Commit | Line | Data |
---|---|---|
252b5132 RH |
1 | \input texinfo @c -*-texinfo-*- |
2 | @setfilename gprof.info | |
3 | @settitle GNU gprof | |
4 | @setchapternewpage odd | |
5 | ||
6 | @ifinfo | |
7 | @c This is a dir.info fragment to support semi-automated addition of | |
8 | @c manuals to an info tree. zoo@cygnus.com is developing this facility. | |
9 | @format | |
10 | START-INFO-DIR-ENTRY | |
11 | * gprof: (gprof). Profiling your program's execution | |
12 | END-INFO-DIR-ENTRY | |
13 | @end format | |
14 | @end ifinfo | |
15 | ||
16 | @ifinfo | |
17 | This file documents the gprof profiler of the GNU system. | |
18 | ||
5af11cab | 19 | Copyright (C) 1988, 92, 97, 98, 99, 2000 Free Software Foundation, Inc. |
252b5132 RH |
20 | |
21 | Permission is granted to make and distribute verbatim copies of | |
22 | this manual provided the copyright notice and this permission notice | |
23 | are preserved on all copies. | |
24 | ||
25 | @ignore | |
26 | Permission is granted to process this file through Tex and print the | |
27 | results, provided the printed document carries copying permission | |
28 | notice identical to this one except for the removal of this paragraph | |
29 | (this paragraph not being relevant to the printed manual). | |
30 | ||
31 | @end ignore | |
32 | Permission is granted to copy and distribute modified versions of this | |
33 | manual under the conditions for verbatim copying, provided that the entire | |
34 | resulting derived work is distributed under the terms of a permission | |
35 | notice identical to this one. | |
36 | ||
37 | Permission is granted to copy and distribute translations of this manual | |
38 | into another language, under the above conditions for modified versions. | |
39 | @end ifinfo | |
40 | ||
41 | @finalout | |
42 | @smallbook | |
43 | ||
44 | @titlepage | |
45 | @title GNU gprof | |
46 | @subtitle The @sc{gnu} Profiler | |
47 | @author Jay Fenlason and Richard Stallman | |
48 | ||
49 | @page | |
50 | ||
51 | This manual describes the @sc{gnu} profiler, @code{gprof}, and how you | |
52 | can use it to determine which parts of a program are taking most of the | |
53 | execution time. We assume that you know how to write, compile, and | |
54 | execute programs. @sc{gnu} @code{gprof} was written by Jay Fenlason. | |
55 | ||
252b5132 | 56 | @vskip 0pt plus 1filll |
5af11cab | 57 | Copyright @copyright{} 1988, 92, 97, 98, 99, 2000 Free Software Foundation, Inc. |
252b5132 RH |
58 | |
59 | Permission is granted to make and distribute verbatim copies of | |
60 | this manual provided the copyright notice and this permission notice | |
61 | are preserved on all copies. | |
62 | ||
63 | @ignore | |
64 | Permission is granted to process this file through TeX and print the | |
65 | results, provided the printed document carries copying permission | |
66 | notice identical to this one except for the removal of this paragraph | |
67 | (this paragraph not being relevant to the printed manual). | |
68 | ||
69 | @end ignore | |
70 | Permission is granted to copy and distribute modified versions of this | |
71 | manual under the conditions for verbatim copying, provided that the entire | |
72 | resulting derived work is distributed under the terms of a permission | |
73 | notice identical to this one. | |
74 | ||
75 | Permission is granted to copy and distribute translations of this manual | |
76 | into another language, under the same conditions as for modified versions. | |
77 | ||
78 | @end titlepage | |
79 | ||
80 | @ifinfo | |
81 | @node Top | |
82 | @top Profiling a Program: Where Does It Spend Its Time? | |
83 | ||
84 | This manual describes the @sc{gnu} profiler, @code{gprof}, and how you | |
85 | can use it to determine which parts of a program are taking most of the | |
86 | execution time. We assume that you know how to write, compile, and | |
87 | execute programs. @sc{gnu} @code{gprof} was written by Jay Fenlason. | |
88 | ||
252b5132 RH |
89 | @menu |
90 | * Introduction:: What profiling means, and why it is useful. | |
91 | ||
92 | * Compiling:: How to compile your program for profiling. | |
93 | * Executing:: Executing your program to generate profile data | |
94 | * Invoking:: How to run @code{gprof}, and its options | |
95 | ||
96 | * Output:: Interpreting @code{gprof}'s output | |
97 | ||
98 | * Inaccuracy:: Potential problems you should be aware of | |
99 | * How do I?:: Answers to common questions | |
100 | * Incompatibilities:: (between @sc{gnu} @code{gprof} and Unix @code{gprof}.) | |
101 | * Details:: Details of how profiling is done | |
102 | @end menu | |
103 | @end ifinfo | |
104 | ||
105 | @node Introduction | |
106 | @chapter Introduction to Profiling | |
107 | ||
108 | Profiling allows you to learn where your program spent its time and which | |
109 | functions called which other functions while it was executing. This | |
110 | information can show you which pieces of your program are slower than you | |
111 | expected, and might be candidates for rewriting to make your program | |
112 | execute faster. It can also tell you which functions are being called more | |
113 | or less often than you expected. This may help you spot bugs that had | |
114 | otherwise been unnoticed. | |
115 | ||
116 | Since the profiler uses information collected during the actual execution | |
117 | of your program, it can be used on programs that are too large or too | |
118 | complex to analyze by reading the source. However, how your program is run | |
119 | will affect the information that shows up in the profile data. If you | |
120 | don't use some feature of your program while it is being profiled, no | |
121 | profile information will be generated for that feature. | |
122 | ||
123 | Profiling has several steps: | |
124 | ||
125 | @itemize @bullet | |
126 | @item | |
127 | You must compile and link your program with profiling enabled. | |
128 | @xref{Compiling}. | |
129 | ||
130 | @item | |
131 | You must execute your program to generate a profile data file. | |
132 | @xref{Executing}. | |
133 | ||
134 | @item | |
135 | You must run @code{gprof} to analyze the profile data. | |
136 | @xref{Invoking}. | |
137 | @end itemize | |
138 | ||
139 | The next three chapters explain these steps in greater detail. | |
140 | ||
141 | Several forms of output are available from the analysis. | |
142 | ||
143 | The @dfn{flat profile} shows how much time your program spent in each function, | |
144 | and how many times that function was called. If you simply want to know | |
145 | which functions burn most of the cycles, it is stated concisely here. | |
146 | @xref{Flat Profile}. | |
147 | ||
148 | The @dfn{call graph} shows, for each function, which functions called it, which | |
149 | other functions it called, and how many times. There is also an estimate | |
150 | of how much time was spent in the subroutines of each function. This can | |
151 | suggest places where you might try to eliminate function calls that use a | |
152 | lot of time. @xref{Call Graph}. | |
153 | ||
154 | The @dfn{annotated source} listing is a copy of the program's | |
155 | source code, labeled with the number of times each line of the | |
