2 * Histogram related operations.
15 static void scale_and_align_entries
PARAMS ((void));
17 /* declarations of automatically generated functions to output blurbs: */
18 extern void flat_blurb
PARAMS ((FILE * fp
));
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!) */
26 char hist_dimension
[sizeof (((struct gmon_hist_hdr
*) 0)->dimen
) + 1] =
28 char hist_dimension_abbrev
= 's';
30 static double accum_time
; /* accumulated time so far for print_line() */
31 static double total_time
; /* total time for all routines */
33 * Table of SI prefixes for powers of 10 (used to automatically
34 * scale some of the values in the flat profile).
90 * Read the histogram from file IFP. FILENAME is the name of IFP and
91 * is provided for formatting error messages only.
94 DEFUN (hist_read_rec
, (ifp
, filename
), FILE * ifp AND
const char *filename
)
96 struct gmon_hist_hdr hdr
;
97 bfd_vma n_lowpc
, n_highpc
;
98 int i
, ncnt
, profrate
;
101 if (fread (&hdr
, sizeof (hdr
), 1, ifp
) != 1)
103 fprintf (stderr
, "%s: %s: unexpected end of file\n",
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
;
119 /* this is the first histogram record: */
123 lowpc
= (bfd_vma
) n_lowpc
/ sizeof (UNIT
);
124 highpc
= (bfd_vma
) n_highpc
/ sizeof (UNIT
);
125 hist_num_bins
= ncnt
;
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",
137 if (n_lowpc
!= s_lowpc
|| n_highpc
!= s_highpc
138 || ncnt
!= hist_num_bins
|| hz
!= profrate
)
140 fprintf (stderr
, "%s: `%s' is incompatible with first gmon file\n",
147 hist_sample
= (int *) xmalloc (hist_num_bins
* sizeof (hist_sample
[0]));
148 memset (hist_sample
, 0, hist_num_bins
* sizeof (hist_sample
[0]));
151 for (i
= 0; i
< hist_num_bins
; ++i
)
153 if (fread (&count
[0], sizeof (count
), 1, ifp
) != 1)
156 "%s: %s: unexpected EOF after reading %d of %d samples\n",
157 whoami
, filename
, i
, hist_num_bins
);
160 hist_sample
[i
] += bfd_get_16 (core_bfd
, (bfd_byte
*) & count
[0]);
166 * Write execution histogram to file OFP. FILENAME is the name
167 * of OFP and is provided for formatting error-messages only.
170 DEFUN (hist_write_hist
, (ofp
, filename
), FILE * ofp AND
const char *filename
)
172 struct gmon_hist_hdr hdr
;
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
;
187 if (fwrite (&tag
, sizeof (tag
), 1, ofp
) != 1
188 || fwrite (&hdr
, sizeof (hdr
), 1, ofp
) != 1)
194 for (i
= 0; i
< hist_num_bins
; ++i
)
196 bfd_put_16 (core_bfd
, hist_sample
[i
], (bfd_byte
*) & count
[0]);
197 if (fwrite (&count
[0], sizeof (count
), 1, ofp
) != 1)
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
215 scale_and_align_entries ()
218 bfd_vma bin_of_entry
;
221 for (sym
= symtab
.base
; sym
< symtab
.limit
; sym
++)
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
)
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
;
239 * Assign samples to the symbol to which they belong.
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.
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
253 * sym_low_pc sym_high_pc
257 * +-----------------------------------------------+
259 * | ->| |<- ->| |<- ->| |<- |
261 * +---------+ +---------+ +---------+
265 * bin_low_pc bin_high_pc bin_low_pc bin_high_pc bin_low_pc bin_high_pc
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
277 DEFUN_VOID (hist_assign_samples
)
279 bfd_vma bin_low_pc
, bin_high_pc
;
280 bfd_vma sym_low_pc
, sym_high_pc
;
281 bfd_vma overlap
, addr
;
285 /* read samples and assign to symbols: */
286 hist_scale
= highpc
- lowpc
;
287 hist_scale
/= hist_num_bins
;
288 scale_and_align_entries ();
290 /* iterate over all sample bins: */
292 for (i
= 0, j
= 1; i
< hist_num_bins
; ++i
)
294 bin_count
= hist_sample
[i
];
299 bin_low_pc
= lowpc
+ (bfd_vma
) (hist_scale
* i
);
300 bin_high_pc
= lowpc
+ (bfd_vma
) (hist_scale
* (i
+ 1));
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
,
309 /* credit all symbols that are covered by bin I: */
311 for (j
= j
- 1; j
< symtab
.len
; ++j
)
313 sym_low_pc
= symtab
.base
[j
].hist
.scaled_addr
;
314 sym_high_pc
= symtab
.base
[j
+ 1].hist
.scaled_addr
;
