2 * Histogram related operations.
15 /* declarations of automatically generated functions to output blurbs: */
16 extern void flat_blurb PARAMS ((FILE * fp));
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!) */
24 char hist_dimension[sizeof (((struct gmon_hist_hdr *) 0)->dimen) + 1] =
26 char hist_dimension_abbrev = 's';
28 static double accum_time; /* accumulated time so far for print_line() */
29 static double total_time; /* total time for all routines */
31 * Table of SI prefixes for powers of 10 (used to automatically
32 * scale some of the values in the flat profile).
88 * Read the histogram from file IFP. FILENAME is the name of IFP and
89 * is provided for formatting error messages only.
92 DEFUN (hist_read_rec, (ifp, filename), FILE * ifp AND const char *filename)
94 struct gmon_hist_hdr hdr;
95 bfd_vma n_lowpc, n_highpc;
96 int i, ncnt, profrate;
99 if (fread (&hdr, sizeof (hdr), 1, ifp) != 1)
101 fprintf (stderr, "%s: %s: unexpected end of file\n",
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;
117 /* this is the first histogram record: */
121 lowpc = (bfd_vma) n_lowpc / sizeof (UNIT);
122 highpc = (bfd_vma) n_highpc / sizeof (UNIT);
123 hist_num_bins = ncnt;
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",
135 if (n_lowpc != s_lowpc || n_highpc != s_highpc
136 || ncnt != hist_num_bins || hz != profrate)
138 fprintf (stderr, "%s: `%s' is incompatible with first gmon file\n",
145 hist_sample = (int *) xmalloc (hist_num_bins * sizeof (hist_sample[0]));
146 memset (hist_sample, 0, hist_num_bins * sizeof (hist_sample[0]));
149 for (i = 0; i < hist_num_bins; ++i)
151 if (fread (&count[0], sizeof (count), 1, ifp) != 1)
154 "%s: %s: unexpected EOF after reading %d of %d samples\n",
155 whoami, filename, i, hist_num_bins);
158 hist_sample[i] += bfd_get_16 (core_bfd, (bfd_byte *) & count[0]);
164 * Write execution histogram to file OFP. FILENAME is the name
165 * of OFP and is provided for formatting error-messages only.
168 DEFUN (hist_write_hist, (ofp, filename), FILE * ofp AND const char *filename)
170 struct gmon_hist_hdr hdr;
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;
185 if (fwrite (&tag, sizeof (tag), 1, ofp) != 1
186 || fwrite (&hdr, sizeof (hdr), 1, ofp) != 1)
192 for (i = 0; i < hist_num_bins; ++i)
194 bfd_put_16 (core_bfd, hist_sample[i], (bfd_byte *) & count[0]);
195 if (fwrite (&count[0], sizeof (count), 1, ofp) != 1)
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
213 scale_and_align_entries ()
216 bfd_vma bin_of_entry;
219 for (sym = symtab.base; sym < symtab.limit; sym++)
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)
227 printf ("[scale_and_align_entries] pushing 0x%lx to 0x%lx\n",
228 sym->hist.scaled_addr, sym->scaled_addr + UNITS_TO_CODE));
229 sym->scaled_addr += UNITS_TO_CODE;
236 * Assign samples to the symbol to which they belong.
238 * Histogram bin I covers some address range [BIN_LOWPC,BIN_HIGH_PC)
239 * which may overlap one more symbol address ranges. If a symbol
240 * overlaps with the bin's address range by O percent, then O percent
241 * of the bin's count is credited to that symbol.
