| 1 | /* hist.c - Histogram related operations. |
| 2 | |
| 3 | Copyright 1999, 2000, 2001, 2002, 2004, 2005, 2007, 2009 |
| 4 | Free Software Foundation, Inc. |
| 5 | |
| 6 | This file is part of GNU Binutils. |
| 7 | |
| 8 | This program is free software; you can redistribute it and/or modify |
| 9 | it under the terms of the GNU General Public License as published by |
| 10 | the Free Software Foundation; either version 3 of the License, or |
| 11 | (at your option) any later version. |
| 12 | |
| 13 | This program is distributed in the hope that it will be useful, |
| 14 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 | GNU General Public License for more details. |
| 17 | |
| 18 | You should have received a copy of the GNU General Public License |
| 19 | along with this program; if not, write to the Free Software |
| 20 | Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston, MA |
| 21 | 02110-1301, USA. */ |
| 22 | \f |
| 23 | #include "gprof.h" |
| 24 | #include "libiberty.h" |
| 25 | #include "search_list.h" |
| 26 | #include "source.h" |
| 27 | #include "symtab.h" |
| 28 | #include "corefile.h" |
| 29 | #include "gmon_io.h" |
| 30 | #include "gmon_out.h" |
| 31 | #include "hist.h" |
| 32 | #include "sym_ids.h" |
| 33 | #include "utils.h" |
| 34 | #include "math.h" |
| 35 | #include "stdio.h" |
| 36 | #include "stdlib.h" |
| 37 | |
| 38 | #define UNITS_TO_CODE (offset_to_code / sizeof(UNIT)) |
| 39 | |
| 40 | static void scale_and_align_entries (void); |
| 41 | static void print_header (int); |
| 42 | static void print_line (Sym *, double); |
| 43 | static int cmp_time (const PTR, const PTR); |
| 44 | |
| 45 | /* Declarations of automatically generated functions to output blurbs. */ |
| 46 | extern void flat_blurb (FILE * fp); |
| 47 | |
| 48 | static histogram *find_histogram (bfd_vma lowpc, bfd_vma highpc); |
| 49 | static histogram *find_histogram_for_pc (bfd_vma pc); |
| 50 | |
| 51 | histogram * histograms; |
| 52 | unsigned num_histograms; |
| 53 | double hist_scale; |
| 54 | static char hist_dimension[16] = "seconds"; |
| 55 | static char hist_dimension_abbrev = 's'; |
| 56 | |
| 57 | static double accum_time; /* Accumulated time so far for print_line(). */ |
| 58 | static double total_time; /* Total time for all routines. */ |
| 59 | |
| 60 | /* Table of SI prefixes for powers of 10 (used to automatically |
| 61 | scale some of the values in the flat profile). */ |
| 62 | const struct |
| 63 | { |
| 64 | char prefix; |
| 65 | double scale; |
| 66 | } |
| 67 | SItab[] = |
| 68 | { |
| 69 | { 'T', 1e-12 }, /* tera */ |
| 70 | { 'G', 1e-09 }, /* giga */ |
| 71 | { 'M', 1e-06 }, /* mega */ |
| 72 | { 'K', 1e-03 }, /* kilo */ |
| 73 | { ' ', 1e-00 }, |
| 74 | { 'm', 1e+03 }, /* milli */ |
| 75 | { 'u', 1e+06 }, /* micro */ |
| 76 | { 'n', 1e+09 }, /* nano */ |
| 77 | { 'p', 1e+12 }, /* pico */ |
| 78 | { 'f', 1e+15 }, /* femto */ |
| 79 | { 'a', 1e+18 } /* ato */ |
| 80 | }; |
| 81 | |
| 82 | /* Reads just the header part of histogram record into |
| 83 | *RECORD from IFP. FILENAME is the name of IFP and |
| 84 | is provided for formatting error messages only. |
| 85 | |
| 86 | If FIRST is non-zero, sets global variables HZ, HIST_DIMENSION, |