156 | program was executed. @xref{Annotated Source}. | |
157 | ||
158 | To better understand how profiling works, you may wish to read | |
159 | a description of its implementation. | |
160 | @xref{Implementation}. | |
161 | ||
162 | @node Compiling | |
163 | @chapter Compiling a Program for Profiling | |
164 | ||
165 | The first step in generating profile information for your program is | |
166 | to compile and link it with profiling enabled. | |
167 | ||
168 | To compile a source file for profiling, specify the @samp{-pg} option when | |
169 | you run the compiler. (This is in addition to the options you normally | |
170 | use.) | |
171 | ||
172 | To link the program for profiling, if you use a compiler such as @code{cc} | |
173 | to do the linking, simply specify @samp{-pg} in addition to your usual | |
174 | options. The same option, @samp{-pg}, alters either compilation or linking | |
175 | to do what is necessary for profiling. Here are examples: | |
176 | ||
177 | @example | |
178 | cc -g -c myprog.c utils.c -pg | |
179 | cc -o myprog myprog.o utils.o -pg | |
180 | @end example | |
181 | ||
182 | The @samp{-pg} option also works with a command that both compiles and links: | |
183 | ||
184 | @example | |
185 | cc -o myprog myprog.c utils.c -g -pg | |
186 | @end example | |
187 | ||
188 | If you run the linker @code{ld} directly instead of through a compiler | |
189 | such as @code{cc}, you may have to specify a profiling startup file | |
190 | @file{gcrt0.o} as the first input file instead of the usual startup | |
191 | file @file{crt0.o}. In addition, you would probably want to | |
192 | specify the profiling C library, @file{libc_p.a}, by writing | |
193 | @samp{-lc_p} instead of the usual @samp{-lc}. This is not absolutely | |
194 | necessary, but doing this gives you number-of-calls information for | |
195 | standard library functions such as @code{read} and @code{open}. For | |
196 | example: | |
197 | ||
198 | @example | |
199 | ld -o myprog /lib/gcrt0.o myprog.o utils.o -lc_p | |
200 | @end example | |
201 | ||
202 | If you compile only some of the modules of the program with @samp{-pg}, you | |
203 | can still profile the program, but you won't get complete information about | |
204 | the modules that were compiled without @samp{-pg}. The only information | |
205 | you get for the functions in those modules is the total time spent in them; | |
206 | there is no record of how many times they were called, or from where. This | |
207 | will not affect the flat profile (except that the @code{calls} field for | |
208 | the functions will be blank), but will greatly reduce the usefulness of the | |
209 | call graph. | |
210 | ||
211 | If you wish to perform line-by-line profiling, | |
212 | you will also need to specify the @samp{-g} option, | |
213 | instructing the compiler to insert debugging symbols into the program | |
214 | that match program addresses to source code lines. | |
215 | @xref{Line-by-line}. | |
216 | ||
217 | In addition to the @samp{-pg} and @samp{-g} options, | |
218 | you may also wish to specify the @samp{-a} option when compiling. | |
219 | This will instrument | |
220 | the program to perform basic-block counting. As the program runs, | |
221 | it will count how many times it executed each branch of each @samp{if} | |
222 | statement, each iteration of each @samp{do} loop, etc. This will | |
223 | enable @code{gprof} to construct an annotated source code | |
224 | listing showing how many times each line of code was executed. | |
225 | ||
226 | @node Executing | |
227 | @chapter Executing the Program | |
228 | ||
229 | Once the program is compiled for profiling, you must run it in order to | |
230 | generate the information that @code{gprof} needs. Simply run the program | |
231 | as usual, using the normal arguments, file names, etc. The program should | |
232 | run normally, producing the same output as usual. It will, however, run | |
233 | somewhat slower than normal because of the time spent collecting and the | |
234 | writing the profile data. | |
235 | ||
236 | The way you run the program---the arguments and input that you give | |
237 | it---may have a dramatic effect on what the profile information shows. The | |
238 | profile data will describe the parts of the program that were activated for | |
239 | the particular input you use. For example, if the first command you give | |
240 | to your program is to quit, the profile data will show the time used in | |
241 | initialization and in cleanup, but not much else. | |
242 | ||
243 | Your program will write the profile data into a file called @file{gmon.out} | |
244 | just before exiting. If there is already a file called @file{gmon.out}, | |
245 | its contents are overwritten. There is currently no way to tell the | |
246 | program to write the profile data under a different name, but you can rename | |
247 | the file afterward if you are concerned that it may be overwritten. | |
248 | ||
249 | In order to write the @file{gmon.out} file properly, your program must exit | |
250 | normally: by returning from @code{main} or by calling @code{exit}. Calling | |
251 | the low-level function @code{_exit} does not write the profile data, and | |
252 | neither does abnormal termination due to an unhandled signal. | |
253 | ||
254 | The @file{gmon.out} file is written in the program's @emph{current working | |
255 | directory} at the time it exits. This means that if your program calls | |
256 | @code{chdir}, the @file{gmon.out} file will be left in the last directory | |
257 | your program @code{chdir}'d to. If you don't have permission to write in | |
258 | this directory, the file is not written, and you will get an error message. | |
259 | ||
260 | Older versions of the @sc{gnu} profiling library may also write a file | |
261 | called @file{bb.out}. This file, if present, contains an human-readable | |
262 | listing of the basic-block execution counts. Unfortunately, the | |
263 | appearance of a human-readable @file{bb.out} means the basic-block | |
264 | counts didn't get written into @file{gmon.out}. | |
265 | The Perl script @code{bbconv.pl}, included with the @code{gprof} | |
266 | source distribution, will convert a @file{bb.out} file into | |
267 | a format readable by @code{gprof}. | |
268 | ||
269 | @node Invoking | |
270 | @chapter @code{gprof} Command Summary | |
271 | ||
272 | After you have a profile data file @file{gmon.out}, you can run @code{gprof} | |
273 | to interpret the information in it. The @code{gprof} program prints a | |
274 | flat profile and a call graph on standard output. Typically you would | |
275 | redirect the output of @code{gprof} into a file with @samp{>}. | |
276 | ||
277 | You run @code{gprof} like this: | |
278 | ||
279 | @smallexample | |
280 | gprof @var{options} [@var{executable-file} [@var{profile-data-files}@dots{}]] [> @var{outfile}] | |
281 | @end smallexample | |
282 | ||
283 | @noindent | |
284 | Here square-brackets indicate optional arguments. | |
285 | ||
286 | If you omit the executable file name, the file @file{a.out} is used. If | |
