1 This is gprof.info, produced by makeinfo version 4.8 from gprof.texi.
4 * gprof: (gprof). Profiling your program's execution
7 This file documents the gprof profiler of the GNU system.
9 Copyright (C) 1988, 92, 97, 98, 99, 2000, 2001, 2003, 2007 Free
10 Software Foundation, Inc.
12 Permission is granted to copy, distribute and/or modify this document
13 under the terms of the GNU Free Documentation License, Version 1.1 or
14 any later version published by the Free Software Foundation; with no
15 Invariant Sections, with no Front-Cover Texts, and with no Back-Cover
16 Texts. A copy of the license is included in the section entitled "GNU
17 Free Documentation License".
20 File: gprof.info, Node: Top, Next: Introduction, Up: (dir)
22 Profiling a Program: Where Does It Spend Its Time?
23 **************************************************
25 This manual describes the GNU profiler, `gprof', and how you can use it
26 to determine which parts of a program are taking most of the execution
27 time. We assume that you know how to write, compile, and execute
28 programs. GNU `gprof' was written by Jay Fenlason.
30 This manual is for `gprof' (GNU Binutils) version 2.18.90.
32 This document is distributed under the terms of the GNU Free
33 Documentation License. A copy of the license is included in the
34 section entitled "GNU Free Documentation License".
38 * Introduction:: What profiling means, and why it is useful.
40 * Compiling:: How to compile your program for profiling.
41 * Executing:: Executing your program to generate profile data
42 * Invoking:: How to run `gprof', and its options
44 * Output:: Interpreting `gprof''s output
46 * Inaccuracy:: Potential problems you should be aware of
47 * How do I?:: Answers to common questions
48 * Incompatibilities:: (between GNU `gprof' and Unix `gprof'.)
49 * Details:: Details of how profiling is done
50 * GNU Free Documentation License:: GNU Free Documentation License
53 File: gprof.info, Node: Introduction, Next: Compiling, Prev: Top, Up: Top
55 1 Introduction to Profiling
56 ***************************
58 Profiling allows you to learn where your program spent its time and
59 which functions called which other functions while it was executing.
60 This information can show you which pieces of your program are slower
61 than you expected, and might be candidates for rewriting to make your
62 program execute faster. It can also tell you which functions are being
63 called more or less often than you expected. This may help you spot
64 bugs that had otherwise been unnoticed.
66 Since the profiler uses information collected during the actual
67 execution of your program, it can be used on programs that are too
68 large or too complex to analyze by reading the source. However, how
69 your program is run will affect the information that shows up in the
70 profile data. If you don't use some feature of your program while it
71 is being profiled, no profile information will be generated for that
74 Profiling has several steps:
76 * You must compile and link your program with profiling enabled.
77 *Note Compiling a Program for Profiling: Compiling.
79 * You must execute your program to generate a profile data file.
80 *Note Executing the Program: Executing.
82 * You must run `gprof' to analyze the profile data. *Note `gprof'
83 Command Summary: Invoking.
85 The next three chapters explain these steps in greater detail.
87 Several forms of output are available from the analysis.
89 The "flat profile" shows how much time your program spent in each
90 function, and how many times that function was called. If you simply
91 want to know which functions burn most of the cycles, it is stated
92 concisely here. *Note The Flat Profile: Flat Profile.
94 The "call graph" shows, for each function, which functions called
95 it, which other functions it called, and how many times. There is also
96 an estimate of how much time was spent in the subroutines of each
97 function. This can suggest places where you might try to eliminate
98 function calls that use a lot of time. *Note The Call Graph: Call
101 The "annotated source" listing is a copy of the program's source
102 code, labeled with the number of times each line of the program was
103 executed. *Note The Annotated Source Listing: Annotated Source.
105 To better understand how profiling works, you may wish to read a
106 description of its implementation. *Note Implementation of Profiling:
110 File: gprof.info, Node: Compiling, Next: Executing, Prev: Introduction, Up: Top
112 2 Compiling a Program for Profiling
113 ***********************************
115 The first step in generating profile information for your program is to
116 compile and link it with profiling enabled.
118 To compile a source file for profiling, specify the `-pg' option when
119 you run the compiler. (This is in addition to the options you normally
122 To link the program for profiling, if you use a compiler such as `cc'
123 to do the linking, simply specify `-pg' in addition to your usual
124 options. The same option, `-pg', alters either compilation or linking
125 to do what is necessary for profiling. Here are examples:
127 cc -g -c myprog.c utils.c -pg
128 cc -o myprog myprog.o utils.o -pg
130 The `-pg' option also works with a command that both compiles and
133 cc -o myprog myprog.c utils.c -g -pg
135 Note: The `-pg' option must be part of your compilation options as
136 well as your link options. If it is not then no call-graph data will
137 be gathered and when you run `gprof' you will get an error message like
140 gprof: gmon.out file is missing call-graph data
142 If you add the `-Q' switch to suppress the printing of the call
143 graph data you will still be able to see the time samples:
147 Each sample counts as 0.01 seconds.
148 % cumulative self self total
149 time seconds seconds calls Ts/call Ts/call name
150 44.12 0.07 0.07 zazLoop
152 20.59 0.17 0.04 bazMillion
154 If you run the linker `ld' directly instead of through a compiler
155 such as `cc', you may have to specify a profiling startup file
156 `gcrt0.o' as the first input file instead of the usual startup file
157 `crt0.o'. In addition, you would probably want to specify the
158 profiling C library, `libc_p.a', by writing `-lc_p' instead of the
159 usual `-lc'. This is not absolutely necessary, but doing this gives
160 you number-of-calls information for standard library functions such as
161 `read' and `open'. For example:
163 ld -o myprog /lib/gcrt0.o myprog.o utils.o -lc_p
165 If you compile only some of the modules of the program with `-pg',
166 you can still profile the program, but you won't get complete
167 information about the modules that were compiled without `-pg'. The
168 only information you get for the functions in those modules is the
169 total time spent in them; there is no record of how many times they
170 were called, or from where. This will not affect the flat profile
171 (except that the `calls' field for the functions will be blank), but
172 will greatly reduce the usefulness of the call graph.
174 If you wish to perform line-by-line profiling you should use the
175 `gcov' tool instead of `gprof'. See that tool's manual or info pages
176 for more details of how to do this.
178 Note, older versions of `gcc' produce line-by-line profiling
179 information that works with `gprof' rather than `gcov' so there is
180 still support for displaying this kind of information in `gprof'. *Note
181 Line-by-line Profiling: Line-by-line.
183 It also worth noting that `gcc' implements a
184 `-finstrument-functions' command line option which will insert calls to
185 special user supplied instrumentation routines at the entry and exit of
186 every function in their program. This can be used to implement an
187 alternative profiling scheme.
190 File: gprof.info, Node: Executing, Next: Invoking, Prev: Compiling, Up: Top
192 3 Executing the Program
193 ***********************
195 Once the program is compiled for profiling, you must run it in order to
196 generate the information that `gprof' needs. Simply run the program as
197 usual, using the normal arguments, file names, etc. The program should
198 run normally, producing the same output as usual. It will, however, run
199 somewhat slower than normal because of the time spent collecting and
200 writing the profile data.
202 The way you run the program--the arguments and input that you give
203 it--may have a dramatic effect on what the profile information shows.
204 The profile data will describe the parts of the program that were
205 activated for the particular input you use. For example, if the first
206 command you give to your program is to quit, the profile data will show
207 the time used in initialization and in cleanup, but not much else.
209 Your program will write the profile data into a file called
210 `gmon.out' just before exiting. If there is already a file called
211 `gmon.out', its contents are overwritten. There is currently no way to
212 tell the program to write the profile data under a different name, but
213 you can rename the file afterwards if you are concerned that it may be
216 In order to write the `gmon.out' file properly, your program must
217 exit normally: by returning from `main' or by calling `exit'. Calling
218 the low-level function `_exit' does not write the profile data, and
219 neither does abnormal termination due to an unhandled signal.
221 The `gmon.out' file is written in the program's _current working
222 directory_ at the time it exits. This means that if your program calls
223 `chdir', the `gmon.out' file will be left in the last directory your
224 program `chdir''d to. If you don't have permission to write in this
225 directory, the file is not written, and you will get an error message.
227 Older versions of the GNU profiling library may also write a file
228 called `bb.out'. This file, if present, contains an human-readable
229 listing of the basic-block execution counts. Unfortunately, the
230 appearance of a human-readable `bb.out' means the basic-block counts
231 didn't get written into `gmon.out'. The Perl script `bbconv.pl',
232 included with the `gprof' source distribution, will convert a `bb.out'
233 file into a format readable by `gprof'. Invoke it like this:
235 bbconv.pl < bb.out > BH-DATA
237 This translates the information in `bb.out' into a form that `gprof'
238 can understand. But you still need to tell `gprof' about the existence
239 of this translated information. To do that, include BB-DATA on the
240 `gprof' command line, _along with `gmon.out'_, like this:
242 gprof OPTIONS EXECUTABLE-FILE gmon.out BB-DATA [YET-MORE-PROFILE-DATA-FILES...] [> OUTFILE]
245 File: gprof.info, Node: Invoking, Next: Output, Prev: Executing, Up: Top
247 4 `gprof' Command Summary
248 *************************
250 After you have a profile data file `gmon.out', you can run `gprof' to
251 interpret the information in it. The `gprof' program prints a flat
252 profile and a call graph on standard output. Typically you would
253 redirect the output of `gprof' into a file with `>'.
255 You run `gprof' like this:
257 gprof OPTIONS [EXECUTABLE-FILE [PROFILE-DATA-FILES...]] [> OUTFILE]
259 Here square-brackets indicate optional arguments.
261 If you omit the executable file name, the file `a.out' is used. If
262 you give no profile data file name, the file `gmon.out' is used. If
263 any file is not in the proper format, or if the profile data file does
264 not appear to belong to the executable file, an error message is
267 You can give more than one profile data file by entering all their
268 names after the executable file name; then the statistics in all the
269 data files are summed together.
271 The order of these options does not matter.
275 * Output Options:: Controlling `gprof''s output style
276 * Analysis Options:: Controlling how `gprof' analyzes its data
277 * Miscellaneous Options::
278 * Deprecated Options:: Options you no longer need to use, but which
279 have been retained for compatibility
280 * Symspecs:: Specifying functions to include or exclude
283 File: gprof.info, Node: Output Options, Next: Analysis Options, Up: Invoking
288 These options specify which of several output formats `gprof' should
291 Many of these options take an optional "symspec" to specify
292 functions to be included or excluded. These options can be specified
293 multiple times, with different symspecs, to include or exclude sets of
294 symbols. *Note Symspecs: Symspecs.
296 Specifying any of these options overrides the default (`-p -q'),
297 which prints a flat profile and call graph analysis for all functions.
300 `--annotated-source[=SYMSPEC]'
301 The `-A' option causes `gprof' to print annotated source code. If
302 SYMSPEC is specified, print output only for matching symbols.
303 *Note The Annotated Source Listing: Annotated Source.