| 87 | HIST_DIMENSION_ABBREV, HIST_SCALE. If FIRST is zero, checks |
| 88 | that the new histogram is compatible with already-set values |
| 89 | of those variables and emits an error if that's not so. */ |
| 90 | static void |
| 91 | read_histogram_header (histogram *record, |
| 92 | FILE *ifp, const char *filename, |
| 93 | int first) |
| 94 | { |
| 95 | unsigned int profrate; |
| 96 | char n_hist_dimension[15]; |
| 97 | char n_hist_dimension_abbrev; |
| 98 | double n_hist_scale; |
| 99 | |
| 100 | if (gmon_io_read_vma (ifp, &record->lowpc) |
| 101 | || gmon_io_read_vma (ifp, &record->highpc) |
| 102 | || gmon_io_read_32 (ifp, &record->num_bins) |
| 103 | || gmon_io_read_32 (ifp, &profrate) |
| 104 | || gmon_io_read (ifp, n_hist_dimension, 15) |
| 105 | || gmon_io_read (ifp, &n_hist_dimension_abbrev, 1)) |
| 106 | { |
| 107 | fprintf (stderr, _("%s: %s: unexpected end of file\n"), |
| 108 | whoami, filename); |
| 109 | |
| 110 | done (1); |
| 111 | } |
| 112 | |
| 113 | n_hist_scale = (double)((record->highpc - record->lowpc) / sizeof (UNIT)) |
| 114 | / record->num_bins; |
| 115 | |
| 116 | if (first) |
| 117 | { |
| 118 | /* We don't try to veryfy profrate is the same for all histogram |
| 119 | records. If we have two histogram records for the same |
| 120 | address range and profiling samples is done as often |
| 121 | as possible as opposed on timer, then the actual profrate will |
| 122 | be slightly different. Most of the time the difference does not |
| 123 | matter and insisting that profiling rate is exactly the same |
| 124 | will only create inconvenient. */ |
| 125 | hz = profrate; |
| 126 | memcpy (hist_dimension, n_hist_dimension, 15); |
| 127 | hist_dimension_abbrev = n_hist_dimension_abbrev; |
| 128 | hist_scale = n_hist_scale; |
| 129 | } |
| 130 | else |
| 131 | { |
| 132 | if (strncmp (n_hist_dimension, hist_dimension, 15) != 0) |
| 133 | { |
| 134 | fprintf (stderr, |
| 135 | _("%s: dimension unit changed between histogram records\n" |
| 136 | "%s: from '%s'\n" |
| 137 | "%s: to '%s'\n"), |
| 138 | whoami, whoami, hist_dimension, whoami, n_hist_dimension); |
| 139 | done (1); |
| 140 | } |
| 141 | |
| 142 | if (n_hist_dimension_abbrev != hist_dimension_abbrev) |
| 143 | { |
| 144 | fprintf (stderr, |
| 145 | _("%s: dimension abbreviation changed between histogram records\n" |
| 146 | "%s: from '%c'\n" |
| 147 | "%s: to '%c'\n"), |
| 148 | whoami, whoami, hist_dimension_abbrev, whoami, n_hist_dimension_abbrev); |
| 149 | done (1); |
| 150 | } |
| 151 | |
| 152 | /* The only reason we require the same scale for histograms is that |
| 153 | there's code (notably printing code), that prints units, |
| 154 | and it would be very confusing to have one unit mean different |
| 155 | things for different functions. */ |
| 156 | if (fabs (hist_scale - n_hist_scale) > 0.000001) |
| 157 | { |
| 158 | fprintf (stderr, |
| 159 | _("%s: different scales in histogram records"), |
| 160 | whoami); |
| 161 | done (1); |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | /* Read the histogram from file IFP. FILENAME is the name of IFP and |
| 167 | is provided for formatting error messages only. */ |
| 168 | |
| 169 | void |
| 170 | hist_read_rec (FILE * ifp, const char *filename) |