287 | you give no profile data file name, the file @file{gmon.out} is used. If | |
288 | any file is not in the proper format, or if the profile data file does not | |
289 | appear to belong to the executable file, an error message is printed. | |
290 | ||
291 | You can give more than one profile data file by entering all their names | |
292 | after the executable file name; then the statistics in all the data files | |
293 | are summed together. | |
294 | ||
295 | The order of these options does not matter. | |
296 | ||
297 | @menu | |
298 | * Output Options:: Controlling @code{gprof}'s output style | |
299 | * Analysis Options:: Controlling how @code{gprof} analyses its data | |
300 | * Miscellaneous Options:: | |
5af11cab | 301 | * Deprecated Options:: Options you no longer need to use, but which |
252b5132 RH |
302 | have been retained for compatibility |
303 | * Symspecs:: Specifying functions to include or exclude | |
304 | @end menu | |
305 | ||
306 | @node Output Options,Analysis Options,,Invoking | |
307 | @section Output Options | |
308 | ||
309 | These options specify which of several output formats | |
310 | @code{gprof} should produce. | |
311 | ||
312 | Many of these options take an optional @dfn{symspec} to specify | |
313 | functions to be included or excluded. These options can be | |
314 | specified multiple times, with different symspecs, to include | |
315 | or exclude sets of symbols. @xref{Symspecs}. | |
316 | ||
317 | Specifying any of these options overrides the default (@samp{-p -q}), | |
318 | which prints a flat profile and call graph analysis | |
319 | for all functions. | |
320 | ||
321 | @table @code | |
322 | ||
323 | @item -A[@var{symspec}] | |
324 | @itemx --annotated-source[=@var{symspec}] | |
325 | The @samp{-A} option causes @code{gprof} to print annotated source code. | |
326 | If @var{symspec} is specified, print output only for matching symbols. | |
327 | @xref{Annotated Source}. | |
328 | ||
329 | @item -b | |
330 | @itemx --brief | |
331 | If the @samp{-b} option is given, @code{gprof} doesn't print the | |
332 | verbose blurbs that try to explain the meaning of all of the fields in | |
333 | the tables. This is useful if you intend to print out the output, or | |
334 | are tired of seeing the blurbs. | |
335 | ||
336 | @item -C[@var{symspec}] | |
337 | @itemx --exec-counts[=@var{symspec}] | |
338 | The @samp{-C} option causes @code{gprof} to | |
339 | print a tally of functions and the number of times each was called. | |
340 | If @var{symspec} is specified, print tally only for matching symbols. | |
341 | ||
5af11cab | 342 | If the profile data file contains basic-block count records, specifying |
252b5132 RH |
343 | the @samp{-l} option, along with @samp{-C}, will cause basic-block |
344 | execution counts to be tallied and displayed. | |
345 | ||
346 | @item -i | |
347 | @itemx --file-info | |
348 | The @samp{-i} option causes @code{gprof} to display summary information | |
349 | about the profile data file(s) and then exit. The number of histogram, | |
350 | call graph, and basic-block count records is displayed. | |
351 | ||
352 | @item -I @var{dirs} | |
353 | @itemx --directory-path=@var{dirs} | |
354 | The @samp{-I} option specifies a list of search directories in | |
355 | which to find source files. Environment variable @var{GPROF_PATH} | |
5af11cab | 356 | can also be used to convey this information. |
252b5132 RH |
357 | Used mostly for annotated source output. |
358 | ||
359 | @item -J[@var{symspec}] | |
360 | @itemx --no-annotated-source[=@var{symspec}] | |
361 | The @samp{-J} option causes @code{gprof} not to | |
362 | print annotated source code. | |
363 | If @var{symspec} is specified, @code{gprof} prints annotated source, | |
364 | but excludes matching symbols. | |
365 | ||
366 | @item -L | |
367 | @itemx --print-path | |
368 | Normally, source filenames are printed with the path | |
369 | component suppressed. The @samp{-L} option causes @code{gprof} | |
370 | to print the full pathname of | |
371 | source filenames, which is determined | |
372 | from symbolic debugging information in the image file | |
373 | and is relative to the directory in which the compiler | |
374 | was invoked. | |
375 | ||
376 | @item -p[@var{symspec}] | |
377 | @itemx --flat-profile[=@var{symspec}] | |
378 | The @samp{-p} option causes @code{gprof} to print a flat profile. | |
379 | If @var{symspec} is specified, print flat profile only for matching symbols. | |
380 | @xref{Flat Profile}. | |
381 | ||
382 | @item -P[@var{symspec}] | |
383 | @itemx --no-flat-profile[=@var{symspec}] | |
384 | The @samp{-P} option causes @code{gprof} to suppress printing a flat profile. | |
385 | If @var{symspec} is specified, @code{gprof} prints a flat profile, | |
386 | but excludes matching symbols. | |
387 | ||
388 | @item -q[@var{symspec}] | |
389 | @itemx --graph[=@var{symspec}] | |
390 | The @samp{-q} option causes @code{gprof} to print the call graph analysis. | |
391 | If @var{symspec} is specified, print call graph only for matching symbols | |
392 | and their children. | |
393 | @xref{Call Graph}. | |
394 | ||
395 | @item -Q[@var{symspec}] | |
396 | @itemx --no-graph[=@var{symspec}] | |
397 | The @samp{-Q} option causes @code{gprof} to suppress printing the | |
398 | call graph. | |
399 | If @var{symspec} is specified, @code{gprof} prints a call graph, | |
400 | but excludes matching symbols. | |
401 | ||
402 | @item -y | |
403 | @itemx --separate-files | |
404 | This option affects annotated source output only. | |
5af11cab | 405 | Normally, @code{gprof} prints annotated source files |
252b5132 | 406 | to standard-output. If this option is specified, |
5af11cab AM |
407 | annotated source for a file named @file{path/@var{filename}} |
408 | is generated in the file @file{@var{filename}-ann}. If the underlying | |
409 | filesystem would truncate @file{@var{filename}-ann} so that it | |
410 | overwrites the original @file{@var{filename}}, @code{gprof} generates | |
411 | annotated source in the file @file{@var{filename}.ann} instead (if the | |
412 | original file name has an extension, that extension is @emph{replaced} | |
413 | with @file{.ann}). | |
252b5132 RH |
414 | |
415 | @item -Z[@var{symspec}] | |
416 | @itemx --no-exec-counts[=@var{symspec}] | |
417 | The @samp{-Z} option causes @code{gprof} not to | |
418 | print a tally of functions and the number of times each was called. | |
419 | If @var{symspec} is specified, print tally, but exclude matching symbols. | |
420 | ||
421 | @item --function-ordering | |
422 | The @samp{--function-ordering} option causes @code{gprof} to print a | |
423 | suggested function ordering for the program based on profiling data. | |
424 | This option suggests an ordering which may improve paging, tlb and | |
425 | cache behavior for the program on systems which support arbitrary | |
426 | ordering of functions in an executable. | |
427 | ||
428 | The exact details of how to force the linker to place functions | |
429 | in a particular order is system dependent and out of the scope of this | |
430 | manual. | |
431 | ||
432 | @item --file-ordering @var{map_file} | |
