gprof/hist.c

   1 /*
   2  * Histogram related operations.
   3  */
   4 #include <stdio.h>
   5 #include "libiberty.h"
   6 #include "gprof.h"
   7 #include "core.h"
   8 #include "gmon_io.h"
   9 #include "gmon_out.h"
  10 #include "hist.h"
  11 #include "symtab.h"
  12 #include "sym_ids.h"
  13 #include "utils.h"
  14
  15 /* declarations of automatically generated functions to output blurbs: */
  16 extern void flat_blurb PARAMS ((FILE * fp));
  17
  18 bfd_vma s_lowpc;                /* lowest address in .text */
  19 bfd_vma s_highpc = 0;           /* highest address in .text */
  20 bfd_vma lowpc, highpc;          /* same, but expressed in UNITs */
  21 int hist_num_bins = 0;          /* number of histogram samples */
  22 int *hist_sample = 0;           /* histogram samples (shorts in the file!) */
  23 double hist_scale;
  24 char hist_dimension[sizeof (((struct gmon_hist_hdr *) 0)->dimen) + 1] =
  25   "seconds";
  26 char hist_dimension_abbrev = 's';
  27
  28 static double accum_time;       /* accumulated time so far for print_line() */
  29 static double total_time;       /* total time for all routines */
  30 /*
  31  * Table of SI prefixes for powers of 10 (used to automatically
  32  * scale some of the values in the flat profile).
  33  */
  34 const struct
  35   {
  36     char prefix;
  37     double scale;
  38   }
  39 SItab[] =
  40 {
  41   {
  42     'T', 1e-12
  43   }
  44   ,                             /* tera */
  45   {
  46     'G', 1e-09
  47   }
  48   ,                             /* giga */
  49   {
  50     'M', 1e-06
  51   }
  52   ,                             /* mega */
  53   {
  54     'K', 1e-03
  55   }
  56   ,                             /* kilo */
  57   {
  58     ' ', 1e-00
  59   }
  60   ,
  61   {
  62     'm', 1e+03
  63   }
  64   ,                             /* milli */
  65   {
  66     'u', 1e+06
  67   }
  68   ,                             /* micro */
  69   {
  70     'n', 1e+09
  71   }
  72   ,                             /* nano */
  73   {
  74     'p', 1e+12
  75   }
  76   ,                             /* pico */
  77   {
  78     'f', 1e+15
  79   }
  80   ,                             /* femto */
  81   {
  82     'a', 1e+18
  83   }
  84   ,                             /* ato */
  85 };
  86
  87 /*
  88  * Read the histogram from file IFP.  FILENAME is the name of IFP and
  89  * is provided for formatting error messages only.
  90  */
  91 void
  92 DEFUN (hist_read_rec, (ifp, filename), FILE * ifp AND const char *filename)
  93 {
  94   struct gmon_hist_hdr hdr;
  95   bfd_vma n_lowpc, n_highpc;
  96   int i, ncnt, profrate;
  97   UNIT count;
  98
  99   if (fread (&hdr, sizeof (hdr), 1, ifp) != 1)
 100     {
 101       fprintf (stderr, "%s: %s: unexpected end of file\n",
 102                whoami, filename);
 103       done (1);
 104     }
 105
 106   n_lowpc = (bfd_vma) get_vma (core_bfd, (bfd_byte *) hdr.low_pc);
 107   n_highpc = (bfd_vma) get_vma (core_bfd, (bfd_byte *) hdr.high_pc);
 108   ncnt = bfd_get_32 (core_bfd, (bfd_byte *) hdr.hist_size);
 109   profrate = bfd_get_32 (core_bfd, (bfd_byte *) hdr.prof_rate);
 110   strncpy (hist_dimension, hdr.dimen, sizeof (hdr.dimen));
 111   hist_dimension[sizeof (hdr.dimen)] = '\0';
 112   hist_dimension_abbrev = hdr.dimen_abbrev;
 113
 114   if (!s_highpc)
 115     {
 116
 117       /* this is the first histogram record: */
 118
 119       s_lowpc = n_lowpc;
 120       s_highpc = n_highpc;
 121       lowpc = (bfd_vma) n_lowpc / sizeof (UNIT);
 122       highpc = (bfd_vma) n_highpc / sizeof (UNIT);
 123       hist_num_bins = ncnt;
 124       hz = profrate;
 125     }
 126
 127   DBG (SAMPLEDEBUG,
 128        printf ("[hist_read_rec] n_lowpc 0x%lx n_highpc 0x%lx ncnt %d\n",