| 171 | { |
| 172 | bfd_vma lowpc, highpc; |
| 173 | histogram n_record; |
| 174 | histogram *record, *existing_record; |
| 175 | unsigned i; |
| 176 | |
| 177 | /* 1. Read the header and see if there's existing record for the |
| 178 | same address range and that there are no overlapping records. */ |
| 179 | read_histogram_header (&n_record, ifp, filename, num_histograms == 0); |
| 180 | |
| 181 | existing_record = find_histogram (n_record.lowpc, n_record.highpc); |
| 182 | if (existing_record) |
| 183 | { |
| 184 | record = existing_record; |
| 185 | } |
| 186 | else |
| 187 | { |
| 188 | /* If this record overlaps, but does not completely match an existing |
| 189 | record, it's an error. */ |
| 190 | lowpc = n_record.lowpc; |
| 191 | highpc = n_record.highpc; |
| 192 | hist_clip_symbol_address (&lowpc, &highpc); |
| 193 | if (lowpc != highpc) |
| 194 | { |
| 195 | fprintf (stderr, |
| 196 | _("%s: overlapping histogram records\n"), |
| 197 | whoami); |
| 198 | done (1); |
| 199 | } |
| 200 | |
| 201 | /* This is new record. Add it to global array and allocate space for |
| 202 | the samples. */ |
| 203 | histograms = (struct histogram *) |
| 204 | xrealloc (histograms, sizeof (histogram) * (num_histograms + 1)); |
| 205 | memcpy (histograms + num_histograms, |
| 206 | &n_record, sizeof (histogram)); |
| 207 | record = &histograms[num_histograms]; |
| 208 | ++num_histograms; |
| 209 | |
| 210 | record->sample = (int *) xmalloc (record->num_bins |
| 211 | * sizeof (record->sample[0])); |
| 212 | memset (record->sample, 0, record->num_bins * sizeof (record->sample[0])); |
| 213 | } |
| 214 | |
| 215 | /* 2. We have either a new record (with zeroed histogram data), or an existing |
| 216 | record with some data in the histogram already. Read new data into the |
| 217 | record, adding hit counts. */ |
| 218 | |
| 219 | DBG (SAMPLEDEBUG, |
| 220 | printf ("[hist_read_rec] n_lowpc 0x%lx n_highpc 0x%lx ncnt %u\n", |
| 221 | (unsigned long) record->lowpc, (unsigned long) record->highpc, |
| 222 | record->num_bins)); |
| 223 | |
| 224 | for (i = 0; i < record->num_bins; ++i) |
| 225 | { |
| 226 | UNIT count; |
| 227 | if (fread (&count[0], sizeof (count), 1, ifp) != 1) |
| 228 | { |
| 229 | fprintf (stderr, |
| 230 | _("%s: %s: unexpected EOF after reading %u of %u samples\n"), |
| 231 | whoami, filename, i, record->num_bins); |
| 232 | done (1); |
| 233 | } |
| 234 | record->sample[i] += bfd_get_16 (core_bfd, (bfd_byte *) & count[0]); |
| 235 | DBG (SAMPLEDEBUG, |
| 236 | printf ("[hist_read_rec] 0x%lx: %u\n", |
| 237 | (unsigned long) (record->lowpc |
| 238 | + i * (record->highpc - record->lowpc) |
| 239 | / record->num_bins), |
| 240 | record->sample[i])); |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | |
| 245 | /* Write all execution histograms file OFP. FILENAME is the name |
| 246 | of OFP and is provided for formatting error-messages only. */ |
| 247 | |
| 248 | void |
| 249 | hist_write_hist (FILE * ofp, const char *filename) |
| 250 | { |
| 251 | UNIT count; |
| 252 | unsigned int i, r; |
| 253 | |
| 254 | for (r = 0; r < num_histograms; ++r) |
| 255 | { |
| 256 | histogram *record = &histograms[r]; |
| 257 | |
