Run DCE after a LoopFlatten test to reduce spurious output [nfc]
[llvm-project.git] / openmp / runtime / src / kmp_utility.cpp
blob3174bbb7a5b75aa3f51f5a5e1fba0343f2c75199
1 /*
2 * kmp_utility.cpp -- Utility routines for the OpenMP support library.
3 */
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
11 //===----------------------------------------------------------------------===//
13 #include "kmp.h"
14 #include "kmp_i18n.h"
15 #include "kmp_str.h"
16 #include "kmp_wrapper_getpid.h"
17 #include <float.h>
19 static const char *unknown = "unknown";
21 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
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. */
27 #ifdef KMP_DEBUG
28 static int trace_level = 5;
29 #endif
31 /* LOG_ID_BITS = ( 1 + floor( log_2( max( log_per_phy - 1, 1 ))))
32 * APIC_ID = (PHY_ID << LOG_ID_BITS) | LOG_ID
33 * PHY_ID = APIC_ID >> LOG_ID_BITS
35 int __kmp_get_physical_id(int log_per_phy, int apic_id) {
36 int index_lsb, index_msb, temp;
38 if (log_per_phy > 1) {
39 index_lsb = 0;
40 index_msb = 31;
42 temp = log_per_phy;
43 while ((temp & 1) == 0) {
44 temp >>= 1;
45 index_lsb++;
48 temp = log_per_phy;
49 while ((temp & 0x80000000) == 0) {
50 temp <<= 1;
51 index_msb--;
54 /* If >1 bits were set in log_per_phy, choose next higher power of 2 */
55 if (index_lsb != index_msb)
56 index_msb++;
58 return ((int)(apic_id >> index_msb));
61 return apic_id;
65 * LOG_ID_BITS = ( 1 + floor( log_2( max( log_per_phy - 1, 1 ))))
66 * APIC_ID = (PHY_ID << LOG_ID_BITS) | LOG_ID
67 * LOG_ID = APIC_ID & (( 1 << LOG_ID_BITS ) - 1 )
69 int __kmp_get_logical_id(int log_per_phy, int apic_id) {
70 unsigned current_bit;
71 int bits_seen;
73 if (log_per_phy <= 1)
74 return (0);
76 bits_seen = 0;
78 for (current_bit = 1; log_per_phy != 0; current_bit <<= 1) {
79 if (log_per_phy & current_bit) {
80 log_per_phy &= ~current_bit;
81 bits_seen++;
85 /* If exactly 1 bit was set in log_per_phy, choose next lower power of 2 */
86 if (bits_seen == 1) {
87 current_bit >>= 1;
90 return ((int)((current_bit - 1) & apic_id));
93 static kmp_uint64 __kmp_parse_frequency( // R: Frequency in Hz.
94 char const *frequency // I: Float number and unit: MHz, GHz, or TGz.
95 ) {
97 double value = 0.0;
98 char *unit = NULL;
99 kmp_uint64 result = 0; /* Zero is a better unknown value than all ones. */
101 if (frequency == NULL) {
102 return result;
104 value = strtod(frequency, &unit);
105 if (0 < value &&
106 value <= DBL_MAX) { // Good value (not overflow, underflow, etc).
107 if (strcmp(unit, "MHz") == 0) {
108 value = value * 1.0E+6;
109 } else if (strcmp(unit, "GHz") == 0) {
110 value = value * 1.0E+9;
111 } else if (strcmp(unit, "THz") == 0) {
112 value = value * 1.0E+12;
113 } else { // Wrong unit.
114 return result;
116 result = (kmp_uint64)value; // rounds down
118 return result;
120 } // func __kmp_parse_cpu_frequency
122 void __kmp_query_cpuid(kmp_cpuinfo_t *p) {
123 struct kmp_cpuid buf;
124 int max_arg;
125 int log_per_phy;
126 #ifdef KMP_DEBUG
127 int cflush_size;
128 #endif
130 p->initialized = 1;
132 p->flags.sse2 = 1; // Assume SSE2 by default.