 129                n_lowpc, n_highpc, ncnt);
 130        printf ("[hist_read_rec] s_lowpc 0x%lx s_highpc 0x%lx nsamples %d\n",
 131                s_lowpc, s_highpc, hist_num_bins);
 132        printf ("[hist_read_rec]   lowpc 0x%lx   highpc 0x%lx\n",
 133                lowpc, highpc));
 134
 135   if (n_lowpc != s_lowpc || n_highpc != s_highpc
 136       || ncnt != hist_num_bins || hz != profrate)
 137     {
 138       fprintf (stderr, "%s: `%s' is incompatible with first gmon file\n",
 139                whoami, filename);
 140       done (1);
 141     }
 142
 143   if (!hist_sample)
 144     {
 145       hist_sample = (int *) xmalloc (hist_num_bins * sizeof (hist_sample[0]));
 146       memset (hist_sample, 0, hist_num_bins * sizeof (hist_sample[0]));
 147     }
 148
 149   for (i = 0; i < hist_num_bins; ++i)
 150     {
 151       if (fread (&count[0], sizeof (count), 1, ifp) != 1)
 152         {
 153           fprintf (stderr,
 154                    "%s: %s: unexpected EOF after reading %d of %d samples\n",
 155                    whoami, filename, i, hist_num_bins);
 156           done (1);
 157         }
 158       hist_sample[i] += bfd_get_16 (core_bfd, (bfd_byte *) & count[0]);
 159     }
 160 }
 161
 162
 163 /*
 164  * Write execution histogram to file OFP.  FILENAME is the name
 165  * of OFP and is provided for formatting error-messages only.
 166  */
 167 void
 168 DEFUN (hist_write_hist, (ofp, filename), FILE * ofp AND const char *filename)
 169 {
 170   struct gmon_hist_hdr hdr;
 171   unsigned char tag;
 172   UNIT count;
 173   int i;
 174
 175   /* write header: */
 176
 177   tag = GMON_TAG_TIME_HIST;
 178   put_vma (core_bfd, s_lowpc, (bfd_byte *) hdr.low_pc);
 179   put_vma (core_bfd, s_highpc, (bfd_byte *) hdr.high_pc);
 180   bfd_put_32 (core_bfd, hist_num_bins, (bfd_byte *) hdr.hist_size);
 181   bfd_put_32 (core_bfd, hz, (bfd_byte *) hdr.prof_rate);
 182   strncpy (hdr.dimen, hist_dimension, sizeof (hdr.dimen));
 183   hdr.dimen_abbrev = hist_dimension_abbrev;
 184
 185   if (fwrite (&tag, sizeof (tag), 1, ofp) != 1
 186       || fwrite (&hdr, sizeof (hdr), 1, ofp) != 1)
 187     {
 188       perror (filename);
 189       done (1);
 190     }
 191
 192   for (i = 0; i < hist_num_bins; ++i)
 193     {
 194       bfd_put_16 (core_bfd, hist_sample[i], (bfd_byte *) & count[0]);
 195       if (fwrite (&count[0], sizeof (count), 1, ofp) != 1)
 196         {
 197           perror (filename);
 198           done (1);
 199         }
 200     }
 201 }
 202
 203
 204 /*
 205  * Calculate scaled entry point addresses (to save time in
 206  * hist_assign_samples), and, on architectures that have procedure
 207  * entry masks at the start of a function, possibly push the scaled
 208  * entry points over the procedure entry mask, if it turns out that
 209  * the entry point is in one bin and the code for a routine is in the
 210  * next bin.
 211  */
 212 static void
 213 scale_and_align_entries ()
 214 {
 215   Sym *sym;
 216   bfd_vma bin_of_entry;
 217   bfd_vma bin_of_code;
 218
 219   for (sym = symtab.base; sym < symtab.limit; sym++)
 220     {
 221       sym->hist.scaled_addr = sym->addr / sizeof (UNIT);
 222       bin_of_entry = (sym->hist.scaled_addr - lowpc) / hist_scale;
 223       bin_of_code = (sym->hist.scaled_addr + UNITS_TO_CODE - lowpc) / hist_scale;
 224       if (bin_of_entry < bin_of_code)
 225         {
 226           DBG (SAMPLEDEBUG,
 227                printf ("[scale_and_align_entries] pushing 0x%lx to 0x%lx\n",
 228                        sym->hist.scaled_addr,
 229                        sym->hist.scaled_addr + UNITS_TO_CODE));
 230           sym->hist.scaled_addr += UNITS_TO_CODE;
 231         }
 232     }
 233 }
 234
 235
 236 /*
 237  * Assign samples to the symbol to which they belong.