| 258 | /* Write header. */ |
| 259 | |
| 260 | if (gmon_io_write_8 (ofp, GMON_TAG_TIME_HIST) |
| 261 | || gmon_io_write_vma (ofp, record->lowpc) |
| 262 | || gmon_io_write_vma (ofp, record->highpc) |
| 263 | || gmon_io_write_32 (ofp, record->num_bins) |
| 264 | || gmon_io_write_32 (ofp, hz) |
| 265 | || gmon_io_write (ofp, hist_dimension, 15) |
| 266 | || gmon_io_write (ofp, &hist_dimension_abbrev, 1)) |
| 267 | { |
| 268 | perror (filename); |
| 269 | done (1); |
| 270 | } |
| 271 | |
| 272 | for (i = 0; i < record->num_bins; ++i) |
| 273 | { |
| 274 | bfd_put_16 (core_bfd, (bfd_vma) record->sample[i], (bfd_byte *) &count[0]); |
| 275 | |
| 276 | if (fwrite (&count[0], sizeof (count), 1, ofp) != 1) |
| 277 | { |
| 278 | perror (filename); |
| 279 | done (1); |
| 280 | } |
| 281 | } |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | /* Calculate scaled entry point addresses (to save time in |
| 286 | hist_assign_samples), and, on architectures that have procedure |
| 287 | entry masks at the start of a function, possibly push the scaled |
| 288 | entry points over the procedure entry mask, if it turns out that |
| 289 | the entry point is in one bin and the code for a routine is in the |
| 290 | next bin. */ |
| 291 | |
| 292 | static void |
| 293 | scale_and_align_entries () |
| 294 | { |
| 295 | Sym *sym; |
| 296 | bfd_vma bin_of_entry; |
| 297 | bfd_vma bin_of_code; |
| 298 | |
| 299 | for (sym = symtab.base; sym < symtab.limit; sym++) |
| 300 | { |
| 301 | histogram *r = find_histogram_for_pc (sym->addr); |
| 302 | |
| 303 | sym->hist.scaled_addr = sym->addr / sizeof (UNIT); |
| 304 | |
| 305 | if (r) |
| 306 | { |
| 307 | bin_of_entry = (sym->hist.scaled_addr - r->lowpc) / hist_scale; |
| 308 | bin_of_code = ((sym->hist.scaled_addr + UNITS_TO_CODE - r->lowpc) |
| 309 | / hist_scale); |
| 310 | if (bin_of_entry < bin_of_code) |
| 311 | { |
| 312 | DBG (SAMPLEDEBUG, |
| 313 | printf ("[scale_and_align_entries] pushing 0x%lx to 0x%lx\n", |
| 314 | (unsigned long) sym->hist.scaled_addr, |
| 315 | (unsigned long) (sym->hist.scaled_addr |
| 316 | + UNITS_TO_CODE))); |
| 317 | sym->hist.scaled_addr += UNITS_TO_CODE; |
| 318 | } |
| 319 | } |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | |
| 324 | /* Assign samples to the symbol to which they belong. |
| 325 | |
| 326 | Histogram bin I covers some address range [BIN_LOWPC,BIN_HIGH_PC) |
| 327 | which may overlap one more symbol address ranges. If a symbol |
| 328 | overlaps with the bin's address range by O percent, then O percent |
| 329 | of the bin's count is credited to that symbol. |
| 330 | |
| 331 | There are three cases as to where BIN_LOW_PC and BIN_HIGH_PC can be |
| 332 | with respect to the symbol's address range [SYM_LOW_PC, |
| 333 | SYM_HIGH_PC) as shown in the following diagram. OVERLAP computes |
| 334 | the distance (in UNITs) between the arrows, the fraction of the |
| 335 | sample that is to be credited to the symbol which starts at |
| 336 | SYM_LOW_PC. |
| 337 | |
| 338 | sym_low_pc sym_high_pc |
| 339 | | | |
| 340 | v v |
| 341 | |
| 342 | +-----------------------------------------------+ |
| 343 | | | |
| 344 | | ->| |<- ->| |<- ->| |<- | |
| 345 | | | | | | | |
| 346 | +---------+ +---------+ +---------+ |