307 If the `-b' option is given, `gprof' doesn't print the verbose
308 blurbs that try to explain the meaning of all of the fields in the
309 tables. This is useful if you intend to print out the output, or
310 are tired of seeing the blurbs.
313 `--exec-counts[=SYMSPEC]'
314 The `-C' option causes `gprof' to print a tally of functions and
315 the number of times each was called. If SYMSPEC is specified,
316 print tally only for matching symbols.
318 If the profile data file contains basic-block count records,
319 specifying the `-l' option, along with `-C', will cause basic-block
320 execution counts to be tallied and displayed.
324 The `-i' option causes `gprof' to display summary information
325 about the profile data file(s) and then exit. The number of
326 histogram, call graph, and basic-block count records is displayed.
329 `--directory-path=DIRS'
330 The `-I' option specifies a list of search directories in which to
331 find source files. Environment variable GPROF_PATH can also be
332 used to convey this information. Used mostly for annotated source
336 `--no-annotated-source[=SYMSPEC]'
337 The `-J' option causes `gprof' not to print annotated source code.
338 If SYMSPEC is specified, `gprof' prints annotated source, but
339 excludes matching symbols.
343 Normally, source filenames are printed with the path component
344 suppressed. The `-L' option causes `gprof' to print the full
345 pathname of source filenames, which is determined from symbolic
346 debugging information in the image file and is relative to the
347 directory in which the compiler was invoked.
350 `--flat-profile[=SYMSPEC]'
351 The `-p' option causes `gprof' to print a flat profile. If
352 SYMSPEC is specified, print flat profile only for matching symbols.
353 *Note The Flat Profile: Flat Profile.
356 `--no-flat-profile[=SYMSPEC]'
357 The `-P' option causes `gprof' to suppress printing a flat profile.
358 If SYMSPEC is specified, `gprof' prints a flat profile, but
359 excludes matching symbols.
363 The `-q' option causes `gprof' to print the call graph analysis.
364 If SYMSPEC is specified, print call graph only for matching symbols
365 and their children. *Note The Call Graph: Call Graph.
368 `--no-graph[=SYMSPEC]'
369 The `-Q' option causes `gprof' to suppress printing the call graph.
370 If SYMSPEC is specified, `gprof' prints a call graph, but excludes
375 The `-t' option causes the NUM most active source lines in each
376 source file to be listed when source annotation is enabled. The
381 This option affects annotated source output only. Normally,
382 `gprof' prints annotated source files to standard-output. If this
383 option is specified, annotated source for a file named
384 `path/FILENAME' is generated in the file `FILENAME-ann'. If the
385 underlying file system would truncate `FILENAME-ann' so that it
386 overwrites the original `FILENAME', `gprof' generates annotated
387 source in the file `FILENAME.ann' instead (if the original file
388 name has an extension, that extension is _replaced_ with `.ann').
391 `--no-exec-counts[=SYMSPEC]'
392 The `-Z' option causes `gprof' not to print a tally of functions
393 and the number of times each was called. If SYMSPEC is specified,
394 print tally, but exclude matching symbols.
397 `--function-ordering'
398 The `--function-ordering' option causes `gprof' to print a
399 suggested function ordering for the program based on profiling
400 data. This option suggests an ordering which may improve paging,
401 tlb and cache behavior for the program on systems which support
402 arbitrary ordering of functions in an executable.
404 The exact details of how to force the linker to place functions in
405 a particular order is system dependent and out of the scope of this
409 `--file-ordering MAP_FILE'
410 The `--file-ordering' option causes `gprof' to print a suggested
411 .o link line ordering for the program based on profiling data.
412 This option suggests an ordering which may improve paging, tlb and
413 cache behavior for the program on systems which do not support
414 arbitrary ordering of functions in an executable.
416 Use of the `-a' argument is highly recommended with this option.
418 The MAP_FILE argument is a pathname to a file which provides
419 function name to object file mappings. The format of the file is
420 similar to the output of the program `nm'.
422 c-parse.o:00000000 T yyparse
423 c-parse.o:00000004 C yyerrflag
424 c-lang.o:00000000 T maybe_objc_method_name
425 c-lang.o:00000000 T print_lang_statistics
426 c-lang.o:00000000 T recognize_objc_keyword
427 c-decl.o:00000000 T print_lang_identifier
428 c-decl.o:00000000 T print_lang_type
431 To create a MAP_FILE with GNU `nm', type a command like `nm
432 --extern-only --defined-only -v --print-file-name program-name'.
436 The `-T' option causes `gprof' to print its output in
437 "traditional" BSD style.
441 Sets width of output lines to WIDTH. Currently only used when
442 printing the function index at the bottom of the call graph.
446 This option affects annotated source output only. By default,
447 only the lines at the beginning of a basic-block are annotated.
448 If this option is specified, every line in a basic-block is
449 annotated by repeating the annotation for the first line. This
450 behavior is similar to `tcov''s `-a'.
454 These options control whether C++ symbol names should be demangled
455 when printing output. The default is to demangle symbols. The
456 `--no-demangle' option may be used to turn off demangling.
457 Different compilers have different mangling styles. The optional
458 demangling style argument can be used to choose an appropriate
459 demangling style for your compiler.
462 File: gprof.info, Node: Analysis Options, Next: Miscellaneous Options, Prev: Output Options, Up: Invoking
469 The `-a' option causes `gprof' to suppress the printing of
470 statically declared (private) functions. (These are functions
471 whose names are not listed as global, and which are not visible
472 outside the file/function/block where they were defined.) Time
473 spent in these functions, calls to/from them, etc., will all be
474 attributed to the function that was loaded directly before it in
475 the executable file. This option affects both the flat profile
479 `--static-call-graph'
480 The `-c' option causes the call graph of the program to be
481 augmented by a heuristic which examines the text space of the
482 object file and identifies function calls in the binary machine
483 code. Since normal call graph records are only generated when
484 functions are entered, this option identifies children that could
485 have been called, but never were. Calls to functions that were
486 not compiled with profiling enabled are also identified, but only
487 if symbol table entries are present for them. Calls to dynamic
488 library routines are typically _not_ found by this option.
489 Parents or children identified via this heuristic are indicated in
490 the call graph with call counts of `0'.
493 `--ignore-non-functions'
494 The `-D' option causes `gprof' to ignore symbols which are not
495 known to be functions. This option will give more accurate
496 profile data on systems where it is supported (Solaris and HPUX for
500 The `-k' option allows you to delete from the call graph any arcs
501 from symbols matching symspec FROM to those matching symspec TO.
505 The `-l' option enables line-by-line profiling, which causes
506 histogram hits to be charged to individual source code lines,
507 instead of functions. This feature only works with programs
508 compiled by older versions of the `gcc' compiler. Newer versions
509 of `gcc' are designed to work with the `gcov' tool instead.
511 If the program was compiled with basic-block counting enabled,
512 this option will also identify how many times each line of code
513 was executed. While line-by-line profiling can help isolate where
514 in a large function a program is spending its time, it also
515 significantly increases the running time of `gprof', and magnifies
516 statistical inaccuracies. *Note Statistical Sampling Error:
521 This option affects execution count output only. Symbols that are
522 executed less than NUM times are suppressed.
526 The `-n' option causes `gprof', in its call graph analysis, to
527 only propagate times for symbols matching SYMSPEC.
531 The `-n' option causes `gprof', in its call graph analysis, not to
532 propagate times for symbols matching SYMSPEC.
535 `--display-unused-functions'
536 If you give the `-z' option, `gprof' will mention all functions in
537 the flat profile, even those that were never called, and that had
538 no time spent in them. This is useful in conjunction with the
539 `-c' option for discovering which routines were never called.
543 File: gprof.info, Node: Miscellaneous Options, Next: Deprecated Options, Prev: Analysis Options, Up: Invoking
545 4.3 Miscellaneous Options
546 =========================
550 The `-d NUM' option specifies debugging options. If NUM is not
551 specified, enable all debugging. *Note Debugging `gprof':
556 The `-h' option prints command line usage.
560 Selects the format of the profile data files. Recognized formats
561 are `auto' (the default), `bsd', `4.4bsd', `magic', and `prof'
566 The `-s' option causes `gprof' to summarize the information in the
567 profile data files it read in, and write out a profile data file
568 called `gmon.sum', which contains all the information from the
569 profile data files that `gprof' read in. The file `gmon.sum' may
570 be one of the specified input files; the effect of this is to
571 merge the data in the other input files into `gmon.sum'.
573 Eventually you can run `gprof' again without `-s' to analyze the
574 cumulative data in the file `gmon.sum'.
578 The `-v' flag causes `gprof' to print the current version number,
583 File: gprof.info, Node: Deprecated Options, Next: Symspecs, Prev: Miscellaneous Options, Up: Invoking
585 4.4 Deprecated Options
586 ======================
588 These options have been replaced with newer versions that use
592 The `-e FUNCTION' option tells `gprof' to not print information
593 about the function FUNCTION_NAME (and its children...) in the call
594 graph. The function will still be listed as a child of any
595 functions that call it, but its index number will be shown as
596 `[not printed]'. More than one `-e' option may be given; only one
597 FUNCTION_NAME may be indicated with each `-e' option.
600 The `-E FUNCTION' option works like the `-e' option, but time
601 spent in the function (and children who were not called from
602 anywhere else), will not be used to compute the
603 percentages-of-time for the call graph. More than one `-E' option
604 may be given; only one FUNCTION_NAME may be indicated with each
608 The `-f FUNCTION' option causes `gprof' to limit the call graph to
609 the function FUNCTION_NAME and its children (and their
610 children...). More than one `-f' option may be given; only one
611 FUNCTION_NAME may be indicated with each `-f' option.
614 The `-F FUNCTION' option works like the `-f' option, but only time
615 spent in the function and its children (and their children...)
616 will be used to determine total-time and percentages-of-time for
617 the call graph. More than one `-F' option may be given; only one
618 FUNCTION_NAME may be indicated with each `-F' option. The `-F'
619 option overrides the `-E' option.
622 Note that only one function can be specified with each `-e', `-E',
623 `-f' or `-F' option. To specify more than one function, use multiple
624 options. For example, this command:
626 gprof -e boring -f foo -f bar myprogram > gprof.output
628 lists in the call graph all functions that were reached from either
629 `foo' or `bar' and were not reachable from `boring'.
632 File: gprof.info, Node: Symspecs, Prev: Deprecated Options, Up: Invoking
637 Many of the output options allow functions to be included or excluded
638 using "symspecs" (symbol specifications), which observe the following
641 filename_containing_a_dot
642 | funcname_not_containing_a_dot
644 | ( [ any_filename ] `:' ( any_funcname | linenumber ) )
646 Here are some sample symspecs:
649 Selects everything in file `main.c'--the dot in the string tells
650 `gprof' to interpret the string as a filename, rather than as a
651 function name. To select a file whose name does not contain a
652 dot, a trailing colon should be specified. For example, `odd:' is
653 interpreted as the file named `odd'.