316 * If high end of bin is below entry address, go for next
319 if (bin_high_pc
< sym_low_pc
)
324 * If low end of bin is above high end of symbol, go for
327 if (bin_low_pc
>= sym_high_pc
)
332 MIN (bin_high_pc
, sym_high_pc
) - MAX (bin_low_pc
, sym_low_pc
);
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
,
341 addr
= symtab
.base
[j
].addr
;
342 credit
= overlap
* time
/ hist_scale
;
344 * Credit symbol if it appears in INCL_FLAT or that
345 * table is empty and it does not appear it in
348 if (sym_lookup (&syms
[INCL_FLAT
], addr
)
349 || (syms
[INCL_FLAT
].len
== 0
350 && !sym_lookup (&syms
[EXCL_FLAT
], addr
)))
352 symtab
.base
[j
].hist
.time
+= credit
;
356 total_time
-= credit
;
361 DBG (SAMPLEDEBUG
, printf ("[assign_samples] total_time %f\n",
367 * Print header for flag histogram profile:
370 DEFUN (print_header
, (prefix
), const char prefix
)
374 sprintf (unit
, "%c%c/call", prefix
, hist_dimension_abbrev
);
376 if (bsd_style_output
)
378 printf ("\ngranularity: each sample hit covers %ld byte(s)",
379 (long) hist_scale
* sizeof (UNIT
));
380 if (total_time
> 0.0)
382 printf (" for %.2f%% of %.2f %s\n\n",
383 100.0 / total_time
, total_time
/ hz
, hist_dimension
);
388 printf ("\nEach sample counts as %g %s.\n", 1.0 / hz
, hist_dimension
);
391 if (total_time
<= 0.0)
393 printf (" no time accumulated\n\n");
394 /* this doesn't hurt since all the numerators will be zero: */
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
,
407 DEFUN (print_line
, (sym
, scale
), Sym
* sym AND
double scale
)
409 if (ignore_zeros
&& sym
->ncalls
== 0 && sym
->hist
.time
== 0)
414 accum_time
+= sym
->hist
.time
;
415 if (bsd_style_output
)
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
);
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
);
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
);
435 printf (" %8.8s %8.8s %8.8s ", "", "", "");
437 if (bsd_style_output
)
443 print_name_only (sym
);
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.
455 DEFUN (cmp_time
, (lp
, rp
), const PTR lp AND
const PTR rp
)
457 const Sym
*left
= *(const Sym
**) lp
;
458 const Sym
*right
= *(const Sym
**) rp
;
462 time_diff
= right
->hist
.time
- left
->hist
.time
;
472 call_diff
= right
->ncalls
- left
->ncalls
;
482 return strcmp (left
->name
, right
->name
);
487 * Print the flat histogram profile.
490 DEFUN_VOID (hist_print
)
492 Sym
**time_sorted_syms
, *top_dog
, *sym
;
493 int index
, log_scale
;
494 double top_time
, time
;
499 first_output
= FALSE
;
507 if (bsd_style_output
)
509 if (print_descriptions
)
511 printf ("\n\n\nflat profile:\n");
517 printf ("Flat profile:\n");
520 * Sort the symbol table by time (call-count and name as secondary
521 * and tertiary keys):
523 time_sorted_syms
= (Sym
**) xmalloc (symtab
.len
* sizeof (Sym
*));
524 for (index
= 0; index
< symtab
.len
; ++index
)
526 time_sorted_syms
[index
] = &symtab
.base
[index
];
528 qsort (time_sorted_syms
, symtab
.len
, sizeof (Sym
*), cmp_time
);
530 if (bsd_style_output
)
532 log_scale
= 5; /* milli-seconds is BSD-default */
537 * Search for symbol with highest per-call execution time and
543 for (index
= 0; index
< symtab
.len
; ++index
)
545 sym
= time_sorted_syms
[index
];
548 time
= (sym
->hist
.time
+ sym
->cg
.child_time
) / sym
->ncalls
;
556 if (top_dog
&& top_dog
->ncalls
&& top_time
> 0.0)
559 while (SItab
[log_scale
].scale
* top_time
< 1000.0
560 && log_scale
< sizeof (SItab
) / sizeof (SItab
[0]) - 1)
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.).
572 print_header (SItab
[log_scale
].prefix
);
573 for (index
= 0; index
< symtab
.len
; ++index
)
575 addr
= time_sorted_syms
[index
]->addr
;
577 * Print symbol if its in INCL_FLAT table or that table
578 * is empty and the symbol is not in EXCL_FLAT.
580 if (sym_lookup (&syms
[INCL_FLAT
], addr
)
581 || (syms
[INCL_FLAT
].len
== 0
582 && !sym_lookup (&syms
[EXCL_FLAT
], addr
)))
584 print_line (time_sorted_syms
[index
], SItab
[log_scale
].scale
);
587 free (time_sorted_syms
);
589 if (print_descriptions
&& !bsd_style_output
)
This page took 0.042407 seconds and 4 git commands to generate.