OpenCores
URL https://opencores.org/ocsvn/openrisc_2011-10-31/openrisc_2011-10-31/trunk

Subversion Repositories openrisc_2011-10-31

[/] [openrisc/] [trunk/] [gnu-src/] [binutils-2.20.1/] [gprof/] [gprof.texi] - Blame information for rev 373

Go to most recent revision | Details | Compare with Previous | View Log

Line No. Rev Author Line
1 205 julius
\input texinfo @c -*-texinfo-*-
2
@setfilename gprof.info
3
@c Copyright 1988, 1992, 1993, 1998, 1999, 2000, 2001, 2002, 2003,
4
@c 2004, 2007, 2008, 2009
5
@c Free Software Foundation, Inc.
6
@settitle GNU gprof
7
@setchapternewpage odd
8
 
9
@c man begin INCLUDE
10
@include bfdver.texi
11
@c man end
12
 
13
@ifinfo
14
@c This is a dir.info fragment to support semi-automated addition of
15
@c manuals to an info tree.  zoo@cygnus.com is developing this facility.
16
@format
17
START-INFO-DIR-ENTRY
18
* gprof: (gprof).                Profiling your program's execution
19
END-INFO-DIR-ENTRY
20
@end format
21
@end ifinfo
22
 
23
@copying
24
This file documents the gprof profiler of the GNU system.
25
 
26
@c man begin COPYRIGHT
27
Copyright @copyright{} 1988, 92, 97, 98, 99, 2000, 2001, 2003, 2007, 2008, 2009 Free Software Foundation, Inc.
28
 
29
Permission is granted to copy, distribute and/or modify this document
30
under the terms of the GNU Free Documentation License, Version 1.3
31
or any later version published by the Free Software Foundation;
32
with no Invariant Sections, with no Front-Cover Texts, and with no
33
Back-Cover Texts.  A copy of the license is included in the
34
section entitled ``GNU Free Documentation License''.
35
 
36
@c man end
37
@end copying
38
 
39
@finalout
40
@smallbook
41
 
42
@titlepage
43
@title GNU gprof
44
@subtitle The @sc{gnu} Profiler
45
@ifset VERSION_PACKAGE
46
@subtitle @value{VERSION_PACKAGE}
47
@end ifset
48
@subtitle Version @value{VERSION}
49
@author Jay Fenlason and Richard Stallman
50
 
51
@page
52
 
53
This manual describes the @sc{gnu} profiler, @code{gprof}, and how you
54
can use it to determine which parts of a program are taking most of the
55
execution time.  We assume that you know how to write, compile, and
56
execute programs.  @sc{gnu} @code{gprof} was written by Jay Fenlason.
57
Eric S. Raymond made some minor corrections and additions in 2003.
58
 
59
@vskip 0pt plus 1filll
60
Copyright @copyright{} 1988, 92, 97, 98, 99, 2000, 2003, 2008, 2009 Free Software Foundation, Inc.
61
 
62
      Permission is granted to copy, distribute and/or modify this document
63
      under the terms of the GNU Free Documentation License, Version 1.3
64
      or any later version published by the Free Software Foundation;
65
      with no Invariant Sections, with no Front-Cover Texts, and with no
66
      Back-Cover Texts.  A copy of the license is included in the
67
      section entitled ``GNU Free Documentation License''.
68
 
69
@end titlepage
70
@contents
71
 
72
@ifnottex
73
@node Top
74
@top Profiling a Program: Where Does It Spend Its Time?
75
 
76
This manual describes the @sc{gnu} profiler, @code{gprof}, and how you
77
can use it to determine which parts of a program are taking most of the
78
execution time.  We assume that you know how to write, compile, and
79
execute programs.  @sc{gnu} @code{gprof} was written by Jay Fenlason.
80
 
81
This manual is for @code{gprof}
82
@ifset VERSION_PACKAGE
83
@value{VERSION_PACKAGE}
84
@end ifset
85
version @value{VERSION}.
86
 
87
This document is distributed under the terms of the GNU Free
88
Documentation License version 1.3.  A copy of the license is included
89
in the section entitled ``GNU Free Documentation License''.
90
 
91
@menu
92
* Introduction::        What profiling means, and why it is useful.
93
 
94
* Compiling::           How to compile your program for profiling.
95
* Executing::           Executing your program to generate profile data
96
* Invoking::            How to run @code{gprof}, and its options
97
 
98
* Output::              Interpreting @code{gprof}'s output
99
 
100
* Inaccuracy::          Potential problems you should be aware of
101
* How do I?::           Answers to common questions
102
* Incompatibilities::   (between @sc{gnu} @code{gprof} and Unix @code{gprof}.)
103
* Details::             Details of how profiling is done
104
* GNU Free Documentation License::  GNU Free Documentation License
105
@end menu
106
@end ifnottex
107
 
108
@node Introduction
109
@chapter Introduction to Profiling
110
 
111
@ifset man
112
@c man title gprof display call graph profile data
113
 
114
@smallexample
115
@c man begin SYNOPSIS
116
gprof [ -[abcDhilLrsTvwxyz] ] [ -[ACeEfFJnNOpPqQZ][@var{name}] ]
117
 [ -I @var{dirs} ] [ -d[@var{num}] ] [ -k @var{from/to} ]
118
 [ -m @var{min-count} ] [ -R @var{map_file} ] [ -t @var{table-length} ]
119
 [ --[no-]annotated-source[=@var{name}] ]
120
 [ --[no-]exec-counts[=@var{name}] ]
121
 [ --[no-]flat-profile[=@var{name}] ] [ --[no-]graph[=@var{name}] ]
122
 [ --[no-]time=@var{name}] [ --all-lines ] [ --brief ]
123
 [ --debug[=@var{level}] ] [ --function-ordering ]
124
 [ --file-ordering @var{map_file} ] [ --directory-path=@var{dirs} ]
125
 [ --display-unused-functions ] [ --file-format=@var{name} ]
126
 [ --file-info ] [ --help ] [ --line ] [ --min-count=@var{n} ]
127
 [ --no-static ] [ --print-path ] [ --separate-files ]
128
 [ --static-call-graph ] [ --sum ] [ --table-length=@var{len} ]
129
 [ --traditional ] [ --version ] [ --width=@var{n} ]
130
 [ --ignore-non-functions ] [ --demangle[=@var{STYLE}] ]
131
 [ --no-demangle ] [--external-symbol-table=name]
132
 [ @var{image-file} ] [ @var{profile-file} @dots{} ]
133
@c man end
134
@end smallexample
135
 
136
@c man begin DESCRIPTION
137
@code{gprof} produces an execution profile of C, Pascal, or Fortran77
138
programs.  The effect of called routines is incorporated in the profile
139
of each caller.  The profile data is taken from the call graph profile file
140
(@file{gmon.out} default) which is created by programs
141
that are compiled with the @samp{-pg} option of
142
@code{cc}, @code{pc}, and @code{f77}.
143
The @samp{-pg} option also links in versions of the library routines
144
that are compiled for profiling.  @code{Gprof} reads the given object
145
file (the default is @code{a.out}) and establishes the relation between
146
its symbol table and the call graph profile from @file{gmon.out}.
147
If more than one profile file is specified, the @code{gprof}
148
output shows the sum of the profile information in the given profile files.
149
 
150
@code{Gprof} calculates the amount of time spent in each routine.
151
Next, these times are propagated along the edges of the call graph.
152
Cycles are discovered, and calls into a cycle are made to share the time
153
of the cycle.
154
 
155
@c man end
156
 
157
@c man begin BUGS
158
The granularity of the sampling is shown, but remains
159
statistical at best.
160
We assume that the time for each execution of a function
161
can be expressed by the total time for the function divided
162
by the number of times the function is called.
163
Thus the time propagated along the call graph arcs to the function's
164
parents is directly proportional to the number of times that
165
arc is traversed.
166
 
167
Parents that are not themselves profiled will have the time of
168
their profiled children propagated to them, but they will appear
169
to be spontaneously invoked in the call graph listing, and will
170
not have their time propagated further.
171
Similarly, signal catchers, even though profiled, will appear
172
to be spontaneous (although for more obscure reasons).
173
Any profiled children of signal catchers should have their times
174
propagated properly, unless the signal catcher was invoked during
175
the execution of the profiling routine, in which case all is lost.
176
 
177
The profiled program must call @code{exit}(2)
178
or return normally for the profiling information to be saved
179
in the @file{gmon.out} file.
180
@c man end
181
 
182
@c man begin FILES
183
@table @code
184
@item @file{a.out}
185
the namelist and text space.
186
@item @file{gmon.out}
187
dynamic call graph and profile.
188
@item @file{gmon.sum}
189
summarized dynamic call graph and profile.
190
@end table
191
@c man end
192
 
193
@c man begin SEEALSO
194
monitor(3), profil(2), cc(1), prof(1), and the Info entry for @file{gprof}.
195
 
196
``An Execution Profiler for Modular Programs'',
197
by S. Graham, P. Kessler, M. McKusick;
198
Software - Practice and Experience,
199
Vol. 13, pp. 671-685, 1983.
200
 
201
``gprof: A Call Graph Execution Profiler'',
202
by S. Graham, P. Kessler, M. McKusick;
203
Proceedings of the SIGPLAN '82 Symposium on Compiler Construction,
204
SIGPLAN Notices, Vol. 17, No  6, pp. 120-126, June 1982.
205
@c man end
206
@end ifset
207
 
208
Profiling allows you to learn where your program spent its time and which
209
functions called which other functions while it was executing.  This
210
information can show you which pieces of your program are slower than you
211
expected, and might be candidates for rewriting to make your program
212
execute faster.  It can also tell you which functions are being called more
213
or less often than you expected.  This may help you spot bugs that had
214
otherwise been unnoticed.
215
 
216
Since the profiler uses information collected during the actual execution
217
of your program, it can be used on programs that are too large or too
218
complex to analyze by reading the source.  However, how your program is run
219
will affect the information that shows up in the profile data.  If you
220
don't use some feature of your program while it is being profiled, no
221
profile information will be generated for that feature.
222
 
223
Profiling has several steps:
224
 
225
@itemize @bullet
226
@item
227
You must compile and link your program with profiling enabled.
228
@xref{Compiling, ,Compiling a Program for Profiling}.
229
 
230
@item
231
You must execute your program to generate a profile data file.
232
@xref{Executing, ,Executing the Program}.
233
 
234
@item
235
You must run @code{gprof} to analyze the profile data.
236
@xref{Invoking, ,@code{gprof} Command Summary}.
237
@end itemize
238
 
239
The next three chapters explain these steps in greater detail.
240
 
241
@c man begin DESCRIPTION
242
 
243
Several forms of output are available from the analysis.
244
 
245
The @dfn{flat profile} shows how much time your program spent in each function,
246
and how many times that function was called.  If you simply want to know
247
which functions burn most of the cycles, it is stated concisely here.
248
@xref{Flat Profile, ,The Flat Profile}.
249
 
250
The @dfn{call graph} shows, for each function, which functions called it, which
251
other functions it called, and how many times.  There is also an estimate
252
of how much time was spent in the subroutines of each function.  This can
253
suggest places where you might try to eliminate function calls that use a
254
lot of time.  @xref{Call Graph, ,The Call Graph}.
255
 
256
The @dfn{annotated source} listing is a copy of the program's
257
source code, labeled with the number of times each line of the
258
program was executed.  @xref{Annotated Source, ,The Annotated Source
259
Listing}.
260
@c man end
261
 
262
To better understand how profiling works, you may wish to read
263
a description of its implementation.
264
@xref{Implementation, ,Implementation of Profiling}.
265
 
266
@node Compiling
267
@chapter Compiling a Program for Profiling
268
 
269
The first step in generating profile information for your program is
270
to compile and link it with profiling enabled.
271
 
272
To compile a source file for profiling, specify the @samp{-pg} option when
273
you run the compiler.  (This is in addition to the options you normally
274
use.)
275
 