433 | The @samp{--file-ordering} option causes @code{gprof} to print a | |
434 | suggested .o link line ordering for the program based on profiling data. | |
435 | This option suggests an ordering which may improve paging, tlb and | |
436 | cache behavior for the program on systems which do not support arbitrary | |
437 | ordering of functions in an executable. | |
438 | ||
439 | Use of the @samp{-a} argument is highly recommended with this option. | |
440 | ||
441 | The @var{map_file} argument is a pathname to a file which provides | |
442 | function name to object file mappings. The format of the file is similar to | |
443 | the output of the program @code{nm}. | |
444 | ||
445 | @smallexample | |
446 | @group | |
447 | c-parse.o:00000000 T yyparse | |
448 | c-parse.o:00000004 C yyerrflag | |
449 | c-lang.o:00000000 T maybe_objc_method_name | |
450 | c-lang.o:00000000 T print_lang_statistics | |
451 | c-lang.o:00000000 T recognize_objc_keyword | |
452 | c-decl.o:00000000 T print_lang_identifier | |
453 | c-decl.o:00000000 T print_lang_type | |
454 | @dots{} | |
455 | ||
456 | @end group | |
457 | @end smallexample | |
458 | ||
5af11cab AM |
459 | To create a @var{map_file} with @sc{gnu} @code{nm}, type a command like |
460 | @kbd{nm --extern-only --defined-only -v --print-file-name program-name}. | |
252b5132 RH |
461 | |
462 | @item -T | |
463 | @itemx --traditional | |
464 | The @samp{-T} option causes @code{gprof} to print its output in | |
465 | ``traditional'' BSD style. | |
466 | ||
467 | @item -w @var{width} | |
468 | @itemx --width=@var{width} | |
469 | Sets width of output lines to @var{width}. | |
470 | Currently only used when printing the function index at the bottom | |
471 | of the call graph. | |
472 | ||
473 | @item -x | |
474 | @itemx --all-lines | |
475 | This option affects annotated source output only. | |
476 | By default, only the lines at the beginning of a basic-block | |
477 | are annotated. If this option is specified, every line in | |
478 | a basic-block is annotated by repeating the annotation for the | |
479 | first line. This behavior is similar to @code{tcov}'s @samp{-a}. | |
480 | ||
28c309a2 | 481 | @item --demangle[=@var{style}] |
252b5132 RH |
482 | @itemx --no-demangle |
483 | These options control whether C++ symbol names should be demangled when | |
484 | printing output. The default is to demangle symbols. The | |
28c309a2 NC |
485 | @code{--no-demangle} option may be used to turn off demangling. Different |
486 | compilers have different mangling styles. The optional demangling style | |
487 | argument can be used to choose an appropriate demangling style for your | |
488 | compiler. | |
252b5132 RH |
489 | @end table |
490 | ||
491 | @node Analysis Options,Miscellaneous Options,Output Options,Invoking | |
492 | @section Analysis Options | |
493 | ||
494 | @table @code | |
495 | ||
496 | @item -a | |
497 | @itemx --no-static | |
498 | The @samp{-a} option causes @code{gprof} to suppress the printing of | |
499 | statically declared (private) functions. (These are functions whose | |
500 | names are not listed as global, and which are not visible outside the | |
501 | file/function/block where they were defined.) Time spent in these | |
502 | functions, calls to/from them, etc, will all be attributed to the | |
503 | function that was loaded directly before it in the executable file. | |
504 | @c This is compatible with Unix @code{gprof}, but a bad idea. | |
505 | This option affects both the flat profile and the call graph. | |
506 | ||
507 | @item -c | |
508 | @itemx --static-call-graph | |
509 | The @samp{-c} option causes the call graph of the program to be | |
510 | augmented by a heuristic which examines the text space of the object | |
511 | file and identifies function calls in the binary machine code. | |
512 | Since normal call graph records are only generated when functions are | |
513 | entered, this option identifies children that could have been called, | |
514 | but never were. Calls to functions that were not compiled with | |
515 | profiling enabled are also identified, but only if symbol table | |
516 | entries are present for them. | |
517 | Calls to dynamic library routines are typically @emph{not} found | |
518 | by this option. | |
519 | Parents or children identified via this heuristic | |
520 | are indicated in the call graph with call counts of @samp{0}. | |
521 | ||
522 | @item -D | |
523 | @itemx --ignore-non-functions | |
524 | The @samp{-D} option causes @code{gprof} to ignore symbols which | |
525 | are not known to be functions. This option will give more accurate | |
526 | profile data on systems where it is supported (Solaris and HPUX for | |
527 | example). | |
528 | ||
529 | @item -k @var{from}/@var{to} | |
530 | The @samp{-k} option allows you to delete from the call graph any arcs from | |
531 | symbols matching symspec @var{from} to those matching symspec @var{to}. | |
532 | ||
533 | @item -l | |
534 | @itemx --line | |
535 | The @samp{-l} option enables line-by-line profiling, which causes | |
536 | histogram hits to be charged to individual source code lines, | |
537 | instead of functions. | |
538 | If the program was compiled with basic-block counting enabled, | |
539 | this option will also identify how many times each line of | |
540 | code was executed. | |
541 | While line-by-line profiling can help isolate where in a large function | |
542 | a program is spending its time, it also significantly increases | |
543 | the running time of @code{gprof}, and magnifies statistical | |
544 | inaccuracies. | |
545 | @xref{Sampling Error}. | |
546 | ||
547 | @item -m @var{num} | |
548 | @itemx --min-count=@var{num} | |
549 | This option affects execution count output only. | |
550 | Symbols that are executed less than @var{num} times are suppressed. | |
551 | ||
552 | @item -n[@var{symspec}] | |
553 | @itemx --time[=@var{symspec}] | |
554 | The @samp{-n} option causes @code{gprof}, in its call graph analysis, | |
555 | to only propagate times for symbols matching @var{symspec}. | |
556 | ||
557 | @item -N[@var{symspec}] | |
558 | @itemx --no-time[=@var{symspec}] | |
559 | The @samp{-n} option causes @code{gprof}, in its call graph analysis, | |
560 | not to propagate times for symbols matching @var{symspec}. | |
561 | ||
562 | @item -z | |
563 | @itemx --display-unused-functions | |
564 | If you give the @samp{-z} option, @code{gprof} will mention all | |
565 | functions in the flat profile, even those that were never called, and | |
566 | that had no time spent in them. This is useful in conjunction with the | |
567 | @samp{-c} option for discovering which routines were never called. | |
568 | ||
569 | @end table | |
570 | ||
5af11cab | 571 | @node Miscellaneous Options,Deprecated Options,Analysis Options,Invoking |
252b5132 RH |
572 | @section Miscellaneous Options |
573 | ||
574 | @table @code | |
575 | ||
576 | @item -d[@var{num}] | |
577 | @itemx --debug[=@var{num}] | |
578 | The @samp{-d @var{num}} option specifies debugging options. | |
579 | If @var{num} is not specified, enable all debugging. | |
580 | @xref{Debugging}. | |
581 | ||
582 | @item -O@var{name} | |
