1 | /* |
2 | * kmp_utility.cpp -- Utility routines for the OpenMP support library. |
3 | */ |
4 | |
5 | //===----------------------------------------------------------------------===// |
6 | // |
7 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
8 | // See https://llvm.org/LICENSE.txt for license information. |
9 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "kmp.h" |
14 | #include "kmp_i18n.h" |
15 | #include "kmp_str.h" |
16 | #include "kmp_wrapper_getpid.h" |
17 | #include <float.h> |
18 | |
19 | static const char *unknown = "unknown" ; |
20 | |
21 | #if KMP_ARCH_X86 || KMP_ARCH_X86_64 |
22 | |
23 | /* NOTE: If called before serial_initialize (i.e. from runtime_initialize), then |
24 | the debugging package has not been initialized yet, and only "0" will print |
25 | debugging output since the environment variables have not been read. */ |
26 | |
27 | #ifdef KMP_DEBUG |
28 | static int trace_level = 5; |
29 | #endif |
30 | |
31 | static kmp_uint64 __kmp_parse_frequency( // R: Frequency in Hz. |
32 | char const *frequency // I: Float number and unit: MHz, GHz, or TGz. |
33 | ) { |
34 | |
35 | double value = 0.0; |
36 | char *unit = NULL; |
37 | kmp_uint64 result = 0; /* Zero is a better unknown value than all ones. */ |
38 | |
39 | if (frequency == NULL) { |
40 | return result; |
41 | } |
42 | value = strtod(nptr: frequency, endptr: &unit); |
43 | if (0 < value && |
44 | value <= DBL_MAX) { // Good value (not overflow, underflow, etc). |
45 | if (strcmp(s1: unit, s2: "MHz" ) == 0) { |
46 | value = value * 1.0E+6; |
47 | } else if (strcmp(s1: unit, s2: "GHz" ) == 0) { |
48 | value = value * 1.0E+9; |
49 | } else if (strcmp(s1: unit, s2: "THz" ) == 0) { |
50 | value = value * 1.0E+12; |
51 | } else { // Wrong unit. |
52 | return result; |
53 | } |
54 | result = (kmp_uint64)value; // rounds down |
55 | } |
56 | return result; |
57 | |
58 | } // func __kmp_parse_cpu_frequency |
59 | |
60 | void __kmp_query_cpuid(kmp_cpuinfo_t *p) { |
61 | struct kmp_cpuid buf; |
62 | int max_arg; |
63 | #ifdef KMP_DEBUG |
64 | int cflush_size; |
65 | #endif |
66 | |
67 | p->initialized = 1; |
68 | |
69 | p->flags.sse2 = 1; // Assume SSE2 by default. |
70 | |
71 | __kmp_x86_cpuid(leaf: 0, subleaf: 0, p: &buf); |
72 | |
73 | KA_TRACE(trace_level, |
74 | ("INFO: CPUID %d: EAX=0x%08X EBX=0x%08X ECX=0x%08X EDX=0x%08X\n" , 0, |
75 | buf.eax, buf.ebx, buf.ecx, buf.edx)); |
76 | |
77 | max_arg = buf.eax; |
78 | |
79 | p->apic_id = -1; |
80 | |
81 | if (max_arg >= 1) { |
82 | int i; |
83 | kmp_uint32 t, data[4]; |
84 | |
85 | __kmp_x86_cpuid(leaf: 1, subleaf: 0, p: &buf); |
86 | KA_TRACE(trace_level, |
87 | ("INFO: CPUID %d: EAX=0x%08X EBX=0x%08X ECX=0x%08X EDX=0x%08X\n" , |
88 | 1, buf.eax, buf.ebx, buf.ecx, buf.edx)); |
89 | |
90 | { |