 238  *
 239  * Histogram bin I covers some address range [BIN_LOWPC,BIN_HIGH_PC)
 240  * which may overlap one more symbol address ranges.  If a symbol
 241  * overlaps with the bin's address range by O percent, then O percent
 242  * of the bin's count is credited to that symbol.
 243  *
 244  * There are three cases as to where BIN_LOW_PC and BIN_HIGH_PC can be
 245  * with respect to the symbol's address range [SYM_LOW_PC,
 246  * SYM_HIGH_PC) as shown in the following diagram.  OVERLAP computes
 247  * the distance (in UNITs) between the arrows, the fraction of the
 248  * sample that is to be credited to the symbol which starts at
 249  * SYM_LOW_PC.
 250  *
 251  *        sym_low_pc                                      sym_high_pc
 252  *             |                                               |
 253  *             v                                               v
 254  *
 255  *             +-----------------------------------------------+
 256  *             |                                               |
 257  *        |  ->|    |<-         ->|         |<-         ->|    |<-  |
 258  *        |         |             |         |             |         |
 259  *        +---------+             +---------+             +---------+
 260  *
 261  *        ^         ^             ^         ^             ^         ^
 262  *        |         |             |         |             |         |
 263  *   bin_low_pc bin_high_pc  bin_low_pc bin_high_pc  bin_low_pc bin_high_pc
 264  *
 265  * For the VAX we assert that samples will never fall in the first two
 266  * bytes of any routine, since that is the entry mask, thus we call
 267  * scale_and_align_entries() to adjust the entry points if the entry
 268  * mask falls in one bin but the code for the routine doesn't start
 269  * until the next bin.  In conjunction with the alignment of routine
 270  * addresses, this should allow us to have only one sample for every
 271  * four bytes of text space and never have any overlap (the two end
 272  * cases, above).
 273  */
 274 void
 275 DEFUN_VOID (hist_assign_samples)
 276 {
 277   bfd_vma bin_low_pc, bin_high_pc;
 278   bfd_vma sym_low_pc, sym_high_pc;
 279   bfd_vma overlap, addr;
 280   int bin_count, i, j;
 281   double time, credit;
 282
 283   /* read samples and assign to symbols: */
 284   hist_scale = highpc - lowpc;
 285   hist_scale /= hist_num_bins;
 286   scale_and_align_entries ();
 287
 288   /* iterate over all sample bins: */
 289
 290   for (i = 0, j = 1; i < hist_num_bins; ++i)
 291     {
 292       bin_count = hist_sample[i];
 293       if (!bin_count)
 294         {
 295           continue;
 296         }
 297       bin_low_pc = lowpc + (bfd_vma) (hist_scale * i);
 298       bin_high_pc = lowpc + (bfd_vma) (hist_scale * (i + 1));
 299       time = bin_count;
 300       DBG (SAMPLEDEBUG,
 301            printf (
 302       "[assign_samples] bin_low_pc=0x%lx, bin_high_pc=0x%lx, bin_count=%d\n",
 303                     sizeof (UNIT) * bin_low_pc, sizeof (UNIT) * bin_high_pc,
 304                     bin_count));
 305       total_time += time;
 306
 307       /* credit all symbols that are covered by bin I: */
 308
 309       for (j = j - 1; j < symtab.len; ++j)
 310         {
 311           sym_low_pc = symtab.base[j].hist.scaled_addr;
 312           sym_high_pc = symtab.base[j + 1].hist.scaled_addr;
 313           /*
 314            * If high end of bin is below entry address, go for next
 315            * bin:
 316            */
 317           if (bin_high_pc < sym_low_pc)
 318             {
 319               break;
 320             }
 321           /*
 322            * If low end of bin is above high end of symbol, go for
 323            * next symbol.
