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 static void scale_and_align_entries PARAMS ((void));
  16
  17 /* declarations of automatically generated functions to output blurbs: */
  18 extern void flat_blurb PARAMS ((FILE * fp));
  19
  20 bfd_vma s_lowpc;                /* lowest address in .text */
  21 bfd_vma s_highpc = 0;           /* highest address in .text */
  22 bfd_vma lowpc, highpc;          /* same, but expressed in UNITs */
  23 int hist_num_bins = 0;          /* number of histogram samples */
  24 int *hist_sample = 0;           /* histogram samples (shorts in the file!) */
  25 double hist_scale;
  26 char hist_dimension[sizeof (((struct gmon_hist_hdr *) 0)->dimen) + 1] =
  27   "seconds";
  28 char hist_dimension_abbrev = 's';
  29
  30 static double accum_time;       /* accumulated time so far for print_line() */
  31 static double total_time;       /* total time for all routines */
  32 /*
  33  * Table of SI prefixes for powers of 10 (used to automatically
  34  * scale some of the values in the flat profile).
  35  */
  36 const struct
  37   {
  38     char prefix;
  39     double scale;
  40   }
  41 SItab[] =
  42 {
  43   {
  44     'T', 1e-12
  45   }
  46   ,                             /* tera */
  47   {
  48     'G', 1e-09
  49   }
  50   ,                             /* giga */
  51   {
  52     'M', 1e-06
  53   }
  54   ,                             /* mega */
  55   {
  56     'K', 1e-03
  57   }
  58   ,                             /* kilo */
  59   {
  60     ' ', 1e-00
  61   }
  62   ,
  63   {
  64     'm', 1e+03
  65   }
  66   ,                             /* milli */
  67   {
  68     'u', 1e+06
  69   }
  70   ,                             /* micro */
  71   {
  72     'n', 1e+09
  73   }
  74   ,                             /* nano */
  75   {
  76     'p', 1e+12
  77   }
  78   ,                             /* pico */
  79   {
  80     'f', 1e+15
  81   }
  82   ,                             /* femto */
  83   {
  84     'a', 1e+18
  85   }
  86   ,                             /* ato */
  87 };
  88
  89 /*
  90  * Read the histogram from file IFP.  FILENAME is the name of IFP and
  91  * is provided for formatting error messages only.
  92  */
  93 void
  94 DEFUN (hist_read_rec, (ifp, filename), FILE * ifp AND const char *filename)
  95 {
  96   struct gmon_hist_hdr hdr;
  97   bfd_vma n_lowpc, n_highpc;
  98   int i, ncnt, profrate;
  99   UNIT count;
 100
 101   if (fread (&hdr, sizeof (hdr), 1, ifp) != 1)
 102     {
 103       fprintf (stderr, "%s: %s: unexpected end of file\n",
 104                whoami, filename);
 105       done (1);
 106     }
 107
 108   n_lowpc = (bfd_vma) get_vma (core_bfd, (bfd_byte *) hdr.low_pc);
 109   n_highpc = (bfd_vma) get_vma (core_bfd, (bfd_byte *) hdr.high_pc);
 110   ncnt = bfd_get_32 (core_bfd, (bfd_byte *) hdr.hist_size);
 111   profrate = bfd_get_32 (core_bfd, (bfd_byte *) hdr.prof_rate);
 112   strncpy (hist_dimension, hdr.dimen, sizeof (hdr.dimen));
 113   hist_dimension[sizeof (hdr.dimen)] = '\0';
 114   hist_dimension_abbrev = hdr.dimen_abbrev;
 115
 116   if (!s_highpc)
 117     {
 118
 119       /* this is the first histogram record: */
 120
 121       s_lowpc = n_lowpc;
 122       s_highpc = n_highpc;
 123       lowpc = (bfd_vma) n_lowpc / sizeof (UNIT);
 124       highpc = (bfd_vma) n_highpc / sizeof (UNIT);
 125       hist_num_bins = ncnt;
 126       hz = profrate;
 127     }
 128
 129   DBG (SAMPLEDEBUG,
 130        printf ("[hist_read_rec] n_lowpc 0x%lx n_highpc 0x%lx ncnt %d\n",
 131                n_lowpc, n_highpc, ncnt);
 132        printf ("[hist_read_rec] s_lowpc 0x%lx s_highpc 0x%lx nsamples %d\n",