| 347 | |
| 348 | ^ ^ ^ ^ ^ ^ |
| 349 | | | | | | | |
| 350 | bin_low_pc bin_high_pc bin_low_pc bin_high_pc bin_low_pc bin_high_pc |
| 351 | |
| 352 | For the VAX we assert that samples will never fall in the first two |
| 353 | bytes of any routine, since that is the entry mask, thus we call |
| 354 | scale_and_align_entries() to adjust the entry points if the entry |
| 355 | mask falls in one bin but the code for the routine doesn't start |
| 356 | until the next bin. In conjunction with the alignment of routine |
| 357 | addresses, this should allow us to have only one sample for every |
| 358 | four bytes of text space and never have any overlap (the two end |
| 359 | cases, above). */ |
| 360 | |
| 361 | static void |
| 362 | hist_assign_samples_1 (histogram *r) |
| 363 | { |
| 364 | bfd_vma bin_low_pc, bin_high_pc; |
| 365 | bfd_vma sym_low_pc, sym_high_pc; |
| 366 | bfd_vma overlap, addr; |
| 367 | unsigned int bin_count; |
| 368 | unsigned int i, j; |
| 369 | double count_time, credit; |
| 370 | |
| 371 | bfd_vma lowpc = r->lowpc / sizeof (UNIT); |
| 372 | |
| 373 | /* Iterate over all sample bins. */ |
| 374 | for (i = 0, j = 1; i < r->num_bins; ++i) |
| 375 | { |
| 376 | bin_count = r->sample[i]; |
| 377 | if (! bin_count) |
| 378 | continue; |
| 379 | |
| 380 | bin_low_pc = lowpc + (bfd_vma) (hist_scale * i); |
| 381 | bin_high_pc = lowpc + (bfd_vma) (hist_scale * (i + 1)); |
| 382 | count_time = bin_count; |
| 383 | |
| 384 | DBG (SAMPLEDEBUG, |
| 385 | printf ( |
| 386 | "[assign_samples] bin_low_pc=0x%lx, bin_high_pc=0x%lx, bin_count=%u\n", |
| 387 | (unsigned long) (sizeof (UNIT) * bin_low_pc), |
| 388 | (unsigned long) (sizeof (UNIT) * bin_high_pc), |
| 389 | bin_count)); |
| 390 | total_time += count_time; |
| 391 | |
| 392 | /* Credit all symbols that are covered by bin I. */ |
| 393 | for (j = j - 1; j < symtab.len; ++j) |
| 394 | { |
| 395 | sym_low_pc = symtab.base[j].hist.scaled_addr; |
| 396 | sym_high_pc = symtab.base[j + 1].hist.scaled_addr; |
| 397 | |
| 398 | /* If high end of bin is below entry address, |
| 399 | go for next bin. */ |
| 400 | if (bin_high_pc < sym_low_pc) |
| 401 | break; |
| 402 | |
| 403 | /* If low end of bin is above high end of symbol, |
| 404 | go for next symbol. */ |
| 405 | if (bin_low_pc >= sym_high_pc) |
| 406 | continue; |
| 407 | |
| 408 | overlap = |
| 409 | MIN (bin_high_pc, sym_high_pc) - MAX (bin_low_pc, sym_low_pc); |
| 410 | if (overlap > 0) |
| 411 | { |
| 412 | DBG (SAMPLEDEBUG, |
| 413 | printf ( |
| 414 | "[assign_samples] [0x%lx,0x%lx) %s gets %f ticks %ld overlap\n", |
| 415 | (unsigned long) symtab.base[j].addr, |
| 416 | (unsigned long) (sizeof (UNIT) * sym_high_pc), |
| 417 | symtab.base[j].name, overlap * count_time / hist_scale, |
| 418 | (long) overlap)); |
| 419 | |
| 420 | addr = symtab.base[j].addr; |
| 421 | credit = overlap * count_time / hist_scale; |
| 422 | |
| 423 | /* Credit symbol if it appears in INCL_FLAT or that |
| 424 | table is empty and it does not appear it in |
| 425 | EXCL_FLAT. */ |
| 426 | if (sym_lookup (&syms[INCL_FLAT], addr) |
| 427 | || (syms[INCL_FLAT].len == 0 |
| 428 | && !sym_lookup (&syms[EXCL_FLAT], addr))) |
| 429 | { |