656 Selects all functions named `main'.
658 Note that there may be multiple instances of the same function name
659 because some of the definitions may be local (i.e., static).
660 Unless a function name is unique in a program, you must use the
661 colon notation explained below to specify a function from a
662 specific source file.
664 Sometimes, function names contain dots. In such cases, it is
665 necessary to add a leading colon to the name. For example,
666 `:.mul' selects function `.mul'.
668 In some object file formats, symbols have a leading underscore.
669 `gprof' will normally not print these underscores. When you name a
670 symbol in a symspec, you should type it exactly as `gprof' prints
671 it in its output. For example, if the compiler produces a symbol
672 `_main' from your `main' function, `gprof' still prints it as
673 `main' in its output, so you should use `main' in symspecs.
676 Selects function `main' in file `main.c'.
679 Selects line 134 in file `main.c'.
682 File: gprof.info, Node: Output, Next: Inaccuracy, Prev: Invoking, Up: Top
684 5 Interpreting `gprof''s Output
685 *******************************
687 `gprof' can produce several different output styles, the most important
688 of which are described below. The simplest output styles (file
689 information, execution count, and function and file ordering) are not
690 described here, but are documented with the respective options that
691 trigger them. *Note Output Options: Output Options.
695 * Flat Profile:: The flat profile shows how much time was spent
696 executing directly in each function.
697 * Call Graph:: The call graph shows which functions called which
698 others, and how much time each function used
699 when its subroutine calls are included.
700 * Line-by-line:: `gprof' can analyze individual source code lines
701 * Annotated Source:: The annotated source listing displays source code
702 labeled with execution counts
705 File: gprof.info, Node: Flat Profile, Next: Call Graph, Up: Output
710 The "flat profile" shows the total amount of time your program spent
711 executing each function. Unless the `-z' option is given, functions
712 with no apparent time spent in them, and no apparent calls to them, are
713 not mentioned. Note that if a function was not compiled for profiling,
714 and didn't run long enough to show up on the program counter histogram,
715 it will be indistinguishable from a function that was never called.
717 This is part of a flat profile for a small program:
721 Each sample counts as 0.01 seconds.
722 % cumulative self self total
723 time seconds seconds calls ms/call ms/call name
724 33.34 0.02 0.02 7208 0.00 0.00 open
725 16.67 0.03 0.01 244 0.04 0.12 offtime
726 16.67 0.04 0.01 8 1.25 1.25 memccpy
727 16.67 0.05 0.01 7 1.43 1.43 write
728 16.67 0.06 0.01 mcount
729 0.00 0.06 0.00 236 0.00 0.00 tzset
730 0.00 0.06 0.00 192 0.00 0.00 tolower
731 0.00 0.06 0.00 47 0.00 0.00 strlen
732 0.00 0.06 0.00 45 0.00 0.00 strchr
733 0.00 0.06 0.00 1 0.00 50.00 main
734 0.00 0.06 0.00 1 0.00 0.00 memcpy
735 0.00 0.06 0.00 1 0.00 10.11 print
736 0.00 0.06 0.00 1 0.00 0.00 profil
737 0.00 0.06 0.00 1 0.00 50.00 report
740 The functions are sorted first by decreasing run-time spent in them,
741 then by decreasing number of calls, then alphabetically by name. The
742 functions `mcount' and `profil' are part of the profiling apparatus and
743 appear in every flat profile; their time gives a measure of the amount
744 of overhead due to profiling.
746 Just before the column headers, a statement appears indicating how
747 much time each sample counted as. This "sampling period" estimates the
748 margin of error in each of the time figures. A time figure that is not
749 much larger than this is not reliable. In this example, each sample
750 counted as 0.01 seconds, suggesting a 100 Hz sampling rate. The
751 program's total execution time was 0.06 seconds, as indicated by the
752 `cumulative seconds' field. Since each sample counted for 0.01
753 seconds, this means only six samples were taken during the run. Two of
754 the samples occurred while the program was in the `open' function, as
755 indicated by the `self seconds' field. Each of the other four samples
756 occurred one each in `offtime', `memccpy', `write', and `mcount'.
757 Since only six samples were taken, none of these values can be regarded
758 as particularly reliable. In another run, the `self seconds' field for
759 `mcount' might well be `0.00' or `0.02'. *Note Statistical Sampling
760 Error: Sampling Error, for a complete discussion.
762 The remaining functions in the listing (those whose `self seconds'
763 field is `0.00') didn't appear in the histogram samples at all.
764 However, the call graph indicated that they were called, so therefore
765 they are listed, sorted in decreasing order by the `calls' field.
766 Clearly some time was spent executing these functions, but the paucity
767 of histogram samples prevents any determination of how much time each
770 Here is what the fields in each line mean:
773 This is the percentage of the total execution time your program
774 spent in this function. These should all add up to 100%.
777 This is the cumulative total number of seconds the computer spent
778 executing this functions, plus the time spent in all the functions
779 above this one in this table.
782 This is the number of seconds accounted for by this function alone.
783 The flat profile listing is sorted first by this number.
786 This is the total number of times the function was called. If the
787 function was never called, or the number of times it was called
788 cannot be determined (probably because the function was not
789 compiled with profiling enabled), the "calls" field is blank.
792 This represents the average number of milliseconds spent in this
793 function per call, if this function is profiled. Otherwise, this
794 field is blank for this function.
797 This represents the average number of milliseconds spent in this
798 function and its descendants per call, if this function is
799 profiled. Otherwise, this field is blank for this function. This
800 is the only field in the flat profile that uses call graph
804 This is the name of the function. The flat profile is sorted by
805 this field alphabetically after the "self seconds" and "calls"
809 File: gprof.info, Node: Call Graph, Next: Line-by-line, Prev: Flat Profile, Up: Output
814 The "call graph" shows how much time was spent in each function and its
815 children. From this information, you can find functions that, while
816 they themselves may not have used much time, called other functions
817 that did use unusual amounts of time.
819 Here is a sample call from a small program. This call came from the
820 same `gprof' run as the flat profile example in the previous section.
822 granularity: each sample hit covers 2 byte(s) for 20.00% of 0.05 seconds
824 index % time self children called name
826 [1] 100.0 0.00 0.05 start [1]
827 0.00 0.05 1/1 main [2]
828 0.00 0.00 1/2 on_exit [28]
829 0.00 0.00 1/1 exit [59]
830 -----------------------------------------------
831 0.00 0.05 1/1 start [1]
832 [2] 100.0 0.00 0.05 1 main [2]
833 0.00 0.05 1/1 report [3]
834 -----------------------------------------------
835 0.00 0.05 1/1 main [2]
836 [3] 100.0 0.00 0.05 1 report [3]
837 0.00 0.03 8/8 timelocal [6]
838 0.00 0.01 1/1 print [9]
839 0.00 0.01 9/9 fgets [12]
840 0.00 0.00 12/34 strncmp <cycle 1> [40]
841 0.00 0.00 8/8 lookup [20]
842 0.00 0.00 1/1 fopen [21]
843 0.00 0.00 8/8 chewtime [24]
844 0.00 0.00 8/16 skipspace [44]
845 -----------------------------------------------
846 [4] 59.8 0.01 0.02 8+472 <cycle 2 as a whole> [4]
847 0.01 0.02 244+260 offtime <cycle 2> [7]
848 0.00 0.00 236+1 tzset <cycle 2> [26]
849 -----------------------------------------------
851 The lines full of dashes divide this table into "entries", one for
852 each function. Each entry has one or more lines.
854 In each entry, the primary line is the one that starts with an index
855 number in square brackets. The end of this line says which function
856 the entry is for. The preceding lines in the entry describe the
857 callers of this function and the following lines describe its
858 subroutines (also called "children" when we speak of the call graph).
860 The entries are sorted by time spent in the function and its
863 The internal profiling function `mcount' (*note The Flat Profile:
864 Flat Profile.) is never mentioned in the call graph.
868 * Primary:: Details of the primary line's contents.
869 * Callers:: Details of caller-lines' contents.
870 * Subroutines:: Details of subroutine-lines' contents.
871 * Cycles:: When there are cycles of recursion,
872 such as `a' calls `b' calls `a'...
875 File: gprof.info, Node: Primary, Next: Callers, Up: Call Graph
877 5.2.1 The Primary Line
878 ----------------------
880 The "primary line" in a call graph entry is the line that describes the
881 function which the entry is about and gives the overall statistics for
884 For reference, we repeat the primary line from the entry for function
885 `report' in our main example, together with the heading line that shows
886 the names of the fields:
888 index % time self children called name
890 [3] 100.0 0.00 0.05 1 report [3]
892 Here is what the fields in the primary line mean:
895 Entries are numbered with consecutive integers. Each function
896 therefore has an index number, which appears at the beginning of
899 Each cross-reference to a function, as a caller or subroutine of
900 another, gives its index number as well as its name. The index
901 number guides you if you wish to look for the entry for that
905 This is the percentage of the total time that was spent in this
906 function, including time spent in subroutines called from this
909 The time spent in this function is counted again for the callers of
910 this function. Therefore, adding up these percentages is
914 This is the total amount of time spent in this function. This
915 should be identical to the number printed in the `seconds' field
916 for this function in the flat profile.
919 This is the total amount of time spent in the subroutine calls
920 made by this function. This should be equal to the sum of all the
921 `self' and `children' entries of the children listed directly
925 This is the number of times the function was called.
927 If the function called itself recursively, there are two numbers,
928 separated by a `+'. The first number counts non-recursive calls,
929 and the second counts recursive calls.
931 In the example above, the function `report' was called once from
935 This is the name of the current function. The index number is
938 If the function is part of a cycle of recursion, the cycle number
939 is printed between the function's name and the index number (*note
940 How Mutually Recursive Functions Are Described: Cycles.). For
941 example, if function `gnurr' is part of cycle number one, and has
942 index number twelve, its primary line would be end like this:
947 File: gprof.info, Node: Callers, Next: Subroutines, Prev: Primary, Up: Call Graph
949 5.2.2 Lines for a Function's Callers
950 ------------------------------------
952 A function's entry has a line for each function it was called by.
953 These lines' fields correspond to the fields of the primary line, but
954 their meanings are different because of the difference in context.
956 For reference, we repeat two lines from the entry for the function
957 `report', the primary line and one caller-line preceding it, together
958 with the heading line that shows the names of the fields:
960 index % time self children called name
962 0.00 0.05 1/1 main [2]
963 [3] 100.0 0.00 0.05 1 report [3]
965 Here are the meanings of the fields in the caller-line for `report'
969 An estimate of the amount of time spent in `report' itself when it
970 was called from `main'.
973 An estimate of the amount of time spent in subroutines of `report'
974 when `report' was called from `main'.
976 The sum of the `self' and `children' fields is an estimate of the
977 amount of time spent within calls to `report' from `main'.
980 Two numbers: the number of times `report' was called from `main',
981 followed by the total number of non-recursive calls to `report'
982 from all its callers.