276
To link the program for profiling, if you use a compiler such as @code{cc}
277
to do the linking, simply specify @samp{-pg} in addition to your usual
278
options.  The same option, @samp{-pg}, alters either compilation or linking
279
to do what is necessary for profiling.  Here are examples:
280
 
281
@example
282
cc -g -c myprog.c utils.c -pg
283
cc -o myprog myprog.o utils.o -pg
284
@end example
285
 
286
The @samp{-pg} option also works with a command that both compiles and links:
287
 
288
@example
289
cc -o myprog myprog.c utils.c -g -pg
290
@end example
291
 
292
Note: The @samp{-pg} option must be part of your compilation options
293
as well as your link options.  If it is not then no call-graph data
294
will be gathered and when you run @code{gprof} you will get an error
295
message like this:
296
 
297
@example
298
gprof: gmon.out file is missing call-graph data
299
@end example
300
 
301
If you add the @samp{-Q} switch to suppress the printing of the call
302
graph data you will still be able to see the time samples:
303
 
304
@example
305
Flat profile:
306
 
307
Each sample counts as 0.01 seconds.
308
  %   cumulative   self              self     total
309
 time   seconds   seconds    calls  Ts/call  Ts/call  name
310
 44.12      0.07     0.07                             zazLoop
311
 35.29      0.14     0.06                             main
312
 20.59      0.17     0.04                             bazMillion
313
@end example
314
 
315
If you run the linker @code{ld} directly instead of through a compiler
316
such as @code{cc}, you may have to specify a profiling startup file
317
@file{gcrt0.o} as the first input file instead of the usual startup
318
file @file{crt0.o}.  In addition, you would probably want to
319
specify the profiling C library, @file{libc_p.a}, by writing
320
@samp{-lc_p} instead of the usual @samp{-lc}.  This is not absolutely
321
necessary, but doing this gives you number-of-calls information for
322
standard library functions such as @code{read} and @code{open}.  For
323
example:
324
 
325
@example
326
ld -o myprog /lib/gcrt0.o myprog.o utils.o -lc_p
327
@end example
328
 
329
If you are running the program on a system which supports shared
330
libraries you may run into problems with the profiling support code in
331
a shared library being called before that library has been fully
332
initialised.  This is usually detected by the program encountering a
333
segmentation fault as soon as it is run.  The solution is to link
334
against a static version of the library containing the profiling
335
support code, which for @code{gcc} users can be done via the
336
@samp{-static} or @samp{-static-libgcc} command line option.  For
337
example:
338
 
339
@example
340
gcc -g -pg -static-libgcc myprog.c utils.c -o myprog
341
@end example
342
 
343
If you compile only some of the modules of the program with @samp{-pg}, you
344
can still profile the program, but you won't get complete information about
345
the modules that were compiled without @samp{-pg}.  The only information
346
you get for the functions in those modules is the total time spent in them;
347
there is no record of how many times they were called, or from where.  This
348
will not affect the flat profile (except that the @code{calls} field for
349
the functions will be blank), but will greatly reduce the usefulness of the
350
call graph.
351
 
352
If you wish to perform line-by-line profiling you should use the
353
@code{gcov} tool instead of @code{gprof}.  See that tool's manual or
354
info pages for more details of how to do this.
355
 
356
Note, older versions of @code{gcc} produce line-by-line profiling
357
information that works with @code{gprof} rather than @code{gcov} so
358
there is still support for displaying this kind of information in
359
@code{gprof}. @xref{Line-by-line, ,Line-by-line Profiling}.
360
 
361
It also worth noting that @code{gcc} implements a
362
@samp{-finstrument-functions} command line option which will insert
363
calls to special user supplied instrumentation routines at the entry
364
and exit of every function in their program.  This can be used to
365
implement an alternative profiling scheme.
366
 
367
@node Executing
368
@chapter Executing the Program
369
 
370
Once the program is compiled for profiling, you must run it in order to
371
generate the information that @code{gprof} needs.  Simply run the program
372
as usual, using the normal arguments, file names, etc.  The program should
373
run normally, producing the same output as usual.  It will, however, run
374
somewhat slower than normal because of the time spent collecting and
375
writing the profile data.
376
 
377
The way you run the program---the arguments and input that you give
378
it---may have a dramatic effect on what the profile information shows.  The
379
profile data will describe the parts of the program that were activated for
380
the particular input you use.  For example, if the first command you give
381
to your program is to quit, the profile data will show the time used in
382
initialization and in cleanup, but not much else.
383
 
384
Your program will write the profile data into a file called @file{gmon.out}
385
just before exiting.  If there is already a file called @file{gmon.out},
386
its contents are overwritten.  There is currently no way to tell the
387
program to write the profile data under a different name, but you can rename
388
the file afterwards if you are concerned that it may be overwritten.
389
 
390
In order to write the @file{gmon.out} file properly, your program must exit
391
normally: by returning from @code{main} or by calling @code{exit}.  Calling
392
the low-level function @code{_exit} does not write the profile data, and
393
neither does abnormal termination due to an unhandled signal.
394
 
395
The @file{gmon.out} file is written in the program's @emph{current working
396
directory} at the time it exits.  This means that if your program calls
397
@code{chdir}, the @file{gmon.out} file will be left in the last directory
398
your program @code{chdir}'d to.  If you don't have permission to write in
399
this directory, the file is not written, and you will get an error message.
400
 
401
Older versions of the @sc{gnu} profiling library may also write a file
402
called @file{bb.out}.  This file, if present, contains an human-readable
403
listing of the basic-block execution counts.  Unfortunately, the
404
appearance of a human-readable @file{bb.out} means the basic-block
405
counts didn't get written into @file{gmon.out}.
406
The Perl script @code{bbconv.pl}, included with the @code{gprof}
407
source distribution, will convert a @file{bb.out} file into
408
a format readable by @code{gprof}.  Invoke it like this:
409
 
410
@smallexample
411
bbconv.pl < bb.out > @var{bh-data}
412
@end smallexample
413
 
414
This translates the information in @file{bb.out} into a form that
415
@code{gprof} can understand.  But you still need to tell @code{gprof}
416
about the existence of this translated information.  To do that, include
417
@var{bb-data} on the @code{gprof} command line, @emph{along with
418
@file{gmon.out}}, like this:
419
 
420
@smallexample
421
gprof @var{options} @var{executable-file} gmon.out @var{bb-data} [@var{yet-more-profile-data-files}@dots{}] [> @var{outfile}]
422
@end smallexample
423
 
424
@node Invoking
425
@chapter @code{gprof} Command Summary
426
 
427
After you have a profile data file @file{gmon.out}, you can run @code{gprof}
428
to interpret the information in it.  The @code{gprof} program prints a
429
flat profile and a call graph on standard output.  Typically you would
430
redirect the output of @code{gprof} into a file with @samp{>}.
431
 
432
You run @code{gprof} like this:
433
 
434
@smallexample
435
gprof @var{options} [@var{executable-file} [@var{profile-data-files}@dots{}]] [> @var{outfile}]
436
@end smallexample
437
 
438
@noindent
439
Here square-brackets indicate optional arguments.
440
 
441
If you omit the executable file name, the file @file{a.out} is used.  If
442
you give no profile data file name, the file @file{gmon.out} is used.  If
443
any file is not in the proper format, or if the profile data file does not
444
appear to belong to the executable file, an error message is printed.
445
 
446
You can give more than one profile data file by entering all their names
447
after the executable file name; then the statistics in all the data files
448
are summed together.
449
 
450
The order of these options does not matter.
451
 
452
@menu
453
* Output Options::      Controlling @code{gprof}'s output style
454
* Analysis Options::    Controlling how @code{gprof} analyzes its data
455
* Miscellaneous Options::
456
* Deprecated Options::  Options you no longer need to use, but which
457
                            have been retained for compatibility
458
* Symspecs::            Specifying functions to include or exclude
459
@end menu
460
 
461
@node Output Options
462
@section Output Options
463
 
464
@c man begin OPTIONS
465
These options specify which of several output formats
466
@code{gprof} should produce.
467
 
468
Many of these options take an optional @dfn{symspec} to specify
469
functions to be included or excluded.  These options can be
470
specified multiple times, with different symspecs, to include
471
or exclude sets of symbols.  @xref{Symspecs, ,Symspecs}.
472
 
473
Specifying any of these options overrides the default (@samp{-p -q}),
474
which prints a flat profile and call graph analysis
475
for all functions.
476
 
477
@table @code
478
 
479
@item -A[@var{symspec}]
480
@itemx --annotated-source[=@var{symspec}]
481
The @samp{-A} option causes @code{gprof} to print annotated source code.
482
If @var{symspec} is specified, print output only for matching symbols.
483
@xref{Annotated Source, ,The Annotated Source Listing}.
484
 
485
@item -b
486
@itemx --brief
487
If the @samp{-b} option is given, @code{gprof} doesn't print the
488
verbose blurbs that try to explain the meaning of all of the fields in
489
the tables.  This is useful if you intend to print out the output, or
490
are tired of seeing the blurbs.
491
 
492
@item -C[@var{symspec}]
493
@itemx --exec-counts[=@var{symspec}]
494
The @samp{-C} option causes @code{gprof} to
495
print a tally of functions and the number of times each was called.
496
If @var{symspec} is specified, print tally only for matching symbols.
497
 
498
If the profile data file contains basic-block count records, specifying
499
the @samp{-l} option, along with @samp{-C}, will cause basic-block
500
execution counts to be tallied and displayed.
501
 
502
@item -i
503
@itemx --file-info
504
The @samp{-i} option causes @code{gprof} to display summary information
505
about the profile data file(s) and then exit.  The number of histogram,
506
call graph, and basic-block count records is displayed.
507
 
508
@item -I @var{dirs}
509
@itemx --directory-path=@var{dirs}
510
The @samp{-I} option specifies a list of search directories in
511
which to find source files.  Environment variable @var{GPROF_PATH}
512
can also be used to convey this information.
513
Used mostly for annotated source output.
514
 
515
@item -J[@var{symspec}]
516
@itemx --no-annotated-source[=@var{symspec}]
517
The @samp{-J} option causes @code{gprof} not to
518
print annotated source code.
519
If @var{symspec} is specified, @code{gprof} prints annotated source,
520
but excludes matching symbols.
521
 
522
@item -L
523
@itemx --print-path
524
Normally, source filenames are printed with the path
525
component suppressed.  The @samp{-L} option causes @code{gprof}
526
to print the full pathname of
527
source filenames, which is determined
528
from symbolic debugging information in the image file
529
and is relative to the directory in which the compiler
530
was invoked.
531
 
532
@item -p[@var{symspec}]
533
@itemx --flat-profile[=@var{symspec}]
534
The @samp{-p} option causes @code{gprof} to print a flat profile.
535
If @var{symspec} is specified, print flat profile only for matching symbols.
536
@xref{Flat Profile, ,The Flat Profile}.
537
 
538
@item -P[@var{symspec}]
539
@itemx --no-flat-profile[=@var{symspec}]
540
The @samp{-P} option causes @code{gprof} to suppress printing a flat profile.
541
If @var{symspec} is specified, @code{gprof} prints a flat profile,
542
but excludes matching symbols.
543
 
544
@item -q[@var{symspec}]
545
@itemx --graph[=@var{symspec}]
546
The @samp{-q} option causes @code{gprof} to print the call graph analysis.
547
If @var{symspec} is specified, print call graph only for matching symbols
548
and their children.
549
@xref{Call Graph, ,The Call Graph}.
550
 
551
@item -Q[@var{symspec}]
552
@itemx --no-graph[=@var{symspec}]
553
The @samp{-Q} option causes @code{gprof} to suppress printing the
554
call graph.
555
If @var{symspec} is specified, @code{gprof} prints a call graph,
556
but excludes matching symbols.
557
 
558
@item -t
559
@itemx --table-length=@var{num}
560
The @samp{-t} option causes the @var{num} most active source lines in
561
each source file to be listed when source annotation is enabled.  The
562
default is 10.
563
 