| 430 | symtab.base[j].hist.time += credit; |
| 431 | } |
| 432 | else |
| 433 | { |
| 434 | total_time -= credit; |
| 435 | } |
| 436 | } |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | DBG (SAMPLEDEBUG, printf ("[assign_samples] total_time %f\n", |
| 441 | total_time)); |
| 442 | } |
| 443 | |
| 444 | /* Calls 'hist_assign_sampes_1' for all histogram records read so far. */ |
| 445 | void |
| 446 | hist_assign_samples () |
| 447 | { |
| 448 | unsigned i; |
| 449 | |
| 450 | scale_and_align_entries (); |
| 451 | |
| 452 | for (i = 0; i < num_histograms; ++i) |
| 453 | hist_assign_samples_1 (&histograms[i]); |
| 454 | |
| 455 | } |
| 456 | |
| 457 | /* Print header for flag histogram profile. */ |
| 458 | |
| 459 | static void |
| 460 | print_header (int prefix) |
| 461 | { |
| 462 | char unit[64]; |
| 463 | |
| 464 | sprintf (unit, _("%c%c/call"), prefix, hist_dimension_abbrev); |
| 465 | |
| 466 | if (bsd_style_output) |
| 467 | { |
| 468 | printf (_("\ngranularity: each sample hit covers %ld byte(s)"), |
| 469 | (long) hist_scale * (long) sizeof (UNIT)); |
| 470 | if (total_time > 0.0) |
| 471 | { |
| 472 | printf (_(" for %.2f%% of %.2f %s\n\n"), |
| 473 | 100.0 / total_time, total_time / hz, hist_dimension); |
| 474 | } |
| 475 | } |
| 476 | else |
| 477 | { |
| 478 | printf (_("\nEach sample counts as %g %s.\n"), 1.0 / hz, hist_dimension); |
| 479 | } |
| 480 | |
| 481 | if (total_time <= 0.0) |
| 482 | { |
| 483 | printf (_(" no time accumulated\n\n")); |
| 484 | |
| 485 | /* This doesn't hurt since all the numerators will be zero. */ |
| 486 | total_time = 1.0; |
| 487 | } |
| 488 | |
| 489 | printf ("%5.5s %10.10s %8.8s %8.8s %8.8s %8.8s %-8.8s\n", |
| 490 | "% ", _("cumulative"), _("self "), "", _("self "), _("total "), |
| 491 | ""); |
| 492 | printf ("%5.5s %9.9s %8.8s %8.8s %8.8s %8.8s %-8.8s\n", |
| 493 | _("time"), hist_dimension, hist_dimension, _("calls"), unit, unit, |
| 494 | _("name")); |
| 495 | } |
| 496 | |
| 497 | |
| 498 | static void |
| 499 | print_line (Sym *sym, double scale) |
| 500 | { |
| 501 | if (ignore_zeros && sym->ncalls == 0 && sym->hist.time == 0) |
| 502 | return; |
| 503 | |
| 504 | accum_time += sym->hist.time; |
| 505 | |
| 506 | if (bsd_style_output) |
| 507 | printf ("%5.1f %10.2f %8.2f", |
| 508 | total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0, |
| 509 | accum_time / hz, sym->hist.time / hz); |
| 510 | else |
| 511 | printf ("%6.2f %9.2f %8.2f", |
| 512 | total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0, |
| 513 | accum_time / hz, sym->hist.time / hz); |
| 514 | |
| 515 | if (sym->ncalls != 0) |
| 516 | printf (" %8lu %8.2f %8.2f ", |
| 517 | sym->ncalls, scale * sym->hist.time / hz / sym->ncalls, |
| 518 | scale * (sym->hist.time + sym->cg.child_time) / hz / sym->ncalls); |
| 519 | else |
| 520 | printf (" %8.8s %8.8s %8.8s ", "", "", ""); |
| 521 | |
| 522 | if (bsd_style_output) |
| 523 | print_name (sym); |
| 524 | else |
| 525 | print_name_only (sym); |
| 526 | |
| 527 | printf ("\n"); |
| 528 | } |
| 529 | |
| 530 | |
| 531 | /* Compare LP and RP. The primary comparison key is execution time, |
| 532 | the secondary is number of invocation, and the tertiary is the |