984 `name and index number'
985 The name of the caller of `report' to which this line applies,
986 followed by the caller's index number.
988 Not all functions have entries in the call graph; some options to
989 `gprof' request the omission of certain functions. When a caller
990 has no entry of its own, it still has caller-lines in the entries
991 of the functions it calls.
993 If the caller is part of a recursion cycle, the cycle number is
994 printed between the name and the index number.
996 If the identity of the callers of a function cannot be determined, a
997 dummy caller-line is printed which has `<spontaneous>' as the "caller's
998 name" and all other fields blank. This can happen for signal handlers.
1001 File: gprof.info, Node: Subroutines, Next: Cycles, Prev: Callers, Up: Call Graph
1003 5.2.3 Lines for a Function's Subroutines
1004 ----------------------------------------
1006 A function's entry has a line for each of its subroutines--in other
1007 words, a line for each other function that it called. These lines'
1008 fields correspond to the fields of the primary line, but their meanings
1009 are different because of the difference in context.
1011 For reference, we repeat two lines from the entry for the function
1012 `main', the primary line and a line for a subroutine, together with the
1013 heading line that shows the names of the fields:
1015 index % time self children called name
1017 [2] 100.0 0.00 0.05 1 main [2]
1018 0.00 0.05 1/1 report [3]
1020 Here are the meanings of the fields in the subroutine-line for `main'
1024 An estimate of the amount of time spent directly within `report'
1025 when `report' was called from `main'.
1028 An estimate of the amount of time spent in subroutines of `report'
1029 when `report' was called from `main'.
1031 The sum of the `self' and `children' fields is an estimate of the
1032 total time spent in calls to `report' from `main'.
1035 Two numbers, the number of calls to `report' from `main' followed
1036 by the total number of non-recursive calls to `report'. This
1037 ratio is used to determine how much of `report''s `self' and
1038 `children' time gets credited to `main'. *Note Estimating
1039 `children' Times: Assumptions.
1042 The name of the subroutine of `main' to which this line applies,
1043 followed by the subroutine's index number.
1045 If the caller is part of a recursion cycle, the cycle number is
1046 printed between the name and the index number.
1049 File: gprof.info, Node: Cycles, Prev: Subroutines, Up: Call Graph
1051 5.2.4 How Mutually Recursive Functions Are Described
1052 ----------------------------------------------------
1054 The graph may be complicated by the presence of "cycles of recursion"
1055 in the call graph. A cycle exists if a function calls another function
1056 that (directly or indirectly) calls (or appears to call) the original
1057 function. For example: if `a' calls `b', and `b' calls `a', then `a'
1058 and `b' form a cycle.
1060 Whenever there are call paths both ways between a pair of functions,
1061 they belong to the same cycle. If `a' and `b' call each other and `b'
1062 and `c' call each other, all three make one cycle. Note that even if
1063 `b' only calls `a' if it was not called from `a', `gprof' cannot
1064 determine this, so `a' and `b' are still considered a cycle.
1066 The cycles are numbered with consecutive integers. When a function
1067 belongs to a cycle, each time the function name appears in the call
1068 graph it is followed by `<cycle NUMBER>'.
1070 The reason cycles matter is that they make the time values in the
1071 call graph paradoxical. The "time spent in children" of `a' should
1072 include the time spent in its subroutine `b' and in `b''s
1073 subroutines--but one of `b''s subroutines is `a'! How much of `a''s
1074 time should be included in the children of `a', when `a' is indirectly
1077 The way `gprof' resolves this paradox is by creating a single entry
1078 for the cycle as a whole. The primary line of this entry describes the
1079 total time spent directly in the functions of the cycle. The
1080 "subroutines" of the cycle are the individual functions of the cycle,
1081 and all other functions that were called directly by them. The
1082 "callers" of the cycle are the functions, outside the cycle, that
1083 called functions in the cycle.
1085 Here is an example portion of a call graph which shows a cycle
1086 containing functions `a' and `b'. The cycle was entered by a call to
1087 `a' from `main'; both `a' and `b' called `c'.
1089 index % time self children called name
1090 ----------------------------------------
1092 [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3]
1093 1.02 0 3 b <cycle 1> [4]
1094 0.75 0 2 a <cycle 1> [5]
1095 ----------------------------------------
1097 [4] 52.85 1.02 0 0 b <cycle 1> [4]
1100 ----------------------------------------
1103 [5] 38.86 0.75 0 1 a <cycle 1> [5]
1106 ----------------------------------------
1108 (The entire call graph for this program contains in addition an entry
1109 for `main', which calls `a', and an entry for `c', with callers `a' and
1112 index % time self children called name
1114 [1] 100.00 0 1.93 0 start [1]
1115 0.16 1.77 1/1 main [2]
1116 ----------------------------------------
1117 0.16 1.77 1/1 start [1]
1118 [2] 100.00 0.16 1.77 1 main [2]
1119 1.77 0 1/1 a <cycle 1> [5]
1120 ----------------------------------------
1122 [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3]
1123 1.02 0 3 b <cycle 1> [4]
1124 0.75 0 2 a <cycle 1> [5]
1126 ----------------------------------------
1128 [4] 52.85 1.02 0 0 b <cycle 1> [4]
1131 ----------------------------------------
1134 [5] 38.86 0.75 0 1 a <cycle 1> [5]
1137 ----------------------------------------
1138 0 0 3/6 b <cycle 1> [4]
1139 0 0 3/6 a <cycle 1> [5]
1140 [6] 0.00 0 0 6 c [6]
1141 ----------------------------------------
1143 The `self' field of the cycle's primary line is the total time spent
1144 in all the functions of the cycle. It equals the sum of the `self'
1145 fields for the individual functions in the cycle, found in the entry in
1146 the subroutine lines for these functions.
1148 The `children' fields of the cycle's primary line and subroutine
1149 lines count only subroutines outside the cycle. Even though `a' calls
1150 `b', the time spent in those calls to `b' is not counted in `a''s
1151 `children' time. Thus, we do not encounter the problem of what to do
1152 when the time in those calls to `b' includes indirect recursive calls
1155 The `children' field of a caller-line in the cycle's entry estimates
1156 the amount of time spent _in the whole cycle_, and its other
1157 subroutines, on the times when that caller called a function in the
1160 The `called' field in the primary line for the cycle has two numbers:
1161 first, the number of times functions in the cycle were called by
1162 functions outside the cycle; second, the number of times they were
1163 called by functions in the cycle (including times when a function in
1164 the cycle calls itself). This is a generalization of the usual split
1165 into non-recursive and recursive calls.
1167 The `called' field of a subroutine-line for a cycle member in the
1168 cycle's entry says how many time that function was called from
1169 functions in the cycle. The total of all these is the second number in
1170 the primary line's `called' field.
1172 In the individual entry for a function in a cycle, the other
1173 functions in the same cycle can appear as subroutines and as callers.
1174 These lines show how many times each function in the cycle called or
1175 was called from each other function in the cycle. The `self' and
1176 `children' fields in these lines are blank because of the difficulty of
1177 defining meanings for them when recursion is going on.
1180 File: gprof.info, Node: Line-by-line, Next: Annotated Source, Prev: Call Graph, Up: Output
1182 5.3 Line-by-line Profiling
1183 ==========================
1185 `gprof''s `-l' option causes the program to perform "line-by-line"
1186 profiling. In this mode, histogram samples are assigned not to
1187 functions, but to individual lines of source code. This only works
1188 with programs compiled with older versions of the `gcc' compiler.
1189 Newer versions of `gcc' use a different program - `gcov' - to display
1190 line-by-line profiling information.
1192 With the older versions of `gcc' the program usually has to be
1193 compiled with a `-g' option, in addition to `-pg', in order to generate
1194 debugging symbols for tracking source code lines. Note, in much older
1195 versions of `gcc' the program had to be compiled with the `-a' command
1196 line option as well.
1198 The flat profile is the most useful output table in line-by-line
1199 mode. The call graph isn't as useful as normal, since the current
1200 version of `gprof' does not propagate call graph arcs from source code
1201 lines to the enclosing function. The call graph does, however, show
1202 each line of code that called each function, along with a count.
1204 Here is a section of `gprof''s output, without line-by-line
1205 profiling. Note that `ct_init' accounted for four histogram hits, and
1206 13327 calls to `init_block'.
1210 Each sample counts as 0.01 seconds.
1211 % cumulative self self total
1212 time seconds seconds calls us/call us/call name
1213 30.77 0.13 0.04 6335 6.31 6.31 ct_init
1216 Call graph (explanation follows)
1219 granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds
1221 index % time self children called name
1223 0.00 0.00 1/13496 name_too_long
1224 0.00 0.00 40/13496 deflate
1225 0.00 0.00 128/13496 deflate_fast
1226 0.00 0.00 13327/13496 ct_init
1227 [7] 0.0 0.00 0.00 13496 init_block
1229 Now let's look at some of `gprof''s output from the same program run,
1230 this time with line-by-line profiling enabled. Note that `ct_init''s
1231 four histogram hits are broken down into four lines of source code--one
1232 hit occurred on each of lines 349, 351, 382 and 385. In the call graph,
1233 note how `ct_init''s 13327 calls to `init_block' are broken down into
1234 one call from line 396, 3071 calls from line 384, 3730 calls from line
1235 385, and 6525 calls from 387.
1239 Each sample counts as 0.01 seconds.
1241 time seconds seconds calls name
1242 7.69 0.10 0.01 ct_init (trees.c:349)
1243 7.69 0.11 0.01 ct_init (trees.c:351)
1244 7.69 0.12 0.01 ct_init (trees.c:382)
1245 7.69 0.13 0.01 ct_init (trees.c:385)
1248 Call graph (explanation follows)
1251 granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds
1253 % time self children called name
1255 0.00 0.00 1/13496 name_too_long (gzip.c:1440)
1256 0.00 0.00 1/13496 deflate (deflate.c:763)
1257 0.00 0.00 1/13496 ct_init (trees.c:396)
1258 0.00 0.00 2/13496 deflate (deflate.c:727)
1259 0.00 0.00 4/13496 deflate (deflate.c:686)
1260 0.00 0.00 5/13496 deflate (deflate.c:675)
1261 0.00 0.00 12/13496 deflate (deflate.c:679)
1262 0.00 0.00 16/13496 deflate (deflate.c:730)
1263 0.00 0.00 128/13496 deflate_fast (deflate.c:654)
1264 0.00 0.00 3071/13496 ct_init (trees.c:384)
1265 0.00 0.00 3730/13496 ct_init (trees.c:385)
1266 0.00 0.00 6525/13496 ct_init (trees.c:387)
1267 [6] 0.0 0.00 0.00 13496 init_block (trees.c:408)
1270 File: gprof.info, Node: Annotated Source, Prev: Line-by-line, Up: Output
1272 5.4 The Annotated Source Listing
1273 ================================
1275 `gprof''s `-A' option triggers an annotated source listing, which lists
1276 the program's source code, each function labeled with the number of
1277 times it was called. You may also need to specify the `-I' option, if
1278 `gprof' can't find the source code files.