564
@item -y
565
@itemx --separate-files
566
This option affects annotated source output only.
567
Normally, @code{gprof} prints annotated source files
568
to standard-output.  If this option is specified,
569
annotated source for a file named @file{path/@var{filename}}
570
is generated in the file @file{@var{filename}-ann}.  If the underlying
571
file system would truncate @file{@var{filename}-ann} so that it
572
overwrites the original @file{@var{filename}}, @code{gprof} generates
573
annotated source in the file @file{@var{filename}.ann} instead (if the
574
original file name has an extension, that extension is @emph{replaced}
575
with @file{.ann}).
576
 
577
@item -Z[@var{symspec}]
578
@itemx --no-exec-counts[=@var{symspec}]
579
The @samp{-Z} option causes @code{gprof} not to
580
print a tally of functions and the number of times each was called.
581
If @var{symspec} is specified, print tally, but exclude matching symbols.
582
 
583
@item -r
584
@itemx --function-ordering
585
The @samp{--function-ordering} option causes @code{gprof} to print a
586
suggested function ordering for the program based on profiling data.
587
This option suggests an ordering which may improve paging, tlb and
588
cache behavior for the program on systems which support arbitrary
589
ordering of functions in an executable.
590
 
591
The exact details of how to force the linker to place functions
592
in a particular order is system dependent and out of the scope of this
593
manual.
594
 
595
@item -R @var{map_file}
596
@itemx --file-ordering @var{map_file}
597
The @samp{--file-ordering} option causes @code{gprof} to print a
598
suggested .o link line ordering for the program based on profiling data.
599
This option suggests an ordering which may improve paging, tlb and
600
cache behavior for the program on systems which do not support arbitrary
601
ordering of functions in an executable.
602
 
603
Use of the @samp{-a} argument is highly recommended with this option.
604
 
605
The @var{map_file} argument is a pathname to a file which provides
606
function name to object file mappings.  The format of the file is similar to
607
the output of the program @code{nm}.
608
 
609
@smallexample
610
@group
611
c-parse.o:00000000 T yyparse
612
c-parse.o:00000004 C yyerrflag
613
c-lang.o:00000000 T maybe_objc_method_name
614
c-lang.o:00000000 T print_lang_statistics
615
c-lang.o:00000000 T recognize_objc_keyword
616
c-decl.o:00000000 T print_lang_identifier
617
c-decl.o:00000000 T print_lang_type
618
@dots{}
619
 
620
@end group
621
@end smallexample
622
 
623
To create a @var{map_file} with @sc{gnu} @code{nm}, type a command like
624
@kbd{nm --extern-only --defined-only -v --print-file-name program-name}.
625
 
626
@item -T
627
@itemx --traditional
628
The @samp{-T} option causes @code{gprof} to print its output in
629
``traditional'' BSD style.
630
 
631
@item -w @var{width}
632
@itemx --width=@var{width}
633
Sets width of output lines to @var{width}.
634
Currently only used when printing the function index at the bottom
635
of the call graph.
636
 
637
@item -x
638
@itemx --all-lines
639
This option affects annotated source output only.
640
By default, only the lines at the beginning of a basic-block
641
are annotated.  If this option is specified, every line in
642
a basic-block is annotated by repeating the annotation for the
643
first line.  This behavior is similar to @code{tcov}'s @samp{-a}.
644
 
645
@item --demangle[=@var{style}]
646
@itemx --no-demangle
647
These options control whether C++ symbol names should be demangled when
648
printing output.  The default is to demangle symbols.  The
649
@code{--no-demangle} option may be used to turn off demangling. Different
650
compilers have different mangling styles.  The optional demangling style
651
argument can be used to choose an appropriate demangling style for your
652
compiler.
653
@end table
654
 
655
@node Analysis Options
656
@section Analysis Options
657
 
658
@table @code
659
 
660
@item -a
661
@itemx --no-static
662
The @samp{-a} option causes @code{gprof} to suppress the printing of
663
statically declared (private) functions.  (These are functions whose
664
names are not listed as global, and which are not visible outside the
665
file/function/block where they were defined.)  Time spent in these
666
functions, calls to/from them, etc., will all be attributed to the
667
function that was loaded directly before it in the executable file.
668
@c This is compatible with Unix @code{gprof}, but a bad idea.
669
This option affects both the flat profile and the call graph.
670
 
671
@item -c
672
@itemx --static-call-graph
673
The @samp{-c} option causes the call graph of the program to be
674
augmented by a heuristic which examines the text space of the object
675
file and identifies function calls in the binary machine code.
676
Since normal call graph records are only generated when functions are
677
entered, this option identifies children that could have been called,
678
but never were.  Calls to functions that were not compiled with
679
profiling enabled are also identified, but only if symbol table
680
entries are present for them.
681
Calls to dynamic library routines are typically @emph{not} found
682
by this option.
683
Parents or children identified via this heuristic
684
are indicated in the call graph with call counts of @samp{0}.
685
 
686
@item -D
687
@itemx --ignore-non-functions
688
The @samp{-D} option causes @code{gprof} to ignore symbols which
689
are not known to be functions.  This option will give more accurate
690
profile data on systems where it is supported (Solaris and HPUX for
691
example).
692
 
693
@item -k @var{from}/@var{to}
694
The @samp{-k} option allows you to delete from the call graph any arcs from
695
symbols matching symspec @var{from} to those matching symspec @var{to}.
696
 
697
@item -l
698
@itemx --line
699
The @samp{-l} option enables line-by-line profiling, which causes
700
histogram hits to be charged to individual source code lines,
701
instead of functions.  This feature only works with programs compiled
702
by older versions of the @code{gcc} compiler.  Newer versions of
703
@code{gcc} are designed to work with the @code{gcov} tool instead.
704
 
705
If the program was compiled with basic-block counting enabled,
706
this option will also identify how many times each line of
707
code was executed.
708
While line-by-line profiling can help isolate where in a large function
709
a program is spending its time, it also significantly increases
710
the running time of @code{gprof}, and magnifies statistical
711
inaccuracies.
712
@xref{Sampling Error, ,Statistical Sampling Error}.
713
 
714
@item -m @var{num}
715
@itemx --min-count=@var{num}
716
This option affects execution count output only.
717
Symbols that are executed less than @var{num} times are suppressed.
718
 
719
@item -n@var{symspec}
720
@itemx --time=@var{symspec}
721
The @samp{-n} option causes @code{gprof}, in its call graph analysis,
722
to only propagate times for symbols matching @var{symspec}.
723
 
724
@item -N@var{symspec}
725
@itemx --no-time=@var{symspec}
726
The @samp{-n} option causes @code{gprof}, in its call graph analysis,
727
not to propagate times for symbols matching @var{symspec}.
728
 
729
@item -S@var{filename}
730
@itemx --external-symbol-table=@var{filename}
731
The @samp{-S} option causes @code{gprof} to read an external symbol table
732
file, such as @file{/proc/kallsyms}, rather than read the symbol table
733
from the given object file (the default is @code{a.out}). This is useful
734
for profiling kernel modules.
735
 
736
@item -z
737
@itemx --display-unused-functions
738
If you give the @samp{-z} option, @code{gprof} will mention all
739
functions in the flat profile, even those that were never called, and
740
that had no time spent in them.  This is useful in conjunction with the
741
@samp{-c} option for discovering which routines were never called.
742
 
743
@end table
744
 
745
@node Miscellaneous Options
746
@section Miscellaneous Options
747
 
748
@table @code
749
 
750
@item -d[@var{num}]
751
@itemx --debug[=@var{num}]
752
The @samp{-d @var{num}} option specifies debugging options.
753
If @var{num} is not specified, enable all debugging.
754
@xref{Debugging, ,Debugging @code{gprof}}.
755
 
756
@item -h
757
@itemx --help
758
The @samp{-h} option prints command line usage.
759
 
760
@item -O@var{name}
761
@itemx --file-format=@var{name}
762
Selects the format of the profile data files.  Recognized formats are
763
@samp{auto} (the default), @samp{bsd}, @samp{4.4bsd}, @samp{magic}, and
764
@samp{prof} (not yet supported).
765
 
766
@item -s
767
@itemx --sum
768
The @samp{-s} option causes @code{gprof} to summarize the information
769
in the profile data files it read in, and write out a profile data
770
file called @file{gmon.sum}, which contains all the information from
771
the profile data files that @code{gprof} read in.  The file @file{gmon.sum}
772
may be one of the specified input files; the effect of this is to
773
merge the data in the other input files into @file{gmon.sum}.
774
 
775
Eventually you can run @code{gprof} again without @samp{-s} to analyze the
776
cumulative data in the file @file{gmon.sum}.
777
 
778
@item -v
779
@itemx --version
780
The @samp{-v} flag causes @code{gprof} to print the current version
781
number, and then exit.
782
 
783
@end table
784
 
785
@node Deprecated Options
786
@section Deprecated Options
787
 
788
@table @code
789
 
790
These options have been replaced with newer versions that use symspecs.
791
 
792
@item -e @var{function_name}
793
The @samp{-e @var{function}} option tells @code{gprof} to not print
794
information about the function @var{function_name} (and its
795
children@dots{}) in the call graph.  The function will still be listed
796
as a child of any functions that call it, but its index number will be
797
shown as @samp{[not printed]}.  More than one @samp{-e} option may be
798
given; only one @var{function_name} may be indicated with each @samp{-e}
799
option.
800
 
801
@item -E @var{function_name}
802
The @code{-E @var{function}} option works like the @code{-e} option, but
803
time spent in the function (and children who were not called from
804
anywhere else), will not be used to compute the percentages-of-time for
805
the call graph.  More than one @samp{-E} option may be given; only one
806
@var{function_name} may be indicated with each @samp{-E} option.
807
 
808
@item -f @var{function_name}
809
The @samp{-f @var{function}} option causes @code{gprof} to limit the
810
call graph to the function @var{function_name} and its children (and
811
their children@dots{}).  More than one @samp{-f} option may be given;
812
only one @var{function_name} may be indicated with each @samp{-f}
813
option.
814
 
815
@item -F @var{function_name}
816
The @samp{-F @var{function}} option works like the @code{-f} option, but
817
only time spent in the function and its children (and their
818
children@dots{}) will be used to determine total-time and
819
percentages-of-time for the call graph.  More than one @samp{-F} option
820
may be given; only one @var{function_name} may be indicated with each
821
@samp{-F} option.  The @samp{-F} option overrides the @samp{-E} option.
822
 
823
@end table
824
 
825
@c man end
826
 
827
Note that only one function can be specified with each @code{-e},
828
@code{-E}, @code{-f} or @code{-F} option.  To specify more than one
829
function, use multiple options.  For example, this command:
830
 
831
@example
832
gprof -e boring -f foo -f bar myprogram > gprof.output
833
@end example
834
 
835
@noindent
836
lists in the call graph all functions that were reached from either
837
@code{foo} or @code{bar} and were not reachable from @code{boring}.
838
 
839
@node Symspecs
840
@section Symspecs
841
 
842
Many of the output options allow functions to be included or excluded
843
using @dfn{symspecs} (symbol specifications), which observe the
844
following syntax:
845
 
846
@example
847
  filename_containing_a_dot
848
| funcname_not_containing_a_dot
849
| linenumber
850
| ( [ any_filename ] `:' ( any_funcname | linenumber ) )
851
@end example
852
 
853
Here are some sample symspecs:
854
 
855
@table @samp
856
@item main.c
857
Selects everything in file @file{main.c}---the
858
dot in the string tells @code{gprof} to interpret
859
the string as a filename, rather than as
860
a function name.  To select a file whose
861
name does not contain a dot, a trailing colon
862
should be specified.  For example, @samp{odd:} is
863
interpreted as the file named @file{odd}.
864
 