| 533 | lexicographic order of the function names. */ |
| 534 | |
| 535 | static int |
| 536 | cmp_time (const PTR lp, const PTR rp) |
| 537 | { |
| 538 | const Sym *left = *(const Sym **) lp; |
| 539 | const Sym *right = *(const Sym **) rp; |
| 540 | double time_diff; |
| 541 | |
| 542 | time_diff = right->hist.time - left->hist.time; |
| 543 | |
| 544 | if (time_diff > 0.0) |
| 545 | return 1; |
| 546 | |
| 547 | if (time_diff < 0.0) |
| 548 | return -1; |
| 549 | |
| 550 | if (right->ncalls > left->ncalls) |
| 551 | return 1; |
| 552 | |
| 553 | if (right->ncalls < left->ncalls) |
| 554 | return -1; |
| 555 | |
| 556 | return strcmp (left->name, right->name); |
| 557 | } |
| 558 | |
| 559 | |
| 560 | /* Print the flat histogram profile. */ |
| 561 | |
| 562 | void |
| 563 | hist_print () |
| 564 | { |
| 565 | Sym **time_sorted_syms, *top_dog, *sym; |
| 566 | unsigned int sym_index; |
| 567 | unsigned log_scale; |
| 568 | double top_time; |
| 569 | bfd_vma addr; |
| 570 | |
| 571 | if (first_output) |
| 572 | first_output = FALSE; |
| 573 | else |
| 574 | printf ("\f\n"); |
| 575 | |
| 576 | accum_time = 0.0; |
| 577 | |
| 578 | if (bsd_style_output) |
| 579 | { |
| 580 | if (print_descriptions) |
| 581 | { |
| 582 | printf (_("\n\n\nflat profile:\n")); |
| 583 | flat_blurb (stdout); |
| 584 | } |
| 585 | } |
| 586 | else |
| 587 | { |
| 588 | printf (_("Flat profile:\n")); |
| 589 | } |
| 590 | |
| 591 | /* Sort the symbol table by time (call-count and name as secondary |
| 592 | and tertiary keys). */ |
| 593 | time_sorted_syms = (Sym **) xmalloc (symtab.len * sizeof (Sym *)); |
| 594 | |
| 595 | for (sym_index = 0; sym_index < symtab.len; ++sym_index) |
| 596 | time_sorted_syms[sym_index] = &symtab.base[sym_index]; |
| 597 | |
| 598 | qsort (time_sorted_syms, symtab.len, sizeof (Sym *), cmp_time); |
| 599 | |
| 600 | if (bsd_style_output) |
| 601 | { |
| 602 | log_scale = 5; /* Milli-seconds is BSD-default. */ |
| 603 | } |
| 604 | else |
| 605 | { |
| 606 | /* Search for symbol with highest per-call |
| 607 | execution time and scale accordingly. */ |
| 608 | log_scale = 0; |
| 609 | top_dog = 0; |
| 610 | top_time = 0.0; |
| 611 | |
| 612 | for (sym_index = 0; sym_index < symtab.len; ++sym_index) |
| 613 | { |
| 614 | sym = time_sorted_syms[sym_index]; |
| 615 | |
| 616 | if (sym->ncalls != 0) |
| 617 | { |
| 618 | double call_time; |
| 619 | |
| 620 | call_time = (sym->hist.time + sym->cg.child_time) / sym->ncalls; |
| 621 | |
| 622 | if (call_time > top_time) |
| 623 | { |
| 624 | top_dog = sym; |
| 625 | top_time = call_time; |
| 626 | } |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | if (top_dog && top_dog->ncalls != 0 && top_time > 0.0) |
| 631 | { |
| 632 | top_time /= hz; |
| 633 | |
| 634 | for (log_scale = 0; log_scale < ARRAY_SIZE (SItab); log_scale ++) |
| 635 | { |
| 636 | double scaled_value = SItab[log_scale].scale * top_time; |
| 637 | |
| 638 | if (scaled_value >= 1.0 && scaled_value < 1000.0) |
| 639 | break; |
| 640 | } |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | /* For now, the dimension is always seconds. In the future, we |
| 645 | may also want to support other (pseudo-)dimensions (such as |
| 646 | I-cache misses etc.). */ |