243 * There are three cases as to where BIN_LOW_PC and BIN_HIGH_PC can be
244 * with respect to the symbol's address range [SYM_LOW_PC,
245 * SYM_HIGH_PC) as shown in the following diagram. OVERLAP computes
246 * the distance (in UNITs) between the arrows, the fraction of the
247 * sample that is to be credited to the symbol which starts at
250 * sym_low_pc sym_high_pc
254 * +-----------------------------------------------+
256 * | ->| |<- ->| |<- ->| |<- |
258 * +---------+ +---------+ +---------+
262 * bin_low_pc bin_high_pc bin_low_pc bin_high_pc bin_low_pc bin_high_pc
264 * For the VAX we assert that samples will never fall in the first two
265 * bytes of any routine, since that is the entry mask, thus we call
266 * scale_and_align_entries() to adjust the entry points if the entry
267 * mask falls in one bin but the code for the routine doesn't start
268 * until the next bin. In conjunction with the alignment of routine
269 * addresses, this should allow us to have only one sample for every
270 * four bytes of text space and never have any overlap (the two end
274 DEFUN_VOID (hist_assign_samples)
276 bfd_vma bin_low_pc, bin_high_pc;
277 bfd_vma sym_low_pc, sym_high_pc;
278 bfd_vma overlap, addr;
282 /* read samples and assign to symbols: */
283 hist_scale = highpc - lowpc;
284 hist_scale /= hist_num_bins;
285 scale_and_align_entries ();
287 /* iterate over all sample bins: */
289 for (i = 0, j = 1; i < hist_num_bins; ++i)
291 bin_count = hist_sample[i];
296 bin_low_pc = lowpc + (bfd_vma) (hist_scale * i);
297 bin_high_pc = lowpc + (bfd_vma) (hist_scale * (i + 1));
301 "[assign_samples] bin_low_pc=0x%lx, bin_high_pc=0x%lx, bin_count=%d\n",
302 sizeof (UNIT) * bin_low_pc, sizeof (UNIT) * bin_high_pc,
306 /* credit all symbols that are covered by bin I: */
308 for (j = j - 1; j < symtab.len; ++j)
310 sym_low_pc = symtab.base[j].hist.scaled_addr;
311 sym_high_pc = symtab.base[j + 1].hist.scaled_addr;
313 * If high end of bin is below entry address, go for next
316 if (bin_high_pc < sym_low_pc)
321 * If low end of bin is above high end of symbol, go for
324 if (bin_low_pc >= sym_high_pc)
329 MIN (bin_high_pc, sym_high_pc) - MAX (bin_low_pc, sym_low_pc);
334 "[assign_samples] [0x%lx,0x%lx) %s gets %f ticks %ld overlap\n",
335 symtab.base[j].addr, sizeof (UNIT) * sym_high_pc,
336 symtab.base[j].name, overlap * time / hist_scale,
338 addr = symtab.base[j].addr;
339 credit = overlap * time / hist_scale;
341 * Credit symbol if it appears in INCL_FLAT or that
342 * table is empty and it does not appear it in
345 if (sym_lookup (&syms[INCL_FLAT], addr)
346 || (syms[INCL_FLAT].len == 0
347 && !sym_lookup (&syms[EXCL_FLAT], addr)))
349 symtab.base[j].hist.time += credit;
353 total_time -= credit;
358 DBG (SAMPLEDEBUG, printf ("[assign_samples] total_time %f\n",
364 * Print header for flag histogram profile:
367 DEFUN (print_header, (prefix), const char prefix)
371 sprintf (unit, "%c%c/call", prefix, hist_dimension_abbrev);
373 if (bsd_style_output)
375 printf ("\ngranularity: each sample hit covers %ld byte(s)",
376 (long) hist_scale * sizeof (UNIT));
377 if (total_time > 0.0)
379 printf (" for %.2f%% of %.2f %s\n\n",
380 100.0 / total_time, total_time / hz, hist_dimension);
385 printf ("\nEach sample counts as %g %s.\n", 1.0 / hz, hist_dimension);
388 if (total_time <= 0.0)