865
@item main
866
Selects all functions named @samp{main}.
867
 
868
Note that there may be multiple instances of the same function name
869
because some of the definitions may be local (i.e., static).  Unless a
870
function name is unique in a program, you must use the colon notation
871
explained below to specify a function from a specific source file.
872
 
873
Sometimes, function names contain dots.  In such cases, it is necessary
874
to add a leading colon to the name.  For example, @samp{:.mul} selects
875
function @samp{.mul}.
876
 
877
In some object file formats, symbols have a leading underscore.
878
@code{gprof} will normally not print these underscores.  When you name a
879
symbol in a symspec, you should type it exactly as @code{gprof} prints
880
it in its output.  For example, if the compiler produces a symbol
881
@samp{_main} from your @code{main} function, @code{gprof} still prints
882
it as @samp{main} in its output, so you should use @samp{main} in
883
symspecs.
884
 
885
@item main.c:main
886
Selects function @samp{main} in file @file{main.c}.
887
 
888
@item main.c:134
889
Selects line 134 in file @file{main.c}.
890
@end table
891
 
892
@node Output
893
@chapter Interpreting @code{gprof}'s Output
894
 
895
@code{gprof} can produce several different output styles, the
896
most important of which are described below.  The simplest output
897
styles (file information, execution count, and function and file ordering)
898
are not described here, but are documented with the respective options
899
that trigger them.
900
@xref{Output Options, ,Output Options}.
901
 
902
@menu
903
* Flat Profile::        The flat profile shows how much time was spent
904
                            executing directly in each function.
905
* Call Graph::          The call graph shows which functions called which
906
                            others, and how much time each function used
907
                            when its subroutine calls are included.
908
* Line-by-line::        @code{gprof} can analyze individual source code lines
909
* Annotated Source::    The annotated source listing displays source code
910
                            labeled with execution counts
911
@end menu
912
 
913
 
914
@node Flat Profile
915
@section The Flat Profile
916
@cindex flat profile
917
 
918
The @dfn{flat profile} shows the total amount of time your program
919
spent executing each function.  Unless the @samp{-z} option is given,
920
functions with no apparent time spent in them, and no apparent calls
921
to them, are not mentioned.  Note that if a function was not compiled
922
for profiling, and didn't run long enough to show up on the program
923
counter histogram, it will be indistinguishable from a function that
924
was never called.
925
 
926
This is part of a flat profile for a small program:
927
 
928
@smallexample
929
@group
930
Flat profile:
931
 
932
Each sample counts as 0.01 seconds.
933
  %   cumulative   self              self     total
934
 time   seconds   seconds    calls  ms/call  ms/call  name
935
 33.34      0.02     0.02     7208     0.00     0.00  open
936
 16.67      0.03     0.01      244     0.04     0.12  offtime
937
 16.67      0.04     0.01        8     1.25     1.25  memccpy
938
 16.67      0.05     0.01        7     1.43     1.43  write
939
 16.67      0.06     0.01                             mcount
940
  0.00      0.06     0.00      236     0.00     0.00  tzset
941
  0.00      0.06     0.00      192     0.00     0.00  tolower
942
  0.00      0.06     0.00       47     0.00     0.00  strlen
943
  0.00      0.06     0.00       45     0.00     0.00  strchr
944
  0.00      0.06     0.00        1     0.00    50.00  main
945
  0.00      0.06     0.00        1     0.00     0.00  memcpy
946
  0.00      0.06     0.00        1     0.00    10.11  print
947
  0.00      0.06     0.00        1     0.00     0.00  profil
948
  0.00      0.06     0.00        1     0.00    50.00  report
949
@dots{}
950
@end group
951
@end smallexample
952
 
953
@noindent
954
The functions are sorted first by decreasing run-time spent in them,
955
then by decreasing number of calls, then alphabetically by name.  The
956
functions @samp{mcount} and @samp{profil} are part of the profiling
957
apparatus and appear in every flat profile; their time gives a measure of
958
the amount of overhead due to profiling.
959
 
960
Just before the column headers, a statement appears indicating
961
how much time each sample counted as.
962
This @dfn{sampling period} estimates the margin of error in each of the time
963
figures.  A time figure that is not much larger than this is not
964
reliable.  In this example, each sample counted as 0.01 seconds,
965
suggesting a 100 Hz sampling rate.
966
The program's total execution time was 0.06
967
seconds, as indicated by the @samp{cumulative seconds} field.  Since
968
each sample counted for 0.01 seconds, this means only six samples
969
were taken during the run.  Two of the samples occurred while the
970
program was in the @samp{open} function, as indicated by the
971
@samp{self seconds} field.  Each of the other four samples
972
occurred one each in @samp{offtime}, @samp{memccpy}, @samp{write},
973
and @samp{mcount}.
974
Since only six samples were taken, none of these values can
975
be regarded as particularly reliable.
976
In another run,
977
the @samp{self seconds} field for
978
@samp{mcount} might well be @samp{0.00} or @samp{0.02}.
979
@xref{Sampling Error, ,Statistical Sampling Error},
980
for a complete discussion.
981
 
982
The remaining functions in the listing (those whose
983
@samp{self seconds} field is @samp{0.00}) didn't appear
984
in the histogram samples at all.  However, the call graph
985
indicated that they were called, so therefore they are listed,
986
sorted in decreasing order by the @samp{calls} field.
987
Clearly some time was spent executing these functions,
988
but the paucity of histogram samples prevents any
989
determination of how much time each took.
990
 
991
Here is what the fields in each line mean:
992
 
993
@table @code
994
@item % time
995
This is the percentage of the total execution time your program spent
996
in this function.  These should all add up to 100%.
997
 
998
@item cumulative seconds
999
This is the cumulative total number of seconds the computer spent
1000
executing this functions, plus the time spent in all the functions
1001
above this one in this table.
1002
 
1003
@item self seconds
1004
This is the number of seconds accounted for by this function alone.
1005
The flat profile listing is sorted first by this number.
1006
 
1007
@item calls
1008
This is the total number of times the function was called.  If the
1009
function was never called, or the number of times it was called cannot
1010
be determined (probably because the function was not compiled with
1011
profiling enabled), the @dfn{calls} field is blank.
1012
 
1013
@item self ms/call
1014
This represents the average number of milliseconds spent in this
1015
function per call, if this function is profiled.  Otherwise, this field
1016
is blank for this function.
1017
 
1018
@item total ms/call
1019
This represents the average number of milliseconds spent in this
1020
function and its descendants per call, if this function is profiled.
1021
Otherwise, this field is blank for this function.
1022
This is the only field in the flat profile that uses call graph analysis.
1023
 
1024
@item name
1025
This is the name of the function.   The flat profile is sorted by this
1026
field alphabetically after the @dfn{self seconds} and @dfn{calls}
1027
fields are sorted.
1028
@end table
1029
 
1030
@node Call Graph
1031
@section The Call Graph
1032
@cindex call graph
1033
 
1034
The @dfn{call graph} shows how much time was spent in each function
1035
and its children.  From this information, you can find functions that,
1036
while they themselves may not have used much time, called other
1037
functions that did use unusual amounts of time.
1038
 
1039
Here is a sample call from a small program.  This call came from the
1040
same @code{gprof} run as the flat profile example in the previous
1041
section.
1042
 
1043
@smallexample
1044
@group
1045
granularity: each sample hit covers 2 byte(s) for 20.00% of 0.05 seconds
1046
 
1047
index % time    self  children    called     name
1048
                                                 <spontaneous>
1049
[1]    100.0    0.00    0.05                 start [1]
1050
                0.00    0.05       1/1           main [2]
1051
                0.00    0.00       1/2           on_exit [28]
1052
                0.00    0.00       1/1           exit [59]
1053
-----------------------------------------------
1054
                0.00    0.05       1/1           start [1]
1055
[2]    100.0    0.00    0.05       1         main [2]
1056
                0.00    0.05       1/1           report [3]
1057
-----------------------------------------------
1058
                0.00    0.05       1/1           main [2]
1059
[3]    100.0    0.00    0.05       1         report [3]
1060
                0.00    0.03       8/8           timelocal [6]
1061
                0.00    0.01       1/1           print [9]
1062
                0.00    0.01       9/9           fgets [12]
1063
                0.00    0.00      12/34          strncmp <cycle 1> [40]
1064
                0.00    0.00       8/8           lookup [20]
1065
                0.00    0.00       1/1           fopen [21]
1066
                0.00    0.00       8/8           chewtime [24]
1067
                0.00    0.00       8/16          skipspace [44]
1068
-----------------------------------------------
1069
[4]     59.8    0.01        0.02       8+472     <cycle 2 as a whole> [4]
1070
                0.01        0.02     244+260         offtime <cycle 2> [7]
1071
                0.00        0.00     236+1           tzset <cycle 2> [26]
1072
-----------------------------------------------
1073
@end group
1074
@end smallexample
1075
 
1076
The lines full of dashes divide this table into @dfn{entries}, one for each
1077
function.  Each entry has one or more lines.
1078
 
1079
In each entry, the primary line is the one that starts with an index number
1080
in square brackets.  The end of this line says which function the entry is
1081
for.  The preceding lines in the entry describe the callers of this
1082
function and the following lines describe its subroutines (also called
1083
@dfn{children} when we speak of the call graph).
1084
 
1085
The entries are sorted by time spent in the function and its subroutines.
1086
 
1087
The internal profiling function @code{mcount} (@pxref{Flat Profile, ,The
1088
Flat Profile}) is never mentioned in the call graph.
1089
 
1090
@menu
1091
* Primary::       Details of the primary line's contents.
1092
* Callers::       Details of caller-lines' contents.
1093
* Subroutines::   Details of subroutine-lines' contents.
1094
* Cycles::        When there are cycles of recursion,
1095
                   such as @code{a} calls @code{b} calls @code{a}@dots{}
1096
@end menu
1097
 
1098
@node Primary
1099
@subsection The Primary Line
1100
 
1101
The @dfn{primary line} in a call graph entry is the line that
1102
describes the function which the entry is about and gives the overall
1103
statistics for this function.
1104
 
1105
For reference, we repeat the primary line from the entry for function
1106
@code{report} in our main example, together with the heading line that
1107
shows the names of the fields:
1108
 
1109
@smallexample
1110
@group
1111
index  % time    self  children called     name
1112
@dots{}
1113
[3]    100.0    0.00    0.05       1         report [3]
1114
@end group
1115
@end smallexample
1116
 
1117
Here is what the fields in the primary line mean:
1118
 
1119
@table @code
1120
@item index
1121
Entries are numbered with consecutive integers.  Each function
1122
therefore has an index number, which appears at the beginning of its
1123
primary line.
1124
 
1125
Each cross-reference to a function, as a caller or subroutine of
1126
another, gives its index number as well as its name.  The index number
1127
guides you if you wish to look for the entry for that function.
1128
 
1129
@item % time
1130
This is the percentage of the total time that was spent in this
1131
function, including time spent in subroutines called from this
1132
function.
1133
 
1134
The time spent in this function is counted again for the callers of
1135
this function.  Therefore, adding up these percentages is meaningless.
1136
 
1137
@item self
1138
This is the total amount of time spent in this function.  This
1139
should be identical to the number printed in the @code{seconds} field
1140
for this function in the flat profile.
1141
 
1142
@item children
1143
This is the total amount of time spent in the subroutine calls made by
1144
this function.  This should be equal to the sum of all the @code{self}
1145
and @code{children} entries of the children listed directly below this
1146
function.
1147
 
1148
@item called
1149
This is the number of times the function was called.
1150
 
1151
If the function called itself recursively, there are two numbers,
1152
separated by a @samp{+}.  The first number counts non-recursive calls,
1153
and the second counts recursive calls.
1154
 
1155
In the example above, the function @code{report} was called once from
1156
@code{main}.
1157
 
1158
@item name
1159
This is the name of the current function.  The index number is
1160
repeated after it.
1161
 