 133                s_lowpc, s_highpc, hist_num_bins);
 134        printf ("[hist_read_rec]   lowpc 0x%lx   highpc 0x%lx\n",
 135                lowpc, highpc));
 136
 137   if (n_lowpc != s_lowpc || n_highpc != s_highpc
 138       || ncnt != hist_num_bins || hz != profrate)
 139     {
 140       fprintf (stderr, "%s: `%s' is incompatible with first gmon file\n",
 141                whoami, filename);
 142       done (1);
 143     }
 144
 145   if (!hist_sample)
 146     {
 147       hist_sample = (int *) xmalloc (hist_num_bins * sizeof (hist_sample[0]));
 148       memset (hist_sample, 0, hist_num_bins * sizeof (hist_sample[0]));
 149     }
 150
 151   for (i = 0; i < hist_num_bins; ++i)
 152     {
 153       if (fread (&count[0], sizeof (count), 1, ifp) != 1)
 154         {
 155           fprintf (stderr,
 156                    "%s: %s: unexpected EOF after reading %d of %d samples\n",
 157                    whoami, filename, i, hist_num_bins);
 158           done (1);
 159         }
 160       hist_sample[i] += bfd_get_16 (core_bfd, (bfd_byte *) & count[0]);
 161     }
 162 }
 163
 164
 165 /*
 166  * Write execution histogram to file OFP.  FILENAME is the name
 167  * of OFP and is provided for formatting error-messages only.
 168  */
 169 void
 170 DEFUN (hist_write_hist, (ofp, filename), FILE * ofp AND const char *filename)
 171 {
 172   struct gmon_hist_hdr hdr;
 173   unsigned char tag;
 174   UNIT count;
 175   int i;
 176
 177   /* write header: */
 178
 179   tag = GMON_TAG_TIME_HIST;
 180   put_vma (core_bfd, s_lowpc, (bfd_byte *) hdr.low_pc);
 181   put_vma (core_bfd, s_highpc, (bfd_byte *) hdr.high_pc);
 182   bfd_put_32 (core_bfd, hist_num_bins, (bfd_byte *) hdr.hist_size);
 183   bfd_put_32 (core_bfd, hz, (bfd_byte *) hdr.prof_rate);
 184   strncpy (hdr.dimen, hist_dimension, sizeof (hdr.dimen));
 185   hdr.dimen_abbrev = hist_dimension_abbrev;
 186
 187   if (fwrite (&tag, sizeof (tag), 1, ofp) != 1
 188       || fwrite (&hdr, sizeof (hdr), 1, ofp) != 1)
 189     {
 190       perror (filename);
 191       done (1);
 192     }
 193
 194   for (i = 0; i < hist_num_bins; ++i)
 195     {
 196       bfd_put_16 (core_bfd, hist_sample[i], (bfd_byte *) & count[0]);
 197       if (fwrite (&count[0], sizeof (count), 1, ofp) != 1)
 198         {
 199           perror (filename);
 200           done (1);
 201         }
 202     }
 203 }
 204
 205
 206 /*
 207  * Calculate scaled entry point addresses (to save time in
 208  * hist_assign_samples), and, on architectures that have procedure
 209  * entry masks at the start of a function, possibly push the scaled
 210  * entry points over the procedure entry mask, if it turns out that
 211  * the entry point is in one bin and the code for a routine is in the
 212  * next bin.
 213  */
 214 static void
 215 scale_and_align_entries ()
 216 {
 217   Sym *sym;
 218   bfd_vma bin_of_entry;
 219   bfd_vma bin_of_code;
 220
 221   for (sym = symtab.base; sym < symtab.limit; sym++)
 222     {
 223       sym->hist.scaled_addr = sym->addr / sizeof (UNIT);
 224       bin_of_entry = (sym->hist.scaled_addr - lowpc) / hist_scale;
 225       bin_of_code = (sym->hist.scaled_addr + UNITS_TO_CODE - lowpc) / hist_scale;
 226       if (bin_of_entry < bin_of_code)
 227         {
 228           DBG (SAMPLEDEBUG,
 229                printf ("[scale_and_align_entries] pushing 0x%lx to 0x%lx\n",
 230                        sym->hist.scaled_addr,
 231                        sym->hist.scaled_addr + UNITS_TO_CODE));
 232           sym->hist.scaled_addr += UNITS_TO_CODE;
 233         }
 234     }
 235 }
 236
 237
 238 /*
 239  * Assign samples to the symbol to which they belong.