134 __kmp_x86_cpuid(0, 0, &buf);
136 KA_TRACE(trace_level,
137 ("INFO: CPUID %d: EAX=0x%08X EBX=0x%08X ECX=0x%08X EDX=0x%08X\n", 0,
138 buf.eax, buf.ebx, buf.ecx, buf.edx));
140 max_arg = buf.eax;
142 p->apic_id = -1;
144 if (max_arg >= 1) {
145 int i;
146 kmp_uint32 t, data[4];
148 __kmp_x86_cpuid(1, 0, &buf);
149 KA_TRACE(trace_level,
150 ("INFO: CPUID %d: EAX=0x%08X EBX=0x%08X ECX=0x%08X EDX=0x%08X\n",
151 1, buf.eax, buf.ebx, buf.ecx, buf.edx));
154 #define get_value(reg, lo, mask) (((reg) >> (lo)) & (mask))
156 p->signature = buf.eax;
157 p->family = get_value(buf.eax, 20, 0xff) + get_value(buf.eax, 8, 0x0f);
158 p->model =
159 (get_value(buf.eax, 16, 0x0f) << 4) + get_value(buf.eax, 4, 0x0f);
160 p->stepping = get_value(buf.eax, 0, 0x0f);
162 #undef get_value
164 KA_TRACE(trace_level, (" family = %d, model = %d, stepping = %d\n",
165 p->family, p->model, p->stepping));
168 for (t = buf.ebx, i = 0; i < 4; t >>= 8, ++i) {
169 data[i] = (t & 0xff);
172 p->flags.sse2 = (buf.edx >> 26) & 1;
174 #ifdef KMP_DEBUG
176 if ((buf.edx >> 4) & 1) {
177 /* TSC - Timestamp Counter Available */
178 KA_TRACE(trace_level, (" TSC"));
180 if ((buf.edx >> 8) & 1) {
181 /* CX8 - CMPXCHG8B Instruction Available */
182 KA_TRACE(trace_level, (" CX8"));
184 if ((buf.edx >> 9) & 1) {
185 /* APIC - Local APIC Present (multi-processor operation support */
186 KA_TRACE(trace_level, (" APIC"));
188 if ((buf.edx >> 15) & 1) {
189 /* CMOV - Conditional MOVe Instruction Available */
190 KA_TRACE(trace_level, (" CMOV"));
192 if ((buf.edx >> 18) & 1) {
193 /* PSN - Processor Serial Number Available */
194 KA_TRACE(trace_level, (" PSN"));
196 if ((buf.edx >> 19) & 1) {
197 /* CLFLUSH - Cache Flush Instruction Available */
198 cflush_size =
199 data[1] * 8; /* Bits 15-08: CLFLUSH line size = 8 (64 bytes) */
200 KA_TRACE(trace_level, (" CLFLUSH(%db)", cflush_size));
202 if ((buf.edx >> 21) & 1) {
203 /* DTES - Debug Trace & EMON Store */
204 KA_TRACE(trace_level, (" DTES"));
206 if ((buf.edx >> 22) & 1) {
207 /* ACPI - ACPI Support Available */
208 KA_TRACE(trace_level, (" ACPI"));
210 if ((buf.edx >> 23) & 1) {
211 /* MMX - Multimedia Extensions */
212 KA_TRACE(trace_level, (" MMX"));
214 if ((buf.edx >> 25) & 1) {
215 /* SSE - SSE Instructions */
216 KA_TRACE(trace_level, (" SSE"));
218 if ((buf.edx >> 26) & 1) {
219 /* SSE2 - SSE2 Instructions */
220 KA_TRACE(trace_level, (" SSE2"));
222 if ((buf.edx >> 27) & 1) {
223 /* SLFSNP - Self-Snooping Cache */
224 KA_TRACE(trace_level, (" SLFSNP"));
226 #endif /* KMP_DEBUG */
228 if ((buf.edx >> 28) & 1) {
229 /* Bits 23-16: Logical Processors per Physical Processor (1 for P4) */
230 log_per_phy = data[2];
231 p->apic_id = data[3]; /* Bits 31-24: Processor Initial APIC ID (X) */
232 KA_TRACE(trace_level, (" HT(%d TPUs)", log_per_phy));
233 p->physical_id = __kmp_get_physical_id(log_per_phy, p->apic_id);
234 p->logical_id = __kmp_get_logical_id(log_per_phy, p->apic_id);
236 #ifdef KMP_DEBUG
237 if ((buf.edx >> 29) & 1) {
238 /* ATHROTL - Automatic Throttle Control */
239 KA_TRACE(trace_level, (" ATHROTL"));
241 KA_TRACE(trace_level, (" ]\n"));
243 for (i = 2; i <= max_arg; ++i) {
244 __kmp_x86_cpuid(i, 0, &buf);
245 KA_TRACE(trace_level,
246 ("INFO: CPUID %d: EAX=0x%08X EBX=0x%08X ECX=0x%08X EDX=0x%08X\n",
247 i, buf.eax, buf.ebx, buf.ecx, buf.edx));
249 #endif
250 p->flags.rtm = 0;
251 p->flags.hybrid = 0;
252 if (max_arg > 7) {
253 /* RTM bit CPUID.07:EBX, bit 11 */
254 /* HYRBID bit CPUID.07:EDX, bit 15 */
255 __kmp_x86_cpuid(7, 0, &buf);
256 p->flags.rtm = (buf.ebx >> 11) & 1;
257 p->flags.hybrid = (buf.edx >> 15) & 1;
258 if (p->flags.rtm) {
259 KA_TRACE(trace_level, (" RTM"));
261 if (p->flags.hybrid) {
262 KA_TRACE(trace_level, (" HYBRID"));
267 { // Parse CPU brand string for frequency, saving the string for later.