91 | #define get_value(reg, lo, mask) (((reg) >> (lo)) & (mask)) |
92 | |
93 | p->signature = buf.eax; |
94 | p->family = get_value(buf.eax, 20, 0xff) + get_value(buf.eax, 8, 0x0f); |
95 | p->model = |
96 | (get_value(buf.eax, 16, 0x0f) << 4) + get_value(buf.eax, 4, 0x0f); |
97 | p->stepping = get_value(buf.eax, 0, 0x0f); |
98 | |
99 | #undef get_value |
100 | |
101 | KA_TRACE(trace_level, (" family = %d, model = %d, stepping = %d\n" , |
102 | p->family, p->model, p->stepping)); |
103 | } |
104 | |
105 | for (t = buf.ebx, i = 0; i < 4; t >>= 8, ++i) { |
106 | data[i] = (t & 0xff); |
107 | } |
108 | |
109 | p->flags.sse2 = (buf.edx >> 26) & 1; |
110 | |
111 | #ifdef KMP_DEBUG |
112 | |
113 | if ((buf.edx >> 4) & 1) { |
114 | /* TSC - Timestamp Counter Available */ |
115 | KA_TRACE(trace_level, (" TSC" )); |
116 | } |
117 | if ((buf.edx >> 8) & 1) { |
118 | /* CX8 - CMPXCHG8B Instruction Available */ |
119 | KA_TRACE(trace_level, (" CX8" )); |
120 | } |
121 | if ((buf.edx >> 9) & 1) { |
122 | /* APIC - Local APIC Present (multi-processor operation support */ |
123 | KA_TRACE(trace_level, (" APIC" )); |
124 | } |
125 | if ((buf.edx >> 15) & 1) { |
126 | /* CMOV - Conditional MOVe Instruction Available */ |
127 | KA_TRACE(trace_level, (" CMOV" )); |
128 | } |
129 | if ((buf.edx >> 18) & 1) { |
130 | /* PSN - Processor Serial Number Available */ |
131 | KA_TRACE(trace_level, (" PSN" )); |
132 | } |
133 | if ((buf.edx >> 19) & 1) { |
134 | /* CLFLUSH - Cache Flush Instruction Available */ |
135 | cflush_size = |
136 | data[1] * 8; /* Bits 15-08: CLFLUSH line size = 8 (64 bytes) */ |
137 | KA_TRACE(trace_level, (" CLFLUSH(%db)" , cflush_size)); |
138 | } |
139 | if ((buf.edx >> 21) & 1) { |
140 | /* DTES - Debug Trace & EMON Store */ |
141 | KA_TRACE(trace_level, (" DTES" )); |
142 | } |
143 | if ((buf.edx >> 22) & 1) { |
144 | /* ACPI - ACPI Support Available */ |
145 | KA_TRACE(trace_level, (" ACPI" )); |
146 | } |
147 | if ((buf.edx >> 23) & 1) { |
148 | /* MMX - Multimedia Extensions */ |
149 | KA_TRACE(trace_level, (" MMX" )); |
150 | } |
151 | if ((buf.edx >> 25) & 1) { |
152 | /* SSE - SSE Instructions */ |
153 | KA_TRACE(trace_level, (" SSE" )); |
154 | } |
155 | if ((buf.edx >> 26) & 1) { |
156 | /* SSE2 - SSE2 Instructions */ |
157 | KA_TRACE(trace_level, (" SSE2" )); |
158 | } |
159 | if ((buf.edx >> 27) & 1) { |
160 | /* SLFSNP - Self-Snooping Cache */ |
161 | KA_TRACE(trace_level, (" SLFSNP" )); |
162 | } |
163 | #endif /* KMP_DEBUG */ |
164 | |
165 | if ((buf.edx >> 28) & 1) { |
166 | /* Bits 23-16: Logical Processors per Physical Processor (1 for P4) */ |
167 | p->apic_id = data[3]; /* Bits 31-24: Processor Initial APIC ID (X) */ |
168 | KA_TRACE(trace_level, (" HT(%d TPUs)" , data[2])); |
169 | } |
170 | #ifdef KMP_DEBUG |