 240  *
 241  * Histogram bin I covers some address range [BIN_LOWPC,BIN_HIGH_PC)
 242  * which may overlap one more symbol address ranges.  If a symbol
 243  * overlaps with the bin's address range by O percent, then O percent
 244  * of the bin's count is credited to that symbol.
 245  *
 246  * There are three cases as to where BIN_LOW_PC and BIN_HIGH_PC can be
 247  * with respect to the symbol's address range [SYM_LOW_PC,
 248  * SYM_HIGH_PC) as shown in the following diagram.  OVERLAP computes
 249  * the distance (in UNITs) between the arrows, the fraction of the
 250  * sample that is to be credited to the symbol which starts at
 251  * SYM_LOW_PC.
 252  *
 253  *        sym_low_pc                                      sym_high_pc
 254  *             |                                               |
 255  *             v                                               v
 256  *
 257  *             +-----------------------------------------------+
 258  *             |                                               |
 259  *        |  ->|    |<-         ->|         |<-         ->|    |<-  |
 260  *        |         |             |         |             |         |
 261  *        +---------+             +---------+             +---------+
 262  *
 263  *        ^         ^             ^         ^             ^         ^
 264  *        |         |             |         |             |         |
 265  *   bin_low_pc bin_high_pc  bin_low_pc bin_high_pc  bin_low_pc bin_high_pc
 266  *
 267  * For the VAX we assert that samples will never fall in the first two
 268  * bytes of any routine, since that is the entry mask, thus we call
 269  * scale_and_align_entries() to adjust the entry points if the entry
 270  * mask falls in one bin but the code for the routine doesn't start
 271  * until the next bin.  In conjunction with the alignment of routine
 272  * addresses, this should allow us to have only one sample for every
 273  * four bytes of text space and never have any overlap (the two end
 274  * cases, above).
 275  */
 276 void
 277 DEFUN_VOID (hist_assign_samples)
 278 {
 279   bfd_vma bin_low_pc, bin_high_pc;
 280   bfd_vma sym_low_pc, sym_high_pc;
 281   bfd_vma overlap, addr;
 282   int bin_count, i, j;
 283   double time, credit;
 284
 285   /* read samples and assign to symbols: */
 286   hist_scale = highpc - lowpc;
 287   hist_scale /= hist_num_bins;
 288   scale_and_align_entries ();
 289
 290   /* iterate over all sample bins: */
 291
 292   for (i = 0, j = 1; i < hist_num_bins; ++i)
 293     {
 294       bin_count = hist_sample[i];
 295       if (!bin_count)
 296         {
 297           continue;
 298         }
 299       bin_low_pc = lowpc + (bfd_vma) (hist_scale * i);
 300       bin_high_pc = lowpc + (bfd_vma) (hist_scale * (i + 1));
 301       time = bin_count;
 302       DBG (SAMPLEDEBUG,
 303            printf (
 304       "[assign_samples] bin_low_pc=0x%lx, bin_high_pc=0x%lx, bin_count=%d\n",
 305                     sizeof (UNIT) * bin_low_pc, sizeof (UNIT) * bin_high_pc,
 306                     bin_count));
 307       total_time += time;
 308
 309       /* credit all symbols that are covered by bin I: */
 310
 311       for (j = j - 1; j < symtab.len; ++j)
 312         {
 313           sym_low_pc = symtab.base[j].hist.scaled_addr;
 314           sym_high_pc = symtab.base[j + 1].hist.scaled_addr;
 315           /*
 316            * If high end of bin is below entry address, go for next
 317            * bin:
 318            */
 319           if (bin_high_pc < sym_low_pc)
 320             {
 321               break;
 322             }
 323           /*
 324            * If low end of bin is above high end of symbol, go for
 325            * next symbol.