1162
If the function is part of a cycle of recursion, the cycle number is
1163
printed between the function's name and the index number
1164
(@pxref{Cycles, ,How Mutually Recursive Functions Are Described}).
1165
For example, if function @code{gnurr} is part of
1166
cycle number one, and has index number twelve, its primary line would
1167
be end like this:
1168
 
1169
@example
1170
gnurr <cycle 1> [12]
1171
@end example
1172
@end table
1173
 
1174
@node Callers
1175
@subsection Lines for a Function's Callers
1176
 
1177
A function's entry has a line for each function it was called by.
1178
These lines' fields correspond to the fields of the primary line, but
1179
their meanings are different because of the difference in context.
1180
 
1181
For reference, we repeat two lines from the entry for the function
1182
@code{report}, the primary line and one caller-line preceding it, together
1183
with the heading line that shows the names of the fields:
1184
 
1185
@smallexample
1186
index  % time    self  children called     name
1187
@dots{}
1188
                0.00    0.05       1/1           main [2]
1189
[3]    100.0    0.00    0.05       1         report [3]
1190
@end smallexample
1191
 
1192
Here are the meanings of the fields in the caller-line for @code{report}
1193
called from @code{main}:
1194
 
1195
@table @code
1196
@item self
1197
An estimate of the amount of time spent in @code{report} itself when it was
1198
called from @code{main}.
1199
 
1200
@item children
1201
An estimate of the amount of time spent in subroutines of @code{report}
1202
when @code{report} was called from @code{main}.
1203
 
1204
The sum of the @code{self} and @code{children} fields is an estimate
1205
of the amount of time spent within calls to @code{report} from @code{main}.
1206
 
1207
@item called
1208
Two numbers: the number of times @code{report} was called from @code{main},
1209
followed by the total number of non-recursive calls to @code{report} from
1210
all its callers.
1211
 
1212
@item name and index number
1213
The name of the caller of @code{report} to which this line applies,
1214
followed by the caller's index number.
1215
 
1216
Not all functions have entries in the call graph; some
1217
options to @code{gprof} request the omission of certain functions.
1218
When a caller has no entry of its own, it still has caller-lines
1219
in the entries of the functions it calls.
1220
 
1221
If the caller is part of a recursion cycle, the cycle number is
1222
printed between the name and the index number.
1223
@end table
1224
 
1225
If the identity of the callers of a function cannot be determined, a
1226
dummy caller-line is printed which has @samp{<spontaneous>} as the
1227
``caller's name'' and all other fields blank.  This can happen for
1228
signal handlers.
1229
@c What if some calls have determinable callers' names but not all?
1230
@c FIXME - still relevant?
1231
 
1232
@node Subroutines
1233
@subsection Lines for a Function's Subroutines
1234
 
1235
A function's entry has a line for each of its subroutines---in other
1236
words, a line for each other function that it called.  These lines'
1237
fields correspond to the fields of the primary line, but their meanings
1238
are different because of the difference in context.
1239
 
1240
For reference, we repeat two lines from the entry for the function
1241
@code{main}, the primary line and a line for a subroutine, together
1242
with the heading line that shows the names of the fields:
1243
 
1244
@smallexample
1245
index  % time    self  children called     name
1246
@dots{}
1247
[2]    100.0    0.00    0.05       1         main [2]
1248
                0.00    0.05       1/1           report [3]
1249
@end smallexample
1250
 
1251
Here are the meanings of the fields in the subroutine-line for @code{main}
1252
calling @code{report}:
1253
 
1254
@table @code
1255
@item self
1256
An estimate of the amount of time spent directly within @code{report}
1257
when @code{report} was called from @code{main}.
1258
 
1259
@item children
1260
An estimate of the amount of time spent in subroutines of @code{report}
1261
when @code{report} was called from @code{main}.
1262
 
1263
The sum of the @code{self} and @code{children} fields is an estimate
1264
of the total time spent in calls to @code{report} from @code{main}.
1265
 
1266
@item called
1267
Two numbers, the number of calls to @code{report} from @code{main}
1268
followed by the total number of non-recursive calls to @code{report}.
1269
This ratio is used to determine how much of @code{report}'s @code{self}
1270
and @code{children} time gets credited to @code{main}.
1271
@xref{Assumptions, ,Estimating @code{children} Times}.
1272
 
1273
@item name
1274
The name of the subroutine of @code{main} to which this line applies,
1275
followed by the subroutine's index number.
1276
 
1277
If the caller is part of a recursion cycle, the cycle number is
1278
printed between the name and the index number.
1279
@end table
1280
 
1281
@node Cycles
1282
@subsection How Mutually Recursive Functions Are Described
1283
@cindex cycle
1284
@cindex recursion cycle
1285
 
1286
The graph may be complicated by the presence of @dfn{cycles of
1287
recursion} in the call graph.  A cycle exists if a function calls
1288
another function that (directly or indirectly) calls (or appears to
1289
call) the original function.  For example: if @code{a} calls @code{b},
1290
and @code{b} calls @code{a}, then @code{a} and @code{b} form a cycle.
1291
 
1292
Whenever there are call paths both ways between a pair of functions, they
1293
belong to the same cycle.  If @code{a} and @code{b} call each other and
1294
@code{b} and @code{c} call each other, all three make one cycle.  Note that
1295
even if @code{b} only calls @code{a} if it was not called from @code{a},
1296
@code{gprof} cannot determine this, so @code{a} and @code{b} are still
1297
considered a cycle.
1298
 
1299
The cycles are numbered with consecutive integers.  When a function
1300
belongs to a cycle, each time the function name appears in the call graph
1301
it is followed by @samp{<cycle @var{number}>}.
1302
 
1303
The reason cycles matter is that they make the time values in the call
1304
graph paradoxical.  The ``time spent in children'' of @code{a} should
1305
include the time spent in its subroutine @code{b} and in @code{b}'s
1306
subroutines---but one of @code{b}'s subroutines is @code{a}!  How much of
1307
@code{a}'s time should be included in the children of @code{a}, when
1308
@code{a} is indirectly recursive?
1309
 
1310
The way @code{gprof} resolves this paradox is by creating a single entry
1311
for the cycle as a whole.  The primary line of this entry describes the
1312
total time spent directly in the functions of the cycle.  The
1313
``subroutines'' of the cycle are the individual functions of the cycle, and
1314
all other functions that were called directly by them.  The ``callers'' of
1315
the cycle are the functions, outside the cycle, that called functions in
1316
the cycle.
1317
 
1318
Here is an example portion of a call graph which shows a cycle containing
1319
functions @code{a} and @code{b}.  The cycle was entered by a call to
1320
@code{a} from @code{main}; both @code{a} and @code{b} called @code{c}.
1321
 
1322
@smallexample
1323
index  % time    self  children called     name
1324
----------------------------------------
1325
                 1.77        0    1/1        main [2]
1326
[3]     91.71    1.77        0    1+5    <cycle 1 as a whole> [3]
1327
                 1.02        0    3          b <cycle 1> [4]
1328
                 0.75        0    2          a <cycle 1> [5]
1329
----------------------------------------
1330
                                  3          a <cycle 1> [5]
1331
[4]     52.85    1.02        0    0      b <cycle 1> [4]
1332
                                  2          a <cycle 1> [5]
1333
 
1334
----------------------------------------
1335
                 1.77        0    1/1        main [2]
1336
                                  2          b <cycle 1> [4]
1337
[5]     38.86    0.75        0    1      a <cycle 1> [5]
1338
                                  3          b <cycle 1> [4]
1339
 
1340
----------------------------------------
1341
@end smallexample
1342
 
1343
@noindent
1344
(The entire call graph for this program contains in addition an entry for
1345
@code{main}, which calls @code{a}, and an entry for @code{c}, with callers
1346
@code{a} and @code{b}.)
1347
 
1348
@smallexample
1349
index  % time    self  children called     name
1350
                                             <spontaneous>
1351
[1]    100.00       0     1.93    0      start [1]
1352
                 0.16     1.77    1/1        main [2]
1353
----------------------------------------
1354
                 0.16     1.77    1/1        start [1]
1355
[2]    100.00    0.16     1.77    1      main [2]
1356
                 1.77        0    1/1        a <cycle 1> [5]
1357
----------------------------------------
1358
                 1.77        0    1/1        main [2]
1359
[3]     91.71    1.77        0    1+5    <cycle 1 as a whole> [3]
1360
                 1.02        0    3          b <cycle 1> [4]
1361
                 0.75        0    2          a <cycle 1> [5]
1362
 
1363
----------------------------------------
1364
                                  3          a <cycle 1> [5]
1365
[4]     52.85    1.02        0    0      b <cycle 1> [4]
1366
                                  2          a <cycle 1> [5]
1367
 
1368
----------------------------------------
1369
                 1.77        0    1/1        main [2]
1370
                                  2          b <cycle 1> [4]
1371
[5]     38.86    0.75        0    1      a <cycle 1> [5]
1372
                                  3          b <cycle 1> [4]
1373
 
1374
----------------------------------------
1375
 
1376
 
1377
[6]      0.00       0        0    6      c [6]
1378
----------------------------------------
1379
@end smallexample
1380
 
1381
The @code{self} field of the cycle's primary line is the total time
1382
spent in all the functions of the cycle.  It equals the sum of the
1383
@code{self} fields for the individual functions in the cycle, found
1384
in the entry in the subroutine lines for these functions.
1385
 
1386
The @code{children} fields of the cycle's primary line and subroutine lines
1387
count only subroutines outside the cycle.  Even though @code{a} calls
1388
@code{b}, the time spent in those calls to @code{b} is not counted in
1389
@code{a}'s @code{children} time.  Thus, we do not encounter the problem of
1390
what to do when the time in those calls to @code{b} includes indirect
1391
recursive calls back to @code{a}.
1392
 
1393
The @code{children} field of a caller-line in the cycle's entry estimates
1394
the amount of time spent @emph{in the whole cycle}, and its other
1395
subroutines, on the times when that caller called a function in the cycle.
1396
 
1397
The @code{called} field in the primary line for the cycle has two numbers:
1398
first, the number of times functions in the cycle were called by functions
1399
outside the cycle; second, the number of times they were called by
1400
functions in the cycle (including times when a function in the cycle calls
1401
itself).  This is a generalization of the usual split into non-recursive and
1402
recursive calls.
1403
 
1404
The @code{called} field of a subroutine-line for a cycle member in the
1405
cycle's entry says how many time that function was called from functions in
1406
the cycle.  The total of all these is the second number in the primary line's
1407
@code{called} field.
1408
 
1409
In the individual entry for a function in a cycle, the other functions in
1410
the same cycle can appear as subroutines and as callers.  These lines show
1411
how many times each function in the cycle called or was called from each other
1412
function in the cycle.  The @code{self} and @code{children} fields in these
1413
lines are blank because of the difficulty of defining meanings for them
1414
when recursion is going on.
1415
 
1416
@node Line-by-line
1417
@section Line-by-line Profiling
1418
 
1419
@code{gprof}'s @samp{-l} option causes the program to perform
1420
@dfn{line-by-line} profiling.  In this mode, histogram
1421
samples are assigned not to functions, but to individual
1422
lines of source code.  This only works with programs compiled with
1423
older versions of the @code{gcc} compiler.  Newer versions of @code{gcc}
1424
use a different program - @code{gcov} - to display line-by-line
1425
profiling information.
1426
 
1427
With the older versions of @code{gcc} the program usually has to be
1428
compiled with a @samp{-g} option, in addition to @samp{-pg}, in order
1429
to generate debugging symbols for tracking source code lines.
1430
Note, in much older versions of @code{gcc} the program had to be
1431
compiled with the @samp{-a} command line option as well.
1432
 
1433
The flat profile is the most useful output table
1434
in line-by-line mode.
1435
The call graph isn't as useful as normal, since
1436
the current version of @code{gprof} does not propagate
1437
call graph arcs from source code lines to the enclosing function.
1438
The call graph does, however, show each line of code
1439
that called each function, along with a count.
1440
 