171 | if ((buf.edx >> 29) & 1) { |
172 | /* ATHROTL - Automatic Throttle Control */ |
173 | KA_TRACE(trace_level, (" ATHROTL" )); |
174 | } |
175 | KA_TRACE(trace_level, (" ]\n" )); |
176 | |
177 | for (i = 2; i <= max_arg; ++i) { |
178 | __kmp_x86_cpuid(leaf: i, subleaf: 0, p: &buf); |
179 | KA_TRACE(trace_level, |
180 | ("INFO: CPUID %d: EAX=0x%08X EBX=0x%08X ECX=0x%08X EDX=0x%08X\n" , |
181 | i, buf.eax, buf.ebx, buf.ecx, buf.edx)); |
182 | } |
183 | #endif |
184 | p->flags.rtm = 0; |
185 | p->flags.hybrid = 0; |
186 | if (max_arg > 7) { |
187 | /* RTM bit CPUID.07:EBX, bit 11 */ |
188 | /* HYRBID bit CPUID.07:EDX, bit 15 */ |
189 | __kmp_x86_cpuid(leaf: 7, subleaf: 0, p: &buf); |
190 | p->flags.rtm = (buf.ebx >> 11) & 1; |
191 | p->flags.hybrid = (buf.edx >> 15) & 1; |
192 | if (p->flags.rtm) { |
193 | KA_TRACE(trace_level, (" RTM" )); |
194 | } |
195 | if (p->flags.hybrid) { |
196 | KA_TRACE(trace_level, (" HYBRID" )); |
197 | } |
198 | } |
199 | } |
200 | |
201 | { // Parse CPU brand string for frequency, saving the string for later. |
202 | int i; |
203 | kmp_cpuid_t *base = (kmp_cpuid_t *)&p->name[0]; |
204 | |
205 | // Get CPU brand string. |
206 | for (i = 0; i < 3; ++i) { |
207 | __kmp_x86_cpuid(leaf: 0x80000002 + i, subleaf: 0, p: base + i); |
208 | } |
209 | p->name[sizeof(p->name) - 1] = 0; // Just in case. ;-) |
210 | KA_TRACE(trace_level, ("cpu brand string: \"%s\"\n" , &p->name[0])); |
211 | |
212 | // Parse frequency. |
213 | p->frequency = __kmp_parse_frequency(frequency: strrchr(s: &p->name[0], c: ' ')); |
214 | KA_TRACE(trace_level, |
215 | ("cpu frequency from brand string: %" KMP_UINT64_SPEC "\n" , |
216 | p->frequency)); |
217 | } |
218 | } |
219 | |
220 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
221 | |
222 | void __kmp_expand_host_name(char *buffer, size_t size) { |
223 | KMP_DEBUG_ASSERT(size >= sizeof(unknown)); |
224 | #if KMP_OS_WINDOWS |
225 | { |
226 | DWORD s = size; |
227 | |
228 | if (!GetComputerNameA(buffer, &s)) |
229 | KMP_STRCPY_S(buffer, size, unknown); |
230 | } |
231 | #elif KMP_OS_WASI |
232 | KMP_STRCPY_S(buffer, size, unknown); |
233 | #else |
234 | buffer[size - 2] = 0; |
235 | if (gethostname(name: buffer, len: size) || buffer[size - 2] != 0) |
236 | KMP_STRCPY_S(buffer, size, unknown); |
237 | #endif |
238 | } |
239 | |
240 | /* Expand the meta characters in the filename: |
241 | * Currently defined characters are: |
242 | * %H the hostname |
243 | * %P the number of threads used. |
244 | * %I the unique identifier for this run. |
245 | */ |
246 | |
247 | void __kmp_expand_file_name(char *result, size_t rlen, char *pattern) { |
248 | char *pos = result, *end = result + rlen - 1; |
249 | char buffer[256]; |
250 | int default_cpu_width = 1; |
251 | int snp_result; |
252 | |