| 647 | print_header (SItab[log_scale].prefix); |
| 648 | |
| 649 | for (sym_index = 0; sym_index < symtab.len; ++sym_index) |
| 650 | { |
| 651 | addr = time_sorted_syms[sym_index]->addr; |
| 652 | |
| 653 | /* Print symbol if its in INCL_FLAT table or that table |
| 654 | is empty and the symbol is not in EXCL_FLAT. */ |
| 655 | if (sym_lookup (&syms[INCL_FLAT], addr) |
| 656 | || (syms[INCL_FLAT].len == 0 |
| 657 | && !sym_lookup (&syms[EXCL_FLAT], addr))) |
| 658 | print_line (time_sorted_syms[sym_index], SItab[log_scale].scale); |
| 659 | } |
| 660 | |
| 661 | free (time_sorted_syms); |
| 662 | |
| 663 | if (print_descriptions && !bsd_style_output) |
| 664 | flat_blurb (stdout); |
| 665 | } |
| 666 | |
| 667 | int |
| 668 | hist_check_address (unsigned address) |
| 669 | { |
| 670 | unsigned i; |
| 671 | |
| 672 | for (i = 0; i < num_histograms; ++i) |
| 673 | if (histograms[i].lowpc <= address && address < histograms[i].highpc) |
| 674 | return 1; |
| 675 | |
| 676 | return 0; |
| 677 | } |
| 678 | |
| 679 | #if ! defined(min) |
| 680 | #define min(a,b) (((a)<(b)) ? (a) : (b)) |
| 681 | #endif |
| 682 | #if ! defined(max) |
| 683 | #define max(a,b) (((a)>(b)) ? (a) : (b)) |
| 684 | #endif |
| 685 | |
| 686 | void |
| 687 | hist_clip_symbol_address (bfd_vma *p_lowpc, bfd_vma *p_highpc) |
| 688 | { |
| 689 | unsigned i; |
| 690 | int found = 0; |
| 691 | |
| 692 | if (num_histograms == 0) |
| 693 | { |
| 694 | *p_highpc = *p_lowpc; |
| 695 | return; |
| 696 | } |
| 697 | |
| 698 | for (i = 0; i < num_histograms; ++i) |
| 699 | { |
| 700 | bfd_vma common_low, common_high; |
| 701 | common_low = max (histograms[i].lowpc, *p_lowpc); |
| 702 | common_high = min (histograms[i].highpc, *p_highpc); |
| 703 | |
| 704 | if (common_low < common_high) |
| 705 | { |
| 706 | if (found) |
| 707 | { |
| 708 | fprintf (stderr, |
| 709 | _("%s: found a symbol that covers " |
| 710 | "several histogram records"), |
| 711 | whoami); |
| 712 | done (1); |
| 713 | } |
| 714 | |
| 715 | found = 1; |
| 716 | *p_lowpc = common_low; |
| 717 | *p_highpc = common_high; |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | if (!found) |
| 722 | *p_highpc = *p_lowpc; |
| 723 | } |
| 724 | |
| 725 | /* Find and return exising histogram record having the same lowpc and |
| 726 | highpc as passed via the parameters. Return NULL if nothing is found. |
| 727 | The return value is valid until any new histogram is read. */ |
| 728 | static histogram * |
| 729 | find_histogram (bfd_vma lowpc, bfd_vma highpc) |
| 730 | { |
| 731 | unsigned i; |
| 732 | for (i = 0; i < num_histograms; ++i) |
| 733 | { |
| 734 | if (histograms[i].lowpc == lowpc && histograms[i].highpc == highpc) |
| 735 | return &histograms[i]; |
| 736 | } |
| 737 | return 0; |
| 738 | } |
| 739 | |
| 740 | /* Given a PC, return histogram record which address range include this PC. |
| 741 | Return NULL if there's no such record. */ |
| 742 | static histogram * |
| 743 | find_histogram_for_pc (bfd_vma pc) |
| 744 | { |
| 745 | unsigned i; |
| 746 | for (i = 0; i < num_histograms; ++i) |
| 747 | { |
| 748 | if (histograms[i].lowpc <= pc && pc < histograms[i].highpc) |
| 749 | return &histograms[i]; |
| 750 | } |
| 751 | return 0; |
| 752 | } |