268 int i;
269 kmp_cpuid_t *base = (kmp_cpuid_t *)&p->name[0];
271 // Get CPU brand string.
272 for (i = 0; i < 3; ++i) {
273 __kmp_x86_cpuid(0x80000002 + i, 0, base + i);
275 p->name[sizeof(p->name) - 1] = 0; // Just in case. ;-)
276 KA_TRACE(trace_level, ("cpu brand string: \"%s\"\n", &p->name[0]));
278 // Parse frequency.
279 p->frequency = __kmp_parse_frequency(strrchr(&p->name[0], ' '));
280 KA_TRACE(trace_level,
281 ("cpu frequency from brand string: %" KMP_UINT64_SPEC "\n",
282 p->frequency));
286 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
288 void __kmp_expand_host_name(char *buffer, size_t size) {
289 KMP_DEBUG_ASSERT(size >= sizeof(unknown));
290 #if KMP_OS_WINDOWS
292 DWORD s = size;
294 if (!GetComputerNameA(buffer, &s))
295 KMP_STRCPY_S(buffer, size, unknown);
297 #else
298 buffer[size - 2] = 0;
299 if (gethostname(buffer, size) || buffer[size - 2] != 0)
300 KMP_STRCPY_S(buffer, size, unknown);
301 #endif
304 /* Expand the meta characters in the filename:
305 * Currently defined characters are:
306 * %H the hostname
307 * %P the number of threads used.
308 * %I the unique identifier for this run.
311 void __kmp_expand_file_name(char *result, size_t rlen, char *pattern) {
312 char *pos = result, *end = result + rlen - 1;
313 char buffer[256];
314 int default_cpu_width = 1;
315 int snp_result;
317 KMP_DEBUG_ASSERT(rlen > 0);
318 *end = 0;
320 int i;
321 for (i = __kmp_xproc; i >= 10; i /= 10, ++default_cpu_width)
325 if (pattern != NULL) {
326 while (*pattern != '\0' && pos < end) {
327 if (*pattern != '%') {
328 *pos++ = *pattern++;
329 } else {
330 char *old_pattern = pattern;
331 int width = 1;
332 int cpu_width = default_cpu_width;
334 ++pattern;
336 if (*pattern >= '0' && *pattern <= '9') {
337 width = 0;
338 do {
339 width = (width * 10) + *pattern++ - '0';
340 } while (*pattern >= '0' && *pattern <= '9');
341 if (width < 0 || width > 1024)
342 width = 1;
344 cpu_width = width;
347 switch (*pattern) {
348 case 'H':
349 case 'h': {
350 __kmp_expand_host_name(buffer, sizeof(buffer));
351 KMP_STRNCPY(pos, buffer, end - pos + 1);
352 if (*end == 0) {
353 while (*pos)
354 ++pos;
355 ++pattern;
356 } else
357 pos = end;
358 } break;
359 case 'P':
360 case 'p': {
361 snp_result = KMP_SNPRINTF(pos, end - pos + 1, "%0*d", cpu_width,
362 __kmp_dflt_team_nth);
363 if (snp_result >= 0 && snp_result <= end - pos) {
364 while (*pos)
365 ++pos;
366 ++pattern;
367 } else
368 pos = end;
369 } break;
370 case 'I':
371 case 'i': {
372 pid_t id = getpid();
373 #if (KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && defined(__MINGW32__)
374 snp_result = KMP_SNPRINTF(pos, end - pos + 1, "%0*lld", width, id);
375 #else
376 snp_result = KMP_SNPRINTF(pos, end - pos + 1, "%0*d", width, id);
377 #endif
378 if (snp_result >= 0 && snp_result <= end - pos) {
379 while (*pos)
380 ++pos;
381 ++pattern;
382 } else
383 pos = end;
384 break;
386 case '%': {
387 *pos++ = '%';
388 ++pattern;
389 break;
391 default: {
392 *pos++ = '%';
393 pattern = old_pattern + 1;
394 break;
399 /* TODO: How do we get rid of this? */
400 if (*pattern != '\0')
401 KMP_FATAL(FileNameTooLong);
404 *pos = '\0';
407 #if !OMPT_SUPPORT
408 extern "C" {
409 typedef struct ompt_start_tool_result_t ompt_start_tool_result_t;
410 // Define symbols expected by VERSION script
411 ompt_start_tool_result_t *ompt_start_tool(unsigned int omp_version,
412 const char *runtime_version) {
413 return nullptr;
416 void ompt_libomp_connect(ompt_start_tool_result_t *result) { result = nullptr; }
418 #endif