253 | KMP_DEBUG_ASSERT(rlen > 0); |
254 | *end = 0; |
255 | { |
256 | int i; |
257 | for (i = __kmp_xproc; i >= 10; i /= 10, ++default_cpu_width) |
258 | ; |
259 | } |
260 | |
261 | if (pattern != NULL) { |
262 | while (*pattern != '\0' && pos < end) { |
263 | if (*pattern != '%') { |
264 | *pos++ = *pattern++; |
265 | } else { |
266 | char *old_pattern = pattern; |
267 | int width = 1; |
268 | int cpu_width = default_cpu_width; |
269 | |
270 | ++pattern; |
271 | |
272 | if (*pattern >= '0' && *pattern <= '9') { |
273 | width = 0; |
274 | do { |
275 | width = (width * 10) + *pattern++ - '0'; |
276 | } while (*pattern >= '0' && *pattern <= '9'); |
277 | if (width < 0 || width > 1024) |
278 | width = 1; |
279 | |
280 | cpu_width = width; |
281 | } |
282 | |
283 | switch (*pattern) { |
284 | case 'H': |
285 | case 'h': { |
286 | __kmp_expand_host_name(buffer, size: sizeof(buffer)); |
287 | KMP_STRNCPY(dest: pos, src: buffer, n: end - pos + 1); |
288 | if (*end == 0) { |
289 | while (*pos) |
290 | ++pos; |
291 | ++pattern; |
292 | } else |
293 | pos = end; |
294 | } break; |
295 | case 'P': |
296 | case 'p': { |
297 | snp_result = KMP_SNPRINTF(s: pos, maxlen: end - pos + 1, format: "%0*d" , cpu_width, |
298 | __kmp_dflt_team_nth); |
299 | if (snp_result >= 0 && snp_result <= end - pos) { |
300 | while (*pos) |
301 | ++pos; |
302 | ++pattern; |
303 | } else |
304 | pos = end; |
305 | } break; |
306 | case 'I': |
307 | case 'i': { |
308 | pid_t id = getpid(); |
309 | #if (KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && defined(__MINGW32__) |
310 | snp_result = KMP_SNPRINTF(pos, end - pos + 1, "%0*lld" , width, id); |
311 | #else |
312 | snp_result = KMP_SNPRINTF(s: pos, maxlen: end - pos + 1, format: "%0*d" , width, id); |
313 | #endif |
314 | if (snp_result >= 0 && snp_result <= end - pos) { |
315 | while (*pos) |
316 | ++pos; |
317 | ++pattern; |
318 | } else |
319 | pos = end; |
320 | break; |
321 | } |
322 | case '%': { |
323 | *pos++ = '%'; |
324 | ++pattern; |
325 | break; |
326 | } |
327 | default: { |
328 | *pos++ = '%'; |
329 | pattern = old_pattern + 1; |
330 | break; |
331 | } |
332 | } |
333 | } |
334 | } |
335 | /* TODO: How do we get rid of this? */ |
336 | if (*pattern != '\0') |
337 | KMP_FATAL(FileNameTooLong); |
338 | } |
339 | |
340 | *pos = '\0'; |
341 | } |
342 | |
343 | #if !OMPT_SUPPORT |
344 | extern "C" { |
345 | typedef struct ompt_start_tool_result_t ompt_start_tool_result_t; |
346 | // Define symbols expected by VERSION script |
347 | ompt_start_tool_result_t *ompt_start_tool(unsigned int omp_version, |
348 | const char *runtime_version) { |
349 | return nullptr; |
350 | } |
351 | |
352 | void ompt_libomp_connect(ompt_start_tool_result_t *result) { result = nullptr; } |
353 | } |
354 | #endif |
355 | |