 326            */
 327           if (bin_low_pc >= sym_high_pc)
 328             {
 329               continue;
 330             }
 331           overlap =
 332             MIN (bin_high_pc, sym_high_pc) - MAX (bin_low_pc, sym_low_pc);
 333           if (overlap > 0)
 334             {
 335               DBG (SAMPLEDEBUG,
 336                    printf (
 337                             "[assign_samples] [0x%lx,0x%lx) %s gets %f ticks %ld overlap\n",
 338                             symtab.base[j].addr, sizeof (UNIT) * sym_high_pc,
 339                             symtab.base[j].name, overlap * time / hist_scale,
 340                             overlap));
 341               addr = symtab.base[j].addr;
 342               credit = overlap * time / hist_scale;
 343               /*
 344                * Credit symbol if it appears in INCL_FLAT or that
 345                * table is empty and it does not appear it in
 346                * EXCL_FLAT.
 347                */
 348               if (sym_lookup (&syms[INCL_FLAT], addr)
 349                   || (syms[INCL_FLAT].len == 0
 350                       && !sym_lookup (&syms[EXCL_FLAT], addr)))
 351                 {
 352                   symtab.base[j].hist.time += credit;
 353                 }
 354               else
 355                 {
 356                   total_time -= credit;
 357                 }
 358             }
 359         }
 360     }
 361   DBG (SAMPLEDEBUG, printf ("[assign_samples] total_time %f\n",
 362                             total_time));
 363 }
 364
 365
 366 /*
 367  * Print header for flag histogram profile:
 368  */
 369 static void
 370 DEFUN (print_header, (prefix), const char prefix)
 371 {
 372   char unit[64];
 373
 374   sprintf (unit, "%c%c/call", prefix, hist_dimension_abbrev);
 375
 376   if (bsd_style_output)
 377     {
 378       printf ("\ngranularity: each sample hit covers %ld byte(s)",
 379               (long) hist_scale * sizeof (UNIT));
 380       if (total_time > 0.0)
 381         {
 382           printf (" for %.2f%% of %.2f %s\n\n",
 383                   100.0 / total_time, total_time / hz, hist_dimension);
 384         }
 385     }
 386   else
 387     {
 388       printf ("\nEach sample counts as %g %s.\n", 1.0 / hz, hist_dimension);
 389     }
 390
 391   if (total_time <= 0.0)
 392     {
 393       printf (" no time accumulated\n\n");
 394       /* this doesn't hurt since all the numerators will be zero: */
 395       total_time = 1.0;
 396     }
 397
 398   printf ("%5.5s %10.10s %8.8s %8.8s %8.8s %8.8s  %-8.8s\n",
 399           "%  ", "cumulative", "self  ", "", "self  ", "total ", "");
 400   printf ("%5.5s %9.9s  %8.8s %8.8s %8.8s %8.8s  %-8.8s\n",
 401           "time", hist_dimension, hist_dimension, "calls", unit, unit,
 402           "name");
 403 }
 404
 405
 406 static void
 407 DEFUN (print_line, (sym, scale), Sym * sym AND double scale)
 408 {
 409   if (ignore_zeros && sym->ncalls == 0 && sym->hist.time == 0)
 410     {
 411       return;
 412     }
 413
 414   accum_time += sym->hist.time;
 415   if (bsd_style_output)
 416     {
 417       printf ("%5.1f %10.2f %8.2f",
 418               total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0,
 419               accum_time / hz, sym->hist.time / hz);
 420     }
 421   else
 422     {
 423       printf ("%6.2f %9.2f %8.2f",
 424               total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0,
 425               accum_time / hz, sym->hist.time / hz);
 426     }
 427   if (sym->ncalls)
 428     {
 429       printf (" %8d %8.2f %8.2f  ",
 430               sym->ncalls, scale * sym->hist.time / hz / sym->ncalls,
 431           scale * (sym->hist.time + sym->cg.child_time) / hz / sym->ncalls);
 432     }
 433   else
 434     {
 435       printf (" %8.8s %8.8s %8.8s  ", "", "", "");
 436     }
 437   if (bsd_style_output)
 438     {
 439       print_name (sym);
 440     }
 441   else
 442     {
 443       print_name_only (sym);
 444     }
 445   printf ("\n");
 446 }
 447
 448
 449 /*
 450  * Compare LP and RP.  The primary comparison key is execution time,
 451  * the secondary is number of invocation, and the tertiary is the
 452  * lexicographic order of the function names.