583 | @itemx --file-format=@var{name} | |
584 | Selects the format of the profile data files. Recognized formats are | |
585 | @samp{auto} (the default), @samp{bsd}, @samp{4.4bsd}, @samp{magic}, and | |
586 | @samp{prof} (not yet supported). | |
587 | ||
588 | @item -s | |
589 | @itemx --sum | |
590 | The @samp{-s} option causes @code{gprof} to summarize the information | |
591 | in the profile data files it read in, and write out a profile data | |
592 | file called @file{gmon.sum}, which contains all the information from | |
593 | the profile data files that @code{gprof} read in. The file @file{gmon.sum} | |
594 | may be one of the specified input files; the effect of this is to | |
595 | merge the data in the other input files into @file{gmon.sum}. | |
596 | ||
597 | Eventually you can run @code{gprof} again without @samp{-s} to analyze the | |
598 | cumulative data in the file @file{gmon.sum}. | |
599 | ||
600 | @item -v | |
601 | @itemx --version | |
602 | The @samp{-v} flag causes @code{gprof} to print the current version | |
603 | number, and then exit. | |
604 | ||
605 | @end table | |
606 | ||
5af11cab AM |
607 | @node Deprecated Options,Symspecs,Miscellaneous Options,Invoking |
608 | @section Deprecated Options | |
252b5132 RH |
609 | |
610 | @table @code | |
611 | ||
612 | These options have been replaced with newer versions that use symspecs. | |
613 | ||
614 | @item -e @var{function_name} | |
615 | The @samp{-e @var{function}} option tells @code{gprof} to not print | |
616 | information about the function @var{function_name} (and its | |
617 | children@dots{}) in the call graph. The function will still be listed | |
618 | as a child of any functions that call it, but its index number will be | |
619 | shown as @samp{[not printed]}. More than one @samp{-e} option may be | |
620 | given; only one @var{function_name} may be indicated with each @samp{-e} | |
621 | option. | |
622 | ||
623 | @item -E @var{function_name} | |
624 | The @code{-E @var{function}} option works like the @code{-e} option, but | |
625 | time spent in the function (and children who were not called from | |
626 | anywhere else), will not be used to compute the percentages-of-time for | |
627 | the call graph. More than one @samp{-E} option may be given; only one | |
628 | @var{function_name} may be indicated with each @samp{-E} option. | |
629 | ||
630 | @item -f @var{function_name} | |
631 | The @samp{-f @var{function}} option causes @code{gprof} to limit the | |
632 | call graph to the function @var{function_name} and its children (and | |
633 | their children@dots{}). More than one @samp{-f} option may be given; | |
634 | only one @var{function_name} may be indicated with each @samp{-f} | |
635 | option. | |
636 | ||
637 | @item -F @var{function_name} | |
638 | The @samp{-F @var{function}} option works like the @code{-f} option, but | |
639 | only time spent in the function and its children (and their | |
640 | children@dots{}) will be used to determine total-time and | |
641 | percentages-of-time for the call graph. More than one @samp{-F} option | |
642 | may be given; only one @var{function_name} may be indicated with each | |
643 | @samp{-F} option. The @samp{-F} option overrides the @samp{-E} option. | |
644 | ||
645 | @end table | |
646 | ||
647 | Note that only one function can be specified with each @code{-e}, | |
648 | @code{-E}, @code{-f} or @code{-F} option. To specify more than one | |
649 | function, use multiple options. For example, this command: | |
650 | ||
651 | @example | |
652 | gprof -e boring -f foo -f bar myprogram > gprof.output | |
653 | @end example | |
654 | ||
655 | @noindent | |
656 | lists in the call graph all functions that were reached from either | |
657 | @code{foo} or @code{bar} and were not reachable from @code{boring}. | |
658 | ||
5af11cab | 659 | @node Symspecs,,Deprecated Options,Invoking |
252b5132 RH |
660 | @section Symspecs |
661 | ||
662 | Many of the output options allow functions to be included or excluded | |
663 | using @dfn{symspecs} (symbol specifications), which observe the | |
664 | following syntax: | |
665 | ||
666 | @example | |
667 | filename_containing_a_dot | |
668 | | funcname_not_containing_a_dot | |
669 | | linenumber | |
670 | | ( [ any_filename ] `:' ( any_funcname | linenumber ) ) | |
671 | @end example | |
672 | ||
673 | Here are some sample symspecs: | |
674 | ||
675 | @table @samp | |
676 | @item main.c | |
677 | Selects everything in file @file{main.c}---the | |
5af11cab | 678 | dot in the string tells @code{gprof} to interpret |
252b5132 RH |
679 | the string as a filename, rather than as |
680 | a function name. To select a file whose | |
681 | name does not contain a dot, a trailing colon | |
682 | should be specified. For example, @samp{odd:} is | |
683 | interpreted as the file named @file{odd}. | |
684 | ||
685 | @item main | |
686 | Selects all functions named @samp{main}. | |
687 | ||
688 | Note that there may be multiple instances of the same function name | |
689 | because some of the definitions may be local (i.e., static). Unless a | |
690 | function name is unique in a program, you must use the colon notation | |
691 | explained below to specify a function from a specific source file. | |
692 | ||
a53f781e | 693 | Sometimes, function names contain dots. In such cases, it is necessary |
252b5132 RH |
694 | to add a leading colon to the name. For example, @samp{:.mul} selects |
695 | function @samp{.mul}. | |
696 | ||
5af11cab AM |
697 | In some object file formats, symbols have a leading underscore. |
698 | @code{gprof} will normally not print these underscores. When you name a | |
699 | symbol in a symspec, you should type it exactly as @code{gprof} prints | |
700 | it in its output. For example, if the compiler produces a symbol | |
701 | @samp{_main} from your @code{main} function, @code{gprof} still prints | |
702 | it as @samp{main} in its output, so you should use @samp{main} in | |
703 | symspecs. | |
252b5132 RH |
704 | |
705 | @item main.c:main | |
706 | Selects function @samp{main} in file @file{main.c}. | |
707 | ||
708 | @item main.c:134 | |
709 | Selects line 134 in file @file{main.c}. | |
710 | @end table | |
711 | ||
712 | @node Output | |
713 | @chapter Interpreting @code{gprof}'s Output | |
714 | ||
715 | @code{gprof} can produce several different output styles, the | |
716 | most important of which are described below. The simplest output | |
717 | styles (file information, execution count, and function and file ordering) | |
718 | are not described here, but are documented with the respective options | |
719 | that trigger them. | |
720 | @xref{Output Options}. | |
721 | ||
722 | @menu | |
723 | * Flat Profile:: The flat profile shows how much time was spent | |
724 | executing directly in each function. | |
725 | * Call Graph:: The call graph shows which functions called which | |
726 | others, and how much time each function used | |
727 | when its subroutine calls are included. | |
728 | * Line-by-line:: @code{gprof} can analyze individual source code lines | |
729 | * Annotated Source:: The annotated source listing displays source code | |
