuineIntel"))
92 return;
93 eax = cpuid_signature();
94 if (cpu_family(eax) != 6 || cpu_model(eax) != 26
95 || cpu_stepping(eax) > 4)
96 return;
97 *ebx |= (1 << 2); /* disable unsupported event */
98 }
99
arch_get_filter(op_cpu cpu_type)100 static inline unsigned arch_get_filter(op_cpu cpu_type)
101 {
102 if (op_cpu_base_type(cpu_type) == CPU_ARCH_PERFMON) {
103 unsigned ebx, eax;
104 #ifdef __PIC__
105 __asm__ __volatile__ (
106 "pushl %%ebx\n"
107 "cpuid\n"
108 "mov %%ebx, %%ecx\n"
109 "popl %%ebx"
110 : "=a" (eax), "=c" (ebx)
111 : "a" (0xa)
112 : "edx"
113 );
114 #else
115 asm("cpuid" : "=a" (eax), "=b" (ebx) : "0" (0xa) : "ecx","edx");
116 #endif
117 workaround_nehalem_aaj79(&ebx);
118 return ebx & num_to_mask(eax >> 24);
119 }
120 return -1U;
121 }
122
arch_num_counters(op_cpu cpu_type)123 static inline int arch_num_counters(op_cpu cpu_type)
124 {
125 if (op_cpu_base_type(cpu_type) == CPU_ARCH_PERFMON) {
126 unsigned v = arch_cpuid_1(0xa);
127 return (v >> 8) & 0xff;
128 }
129 return -1;
130 }
131
arch_get_counter_mask(void)132 static inline unsigned arch_get_counter_mask(void)
133 {
134 unsigned v = arch_cpuid_1(0xa);
135 return num_to_mask((v >> 8) & 0xff);
136 }
137
op_cpu_specific_type(op_cpu cpu_type)138 static inline op_cpu op_cpu_specific_type(op_cpu cpu_type)
139 {
140 if (cpu_type == CPU_ARCH_PERFMON) {
141 /* Already know is Intel family 6, so just check the model. */
142 int model = cpu_model(cpuid_signature());
143 switch(model) {
144 case 0x0f:
145 case 0x16:
146 case 0x17:
147 case 0x1d:
148 return CPU_CORE_2;
149 case 0x1a:
150 case 0x1e:
151 case 0x2e:
152 return CPU_CORE_I7;
153 case 0x1c:
154 return CPU_ATOM;
155 case 0x25:
156 return CPU_WESTMERE;
157 }
158 }
159 return cpu_type;
160 }
161
162 #else
163
arch_get_filter(op_cpu cpu_type)164 static inline unsigned arch_get_filter(op_cpu cpu_type)
165 {
166 /* Do something with passed arg to shut up the compiler warning */
167 if (cpu_type != CPU_NO_GOOD)
168 return 0;
169 return 0;
170 }
171
arch_num_counters(op_cpu cpu_type)172 static inline int arch_num_counters(op_cpu cpu_type)
173 {
174 /* Do something with passed arg to shut up the compiler warning */
175 if (cpu_type != CPU_NO_GOOD)
176 return -1;
177 return -1;
178 }
179
arch_get_counter_mask(void)180 static inline unsigned arch_get_counter_mask(void)
181 {
182 return 0;
183 }
184
op_cpu_specific_type(op_cpu cpu_type)185 static inline op_cpu op_cpu_specific_type(op_cpu cpu_type)
186 {
187 return cpu_type;
188 }
189 #endif
190