 453  */
 454 static int
 455 DEFUN (cmp_time, (lp, rp), const PTR lp AND const PTR rp)
 456 {
 457   const Sym *left = *(const Sym **) lp;
 458   const Sym *right = *(const Sym **) rp;
 459   double time_diff;
 460   long call_diff;
 461
 462   time_diff = right->hist.time - left->hist.time;
 463   if (time_diff > 0.0)
 464     {
 465       return 1;
 466     }
 467   if (time_diff < 0.0)
 468     {
 469       return -1;
 470     }
 471
 472   call_diff = right->ncalls - left->ncalls;
 473   if (call_diff > 0)
 474     {
 475       return 1;
 476     }
 477   if (call_diff < 0)
 478     {
 479       return -1;
 480     }
 481
 482   return strcmp (left->name, right->name);
 483 }
 484
 485
 486 /*
 487  * Print the flat histogram profile.
 488  */
 489 void
 490 DEFUN_VOID (hist_print)
 491 {
 492   Sym **time_sorted_syms, *top_dog, *sym;
 493   int index, log_scale;
 494   double top_time, time;
 495   bfd_vma addr;
 496
 497   if (first_output)
 498     {
 499       first_output = FALSE;
 500     }
 501   else
 502     {
 503       printf ("\f\n");
 504     }
 505
 506   accum_time = 0.0;
 507   if (bsd_style_output)
 508     {
 509       if (print_descriptions)
 510         {
 511           printf ("\n\n\nflat profile:\n");
 512           flat_blurb (stdout);
 513         }
 514     }
 515   else
 516     {
 517       printf ("Flat profile:\n");
 518     }
 519   /*
 520    * Sort the symbol table by time (call-count and name as secondary
 521    * and tertiary keys):
 522    */
 523   time_sorted_syms = (Sym **) xmalloc (symtab.len * sizeof (Sym *));
 524   for (index = 0; index < symtab.len; ++index)
 525     {
 526       time_sorted_syms[index] = &symtab.base[index];
 527     }
 528   qsort (time_sorted_syms, symtab.len, sizeof (Sym *), cmp_time);
 529
 530   if (bsd_style_output)
 531     {
 532       log_scale = 5;            /* milli-seconds is BSD-default */
 533     }
 534   else
 535     {
 536       /*
 537        * Search for symbol with highest per-call execution time and
 538        * scale accordingly:
 539        */
 540       log_scale = 0;
 541       top_dog = 0;
 542       top_time = 0.0;
 543       for (index = 0; index < symtab.len; ++index)
 544         {
 545           sym = time_sorted_syms[index];
 546           if (sym->ncalls)
 547             {
 548               time = (sym->hist.time + sym->cg.child_time) / sym->ncalls;
 549               if (time > top_time)
 550                 {
 551                   top_dog = sym;
 552                   top_time = time;
 553                 }
 554             }
 555         }
 556       if (top_dog && top_dog->ncalls && top_time > 0.0)
 557         {
 558           top_time /= hz;
 559           while (SItab[log_scale].scale * top_time < 1000.0
 560                  && log_scale < sizeof (SItab) / sizeof (SItab[0]) - 1)
 561             {
 562               ++log_scale;
 563             }
 564         }
 565     }
 566
 567   /*
 568    * For now, the dimension is always seconds.  In the future, we
 569    * may also want to support other (pseudo-)dimensions (such as
 570    * I-cache misses etc.).
 571    */
 572   print_header (SItab[log_scale].prefix);
 573   for (index = 0; index < symtab.len; ++index)
 574     {
 575       addr = time_sorted_syms[index]->addr;
 576       /*
 577        * Print symbol if its in INCL_FLAT table or that table
 578        * is empty and the symbol is not in EXCL_FLAT.
 579        */
 580       if (sym_lookup (&syms[INCL_FLAT], addr)
 581           || (syms[INCL_FLAT].len == 0
 582               && !sym_lookup (&syms[EXCL_FLAT], addr)))
 583         {
 584           print_line (time_sorted_syms[index], SItab[log_scale].scale);
 585         }
 586     }
 587   free (time_sorted_syms);
 588
 589   if (print_descriptions && !bsd_style_output)
 590     {
 591       flat_blurb (stdout);
 592     }
 593 }