730 | labeled with execution counts | |
731 | @end menu | |
732 | ||
733 | ||
734 | @node Flat Profile,Call Graph,,Output | |
735 | @section The Flat Profile | |
736 | @cindex flat profile | |
737 | ||
738 | The @dfn{flat profile} shows the total amount of time your program | |
739 | spent executing each function. Unless the @samp{-z} option is given, | |
740 | functions with no apparent time spent in them, and no apparent calls | |
741 | to them, are not mentioned. Note that if a function was not compiled | |
742 | for profiling, and didn't run long enough to show up on the program | |
743 | counter histogram, it will be indistinguishable from a function that | |
744 | was never called. | |
745 | ||
746 | This is part of a flat profile for a small program: | |
747 | ||
748 | @smallexample | |
749 | @group | |
750 | Flat profile: | |
751 | ||
752 | Each sample counts as 0.01 seconds. | |
753 | % cumulative self self total | |
754 | time seconds seconds calls ms/call ms/call name | |
755 | 33.34 0.02 0.02 7208 0.00 0.00 open | |
756 | 16.67 0.03 0.01 244 0.04 0.12 offtime | |
757 | 16.67 0.04 0.01 8 1.25 1.25 memccpy | |
758 | 16.67 0.05 0.01 7 1.43 1.43 write | |
759 | 16.67 0.06 0.01 mcount | |
760 | 0.00 0.06 0.00 236 0.00 0.00 tzset | |
761 | 0.00 0.06 0.00 192 0.00 0.00 tolower | |
762 | 0.00 0.06 0.00 47 0.00 0.00 strlen | |
763 | 0.00 0.06 0.00 45 0.00 0.00 strchr | |
764 | 0.00 0.06 0.00 1 0.00 50.00 main | |
765 | 0.00 0.06 0.00 1 0.00 0.00 memcpy | |
766 | 0.00 0.06 0.00 1 0.00 10.11 print | |
767 | 0.00 0.06 0.00 1 0.00 0.00 profil | |
768 | 0.00 0.06 0.00 1 0.00 50.00 report | |
769 | @dots{} | |
770 | @end group | |
771 | @end smallexample | |
772 | ||
773 | @noindent | |
774 | The functions are sorted by first by decreasing run-time spent in them, | |
775 | then by decreasing number of calls, then alphabetically by name. The | |
776 | functions @samp{mcount} and @samp{profil} are part of the profiling | |
5af11cab | 777 | apparatus and appear in every flat profile; their time gives a measure of |
252b5132 RH |
778 | the amount of overhead due to profiling. |
779 | ||
780 | Just before the column headers, a statement appears indicating | |
781 | how much time each sample counted as. | |
782 | This @dfn{sampling period} estimates the margin of error in each of the time | |
783 | figures. A time figure that is not much larger than this is not | |
784 | reliable. In this example, each sample counted as 0.01 seconds, | |
785 | suggesting a 100 Hz sampling rate. | |
786 | The program's total execution time was 0.06 | |
787 | seconds, as indicated by the @samp{cumulative seconds} field. Since | |
788 | each sample counted for 0.01 seconds, this means only six samples | |
5af11cab | 789 | were taken during the run. Two of the samples occurred while the |
252b5132 RH |
790 | program was in the @samp{open} function, as indicated by the |
791 | @samp{self seconds} field. Each of the other four samples | |
5af11cab | 792 | occurred one each in @samp{offtime}, @samp{memccpy}, @samp{write}, |
252b5132 RH |
793 | and @samp{mcount}. |
794 | Since only six samples were taken, none of these values can | |
795 | be regarded as particularly reliable. | |
796 | In another run, | |
797 | the @samp{self seconds} field for | |
798 | @samp{mcount} might well be @samp{0.00} or @samp{0.02}. | |
799 | @xref{Sampling Error}, for a complete discussion. | |
800 | ||
801 | The remaining functions in the listing (those whose | |
802 | @samp{self seconds} field is @samp{0.00}) didn't appear | |
803 | in the histogram samples at all. However, the call graph | |
804 | indicated that they were called, so therefore they are listed, | |
805 | sorted in decreasing order by the @samp{calls} field. | |
806 | Clearly some time was spent executing these functions, | |
807 | but the paucity of histogram samples prevents any | |
808 | determination of how much time each took. | |
809 | ||
810 | Here is what the fields in each line mean: | |
811 | ||
812 | @table @code | |
813 | @item % time | |
814 | This is the percentage of the total execution time your program spent | |
815 | in this function. These should all add up to 100%. | |
816 | ||
817 | @item cumulative seconds | |
818 | This is the cumulative total number of seconds the computer spent | |
819 | executing this functions, plus the time spent in all the functions | |
820 | above this one in this table. | |
821 | ||
822 | @item self seconds | |
823 | This is the number of seconds accounted for by this function alone. | |
824 | The flat profile listing is sorted first by this number. | |
825 | ||
826 | @item calls | |
827 | This is the total number of times the function was called. If the | |
828 | function was never called, or the number of times it was called cannot | |
829 | be determined (probably because the function was not compiled with | |
830 | profiling enabled), the @dfn{calls} field is blank. | |
831 | ||
832 | @item self ms/call | |
833 | This represents the average number of milliseconds spent in this | |
834 | function per call, if this function is profiled. Otherwise, this field | |
835 | is blank for this function. | |
836 | ||
837 | @item total ms/call | |
838 | This represents the average number of milliseconds spent in this | |
839 | function and its descendants per call, if this function is profiled. | |
840 | Otherwise, this field is blank for this function. | |
841 | This is the only field in the flat profile that uses call graph analysis. | |
842 | ||
843 | @item name | |
844 | This is the name of the function. The flat profile is sorted by this | |
845 | field alphabetically after the @dfn{self seconds} and @dfn{calls} | |
846 | fields are sorted. | |
847 | @end table | |
848 | ||
849 | @node Call Graph,Line-by-line,Flat Profile,Output | |
850 | @section The Call Graph | |
851 | @cindex call graph | |
852 | ||
853 | The @dfn{call graph} shows how much time was spent in each function | |
854 | and its children. From this information, you can find functions that, | |
855 | while they themselves may not have used much time, called other | |
856 | functions that did use unusual amounts of time. | |
857 | ||
858 | Here is a sample call from a small program. This call came from the | |
859 | same @code{gprof} run as the flat profile example in the previous | |
860 | chapter. | |
861 | ||
862 | @smallexample | |
863 | @group | |
864 | granularity: each sample hit covers 2 byte(s) for 20.00% of 0.05 seconds | |
865 | ||
866 | index % time self children called name | |
867 | <spontaneous> | |
868 | [1] 100.0 0.00 0.05 start [1] | |
869 | 0.00 0.05 1/1 main [2] | |
870 | 0.00 0.00 1/2 on_exit [28] | |
871 | 0.00 0.00 1/1 exit [59] | |
872 | ----------------------------------------------- | |
873 | 0.00 0.05 1/1 start [1] | |
874 | [2] 100.0 0.00 0.05 1 main [2] | |
875 | 0.00 0.05 1/1 report [3] | |
876 | ----------------------------------------------- | |
877 | 0.00 0.05 1/1 main [2] | |
878 | [3] 100.0 0.00 0.05 1 report [3] | |
879 | 0.00 0.03 8/8 timelocal [6] | |
880 | 0.00 0.01 1/1 print [9] | |
881 | 0.00 0.01 9/9 fgets [12] | |
882 | 0.00 0.00 12/34 strncmp <cycle 1> [40] | |