1280 With older versions of `gcc' compiling with `gcc ... -g -pg -a'
1281 augments your program with basic-block counting code, in addition to
1282 function counting code. This enables `gprof' to determine how many
1283 times each line of code was executed. With newer versions of `gcc'
1284 support for displaying basic-block counts is provided by the `gcov'
1287 For example, consider the following function, taken from gzip, with
1296 7 static ulg crc = (ulg)0xffffffffL;
1303 14 c = crc_32_tab[...];
1307 18 return c ^ 0xffffffffL;
1310 `updcrc' has at least five basic-blocks. One is the function
1311 itself. The `if' statement on line 9 generates two more basic-blocks,
1312 one for each branch of the `if'. A fourth basic-block results from the
1313 `if' on line 13, and the contents of the `do' loop form the fifth
1314 basic-block. The compiler may also generate additional basic-blocks to
1315 handle various special cases.
1317 A program augmented for basic-block counting can be analyzed with
1318 `gprof -l -A'. The `-x' option is also helpful, to ensure that each
1319 line of code is labeled at least once. Here is `updcrc''s annotated
1320 source listing for a sample `gzip' run:
1328 static ulg crc = (ulg)0xffffffffL;
1330 2 -> if (s == NULL) {
1331 1 -> c = 0xffffffffL;
1335 26312 -> c = crc_32_tab[...];
1336 26312,1,26311 -> } while (--n);
1339 2 -> return c ^ 0xffffffffL;
1342 In this example, the function was called twice, passing once through
1343 each branch of the `if' statement. The body of the `do' loop was
1344 executed a total of 26312 times. Note how the `while' statement is
1345 annotated. It began execution 26312 times, once for each iteration
1346 through the loop. One of those times (the last time) it exited, while
1347 it branched back to the beginning of the loop 26311 times.
1350 File: gprof.info, Node: Inaccuracy, Next: How do I?, Prev: Output, Up: Top
1352 6 Inaccuracy of `gprof' Output
1353 ******************************
1357 * Sampling Error:: Statistical margins of error
1358 * Assumptions:: Estimating children times
1361 File: gprof.info, Node: Sampling Error, Next: Assumptions, Up: Inaccuracy
1363 6.1 Statistical Sampling Error
1364 ==============================
1366 The run-time figures that `gprof' gives you are based on a sampling
1367 process, so they are subject to statistical inaccuracy. If a function
1368 runs only a small amount of time, so that on the average the sampling
1369 process ought to catch that function in the act only once, there is a
1370 pretty good chance it will actually find that function zero times, or
1373 By contrast, the number-of-calls and basic-block figures are derived
1374 by counting, not sampling. They are completely accurate and will not
1375 vary from run to run if your program is deterministic.
1377 The "sampling period" that is printed at the beginning of the flat
1378 profile says how often samples are taken. The rule of thumb is that a
1379 run-time figure is accurate if it is considerably bigger than the
1382 The actual amount of error can be predicted. For N samples, the
1383 _expected_ error is the square-root of N. For example, if the sampling
1384 period is 0.01 seconds and `foo''s run-time is 1 second, N is 100
1385 samples (1 second/0.01 seconds), sqrt(N) is 10 samples, so the expected
1386 error in `foo''s run-time is 0.1 seconds (10*0.01 seconds), or ten
1387 percent of the observed value. Again, if the sampling period is 0.01
1388 seconds and `bar''s run-time is 100 seconds, N is 10000 samples,
1389 sqrt(N) is 100 samples, so the expected error in `bar''s run-time is 1
1390 second, or one percent of the observed value. It is likely to vary
1391 this much _on the average_ from one profiling run to the next.
1392 (_Sometimes_ it will vary more.)
1394 This does not mean that a small run-time figure is devoid of
1395 information. If the program's _total_ run-time is large, a small
1396 run-time for one function does tell you that that function used an
1397 insignificant fraction of the whole program's time. Usually this means
1398 it is not worth optimizing.
1400 One way to get more accuracy is to give your program more (but
1401 similar) input data so it will take longer. Another way is to combine
1402 the data from several runs, using the `-s' option of `gprof'. Here is
1405 1. Run your program once.
1407 2. Issue the command `mv gmon.out gmon.sum'.
1409 3. Run your program again, the same as before.
1411 4. Merge the new data in `gmon.out' into `gmon.sum' with this command:
1413 gprof -s EXECUTABLE-FILE gmon.out gmon.sum
1415 5. Repeat the last two steps as often as you wish.
1417 6. Analyze the cumulative data using this command:
1419 gprof EXECUTABLE-FILE gmon.sum > OUTPUT-FILE
1422 File: gprof.info, Node: Assumptions, Prev: Sampling Error, Up: Inaccuracy
1424 6.2 Estimating `children' Times
1425 ===============================
1427 Some of the figures in the call graph are estimates--for example, the
1428 `children' time values and all the time figures in caller and
1431 There is no direct information about these measurements in the
1432 profile data itself. Instead, `gprof' estimates them by making an
1433 assumption about your program that might or might not be true.
1435 The assumption made is that the average time spent in each call to
1436 any function `foo' is not correlated with who called `foo'. If `foo'
1437 used 5 seconds in all, and 2/5 of the calls to `foo' came from `a',
1438 then `foo' contributes 2 seconds to `a''s `children' time, by
1441 This assumption is usually true enough, but for some programs it is
1442 far from true. Suppose that `foo' returns very quickly when its
1443 argument is zero; suppose that `a' always passes zero as an argument,
1444 while other callers of `foo' pass other arguments. In this program,
1445 all the time spent in `foo' is in the calls from callers other than `a'.
1446 But `gprof' has no way of knowing this; it will blindly and incorrectly
1447 charge 2 seconds of time in `foo' to the children of `a'.
1449 We hope some day to put more complete data into `gmon.out', so that
1450 this assumption is no longer needed, if we can figure out how. For the
1451 novice, the estimated figures are usually more useful than misleading.
1454 File: gprof.info, Node: How do I?, Next: Incompatibilities, Prev: Inaccuracy, Up: Top
1456 7 Answers to Common Questions
1457 *****************************
1459 How can I get more exact information about hot spots in my program?
1460 Looking at the per-line call counts only tells part of the story.
1461 Because `gprof' can only report call times and counts by function,
1462 the best way to get finer-grained information on where the program
1463 is spending its time is to re-factor large functions into sequences
1464 of calls to smaller ones. Beware however that this can introduce
1465 artificial hot spots since compiling with `-pg' adds a significant
1466 overhead to function calls. An alternative solution is to use a
1467 non-intrusive profiler, e.g. oprofile.
1469 How do I find which lines in my program were executed the most times?
1470 Use the `gcov' program.
1472 How do I find which lines in my program called a particular function?
1473 Use `gprof -l' and lookup the function in the call graph. The
1474 callers will be broken down by function and line number.
1476 How do I analyze a program that runs for less than a second?
1477 Try using a shell script like this one:
1479 for i in `seq 1 100`; do
1481 mv gmon.out gmon.out.$i
1484 gprof -s fastprog gmon.out.*
1486 gprof fastprog gmon.sum
1488 If your program is completely deterministic, all the call counts
1489 will be simple multiples of 100 (i.e., a function called once in
1490 each run will appear with a call count of 100).
1494 File: gprof.info, Node: Incompatibilities, Next: Details, Prev: How do I?, Up: Top
1496 8 Incompatibilities with Unix `gprof'
1497 *************************************
1499 GNU `gprof' and Berkeley Unix `gprof' use the same data file
1500 `gmon.out', and provide essentially the same information. But there
1501 are a few differences.
1503 * GNU `gprof' uses a new, generalized file format with support for
1504 basic-block execution counts and non-realtime histograms. A magic
1505 cookie and version number allows `gprof' to easily identify new
1506 style files. Old BSD-style files can still be read. *Note
1507 Profiling Data File Format: File Format.
1509 * For a recursive function, Unix `gprof' lists the function as a
1510 parent and as a child, with a `calls' field that lists the number
1511 of recursive calls. GNU `gprof' omits these lines and puts the
1512 number of recursive calls in the primary line.
1514 * When a function is suppressed from the call graph with `-e', GNU
1515 `gprof' still lists it as a subroutine of functions that call it.
1517 * GNU `gprof' accepts the `-k' with its argument in the form
1518 `from/to', instead of `from to'.
1520 * In the annotated source listing, if there are multiple basic
1521 blocks on the same line, GNU `gprof' prints all of their counts,
1522 separated by commas.
1524 * The blurbs, field widths, and output formats are different. GNU
1525 `gprof' prints blurbs after the tables, so that you can see the
1526 tables without skipping the blurbs.
1529 File: gprof.info, Node: Details, Next: GNU Free Documentation License, Prev: Incompatibilities, Up: Top
1531 9 Details of Profiling
1532 **********************
1536 * Implementation:: How a program collects profiling information
1537 * File Format:: Format of `gmon.out' files
1538 * Internals:: `gprof''s internal operation
1539 * Debugging:: Using `gprof''s `-d' option
1542 File: gprof.info, Node: Implementation, Next: File Format, Up: Details
1544 9.1 Implementation of Profiling
1545 ===============================
1547 Profiling works by changing how every function in your program is
1548 compiled so that when it is called, it will stash away some information
1549 about where it was called from. From this, the profiler can figure out
1550 what function called it, and can count how many times it was called.
1551 This change is made by the compiler when your program is compiled with
1552 the `-pg' option, which causes every function to call `mcount' (or
1553 `_mcount', or `__mcount', depending on the OS and compiler) as one of
1554 its first operations.
1556 The `mcount' routine, included in the profiling library, is
1557 responsible for recording in an in-memory call graph table both its
1558 parent routine (the child) and its parent's parent. This is typically
1559 done by examining the stack frame to find both the address of the
1560 child, and the return address in the original parent. Since this is a
1561 very machine-dependent operation, `mcount' itself is typically a short
1562 assembly-language stub routine that extracts the required information,
1563 and then calls `__mcount_internal' (a normal C function) with two
1564 arguments--`frompc' and `selfpc'. `__mcount_internal' is responsible
1565 for maintaining the in-memory call graph, which records `frompc',
1566 `selfpc', and the number of times each of these call arcs was traversed.
1568 GCC Version 2 provides a magical function
1569 (`__builtin_return_address'), which allows a generic `mcount' function
1570 to extract the required information from the stack frame. However, on
1571 some architectures, most notably the SPARC, using this builtin can be
1572 very computationally expensive, and an assembly language version of
1573 `mcount' is used for performance reasons.
1575 Number-of-calls information for library routines is collected by
1576 using a special version of the C library. The programs in it are the
1577 same as in the usual C library, but they were compiled with `-pg'. If
1578 you link your program with `gcc ... -pg', it automatically uses the
1579 profiling version of the library.