1441
Here is a section of @code{gprof}'s output, without line-by-line profiling.
1442
Note that @code{ct_init} accounted for four histogram hits, and
1443
13327 calls to @code{init_block}.
1444
 
1445
@smallexample
1446
Flat profile:
1447
 
1448
Each sample counts as 0.01 seconds.
1449
  %   cumulative   self              self     total
1450
 time   seconds   seconds    calls  us/call  us/call  name
1451
 30.77      0.13     0.04     6335     6.31     6.31  ct_init
1452
 
1453
 
1454
                     Call graph (explanation follows)
1455
 
1456
 
1457
granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds
1458
 
1459
index % time    self  children    called     name
1460
 
1461
                0.00    0.00       1/13496       name_too_long
1462
                0.00    0.00      40/13496       deflate
1463
                0.00    0.00     128/13496       deflate_fast
1464
                0.00    0.00   13327/13496       ct_init
1465
[7]      0.0    0.00    0.00   13496         init_block
1466
 
1467
@end smallexample
1468
 
1469
Now let's look at some of @code{gprof}'s output from the same program run,
1470
this time with line-by-line profiling enabled.  Note that @code{ct_init}'s
1471
four histogram hits are broken down into four lines of source code---one hit
1472
occurred on each of lines 349, 351, 382 and 385.  In the call graph,
1473
note how
1474
@code{ct_init}'s 13327 calls to @code{init_block} are broken down
1475
into one call from line 396, 3071 calls from line 384, 3730 calls
1476
from line 385, and 6525 calls from 387.
1477
 
1478
@smallexample
1479
Flat profile:
1480
 
1481
Each sample counts as 0.01 seconds.
1482
  %   cumulative   self
1483
 time   seconds   seconds    calls  name
1484
  7.69      0.10     0.01           ct_init (trees.c:349)
1485
  7.69      0.11     0.01           ct_init (trees.c:351)
1486
  7.69      0.12     0.01           ct_init (trees.c:382)
1487
  7.69      0.13     0.01           ct_init (trees.c:385)
1488
 
1489
 
1490
                     Call graph (explanation follows)
1491
 
1492
 
1493
granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds
1494
 
1495
  % time    self  children    called     name
1496
 
1497
            0.00    0.00       1/13496       name_too_long (gzip.c:1440)
1498
            0.00    0.00       1/13496       deflate (deflate.c:763)
1499
            0.00    0.00       1/13496       ct_init (trees.c:396)
1500
            0.00    0.00       2/13496       deflate (deflate.c:727)
1501
            0.00    0.00       4/13496       deflate (deflate.c:686)
1502
            0.00    0.00       5/13496       deflate (deflate.c:675)
1503
            0.00    0.00      12/13496       deflate (deflate.c:679)
1504
            0.00    0.00      16/13496       deflate (deflate.c:730)
1505
            0.00    0.00     128/13496       deflate_fast (deflate.c:654)
1506
            0.00    0.00    3071/13496       ct_init (trees.c:384)
1507
            0.00    0.00    3730/13496       ct_init (trees.c:385)
1508
            0.00    0.00    6525/13496       ct_init (trees.c:387)
1509
[6]  0.0    0.00    0.00   13496         init_block (trees.c:408)
1510
 
1511
@end smallexample
1512
 
1513
 
1514
@node Annotated Source
1515
@section The Annotated Source Listing
1516
 
1517
@code{gprof}'s @samp{-A} option triggers an annotated source listing,
1518
which lists the program's source code, each function labeled with the
1519
number of times it was called.  You may also need to specify the
1520
@samp{-I} option, if @code{gprof} can't find the source code files.
1521
 
1522
With older versions of @code{gcc} compiling with @samp{gcc @dots{} -g
1523
-pg -a} augments your program with basic-block counting code, in
1524
addition to function counting code.  This enables @code{gprof} to
1525
determine how many times each line of code was executed.  With newer
1526
versions of @code{gcc} support for displaying basic-block counts is
1527
provided by the @code{gcov} program.
1528
 
1529
For example, consider the following function, taken from gzip,
1530
with line numbers added:
1531
 
1532
@smallexample
1533
 1 ulg updcrc(s, n)
1534
 2     uch *s;
1535
 3     unsigned n;
1536
 4 @{
1537
 5     register ulg c;
1538
 6
1539
 7     static ulg crc = (ulg)0xffffffffL;
1540
 8
1541
 9     if (s == NULL) @{
1542
10         c = 0xffffffffL;
1543
11     @} else @{
1544
12         c = crc;
1545
13         if (n) do @{
1546
14             c = crc_32_tab[...];
1547
15         @} while (--n);
1548
16     @}
1549
17     crc = c;
1550
18     return c ^ 0xffffffffL;
1551
19 @}
1552
 
1553
@end smallexample
1554
 
1555
@code{updcrc} has at least five basic-blocks.
1556
One is the function itself.  The
1557
@code{if} statement on line 9 generates two more basic-blocks, one
1558
for each branch of the @code{if}.  A fourth basic-block results from
1559
the @code{if} on line 13, and the contents of the @code{do} loop form
1560
the fifth basic-block.  The compiler may also generate additional
1561
basic-blocks to handle various special cases.
1562
 
1563
A program augmented for basic-block counting can be analyzed with
1564
@samp{gprof -l -A}.
1565
The @samp{-x} option is also helpful,
1566
to ensure that each line of code is labeled at least once.
1567
Here is @code{updcrc}'s
1568
annotated source listing for a sample @code{gzip} run:
1569
 
1570
@smallexample
1571
                ulg updcrc(s, n)
1572
                    uch *s;
1573
                    unsigned n;
1574
            2 ->@{
1575
                    register ulg c;
1576
 
1577
                    static ulg crc = (ulg)0xffffffffL;
1578
 
1579
            2 ->    if (s == NULL) @{
1580
            1 ->        c = 0xffffffffL;
1581
            1 ->    @} else @{
1582
            1 ->        c = crc;
1583
            1 ->        if (n) do @{
1584
        26312 ->            c = crc_32_tab[...];
1585
26312,1,26311 ->        @} while (--n);
1586
                    @}
1587
            2 ->    crc = c;
1588
            2 ->    return c ^ 0xffffffffL;
1589
            2 ->@}
1590
@end smallexample
1591
 
1592
In this example, the function was called twice, passing once through
1593
each branch of the @code{if} statement.  The body of the @code{do}
1594
loop was executed a total of 26312 times.  Note how the @code{while}
1595
statement is annotated.  It began execution 26312 times, once for
1596
each iteration through the loop.  One of those times (the last time)
1597
it exited, while it branched back to the beginning of the loop 26311 times.
1598
 
1599
@node Inaccuracy
1600
@chapter Inaccuracy of @code{gprof} Output
1601
 
1602
@menu
1603
* Sampling Error::      Statistical margins of error
1604
* Assumptions::         Estimating children times
1605
@end menu
1606
 
1607
@node Sampling Error
1608
@section Statistical Sampling Error
1609
 
1610
The run-time figures that @code{gprof} gives you are based on a sampling
1611
process, so they are subject to statistical inaccuracy.  If a function runs
1612
only a small amount of time, so that on the average the sampling process
1613
ought to catch that function in the act only once, there is a pretty good
1614
chance it will actually find that function zero times, or twice.
1615
 
1616
By contrast, the number-of-calls and basic-block figures
1617
are derived by counting, not
1618
sampling.  They are completely accurate and will not vary from run to run
1619
if your program is deterministic.
1620
 
1621
The @dfn{sampling period} that is printed at the beginning of the flat
1622
profile says how often samples are taken.  The rule of thumb is that a
1623
run-time figure is accurate if it is considerably bigger than the sampling
1624
period.
1625
 
1626
The actual amount of error can be predicted.
1627
For @var{n} samples, the @emph{expected} error
1628
is the square-root of @var{n}.  For example,
1629
if the sampling period is 0.01 seconds and @code{foo}'s run-time is 1 second,
1630
@var{n} is 100 samples (1 second/0.01 seconds), sqrt(@var{n}) is 10 samples, so
1631
the expected error in @code{foo}'s run-time is 0.1 seconds (10*0.01 seconds),
1632
or ten percent of the observed value.
1633
Again, if the sampling period is 0.01 seconds and @code{bar}'s run-time is
1634
100 seconds, @var{n} is 10000 samples, sqrt(@var{n}) is 100 samples, so
1635
the expected error in @code{bar}'s run-time is 1 second,
1636
or one percent of the observed value.
1637
It is likely to
1638
vary this much @emph{on the average} from one profiling run to the next.
1639
(@emph{Sometimes} it will vary more.)
1640
 
1641
This does not mean that a small run-time figure is devoid of information.
1642
If the program's @emph{total} run-time is large, a small run-time for one
1643
function does tell you that that function used an insignificant fraction of
1644
the whole program's time.  Usually this means it is not worth optimizing.
1645
 
1646
One way to get more accuracy is to give your program more (but similar)
1647
input data so it will take longer.  Another way is to combine the data from
1648
several runs, using the @samp{-s} option of @code{gprof}.  Here is how:
1649
 
1650
@enumerate
1651
@item
1652
Run your program once.
1653
 
1654
@item
1655
Issue the command @samp{mv gmon.out gmon.sum}.
1656
 
1657
@item
1658
Run your program again, the same as before.
1659
 
1660
@item
1661
Merge the new data in @file{gmon.out} into @file{gmon.sum} with this command:
1662
 
1663
@example
1664
gprof -s @var{executable-file} gmon.out gmon.sum
1665
@end example
1666
 
1667
@item
1668
Repeat the last two steps as often as you wish.
1669
 
1670
@item
1671
Analyze the cumulative data using this command:
1672
 
1673
@example
1674
gprof @var{executable-file} gmon.sum > @var{output-file}
1675
@end example
1676
@end enumerate
1677
 
1678
@node Assumptions
1679
@section Estimating @code{children} Times
1680
 
1681
Some of the figures in the call graph are estimates---for example, the
1682
@code{children} time values and all the time figures in caller and
1683
subroutine lines.
1684
 
1685
There is no direct information about these measurements in the profile
1686
data itself.  Instead, @code{gprof} estimates them by making an assumption
1687
about your program that might or might not be true.
1688
 
1689
The assumption made is that the average time spent in each call to any
1690
function @code{foo} is not correlated with who called @code{foo}.  If
1691
@code{foo} used 5 seconds in all, and 2/5 of the calls to @code{foo} came
1692
from @code{a}, then @code{foo} contributes 2 seconds to @code{a}'s
1693
@code{children} time, by assumption.
1694
 
1695
This assumption is usually true enough, but for some programs it is far
1696
from true.  Suppose that @code{foo} returns very quickly when its argument
1697
is zero; suppose that @code{a} always passes zero as an argument, while
1698
other callers of @code{foo} pass other arguments.  In this program, all the
1699
time spent in @code{foo} is in the calls from callers other than @code{a}.
1700
But @code{gprof} has no way of knowing this; it will blindly and
1701
incorrectly charge 2 seconds of time in @code{foo} to the children of
1702
@code{a}.
1703
 
1704
@c FIXME - has this been fixed?
1705
We hope some day to put more complete data into @file{gmon.out}, so that
1706
this assumption is no longer needed, if we can figure out how.  For the
1707
novice, the estimated figures are usually more useful than misleading.
1708
 
1709
@node How do I?
1710
@chapter Answers to Common Questions
1711
 
1712
@table @asis
1713
@item How can I get more exact information about hot spots in my program?
1714
 
1715
Looking at the per-line call counts only tells part of the story.
1716
Because @code{gprof} can only report call times and counts by function,
1717
the best way to get finer-grained information on where the program
1718
is spending its time is to re-factor large functions into sequences
1719
of calls to smaller ones.  Beware however that this can introduce
1720
artificial hot spots since compiling with @samp{-pg} adds a significant
1721
overhead to function calls.  An alternative solution is to use a
1722
non-intrusive profiler, e.g.@: oprofile.
1723
 