883 | 0.00 0.00 8/8 lookup [20] | |
884 | 0.00 0.00 1/1 fopen [21] | |
885 | 0.00 0.00 8/8 chewtime [24] | |
886 | 0.00 0.00 8/16 skipspace [44] | |
887 | ----------------------------------------------- | |
888 | [4] 59.8 0.01 0.02 8+472 <cycle 2 as a whole> [4] | |
889 | 0.01 0.02 244+260 offtime <cycle 2> [7] | |
890 | 0.00 0.00 236+1 tzset <cycle 2> [26] | |
891 | ----------------------------------------------- | |
892 | @end group | |
893 | @end smallexample | |
894 | ||
895 | The lines full of dashes divide this table into @dfn{entries}, one for each | |
896 | function. Each entry has one or more lines. | |
897 | ||
898 | In each entry, the primary line is the one that starts with an index number | |
899 | in square brackets. The end of this line says which function the entry is | |
900 | for. The preceding lines in the entry describe the callers of this | |
901 | function and the following lines describe its subroutines (also called | |
902 | @dfn{children} when we speak of the call graph). | |
903 | ||
904 | The entries are sorted by time spent in the function and its subroutines. | |
905 | ||
906 | The internal profiling function @code{mcount} (@pxref{Flat Profile}) | |
907 | is never mentioned in the call graph. | |
908 | ||
909 | @menu | |
910 | * Primary:: Details of the primary line's contents. | |
911 | * Callers:: Details of caller-lines' contents. | |
912 | * Subroutines:: Details of subroutine-lines' contents. | |
913 | * Cycles:: When there are cycles of recursion, | |
914 | such as @code{a} calls @code{b} calls @code{a}@dots{} | |
915 | @end menu | |
916 | ||
917 | @node Primary | |
918 | @subsection The Primary Line | |
919 | ||
920 | The @dfn{primary line} in a call graph entry is the line that | |
921 | describes the function which the entry is about and gives the overall | |
922 | statistics for this function. | |
923 | ||
924 | For reference, we repeat the primary line from the entry for function | |
925 | @code{report} in our main example, together with the heading line that | |
926 | shows the names of the fields: | |
927 | ||
928 | @smallexample | |
929 | @group | |
930 | index % time self children called name | |
931 | @dots{} | |
932 | [3] 100.0 0.00 0.05 1 report [3] | |
933 | @end group | |
934 | @end smallexample | |
935 | ||
936 | Here is what the fields in the primary line mean: | |
937 | ||
938 | @table @code | |
939 | @item index | |
940 | Entries are numbered with consecutive integers. Each function | |
941 | therefore has an index number, which appears at the beginning of its | |
942 | primary line. | |
943 | ||
944 | Each cross-reference to a function, as a caller or subroutine of | |
945 | another, gives its index number as well as its name. The index number | |
946 | guides you if you wish to look for the entry for that function. | |
947 | ||
948 | @item % time | |
949 | This is the percentage of the total time that was spent in this | |
950 | function, including time spent in subroutines called from this | |
951 | function. | |
952 | ||
953 | The time spent in this function is counted again for the callers of | |
954 | this function. Therefore, adding up these percentages is meaningless. | |
955 | ||
956 | @item self | |
957 | This is the total amount of time spent in this function. This | |
958 | should be identical to the number printed in the @code{seconds} field | |
959 | for this function in the flat profile. | |
960 | ||
961 | @item children | |
962 | This is the total amount of time spent in the subroutine calls made by | |
963 | this function. This should be equal to the sum of all the @code{self} | |
964 | and @code{children} entries of the children listed directly below this | |
965 | function. | |
966 | ||
967 | @item called | |
968 | This is the number of times the function was called. | |
969 | ||
970 | If the function called itself recursively, there are two numbers, | |
971 | separated by a @samp{+}. The first number counts non-recursive calls, | |
972 | and the second counts recursive calls. | |
973 | ||
974 | In the example above, the function @code{report} was called once from | |
975 | @code{main}. | |
976 | ||
977 | @item name | |
978 | This is the name of the current function. The index number is | |
979 | repeated after it. | |
980 | ||
981 | If the function is part of a cycle of recursion, the cycle number is | |
982 | printed between the function's name and the index number | |
983 | (@pxref{Cycles}). For example, if function @code{gnurr} is part of | |
984 | cycle number one, and has index number twelve, its primary line would | |
985 | be end like this: | |
986 | ||
987 | @example | |
988 | gnurr <cycle 1> [12] | |
989 | @end example | |
990 | @end table | |
991 | ||
992 | @node Callers, Subroutines, Primary, Call Graph | |
993 | @subsection Lines for a Function's Callers | |
994 | ||
995 | A function's entry has a line for each function it was called by. | |
996 | These lines' fields correspond to the fields of the primary line, but | |
997 | their meanings are different because of the difference in context. | |
998 | ||
999 | For reference, we repeat two lines from the entry for the function | |
1000 | @code{report}, the primary line and one caller-line preceding it, together | |
1001 | with the heading line that shows the names of the fields: | |
1002 | ||
1003 | @smallexample | |
1004 | index % time self children called name | |
1005 | @dots{} | |
1006 | 0.00 0.05 1/1 main [2] | |
1007 | [3] 100.0 0.00 0.05 1 report [3] | |
1008 | @end smallexample | |
1009 | ||
1010 | Here are the meanings of the fields in the caller-line for @code{report} | |
1011 | called from @code{main}: | |
1012 | ||
1013 | @table @code | |
1014 | @item self | |
1015 | An estimate of the amount of time spent in @code{report} itself when it was | |
1016 | called from @code{main}. | |
1017 | ||
1018 | @item children | |
1019 | An estimate of the amount of time spent in subroutines of @code{report} | |
1020 | when @code{report} was called from @code{main}. | |
1021 | ||
1022 | The sum of the @code{self} and @code{children} fields is an estimate | |
1023 | of the amount of time spent within calls to @code{report} from @code{main}. | |
1024 | ||
1025 | @item called | |
1026 | Two numbers: the number of times @code{report} was called from @code{main}, | |
5af11cab | 1027 | followed by the total number of non-recursive calls to @code{report} from |
252b5132 RH |
1028 | all its callers. |
1029 | ||
1030 | @item name and index number | |
1031 | The name of the caller of @code{report} to which this line applies, | |
1032 | followed by the caller's index number. | |
1033 | ||
1034 | Not all functions have entries in the call graph; some | |
1035 | options to @code{gprof} request the omission of certain functions. | |
1036 | When a caller has no entry of its own, it still has caller-lines | |
1037 | in the entries of the functions it calls. | |
1038 | ||
1039 | If the caller is part of a recursion cycle, the cycle number is | |
1040 | printed between the name and the index number. | |
1041 | @end table | |
1042 | ||
1043 | If the identity of the callers of a function cannot be determined, a | |