1581 Profiling also involves watching your program as it runs, and
1582 keeping a histogram of where the program counter happens to be every
1583 now and then. Typically the program counter is looked at around 100
1584 times per second of run time, but the exact frequency may vary from
1587 This is done is one of two ways. Most UNIX-like operating systems
1588 provide a `profil()' system call, which registers a memory array with
1589 the kernel, along with a scale factor that determines how the program's
1590 address space maps into the array. Typical scaling values cause every
1591 2 to 8 bytes of address space to map into a single array slot. On
1592 every tick of the system clock (assuming the profiled program is
1593 running), the value of the program counter is examined and the
1594 corresponding slot in the memory array is incremented. Since this is
1595 done in the kernel, which had to interrupt the process anyway to handle
1596 the clock interrupt, very little additional system overhead is required.
1598 However, some operating systems, most notably Linux 2.0 (and
1599 earlier), do not provide a `profil()' system call. On such a system,
1600 arrangements are made for the kernel to periodically deliver a signal
1601 to the process (typically via `setitimer()'), which then performs the
1602 same operation of examining the program counter and incrementing a slot
1603 in the memory array. Since this method requires a signal to be
1604 delivered to user space every time a sample is taken, it uses
1605 considerably more overhead than kernel-based profiling. Also, due to
1606 the added delay required to deliver the signal, this method is less
1609 A special startup routine allocates memory for the histogram and
1610 either calls `profil()' or sets up a clock signal handler. This
1611 routine (`monstartup') can be invoked in several ways. On Linux
1612 systems, a special profiling startup file `gcrt0.o', which invokes
1613 `monstartup' before `main', is used instead of the default `crt0.o'.
1614 Use of this special startup file is one of the effects of using `gcc
1615 ... -pg' to link. On SPARC systems, no special startup files are used.
1616 Rather, the `mcount' routine, when it is invoked for the first time
1617 (typically when `main' is called), calls `monstartup'.
1619 If the compiler's `-a' option was used, basic-block counting is also
1620 enabled. Each object file is then compiled with a static array of
1621 counts, initially zero. In the executable code, every time a new
1622 basic-block begins (i.e., when an `if' statement appears), an extra
1623 instruction is inserted to increment the corresponding count in the
1624 array. At compile time, a paired array was constructed that recorded
1625 the starting address of each basic-block. Taken together, the two
1626 arrays record the starting address of every basic-block, along with the
1627 number of times it was executed.
1629 The profiling library also includes a function (`mcleanup') which is
1630 typically registered using `atexit()' to be called as the program
1631 exits, and is responsible for writing the file `gmon.out'. Profiling
1632 is turned off, various headers are output, and the histogram is
1633 written, followed by the call-graph arcs and the basic-block counts.
1635 The output from `gprof' gives no indication of parts of your program
1636 that are limited by I/O or swapping bandwidth. This is because samples
1637 of the program counter are taken at fixed intervals of the program's
1638 run time. Therefore, the time measurements in `gprof' output say
1639 nothing about time that your program was not running. For example, a
1640 part of the program that creates so much data that it cannot all fit in
1641 physical memory at once may run very slowly due to thrashing, but
1642 `gprof' will say it uses little time. On the other hand, sampling by
1643 run time has the advantage that the amount of load due to other users
1644 won't directly affect the output you get.
1647 File: gprof.info, Node: File Format, Next: Internals, Prev: Implementation, Up: Details
1649 9.2 Profiling Data File Format
1650 ==============================
1652 The old BSD-derived file format used for profile data does not contain a
1653 magic cookie that allows to check whether a data file really is a
1654 `gprof' file. Furthermore, it does not provide a version number, thus
1655 rendering changes to the file format almost impossible. GNU `gprof'
1656 uses a new file format that provides these features. For backward
1657 compatibility, GNU `gprof' continues to support the old BSD-derived
1658 format, but not all features are supported with it. For example,
1659 basic-block execution counts cannot be accommodated by the old file
1662 The new file format is defined in header file `gmon_out.h'. It
1663 consists of a header containing the magic cookie and a version number,
1664 as well as some spare bytes available for future extensions. All data
1665 in a profile data file is in the native format of the target for which
1666 the profile was collected. GNU `gprof' adapts automatically to the
1669 In the new file format, the header is followed by a sequence of
1670 records. Currently, there are three different record types: histogram
1671 records, call-graph arc records, and basic-block execution count
1672 records. Each file can contain any number of each record type. When
1673 reading a file, GNU `gprof' will ensure records of the same type are
1674 compatible with each other and compute the union of all records. For
1675 example, for basic-block execution counts, the union is simply the sum
1676 of all execution counts for each basic-block.
1678 9.2.1 Histogram Records
1679 -----------------------
1681 Histogram records consist of a header that is followed by an array of
1682 bins. The header contains the text-segment range that the histogram
1683 spans, the size of the histogram in bytes (unlike in the old BSD
1684 format, this does not include the size of the header), the rate of the
1685 profiling clock, and the physical dimension that the bin counts
1686 represent after being scaled by the profiling clock rate. The physical
1687 dimension is specified in two parts: a long name of up to 15 characters
1688 and a single character abbreviation. For example, a histogram
1689 representing real-time would specify the long name as "seconds" and the
1690 abbreviation as "s". This feature is useful for architectures that
1691 support performance monitor hardware (which, fortunately, is becoming
1692 increasingly common). For example, under DEC OSF/1, the "uprofile"
1693 command can be used to produce a histogram of, say, instruction cache
1694 misses. In this case, the dimension in the histogram header could be
1695 set to "i-cache misses" and the abbreviation could be set to "1"
1696 (because it is simply a count, not a physical dimension). Also, the
1697 profiling rate would have to be set to 1 in this case.
1699 Histogram bins are 16-bit numbers and each bin represent an equal
1700 amount of text-space. For example, if the text-segment is one thousand
1701 bytes long and if there are ten bins in the histogram, each bin
1702 represents one hundred bytes.
1704 9.2.2 Call-Graph Records
1705 ------------------------
1707 Call-graph records have a format that is identical to the one used in
1708 the BSD-derived file format. It consists of an arc in the call graph
1709 and a count indicating the number of times the arc was traversed during
1710 program execution. Arcs are specified by a pair of addresses: the
1711 first must be within caller's function and the second must be within
1712 the callee's function. When performing profiling at the function
1713 level, these addresses can point anywhere within the respective
1714 function. However, when profiling at the line-level, it is better if
1715 the addresses are as close to the call-site/entry-point as possible.
1716 This will ensure that the line-level call-graph is able to identify
1717 exactly which line of source code performed calls to a function.
1719 9.2.3 Basic-Block Execution Count Records
1720 -----------------------------------------
1722 Basic-block execution count records consist of a header followed by a
1723 sequence of address/count pairs. The header simply specifies the
1724 length of the sequence. In an address/count pair, the address
1725 identifies a basic-block and the count specifies the number of times
1726 that basic-block was executed. Any address within the basic-address can
1730 File: gprof.info, Node: Internals, Next: Debugging, Prev: File Format, Up: Details
1732 9.3 `gprof''s Internal Operation
1733 ================================
1735 Like most programs, `gprof' begins by processing its options. During
1736 this stage, it may building its symspec list (`sym_ids.c:sym_id_add'),
1737 if options are specified which use symspecs. `gprof' maintains a
1738 single linked list of symspecs, which will eventually get turned into
1739 12 symbol tables, organized into six include/exclude pairs--one pair
1740 each for the flat profile (INCL_FLAT/EXCL_FLAT), the call graph arcs
1741 (INCL_ARCS/EXCL_ARCS), printing in the call graph
1742 (INCL_GRAPH/EXCL_GRAPH), timing propagation in the call graph
1743 (INCL_TIME/EXCL_TIME), the annotated source listing
1744 (INCL_ANNO/EXCL_ANNO), and the execution count listing
1745 (INCL_EXEC/EXCL_EXEC).
1747 After option processing, `gprof' finishes building the symspec list
1748 by adding all the symspecs in `default_excluded_list' to the exclude
1749 lists EXCL_TIME and EXCL_GRAPH, and if line-by-line profiling is
1750 specified, EXCL_FLAT as well. These default excludes are not added to
1751 EXCL_ANNO, EXCL_ARCS, and EXCL_EXEC.
1753 Next, the BFD library is called to open the object file, verify that
1754 it is an object file, and read its symbol table (`core.c:core_init'),
1755 using `bfd_canonicalize_symtab' after mallocing an appropriately sized
1756 array of symbols. At this point, function mappings are read (if the
1757 `--file-ordering' option has been specified), and the core text space
1758 is read into memory (if the `-c' option was given).
1760 `gprof''s own symbol table, an array of Sym structures, is now built.
1761 This is done in one of two ways, by one of two routines, depending on
1762 whether line-by-line profiling (`-l' option) has been enabled. For
1763 normal profiling, the BFD canonical symbol table is scanned. For
1764 line-by-line profiling, every text space address is examined, and a new
1765 symbol table entry gets created every time the line number changes. In
1766 either case, two passes are made through the symbol table--one to count
1767 the size of the symbol table required, and the other to actually read
1768 the symbols. In between the two passes, a single array of type `Sym'
1769 is created of the appropriate length. Finally,
1770 `symtab.c:symtab_finalize' is called to sort the symbol table and
1771 remove duplicate entries (entries with the same memory address).
1773 The symbol table must be a contiguous array for two reasons. First,
1774 the `qsort' library function (which sorts an array) will be used to
1775 sort the symbol table. Also, the symbol lookup routine
1776 (`symtab.c:sym_lookup'), which finds symbols based on memory address,
1777 uses a binary search algorithm which requires the symbol table to be a
1778 sorted array. Function symbols are indicated with an `is_func' flag.
1779 Line number symbols have no special flags set. Additionally, a symbol
1780 can have an `is_static' flag to indicate that it is a local symbol.
1782 With the symbol table read, the symspecs can now be translated into
1783 Syms (`sym_ids.c:sym_id_parse'). Remember that a single symspec can
1784 match multiple symbols. An array of symbol tables (`syms') is created,
1785 each entry of which is a symbol table of Syms to be included or
1786 excluded from a particular listing. The master symbol table and the
1787 symspecs are examined by nested loops, and every symbol that matches a
1788 symspec is inserted into the appropriate syms table. This is done
1789 twice, once to count the size of each required symbol table, and again
1790 to build the tables, which have been malloced between passes. From now
1791 on, to determine whether a symbol is on an include or exclude symspec
1792 list, `gprof' simply uses its standard symbol lookup routine on the
1793 appropriate table in the `syms' array.
1795 Now the profile data file(s) themselves are read
1796 (`gmon_io.c:gmon_out_read'), first by checking for a new-style
1797 `gmon.out' header, then assuming this is an old-style BSD `gmon.out' if
1798 the magic number test failed.
1800 New-style histogram records are read by `hist.c:hist_read_rec'. For
1801 the first histogram record, allocate a memory array to hold all the
1802 bins, and read them in. When multiple profile data files (or files
1803 with multiple histogram records) are read, the memory ranges of each
1804 pair of histogram records must be either equal, or non-overlapping.