 324            */
 325           if (bin_low_pc >= sym_high_pc)
 326             {
 327               continue;
 328             }
 329           overlap =
 330             MIN (bin_high_pc, sym_high_pc) - MAX (bin_low_pc, sym_low_pc);
 331           if (overlap > 0)
 332             {
 333               DBG (SAMPLEDEBUG,
 334                    printf (
 335                             "[assign_samples] [0x%lx,0x%lx) %s gets %f ticks %ld overlap\n",
 336                             symtab.base[j].addr, sizeof (UNIT) * sym_high_pc,
 337                             symtab.base[j].name, overlap * time / hist_scale,
 338                             overlap));
 339               addr = symtab.base[j].addr;
 340               credit = overlap * time / hist_scale;
 341               /*
 342                * Credit symbol if it appears in INCL_FLAT or that
 343                * table is empty and it does not appear it in
 344                * EXCL_FLAT.
 345                */
 346               if (sym_lookup (&syms[INCL_FLAT], addr)
 347                   || (syms[INCL_FLAT].len == 0
 348                       && !sym_lookup (&syms[EXCL_FLAT], addr)))
 349                 {
 350                   symtab.base[j].hist.time += credit;
 351                 }
 352               else
 353                 {
 354                   total_time -= credit;
 355                 }
 356             }
 357         }
 358     }
 359   DBG (SAMPLEDEBUG, printf ("[assign_samples] total_time %f\n",
 360                             total_time));
 361 }
 362
 363
 364 /*
 365  * Print header for flag histogram profile:
 366  */
 367 static void
 368 DEFUN (print_header, (prefix), const char prefix)
 369 {
 370   char unit[64];
 371
 372   sprintf (unit, "%c%c/call", prefix, hist_dimension_abbrev);
 373
 374   if (bsd_style_output)
 375     {
 376       printf ("\ngranularity: each sample hit covers %ld byte(s)",
 377               (long) hist_scale * sizeof (UNIT));
 378       if (total_time > 0.0)
 379         {
 380           printf (" for %.2f%% of %.2f %s\n\n",
 381                   100.0 / total_time, total_time / hz, hist_dimension);
 382         }
 383     }
 384   else
 385     {
 386       printf ("\nEach sample counts as %g %s.\n", 1.0 / hz, hist_dimension);
 387     }
 388
 389   if (total_time <= 0.0)
 390     {
 391       printf (" no time accumulated\n\n");
 392       /* this doesn't hurt since all the numerators will be zero: */
 393       total_time = 1.0;
 394     }
 395
 396   printf ("%5.5s %10.10s %8.8s %8.8s %8.8s %8.8s  %-8.8s\n",
 397           "%  ", "cumulative", "self  ", "", "self  ", "total ", "");
 398   printf ("%5.5s %9.9s  %8.8s %8.8s %8.8s %8.8s  %-8.8s\n",
 399           "time", hist_dimension, hist_dimension, "calls", unit, unit,
 400           "name");
 401 }
 402
 403
 404 static void
 405 DEFUN (print_line, (sym, scale), Sym * sym AND double scale)
 406 {
 407   if (ignore_zeros && sym->ncalls == 0 && sym->hist.time == 0)
 408     {
 409       return;
 410     }
 411
 412   accum_time += sym->hist.time;
 413   if (bsd_style_output)
 414     {
 415       printf ("%5.1f %10.2f %8.2f",
 416               total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0,
 417               accum_time / hz, sym->hist.time / hz);
 418     }
 419   else
 420     {
 421       printf ("%6.2f %9.2f %8.2f",
 422               total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0,
 423               accum_time / hz, sym->hist.time / hz);
 424     }
 425   if (sym->ncalls)
 426     {
 427       printf (" %8d %8.2f %8.2f  ",
 428               sym->ncalls, scale * sym->hist.time / hz / sym->ncalls,
 429           scale * (sym->hist.time + sym->cg.child_time) / hz / sym->ncalls);
 430     }
 431   else
 432     {
 433       printf (" %8.8s %8.8s %8.8s  ", "", "", "");
 434     }
 435   if (bsd_style_output)
 436     {
 437       print_name (sym);
 438     }
 439   else
 440     {
 441       print_name_only (sym);
 442     }
 443   printf ("\n");
 444 }
 445
 446
 447 /*
 448  * Compare LP and RP.  The primary comparison key is execution time,
 449  * the secondary is number of invocation, and the tertiary is the
 450  * lexicographic order of the function names.