1724
@item How do I find which lines in my program were executed the most times?
1725
 
1726
Use the @code{gcov} program.
1727
 
1728
@item How do I find which lines in my program called a particular function?
1729
 
1730
Use @samp{gprof -l} and lookup the function in the call graph.
1731
The callers will be broken down by function and line number.
1732
 
1733
@item How do I analyze a program that runs for less than a second?
1734
 
1735
Try using a shell script like this one:
1736
 
1737
@example
1738
for i in `seq 1 100`; do
1739
  fastprog
1740
  mv gmon.out gmon.out.$i
1741
done
1742
 
1743
gprof -s fastprog gmon.out.*
1744
 
1745
gprof fastprog gmon.sum
1746
@end example
1747
 
1748
If your program is completely deterministic, all the call counts
1749
will be simple multiples of 100 (i.e., a function called once in
1750
each run will appear with a call count of 100).
1751
 
1752
@end table
1753
 
1754
@node Incompatibilities
1755
@chapter Incompatibilities with Unix @code{gprof}
1756
 
1757
@sc{gnu} @code{gprof} and Berkeley Unix @code{gprof} use the same data
1758
file @file{gmon.out}, and provide essentially the same information.  But
1759
there are a few differences.
1760
 
1761
@itemize @bullet
1762
@item
1763
@sc{gnu} @code{gprof} uses a new, generalized file format with support
1764
for basic-block execution counts and non-realtime histograms.  A magic
1765
cookie and version number allows @code{gprof} to easily identify
1766
new style files.  Old BSD-style files can still be read.
1767
@xref{File Format, ,Profiling Data File Format}.
1768
 
1769
@item
1770
For a recursive function, Unix @code{gprof} lists the function as a
1771
parent and as a child, with a @code{calls} field that lists the number
1772
of recursive calls.  @sc{gnu} @code{gprof} omits these lines and puts
1773
the number of recursive calls in the primary line.
1774
 
1775
@item
1776
When a function is suppressed from the call graph with @samp{-e}, @sc{gnu}
1777
@code{gprof} still lists it as a subroutine of functions that call it.
1778
 
1779
@item
1780
@sc{gnu} @code{gprof} accepts the @samp{-k} with its argument
1781
in the form @samp{from/to}, instead of @samp{from to}.
1782
 
1783
@item
1784
In the annotated source listing,
1785
if there are multiple basic blocks on the same line,
1786
@sc{gnu} @code{gprof} prints all of their counts, separated by commas.
1787
 
1788
@ignore - it does this now
1789
@item
1790
The function names printed in @sc{gnu} @code{gprof} output do not include
1791
the leading underscores that are added internally to the front of all
1792
C identifiers on many operating systems.
1793
@end ignore
1794
 
1795
@item
1796
The blurbs, field widths, and output formats are different.  @sc{gnu}
1797
@code{gprof} prints blurbs after the tables, so that you can see the
1798
tables without skipping the blurbs.
1799
@end itemize
1800
 
1801
@node Details
1802
@chapter Details of Profiling
1803
 
1804
@menu
1805
* Implementation::      How a program collects profiling information
1806
* File Format::         Format of @samp{gmon.out} files
1807
* Internals::           @code{gprof}'s internal operation
1808
* Debugging::           Using @code{gprof}'s @samp{-d} option
1809
@end menu
1810
 
1811
@node Implementation
1812
@section Implementation of Profiling
1813
 
1814
Profiling works by changing how every function in your program is compiled
1815
so that when it is called, it will stash away some information about where
1816
it was called from.  From this, the profiler can figure out what function
1817
called it, and can count how many times it was called.  This change is made
1818
by the compiler when your program is compiled with the @samp{-pg} option,
1819
which causes every function to call @code{mcount}
1820
(or @code{_mcount}, or @code{__mcount}, depending on the OS and compiler)
1821
as one of its first operations.
1822
 
1823
The @code{mcount} routine, included in the profiling library,
1824
is responsible for recording in an in-memory call graph table
1825
both its parent routine (the child) and its parent's parent.  This is
1826
typically done by examining the stack frame to find both
1827
the address of the child, and the return address in the original parent.
1828
Since this is a very machine-dependent operation, @code{mcount}
1829
itself is typically a short assembly-language stub routine
1830
that extracts the required
1831
information, and then calls @code{__mcount_internal}
1832
(a normal C function) with two arguments---@code{frompc} and @code{selfpc}.
1833
@code{__mcount_internal} is responsible for maintaining
1834
the in-memory call graph, which records @code{frompc}, @code{selfpc},
1835
and the number of times each of these call arcs was traversed.
1836
 
1837
GCC Version 2 provides a magical function (@code{__builtin_return_address}),
1838
which allows a generic @code{mcount} function to extract the
1839
required information from the stack frame.  However, on some
1840
architectures, most notably the SPARC, using this builtin can be
1841
very computationally expensive, and an assembly language version
1842
of @code{mcount} is used for performance reasons.
1843
 
1844
Number-of-calls information for library routines is collected by using a
1845
special version of the C library.  The programs in it are the same as in
1846
the usual C library, but they were compiled with @samp{-pg}.  If you
1847
link your program with @samp{gcc @dots{} -pg}, it automatically uses the
1848
profiling version of the library.
1849
 
1850
Profiling also involves watching your program as it runs, and keeping a
1851
histogram of where the program counter happens to be every now and then.
1852
Typically the program counter is looked at around 100 times per second of
1853
run time, but the exact frequency may vary from system to system.
1854
 
1855
This is done is one of two ways.  Most UNIX-like operating systems
1856
provide a @code{profil()} system call, which registers a memory
1857
array with the kernel, along with a scale
1858
factor that determines how the program's address space maps
1859
into the array.
1860
Typical scaling values cause every 2 to 8 bytes of address space
1861
to map into a single array slot.
1862
On every tick of the system clock
1863
(assuming the profiled program is running), the value of the
1864
program counter is examined and the corresponding slot in
1865
the memory array is incremented.  Since this is done in the kernel,
1866
which had to interrupt the process anyway to handle the clock
1867
interrupt, very little additional system overhead is required.
1868
 
1869
However, some operating systems, most notably Linux 2.0 (and earlier),
1870
do not provide a @code{profil()} system call.  On such a system,
1871
arrangements are made for the kernel to periodically deliver
1872
a signal to the process (typically via @code{setitimer()}),
1873
which then performs the same operation of examining the
1874
program counter and incrementing a slot in the memory array.
1875
Since this method requires a signal to be delivered to
1876
user space every time a sample is taken, it uses considerably
1877
more overhead than kernel-based profiling.  Also, due to the
1878
added delay required to deliver the signal, this method is
1879
less accurate as well.
1880
 
1881
A special startup routine allocates memory for the histogram and
1882
either calls @code{profil()} or sets up
1883
a clock signal handler.
1884
This routine (@code{monstartup}) can be invoked in several ways.
1885
On Linux systems, a special profiling startup file @code{gcrt0.o},
1886
which invokes @code{monstartup} before @code{main},
1887
is used instead of the default @code{crt0.o}.
1888
Use of this special startup file is one of the effects
1889
of using @samp{gcc @dots{} -pg} to link.
1890
On SPARC systems, no special startup files are used.
1891
Rather, the @code{mcount} routine, when it is invoked for
1892
the first time (typically when @code{main} is called),
1893
calls @code{monstartup}.
1894
 
1895
If the compiler's @samp{-a} option was used, basic-block counting
1896
is also enabled.  Each object file is then compiled with a static array
1897
of counts, initially zero.
1898
In the executable code, every time a new basic-block begins
1899
(i.e., when an @code{if} statement appears), an extra instruction
1900
is inserted to increment the corresponding count in the array.
1901
At compile time, a paired array was constructed that recorded
1902
the starting address of each basic-block.  Taken together,
1903
the two arrays record the starting address of every basic-block,
1904
along with the number of times it was executed.
1905
 
1906
The profiling library also includes a function (@code{mcleanup}) which is
1907
typically registered using @code{atexit()} to be called as the
1908
program exits, and is responsible for writing the file @file{gmon.out}.
1909
Profiling is turned off, various headers are output, and the histogram
1910
is written, followed by the call-graph arcs and the basic-block counts.
1911
 
1912
The output from @code{gprof} gives no indication of parts of your program that
1913
are limited by I/O or swapping bandwidth.  This is because samples of the
1914
program counter are taken at fixed intervals of the program's run time.
1915
Therefore, the
1916
time measurements in @code{gprof} output say nothing about time that your
1917
program was not running.  For example, a part of the program that creates
1918
so much data that it cannot all fit in physical memory at once may run very
1919
slowly due to thrashing, but @code{gprof} will say it uses little time.  On
1920
the other hand, sampling by run time has the advantage that the amount of
1921
load due to other users won't directly affect the output you get.
1922
 
1923
@node File Format
1924
@section Profiling Data File Format
1925
 
1926
The old BSD-derived file format used for profile data does not contain a
1927
magic cookie that allows to check whether a data file really is a
1928
@code{gprof} file.  Furthermore, it does not provide a version number, thus
1929
rendering changes to the file format almost impossible.  @sc{gnu} @code{gprof}
1930
uses a new file format that provides these features.  For backward
1931
compatibility, @sc{gnu} @code{gprof} continues to support the old BSD-derived
1932
format, but not all features are supported with it.  For example,
1933
basic-block execution counts cannot be accommodated by the old file
1934
format.
1935
 
1936
The new file format is defined in header file @file{gmon_out.h}.  It
1937
consists of a header containing the magic cookie and a version number,
1938
as well as some spare bytes available for future extensions.  All data
1939
in a profile data file is in the native format of the target for which
1940
the profile was collected.  @sc{gnu} @code{gprof} adapts automatically
1941
to the byte-order in use.
1942
 
1943
In the new file format, the header is followed by a sequence of
1944
records.  Currently, there are three different record types: histogram
1945
records, call-graph arc records, and basic-block execution count
1946
records.  Each file can contain any number of each record type.  When
1947
reading a file, @sc{gnu} @code{gprof} will ensure records of the same type are
1948
compatible with each other and compute the union of all records.  For
1949
example, for basic-block execution counts, the union is simply the sum
1950
of all execution counts for each basic-block.
1951
 
1952
@subsection Histogram Records
1953
 
1954
Histogram records consist of a header that is followed by an array of
1955
bins.  The header contains the text-segment range that the histogram
1956
spans, the size of the histogram in bytes (unlike in the old BSD
1957
format, this does not include the size of the header), the rate of the
1958
profiling clock, and the physical dimension that the bin counts
1959
represent after being scaled by the profiling clock rate.  The
1960
physical dimension is specified in two parts: a long name of up to 15
1961
characters and a single character abbreviation.  For example, a
1962
histogram representing real-time would specify the long name as
1963
``seconds'' and the abbreviation as ``s''.  This feature is useful for
1964
architectures that support performance monitor hardware (which,
1965
fortunately, is becoming increasingly common).  For example, under DEC
1966
OSF/1, the ``uprofile'' command can be used to produce a histogram of,
1967
say, instruction cache misses.  In this case, the dimension in the
1968
histogram header could be set to ``i-cache misses'' and the abbreviation
1969
could be set to ``1'' (because it is simply a count, not a physical
1970
dimension).  Also, the profiling rate would have to be set to 1 in
1971
this case.
1972
 
1973
Histogram bins are 16-bit numbers and each bin represent an equal
1974
amount of text-space.  For example, if the text-segment is one
1975
thousand bytes long and if there are ten bins in the histogram, each
1976
bin represents one hundred bytes.
1977
 
1978
 
1979
@subsection Call-Graph Records
1980
 
1981
Call-graph records have a format that is identical to the one used in
1982
the BSD-derived file format.  It consists of an arc in the call graph
1983
and a count indicating the number of times the arc was traversed
1984
during program execution.  Arcs are specified by a pair of addresses:
1985
the first must be within caller's function and the second must be
1986
within the callee's function.  When performing profiling at the
1987
function level, these addresses can point anywhere within the
1988
respective function.  However, when profiling at the line-level, it is
1989
better if the addresses are as close to the call-site/entry-point as
1990
possible.  This will ensure that the line-level call-graph is able to
1991
identify exactly which line of source code performed calls to a
1992
function.
1993
 