1805 For each pair of histogram records, the resolution (memory region size
1806 divided by the number of bins) must be the same. The time unit must be
1807 the same for all histogram records. If the above containts are met, all
1808 histograms for the same memory range are merged.
1810 As each call graph record is read (`call_graph.c:cg_read_rec'), the
1811 parent and child addresses are matched to symbol table entries, and a
1812 call graph arc is created by `cg_arcs.c:arc_add', unless the arc fails
1813 a symspec check against INCL_ARCS/EXCL_ARCS. As each arc is added, a
1814 linked list is maintained of the parent's child arcs, and of the child's
1815 parent arcs. Both the child's call count and the arc's call count are
1816 incremented by the record's call count.
1818 Basic-block records are read (`basic_blocks.c:bb_read_rec'), but
1819 only if line-by-line profiling has been selected. Each basic-block
1820 address is matched to a corresponding line symbol in the symbol table,
1821 and an entry made in the symbol's bb_addr and bb_calls arrays. Again,
1822 if multiple basic-block records are present for the same address, the
1823 call counts are cumulative.
1825 A gmon.sum file is dumped, if requested (`gmon_io.c:gmon_out_write').
1827 If histograms were present in the data files, assign them to symbols
1828 (`hist.c:hist_assign_samples') by iterating over all the sample bins
1829 and assigning them to symbols. Since the symbol table is sorted in
1830 order of ascending memory addresses, we can simple follow along in the
1831 symbol table as we make our pass over the sample bins. This step
1832 includes a symspec check against INCL_FLAT/EXCL_FLAT. Depending on the
1833 histogram scale factor, a sample bin may span multiple symbols, in
1834 which case a fraction of the sample count is allocated to each symbol,
1835 proportional to the degree of overlap. This effect is rare for normal
1836 profiling, but overlaps are more common during line-by-line profiling,
1837 and can cause each of two adjacent lines to be credited with half a
1840 If call graph data is present, `cg_arcs.c:cg_assemble' is called.
1841 First, if `-c' was specified, a machine-dependent routine (`find_call')
1842 scans through each symbol's machine code, looking for subroutine call
1843 instructions, and adding them to the call graph with a zero call count.
1844 A topological sort is performed by depth-first numbering all the
1845 symbols (`cg_dfn.c:cg_dfn'), so that children are always numbered less
1846 than their parents, then making a array of pointers into the symbol
1847 table and sorting it into numerical order, which is reverse topological
1848 order (children appear before parents). Cycles are also detected at
1849 this point, all members of which are assigned the same topological
1850 number. Two passes are now made through this sorted array of symbol
1851 pointers. The first pass, from end to beginning (parents to children),
1852 computes the fraction of child time to propagate to each parent and a
1853 print flag. The print flag reflects symspec handling of
1854 INCL_GRAPH/EXCL_GRAPH, with a parent's include or exclude (print or no
1855 print) property being propagated to its children, unless they
1856 themselves explicitly appear in INCL_GRAPH or EXCL_GRAPH. A second
1857 pass, from beginning to end (children to parents) actually propagates
1858 the timings along the call graph, subject to a check against
1859 INCL_TIME/EXCL_TIME. With the print flag, fractions, and timings now
1860 stored in the symbol structures, the topological sort array is now
1861 discarded, and a new array of pointers is assembled, this time sorted
1864 Finally, print the various outputs the user requested, which is now
1865 fairly straightforward. The call graph (`cg_print.c:cg_print') and
1866 flat profile (`hist.c:hist_print') are regurgitations of values already
1867 computed. The annotated source listing
1868 (`basic_blocks.c:print_annotated_source') uses basic-block information,
1869 if present, to label each line of code with call counts, otherwise only
1870 the function call counts are presented.
1872 The function ordering code is marginally well documented in the
1873 source code itself (`cg_print.c'). Basically, the functions with the
1874 most use and the most parents are placed first, followed by other
1875 functions with the most use, followed by lower use functions, followed
1876 by unused functions at the end.
1879 File: gprof.info, Node: Debugging, Prev: Internals, Up: Details
1881 9.4 Debugging `gprof'
1882 =====================
1884 If `gprof' was compiled with debugging enabled, the `-d' option
1885 triggers debugging output (to stdout) which can be helpful in
1886 understanding its operation. The debugging number specified is
1887 interpreted as a sum of the following options:
1889 2 - Topological sort
1890 Monitor depth-first numbering of symbols during call graph analysis
1893 Shows symbols as they are identified as cycle heads
1896 As the call graph arcs are read, show each arc and how the total
1897 calls to each function are tallied
1899 32 - Call graph arc sorting
1900 Details sorting individual parents/children within each call graph
1903 64 - Reading histogram and call graph records
1904 Shows address ranges of histograms as they are read, and each call
1908 Reading, classifying, and sorting the symbol table from the object
1909 file. For line-by-line profiling (`-l' option), also shows line
1910 numbers being assigned to memory addresses.
1912 256 - Static call graph
1913 Trace operation of `-c' option
1915 512 - Symbol table and arc table lookups
1916 Detail operation of lookup routines
1918 1024 - Call graph propagation
1919 Shows how function times are propagated along the call graph
1922 Shows basic-block records as they are read from profile data (only
1923 meaningful with `-l' option)
1926 Shows symspec-to-symbol pattern matching operation
1928 8192 - Annotate source
1929 Tracks operation of `-A' option
1932 File: gprof.info, Node: GNU Free Documentation License, Prev: Details, Up: Top
1934 Appendix A GNU Free Documentation License
1935 *****************************************
1937 Version 1.1, March 2000
1939 Copyright (C) 2000, 2003 Free Software Foundation, Inc.
1940 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
1942 Everyone is permitted to copy and distribute verbatim copies
1943 of this license document, but changing it is not allowed.
1948 The purpose of this License is to make a manual, textbook, or other
1949 written document "free" in the sense of freedom: to assure everyone
1950 the effective freedom to copy and redistribute it, with or without
1951 modifying it, either commercially or noncommercially. Secondarily,
1952 this License preserves for the author and publisher a way to get
1953 credit for their work, while not being considered responsible for
1954 modifications made by others.
1956 This License is a kind of "copyleft", which means that derivative
1957 works of the document must themselves be free in the same sense.
1958 It complements the GNU General Public License, which is a copyleft
1959 license designed for free software.
1961 We have designed this License in order to use it for manuals for
1962 free software, because free software needs free documentation: a
1963 free program should come with manuals providing the same freedoms
1964 that the software does. But this License is not limited to
1965 software manuals; it can be used for any textual work, regardless
1966 of subject matter or whether it is published as a printed book.
1967 We recommend this License principally for works whose purpose is
1968 instruction or reference.
1971 1. APPLICABILITY AND DEFINITIONS
1973 This License applies to any manual or other work that contains a
1974 notice placed by the copyright holder saying it can be distributed
1975 under the terms of this License. The "Document", below, refers to
1976 any such manual or work. Any member of the public is a licensee,
1977 and is addressed as "you."
1979 A "Modified Version" of the Document means any work containing the
1980 Document or a portion of it, either copied verbatim, or with
1981 modifications and/or translated into another language.
1983 A "Secondary Section" is a named appendix or a front-matter
1984 section of the Document that deals exclusively with the
1985 relationship of the publishers or authors of the Document to the
1986 Document's overall subject (or to related matters) and contains
1987 nothing that could fall directly within that overall subject.
1988 (For example, if the Document is in part a textbook of
1989 mathematics, a Secondary Section may not explain any mathematics.)
1990 The relationship could be a matter of historical connection with
1991 the subject or with related matters, or of legal, commercial,
1992 philosophical, ethical or political position regarding them.
1994 The "Invariant Sections" are certain Secondary Sections whose
1995 titles are designated, as being those of Invariant Sections, in
1996 the notice that says that the Document is released under this
1999 The "Cover Texts" are certain short passages of text that are
2000 listed, as Front-Cover Texts or Back-Cover Texts, in the notice
2001 that says that the Document is released under this License.
2003 A "Transparent" copy of the Document means a machine-readable copy,
2004 represented in a format whose specification is available to the
2005 general public, whose contents can be viewed and edited directly
2006 and straightforwardly with generic text editors or (for images
2007 composed of pixels) generic paint programs or (for drawings) some
2008 widely available drawing editor, and that is suitable for input to
2009 text formatters or for automatic translation to a variety of
2010 formats suitable for input to text formatters. A copy made in an
2011 otherwise Transparent file format whose markup has been designed
2012 to thwart or discourage subsequent modification by readers is not
2013 Transparent. A copy that is not "Transparent" is called "Opaque."
2015 Examples of suitable formats for Transparent copies include plain
2016 ASCII without markup, Texinfo input format, LaTeX input format,
2017 SGML or XML using a publicly available DTD, and
2018 standard-conforming simple HTML designed for human modification.
2019 Opaque formats include PostScript, PDF, proprietary formats that
2020 can be read and edited only by proprietary word processors, SGML
2021 or XML for which the DTD and/or processing tools are not generally
2022 available, and the machine-generated HTML produced by some word
2023 processors for output purposes only.
2025 The "Title Page" means, for a printed book, the title page itself,
2026 plus such following pages as are needed to hold, legibly, the
2027 material this License requires to appear in the title page. For
2028 works in formats which do not have any title page as such, "Title
2029 Page" means the text near the most prominent appearance of the
2030 work's title, preceding the beginning of the body of the text.
2034 You may copy and distribute the Document in any medium, either
2035 commercially or noncommercially, provided that this License, the
2036 copyright notices, and the license notice saying this License
2037 applies to the Document are reproduced in all copies, and that you
2038 add no other conditions whatsoever to those of this License. You
2039 may not use technical measures to obstruct or control the reading
2040 or further copying of the copies you make or distribute. However,
2041 you may accept compensation in exchange for copies. If you
2042 distribute a large enough number of copies you must also follow
2043 the conditions in section 3.
2045 You may also lend copies, under the same conditions stated above,
2046 and you may publicly display copies.
2048 3. COPYING IN QUANTITY
2050 If you publish printed copies of the Document numbering more than
2051 100, and the Document's license notice requires Cover Texts, you
2052 must enclose the copies in covers that carry, clearly and legibly,
2053 all these Cover Texts: Front-Cover Texts on the front cover, and
2054 Back-Cover Texts on the back cover. Both covers must also clearly
2055 and legibly identify you as the publisher of these copies. The
2056 front cover must present the full title with all words of the
2057 title equally prominent and visible. You may add other material
2058 on the covers in addition. Copying with changes limited to the
2059 covers, as long as they preserve the title of the Document and
2060 satisfy these conditions, can be treated as verbatim copying in
2063 If the required texts for either cover are too voluminous to fit
2064 legibly, you should put the first ones listed (as many as fit
2065 reasonably) on the actual cover, and continue the rest onto
2068 If you publish or distribute Opaque copies of the Document
2069 numbering more than 100, you must either include a
2070 machine-readable Transparent copy along with each Opaque copy, or
2071 state in or with each Opaque copy a publicly-accessible
2072 computer-network location containing a complete Transparent copy
2073 of the Document, free of added material, which the general
2074 network-using public has access to download anonymously at no
2075 charge using public-standard network protocols. If you use the
2076 latter option, you must take reasonably prudent steps, when you
2077 begin distribution of Opaque copies in quantity, to ensure that
2078 this Transparent copy will remain thus accessible at the stated
2079 location until at least one year after the last time you
2080 distribute an Opaque copy (directly or through your agents or
2081 retailers) of that edition to the public.