 451  */
 452 static int
 453 DEFUN (cmp_time, (lp, rp), const PTR lp AND const PTR rp)
 454 {
 455   const Sym *left = *(const Sym **) lp;
 456   const Sym *right = *(const Sym **) rp;
 457   double time_diff;
 458   long call_diff;
 459
 460   time_diff = right->hist.time - left->hist.time;
 461   if (time_diff > 0.0)
 462     {
 463       return 1;
 464     }
 465   if (time_diff < 0.0)
 466     {
 467       return -1;
 468     }
 469
 470   call_diff = right->ncalls - left->ncalls;
 471   if (call_diff > 0)
 472     {
 473       return 1;
 474     }
 475   if (call_diff < 0)
 476     {
 477       return -1;
 478     }
 479
 480   return strcmp (left->name, right->name);
 481 }
 482
 483
 484 /*
 485  * Print the flat histogram profile.
 486  */
 487 void
 488 DEFUN_VOID (hist_print)
 489 {
 490   Sym **time_sorted_syms, *top_dog, *sym;
 491   int index, log_scale;
 492   double top_time, time;
 493   bfd_vma addr;
 494
 495   if (first_output)
 496     {
 497       first_output = FALSE;
 498     }
 499   else
 500     {
 501       printf ("\f\n");
 502     }
 503
 504   accum_time = 0.0;
 505   if (bsd_style_output)
 506     {
 507       if (print_descriptions)
 508         {
 509           printf ("\n\n\nflat profile:\n");
 510           flat_blurb (stdout);
 511         }
 512     }
 513   else
 514     {
 515       printf ("Flat profile:\n");
 516     }
 517   /*
 518    * Sort the symbol table by time (call-count and name as secondary
 519    * and tertiary keys):
 520    */
 521   time_sorted_syms = (Sym **) xmalloc (symtab.len * sizeof (Sym *));
 522   for (index = 0; index < symtab.len; ++index)
 523     {
 524       time_sorted_syms[index] = &symtab.base[index];
 525     }
 526   qsort (time_sorted_syms, symtab.len, sizeof (Sym *), cmp_time);
 527
 528   if (bsd_style_output)
 529     {
 530       log_scale = 5;            /* milli-seconds is BSD-default */
 531     }
 532   else
 533     {
 534       /*
 535        * Search for symbol with highest per-call execution time and
 536        * scale accordingly:
 537        */
 538       log_scale = 0;
 539       top_dog = 0;
 540       top_time = 0.0;
 541       for (index = 0; index < symtab.len; ++index)
 542         {
 543           sym = time_sorted_syms[index];
 544           if (sym->ncalls)
 545             {
 546               time = (sym->hist.time + sym->cg.child_time) / sym->ncalls;
 547               if (time > top_time)
 548                 {
 549                   top_dog = sym;
 550                   top_time = time;
 551                 }
 552             }
 553         }
 554       if (top_dog && top_dog->ncalls && top_time > 0.0)
 555         {
 556           top_time /= hz;
 557           while (SItab[log_scale].scale * top_time < 1000.0
 558                  && log_scale < sizeof (SItab) / sizeof (SItab[0]) - 1)
 559             {
 560               ++log_scale;
 561             }
 562         }
 563     }
 564
 565   /*
 566    * For now, the dimension is always seconds.  In the future, we
 567    * may also want to support other (pseudo-)dimensions (such as
 568    * I-cache misses etc.).
 569    */
 570   print_header (SItab[log_scale].prefix);
 571   for (index = 0; index < symtab.len; ++index)
 572     {
 573       addr = time_sorted_syms[index]->addr;
 574       /*
 575        * Print symbol if its in INCL_FLAT table or that table
 576        * is empty and the symbol is not in EXCL_FLAT.
 577        */
 578       if (sym_lookup (&syms[INCL_FLAT], addr)
 579           || (syms[INCL_FLAT].len == 0
 580               && !sym_lookup (&syms[EXCL_FLAT], addr)))
 581         {
 582           print_line (time_sorted_syms[index], SItab[log_scale].scale);
 583         }
 584     }
 585   free (time_sorted_syms);
 586
 587   if (print_descriptions && !bsd_style_output)
 588     {
 589       flat_blurb (stdout);
 590     }
 591 }