390 printf (" no time accumulated\n\n");
391 /* this doesn't hurt since all the numerators will be zero: */
395 printf ("%5.5s %10.10s %8.8s %8.8s %8.8s %8.8s %-8.8s\n",
396 "% ", "cumulative", "self ", "", "self ", "total ", "");
397 printf ("%5.5s %9.9s %8.8s %8.8s %8.8s %8.8s %-8.8s\n",
398 "time", hist_dimension, hist_dimension, "calls", unit, unit,
404 DEFUN (print_line, (sym, scale), Sym * sym AND double scale)
406 if (ignore_zeros && sym->ncalls == 0 && sym->hist.time == 0)
411 accum_time += sym->hist.time;
412 if (bsd_style_output)
414 printf ("%5.1f %10.2f %8.2f",
415 total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0,
416 accum_time / hz, sym->hist.time / hz);
420 printf ("%6.2f %9.2f %8.2f",
421 total_time > 0.0 ? 100 * sym->hist.time / total_time : 0.0,
422 accum_time / hz, sym->hist.time / hz);
426 printf (" %8d %8.2f %8.2f ",
427 sym->ncalls, scale * sym->hist.time / hz / sym->ncalls,
428 scale * (sym->hist.time + sym->cg.child_time) / hz / sym->ncalls);
432 printf (" %8.8s %8.8s %8.8s ", "", "", "");
434 if (bsd_style_output)
440 print_name_only (sym);
447 * Compare LP and RP. The primary comparison key is execution time,
448 * the secondary is number of invocation, and the tertiary is the
449 * lexicographic order of the function names.
452 DEFUN (cmp_time, (lp, rp), const PTR lp AND const PTR rp)
454 const Sym *left = *(const Sym **) lp;
455 const Sym *right = *(const Sym **) rp;
459 time_diff = right->hist.time - left->hist.time;
469 call_diff = right->ncalls - left->ncalls;
479 return strcmp (left->name, right->name);
484 * Print the flat histogram profile.
487 DEFUN_VOID (hist_print)
489 Sym **time_sorted_syms, *top_dog, *sym;
490 int index, log_scale;
491 double top_time, time;
496 first_output = FALSE;
504 if (bsd_style_output)
506 if (print_descriptions)
508 printf ("\n\n\nflat profile:\n");
514 printf ("Flat profile:\n");
517 * Sort the symbol table by time (call-count and name as secondary
518 * and tertiary keys):
520 time_sorted_syms = (Sym **) xmalloc (symtab.len * sizeof (Sym *));
521 for (index = 0; index < symtab.len; ++index)
523 time_sorted_syms[index] = &symtab.base[index];
525 qsort (time_sorted_syms, symtab.len, sizeof (Sym *), cmp_time);
527 if (bsd_style_output)
529 log_scale = 5; /* milli-seconds is BSD-default */
534 * Search for symbol with highest per-call execution time and
540 for (index = 0; index < symtab.len; ++index)
542 sym = time_sorted_syms[index];
545 time = (sym->hist.time + sym->cg.child_time) / sym->ncalls;
553 if (top_dog && top_dog->ncalls && top_time > 0.0)
556 while (SItab[log_scale].scale * top_time < 1000.0
557 && log_scale < sizeof (SItab) / sizeof (SItab[0]) - 1)
565 * For now, the dimension is always seconds. In the future, we
566 * may also want to support other (pseudo-)dimensions (such as
567 * I-cache misses etc.).
569 print_header (SItab[log_scale].prefix);
570 for (index = 0; index < symtab.len; ++index)
572 addr = time_sorted_syms[index]->addr;
574 * Print symbol if its in INCL_FLAT table or that table
575 * is empty and the symbol is not in EXCL_FLAT.
577 if (sym_lookup (&syms[INCL_FLAT], addr)
578 || (syms[INCL_FLAT].len == 0
579 && !sym_lookup (&syms[EXCL_FLAT], addr)))
581 print_line (time_sorted_syms[index], SItab[log_scale].scale);
584 free (time_sorted_syms);
586 if (print_descriptions && !bsd_style_output)