2083 It is requested, but not required, that you contact the authors of
2084 the Document well before redistributing any large number of
2085 copies, to give them a chance to provide you with an updated
2086 version of the Document.
2090 You may copy and distribute a Modified Version of the Document
2091 under the conditions of sections 2 and 3 above, provided that you
2092 release the Modified Version under precisely this License, with
2093 the Modified Version filling the role of the Document, thus
2094 licensing distribution and modification of the Modified Version to
2095 whoever possesses a copy of it. In addition, you must do these
2096 things in the Modified Version:
2098 A. Use in the Title Page (and on the covers, if any) a title
2099 distinct from that of the Document, and from those of previous
2100 versions (which should, if there were any, be listed in the
2101 History section of the Document). You may use the same title
2102 as a previous version if the original publisher of that version
2104 B. List on the Title Page, as authors, one or more persons or
2105 entities responsible for authorship of the modifications in the
2106 Modified Version, together with at least five of the principal
2107 authors of the Document (all of its principal authors, if it
2108 has less than five).
2109 C. State on the Title page the name of the publisher of the
2110 Modified Version, as the publisher.
2111 D. Preserve all the copyright notices of the Document.
2112 E. Add an appropriate copyright notice for your modifications
2113 adjacent to the other copyright notices.
2114 F. Include, immediately after the copyright notices, a license
2115 notice giving the public permission to use the Modified Version
2116 under the terms of this License, in the form shown in the
2118 G. Preserve in that license notice the full lists of Invariant
2119 Sections and required Cover Texts given in the Document's
2121 H. Include an unaltered copy of this License.
2122 I. Preserve the section entitled "History", and its title, and add
2123 to it an item stating at least the title, year, new authors, and
2124 publisher of the Modified Version as given on the Title Page.
2125 If there is no section entitled "History" in the Document,
2126 create one stating the title, year, authors, and publisher of
2127 the Document as given on its Title Page, then add an item
2128 describing the Modified Version as stated in the previous
2130 J. Preserve the network location, if any, given in the Document for
2131 public access to a Transparent copy of the Document, and
2132 likewise the network locations given in the Document for
2133 previous versions it was based on. These may be placed in the
2134 "History" section. You may omit a network location for a work
2135 that was published at least four years before the Document
2136 itself, or if the original publisher of the version it refers
2137 to gives permission.
2138 K. In any section entitled "Acknowledgements" or "Dedications",
2139 preserve the section's title, and preserve in the section all the
2140 substance and tone of each of the contributor acknowledgements
2141 and/or dedications given therein.
2142 L. Preserve all the Invariant Sections of the Document,
2143 unaltered in their text and in their titles. Section numbers
2144 or the equivalent are not considered part of the section titles.
2145 M. Delete any section entitled "Endorsements." Such a section
2146 may not be included in the Modified Version.
2147 N. Do not retitle any existing section as "Endorsements" or to
2148 conflict in title with any Invariant Section.
2150 If the Modified Version includes new front-matter sections or
2151 appendices that qualify as Secondary Sections and contain no
2152 material copied from the Document, you may at your option
2153 designate some or all of these sections as invariant. To do this,
2154 add their titles to the list of Invariant Sections in the Modified
2155 Version's license notice. These titles must be distinct from any
2156 other section titles.
2158 You may add a section entitled "Endorsements", provided it contains
2159 nothing but endorsements of your Modified Version by various
2160 parties-for example, statements of peer review or that the text has
2161 been approved by an organization as the authoritative definition
2164 You may add a passage of up to five words as a Front-Cover Text,
2165 and a passage of up to 25 words as a Back-Cover Text, to the end
2166 of the list of Cover Texts in the Modified Version. Only one
2167 passage of Front-Cover Text and one of Back-Cover Text may be
2168 added by (or through arrangements made by) any one entity. If the
2169 Document already includes a cover text for the same cover,
2170 previously added by you or by arrangement made by the same entity
2171 you are acting on behalf of, you may not add another; but you may
2172 replace the old one, on explicit permission from the previous
2173 publisher that added the old one.
2175 The author(s) and publisher(s) of the Document do not by this
2176 License give permission to use their names for publicity for or to
2177 assert or imply endorsement of any Modified Version.
2179 5. COMBINING DOCUMENTS
2181 You may combine the Document with other documents released under
2182 this License, under the terms defined in section 4 above for
2183 modified versions, provided that you include in the combination
2184 all of the Invariant Sections of all of the original documents,
2185 unmodified, and list them all as Invariant Sections of your
2186 combined work in its license notice.
2188 The combined work need only contain one copy of this License, and
2189 multiple identical Invariant Sections may be replaced with a single
2190 copy. If there are multiple Invariant Sections with the same name
2191 but different contents, make the title of each such section unique
2192 by adding at the end of it, in parentheses, the name of the
2193 original author or publisher of that section if known, or else a
2194 unique number. Make the same adjustment to the section titles in
2195 the list of Invariant Sections in the license notice of the
2198 In the combination, you must combine any sections entitled
2199 "History" in the various original documents, forming one section
2200 entitled "History"; likewise combine any sections entitled
2201 "Acknowledgements", and any sections entitled "Dedications." You
2202 must delete all sections entitled "Endorsements."
2204 6. COLLECTIONS OF DOCUMENTS
2206 You may make a collection consisting of the Document and other
2207 documents released under this License, and replace the individual
2208 copies of this License in the various documents with a single copy
2209 that is included in the collection, provided that you follow the
2210 rules of this License for verbatim copying of each of the
2211 documents in all other respects.
2213 You may extract a single document from such a collection, and
2214 distribute it individually under this License, provided you insert
2215 a copy of this License into the extracted document, and follow
2216 this License in all other respects regarding verbatim copying of
2219 7. AGGREGATION WITH INDEPENDENT WORKS
2221 A compilation of the Document or its derivatives with other
2222 separate and independent documents or works, in or on a volume of
2223 a storage or distribution medium, does not as a whole count as a
2224 Modified Version of the Document, provided no compilation
2225 copyright is claimed for the compilation. Such a compilation is
2226 called an "aggregate", and this License does not apply to the
2227 other self-contained works thus compiled with the Document, on
2228 account of their being thus compiled, if they are not themselves
2229 derivative works of the Document.
2231 If the Cover Text requirement of section 3 is applicable to these
2232 copies of the Document, then if the Document is less than one
2233 quarter of the entire aggregate, the Document's Cover Texts may be
2234 placed on covers that surround only the Document within the
2235 aggregate. Otherwise they must appear on covers around the whole
2240 Translation is considered a kind of modification, so you may
2241 distribute translations of the Document under the terms of section
2242 4. Replacing Invariant Sections with translations requires special
2243 permission from their copyright holders, but you may include
2244 translations of some or all Invariant Sections in addition to the
2245 original versions of these Invariant Sections. You may include a
2246 translation of this License provided that you also include the
2247 original English version of this License. In case of a
2248 disagreement between the translation and the original English
2249 version of this License, the original English version will prevail.
2253 You may not copy, modify, sublicense, or distribute the Document
2254 except as expressly provided for under this License. Any other
2255 attempt to copy, modify, sublicense or distribute the Document is
2256 void, and will automatically terminate your rights under this
2257 License. However, parties who have received copies, or rights,
2258 from you under this License will not have their licenses
2259 terminated so long as such parties remain in full compliance.
2261 10. FUTURE REVISIONS OF THIS LICENSE
2263 The Free Software Foundation may publish new, revised versions of
2264 the GNU Free Documentation License from time to time. Such new
2265 versions will be similar in spirit to the present version, but may
2266 differ in detail to address new problems or concerns. See
2267 http://www.gnu.org/copyleft/.
2269 Each version of the License is given a distinguishing version
2270 number. If the Document specifies that a particular numbered
2271 version of this License "or any later version" applies to it, you
2272 have the option of following the terms and conditions either of
2273 that specified version or of any later version that has been
2274 published (not as a draft) by the Free Software Foundation. If
2275 the Document does not specify a version number of this License,
2276 you may choose any version ever published (not as a draft) by the
2277 Free Software Foundation.
2280 ADDENDUM: How to use this License for your documents
2281 ====================================================
2283 To use this License in a document you have written, include a copy of
2284 the License in the document and put the following copyright and license
2285 notices just after the title page:
2287 Copyright (C) YEAR YOUR NAME.
2288 Permission is granted to copy, distribute and/or modify this document
2289 under the terms of the GNU Free Documentation License, Version 1.1
2290 or any later version published by the Free Software Foundation;
2291 with the Invariant Sections being LIST THEIR TITLES, with the
2292 Front-Cover Texts being LIST, and with the Back-Cover Texts being LIST.
2293 A copy of the license is included in the section entitled "GNU
2294 Free Documentation License."
2296 If you have no Invariant Sections, write "with no Invariant Sections"
2297 instead of saying which ones are invariant. If you have no Front-Cover
2298 Texts, write "no Front-Cover Texts" instead of "Front-Cover Texts being
2299 LIST"; likewise for Back-Cover Texts.
2301 If your document contains nontrivial examples of program code, we
2302 recommend releasing these examples in parallel under your choice of
2303 free software license, such as the GNU General Public License, to
2304 permit their use in free software.
2310 Node: Introduction
\7f2033
2311 Node: Compiling
\7f4525
2312 Node: Executing
\7f7996
2313 Node: Invoking
\7f10784
2314 Node: Output Options
\7f12199
2315 Node: Analysis Options
\7f19288
2316 Node: Miscellaneous Options
\7f22689
2317 Node: Deprecated Options
\7f23944
2318 Node: Symspecs
\7f26023
2319 Node: Output
\7f27849
2320 Node: Flat Profile
\7f28889
2321 Node: Call Graph
\7f33842
2322 Node: Primary
\7f37074
2323 Node: Callers
\7f39662
2324 Node: Subroutines
\7f41779
2325 Node: Cycles
\7f43620
2326 Node: Line-by-line
\7f50397
2327 Node: Annotated Source
\7f54470
2328 Node: Inaccuracy
\7f57469
2329 Node: Sampling Error
\7f57727
2330 Node: Assumptions
\7f60297
2331 Node: How do I?
\7f61767
2332 Node: Incompatibilities
\7f63321
2333 Node: Details
\7f64815
2334 Node: Implementation
\7f65208
2335 Node: File Format
\7f71105
2336 Node: Internals
\7f75395
2337 Node: Debugging
\7f83890
2338 Node: GNU Free Documentation License
\7f85491