1994
@subsection Basic-Block Execution Count Records
1995
 
1996
Basic-block execution count records consist of a header followed by a
1997
sequence of address/count pairs.  The header simply specifies the
1998
length of the sequence.  In an address/count pair, the address
1999
identifies a basic-block and the count specifies the number of times
2000
that basic-block was executed.  Any address within the basic-address can
2001
be used.
2002
 
2003
@node Internals
2004
@section @code{gprof}'s Internal Operation
2005
 
2006
Like most programs, @code{gprof} begins by processing its options.
2007
During this stage, it may building its symspec list
2008
(@code{sym_ids.c:@-sym_id_add}), if
2009
options are specified which use symspecs.
2010
@code{gprof} maintains a single linked list of symspecs,
2011
which will eventually get turned into 12 symbol tables,
2012
organized into six include/exclude pairs---one
2013
pair each for the flat profile (INCL_FLAT/EXCL_FLAT),
2014
the call graph arcs (INCL_ARCS/EXCL_ARCS),
2015
printing in the call graph (INCL_GRAPH/EXCL_GRAPH),
2016
timing propagation in the call graph (INCL_TIME/EXCL_TIME),
2017
the annotated source listing (INCL_ANNO/EXCL_ANNO),
2018
and the execution count listing (INCL_EXEC/EXCL_EXEC).
2019
 
2020
After option processing, @code{gprof} finishes
2021
building the symspec list by adding all the symspecs in
2022
@code{default_excluded_list} to the exclude lists
2023
EXCL_TIME and EXCL_GRAPH, and if line-by-line profiling is specified,
2024
EXCL_FLAT as well.
2025
These default excludes are not added to EXCL_ANNO, EXCL_ARCS, and EXCL_EXEC.
2026
 
2027
Next, the BFD library is called to open the object file,
2028
verify that it is an object file,
2029
and read its symbol table (@code{core.c:@-core_init}),
2030
using @code{bfd_canonicalize_symtab} after mallocing
2031
an appropriately sized array of symbols.  At this point,
2032
function mappings are read (if the @samp{--file-ordering} option
2033
has been specified), and the core text space is read into
2034
memory (if the @samp{-c} option was given).
2035
 
2036
@code{gprof}'s own symbol table, an array of Sym structures,
2037
is now built.
2038
This is done in one of two ways, by one of two routines, depending
2039
on whether line-by-line profiling (@samp{-l} option) has been
2040
enabled.
2041
For normal profiling, the BFD canonical symbol table is scanned.
2042
For line-by-line profiling, every
2043
text space address is examined, and a new symbol table entry
2044
gets created every time the line number changes.
2045
In either case, two passes are made through the symbol
2046
table---one to count the size of the symbol table required,
2047
and the other to actually read the symbols.  In between the
2048
two passes, a single array of type @code{Sym} is created of
2049
the appropriate length.
2050
Finally, @code{symtab.c:@-symtab_finalize}
2051
is called to sort the symbol table and remove duplicate entries
2052
(entries with the same memory address).
2053
 
2054
The symbol table must be a contiguous array for two reasons.
2055
First, the @code{qsort} library function (which sorts an array)
2056
will be used to sort the symbol table.
2057
Also, the symbol lookup routine (@code{symtab.c:@-sym_lookup}),
2058
which finds symbols
2059
based on memory address, uses a binary search algorithm
2060
which requires the symbol table to be a sorted array.
2061
Function symbols are indicated with an @code{is_func} flag.
2062
Line number symbols have no special flags set.
2063
Additionally, a symbol can have an @code{is_static} flag
2064
to indicate that it is a local symbol.
2065
 
2066
With the symbol table read, the symspecs can now be translated
2067
into Syms (@code{sym_ids.c:@-sym_id_parse}).  Remember that a single
2068
symspec can match multiple symbols.
2069
An array of symbol tables
2070
(@code{syms}) is created, each entry of which is a symbol table
2071
of Syms to be included or excluded from a particular listing.
2072
The master symbol table and the symspecs are examined by nested
2073
loops, and every symbol that matches a symspec is inserted
2074
into the appropriate syms table.  This is done twice, once to
2075
count the size of each required symbol table, and again to build
2076
the tables, which have been malloced between passes.
2077
From now on, to determine whether a symbol is on an include
2078
or exclude symspec list, @code{gprof} simply uses its
2079
standard symbol lookup routine on the appropriate table
2080
in the @code{syms} array.
2081
 
2082
Now the profile data file(s) themselves are read
2083
(@code{gmon_io.c:@-gmon_out_read}),
2084
first by checking for a new-style @samp{gmon.out} header,
2085
then assuming this is an old-style BSD @samp{gmon.out}
2086
if the magic number test failed.
2087
 
2088
New-style histogram records are read by @code{hist.c:@-hist_read_rec}.
2089
For the first histogram record, allocate a memory array to hold
2090
all the bins, and read them in.
2091
When multiple profile data files (or files with multiple histogram
2092
records) are read, the memory ranges of each pair of histogram records
2093
must be either equal, or non-overlapping.  For each pair of histogram
2094
records, the resolution (memory region size divided by the number of
2095
bins) must be the same.  The time unit must be the same for all
2096
histogram records. If the above containts are met, all histograms
2097
for the same memory range are merged.
2098
 
2099
As each call graph record is read (@code{call_graph.c:@-cg_read_rec}),
2100
the parent and child addresses
2101
are matched to symbol table entries, and a call graph arc is
2102
created by @code{cg_arcs.c:@-arc_add}, unless the arc fails a symspec
2103
check against INCL_ARCS/EXCL_ARCS.  As each arc is added,
2104
a linked list is maintained of the parent's child arcs, and of the child's
2105
parent arcs.
2106
Both the child's call count and the arc's call count are
2107
incremented by the record's call count.
2108
 
2109
Basic-block records are read (@code{basic_blocks.c:@-bb_read_rec}),
2110
but only if line-by-line profiling has been selected.
2111
Each basic-block address is matched to a corresponding line
2112
symbol in the symbol table, and an entry made in the symbol's
2113
bb_addr and bb_calls arrays.  Again, if multiple basic-block
2114
records are present for the same address, the call counts
2115
are cumulative.
2116
 
2117
A gmon.sum file is dumped, if requested (@code{gmon_io.c:@-gmon_out_write}).
2118
 
2119
If histograms were present in the data files, assign them to symbols
2120
(@code{hist.c:@-hist_assign_samples}) by iterating over all the sample
2121
bins and assigning them to symbols.  Since the symbol table
2122
is sorted in order of ascending memory addresses, we can
2123
simple follow along in the symbol table as we make our pass
2124
over the sample bins.
2125
This step includes a symspec check against INCL_FLAT/EXCL_FLAT.
2126
Depending on the histogram
2127
scale factor, a sample bin may span multiple symbols,
2128
in which case a fraction of the sample count is allocated
2129
to each symbol, proportional to the degree of overlap.
2130
This effect is rare for normal profiling, but overlaps
2131
are more common during line-by-line profiling, and can
2132
cause each of two adjacent lines to be credited with half
2133
a hit, for example.
2134
 
2135
If call graph data is present, @code{cg_arcs.c:@-cg_assemble} is called.
2136
First, if @samp{-c} was specified, a machine-dependent
2137
routine (@code{find_call}) scans through each symbol's machine code,
2138
looking for subroutine call instructions, and adding them
2139
to the call graph with a zero call count.
2140
A topological sort is performed by depth-first numbering
2141
all the symbols (@code{cg_dfn.c:@-cg_dfn}), so that
2142
children are always numbered less than their parents,
2143
then making a array of pointers into the symbol table and sorting it into
2144
numerical order, which is reverse topological
2145
order (children appear before parents).
2146
Cycles are also detected at this point, all members
2147
of which are assigned the same topological number.
2148
Two passes are now made through this sorted array of symbol pointers.
2149
The first pass, from end to beginning (parents to children),
2150
computes the fraction of child time to propagate to each parent
2151
and a print flag.
2152
The print flag reflects symspec handling of INCL_GRAPH/EXCL_GRAPH,
2153
with a parent's include or exclude (print or no print) property
2154
being propagated to its children, unless they themselves explicitly appear
2155
in INCL_GRAPH or EXCL_GRAPH.
2156
A second pass, from beginning to end (children to parents) actually
2157
propagates the timings along the call graph, subject
2158
to a check against INCL_TIME/EXCL_TIME.
2159
With the print flag, fractions, and timings now stored in the symbol
2160
structures, the topological sort array is now discarded, and a
2161
new array of pointers is assembled, this time sorted by propagated time.
2162
 
2163
Finally, print the various outputs the user requested, which is now fairly
2164
straightforward.  The call graph (@code{cg_print.c:@-cg_print}) and
2165
flat profile (@code{hist.c:@-hist_print}) are regurgitations of values
2166
already computed.  The annotated source listing
2167
(@code{basic_blocks.c:@-print_annotated_source}) uses basic-block
2168
information, if present, to label each line of code with call counts,
2169
otherwise only the function call counts are presented.
2170
 
2171
The function ordering code is marginally well documented
2172
in the source code itself (@code{cg_print.c}).  Basically,
2173
the functions with the most use and the most parents are
2174
placed first, followed by other functions with the most use,
2175
followed by lower use functions, followed by unused functions
2176
at the end.
2177
 
2178
@node Debugging
2179
@section Debugging @code{gprof}
2180
 
2181
If @code{gprof} was compiled with debugging enabled,
2182
the @samp{-d} option triggers debugging output
2183
(to stdout) which can be helpful in understanding its operation.
2184
The debugging number specified is interpreted as a sum of the following
2185
options:
2186
 
2187
@table @asis
2188
@item 2 - Topological sort
2189
Monitor depth-first numbering of symbols during call graph analysis
2190
@item 4 - Cycles
2191
Shows symbols as they are identified as cycle heads
2192
@item 16 - Tallying
2193
As the call graph arcs are read, show each arc and how
2194
the total calls to each function are tallied
2195
@item 32 - Call graph arc sorting
2196
Details sorting individual parents/children within each call graph entry
2197
@item 64 - Reading histogram and call graph records
2198
Shows address ranges of histograms as they are read, and each
2199
call graph arc
2200
@item 128 - Symbol table
2201
Reading, classifying, and sorting the symbol table from the object file.
2202
For line-by-line profiling (@samp{-l} option), also shows line numbers
2203
being assigned to memory addresses.
2204
@item 256 - Static call graph
2205
Trace operation of @samp{-c} option
2206
@item 512 - Symbol table and arc table lookups
2207
Detail operation of lookup routines
2208
@item 1024 - Call graph propagation
2209
Shows how function times are propagated along the call graph
2210
@item 2048 - Basic-blocks
2211
Shows basic-block records as they are read from profile data
2212
(only meaningful with @samp{-l} option)
2213
@item 4096 - Symspecs
2214
Shows symspec-to-symbol pattern matching operation
2215
@item 8192 - Annotate source
2216
Tracks operation of @samp{-A} option
2217
@end table
2218
 
2219
@node GNU Free Documentation License
2220
@appendix GNU Free Documentation License
2221
@include fdl.texi
2222
 
2223
@bye
2224
 
2225
NEEDS AN INDEX
2226
 
2227
-T - "traditional BSD style": How is it different?  Should the
2228
differences be documented?
2229
 
2230
example flat file adds up to 100.01%...
2231
 
2232
note: time estimates now only go out to one decimal place (0.0), where
2233
they used to extend two (78.67).

powered by: WebSVN 2.1.0

© copyright 1999-2024 OpenCores.org, equivalent to Oliscience, all rights reserved. OpenCores®, registered trademark.