1044 | dummy caller-line is printed which has @samp{<spontaneous>} as the | |
1045 | ``caller's name'' and all other fields blank. This can happen for | |
1046 | signal handlers. | |
1047 | @c What if some calls have determinable callers' names but not all? | |
1048 | @c FIXME - still relevant? | |
1049 | ||
1050 | @node Subroutines, Cycles, Callers, Call Graph | |
1051 | @subsection Lines for a Function's Subroutines | |
1052 | ||
1053 | A function's entry has a line for each of its subroutines---in other | |
1054 | words, a line for each other function that it called. These lines' | |
1055 | fields correspond to the fields of the primary line, but their meanings | |
1056 | are different because of the difference in context. | |
1057 | ||
1058 | For reference, we repeat two lines from the entry for the function | |
1059 | @code{main}, the primary line and a line for a subroutine, together | |
1060 | with the heading line that shows the names of the fields: | |
1061 | ||
1062 | @smallexample | |
1063 | index % time self children called name | |
1064 | @dots{} | |
1065 | [2] 100.0 0.00 0.05 1 main [2] | |
1066 | 0.00 0.05 1/1 report [3] | |
1067 | @end smallexample | |
1068 | ||
1069 | Here are the meanings of the fields in the subroutine-line for @code{main} | |
1070 | calling @code{report}: | |
1071 | ||
1072 | @table @code | |
1073 | @item self | |
1074 | An estimate of the amount of time spent directly within @code{report} | |
1075 | when @code{report} was called from @code{main}. | |
1076 | ||
1077 | @item children | |
1078 | An estimate of the amount of time spent in subroutines of @code{report} | |
1079 | when @code{report} was called from @code{main}. | |
1080 | ||
1081 | The sum of the @code{self} and @code{children} fields is an estimate | |
1082 | of the total time spent in calls to @code{report} from @code{main}. | |
1083 | ||
1084 | @item called | |
1085 | Two numbers, the number of calls to @code{report} from @code{main} | |
5af11cab | 1086 | followed by the total number of non-recursive calls to @code{report}. |
252b5132 RH |
1087 | This ratio is used to determine how much of @code{report}'s @code{self} |
1088 | and @code{children} time gets credited to @code{main}. | |
1089 | @xref{Assumptions}. | |
1090 | ||
1091 | @item name | |
1092 | The name of the subroutine of @code{main} to which this line applies, | |
1093 | followed by the subroutine's index number. | |
1094 | ||
1095 | If the caller is part of a recursion cycle, the cycle number is | |
1096 | printed between the name and the index number. | |
1097 | @end table | |
1098 | ||
1099 | @node Cycles,, Subroutines, Call Graph | |
1100 | @subsection How Mutually Recursive Functions Are Described | |
1101 | @cindex cycle | |
1102 | @cindex recursion cycle | |
1103 | ||
1104 | The graph may be complicated by the presence of @dfn{cycles of | |
1105 | recursion} in the call graph. A cycle exists if a function calls | |
1106 | another function that (directly or indirectly) calls (or appears to | |
1107 | call) the original function. For example: if @code{a} calls @code{b}, | |
1108 | and @code{b} calls @code{a}, then @code{a} and @code{b} form a cycle. | |
1109 | ||
1110 | Whenever there are call paths both ways between a pair of functions, they | |
1111 | belong to the same cycle. If @code{a} and @code{b} call each other and | |
1112 | @code{b} and @code{c} call each other, all three make one cycle. Note that | |
1113 | even if @code{b} only calls @code{a} if it was not called from @code{a}, | |
1114 | @code{gprof} cannot determine this, so @code{a} and @code{b} are still | |
1115 | considered a cycle. | |
1116 | ||
1117 | The cycles are numbered with consecutive integers. When a function | |
1118 | belongs to a cycle, each time the function name appears in the call graph | |
1119 | it is followed by @samp{<cycle @var{number}>}. | |
1120 | ||
1121 | The reason cycles matter is that they make the time values in the call | |
1122 | graph paradoxical. The ``time spent in children'' of @code{a} should | |
1123 | include the time spent in its subroutine @code{b} and in @code{b}'s | |
1124 | subroutines---but one of @code{b}'s subroutines is @code{a}! How much of | |
1125 | @code{a}'s time should be included in the children of @code{a}, when | |
1126 | @code{a} is indirectly recursive? | |
1127 | ||
1128 | The way @code{gprof} resolves this paradox is by creating a single entry | |
1129 | for the cycle as a whole. The primary line of this entry describes the | |
1130 | total time spent directly in the functions of the cycle. The | |
1131 | ``subroutines'' of the cycle are the individual functions of the cycle, and | |
1132 | all other functions that were called directly by them. The ``callers'' of | |
1133 | the cycle are the functions, outside the cycle, that called functions in | |
1134 | the cycle. | |
1135 | ||
1136 | Here is an example portion of a call graph which shows a cycle containing | |
1137 | functions @code{a} and @code{b}. The cycle was entered by a call to | |
1138 | @code{a} from @code{main}; both @code{a} and @code{b} called @code{c}. | |
1139 | ||
1140 | @smallexample | |
1141 | index % time self children called name | |
1142 | ---------------------------------------- | |
1143 | 1.77 0 1/1 main [2] | |
1144 | [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3] | |
1145 | 1.02 0 3 b <cycle 1> [4] | |
1146 | 0.75 0 2 a <cycle 1> [5] | |
1147 | ---------------------------------------- | |
1148 | 3 a <cycle 1> [5] | |
1149 | [4] 52.85 1.02 0 0 b <cycle 1> [4] | |
1150 | 2 a <cycle 1> [5] | |
1151 | 0 0 3/6 c [6] | |
1152 | ---------------------------------------- | |
1153 | 1.77 0 1/1 main [2] | |
1154 | 2 b <cycle 1> [4] | |
1155 | [5] 38.86 0.75 0 1 a <cycle 1> [5] | |
1156 | 3 b <cycle 1> [4] | |
1157 | 0 0 3/6 c [6] | |
1158 | ---------------------------------------- | |
1159 | @end smallexample | |
1160 | ||
1161 | @noindent | |
1162 | (The entire call graph for this program contains in addition an entry for | |
1163 | @code{main}, which calls @code{a}, and an entry for @code{c}, with callers | |
1164 | @code{a} and @code{b}.) | |
1165 | ||
1166 | @smallexample | |
1167 | index % time self children called name | |
1168 | <spontaneous> | |
1169 | [1] 100.00 0 1.93 0 start [1] | |
1170 | 0.16 1.77 1/1 main [2] | |
1171 | ---------------------------------------- | |
1172 | 0.16 1.77 1/1 start [1] | |
1173 | [2] 100.00 0.16 1.77 1 main [2] | |
1174 | 1.77 0 1/1 a <cycle 1> [5] | |
1175 | ---------------------------------------- | |
1176 | 1.77 0 1/1 main [2] | |
1177 | [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3] | |
1178 | 1.02 0 3 b <cycle 1> [4] | |
1179 | 0.75 0 2 a <cycle 1> [5] | |
1180 | 0 0 6/6 c [6] | |
1181 | ---------------------------------------- | |
1182 | 3 a <cycle 1> [5] | |
1183 | [4] 52.85 1.02 0 0 b <cycle 1> [4] | |
1184 | 2 a <cycle 1> [5] | |
1185 | 0 0 3/6 c [6] | |
1186 | ---------------------------------------- | |
1187 | 1.77 0 1/1 main [2] | |
1188 | 2 b <cycle 1> [4] | |
1189 | [5] 38.86 0.75 0 1 a <cycle 1> [5] | |
1190 | 3 b <cycle 1> [4] | |
1191 | 0 0 3/6 c [6] | |
1192 | ---------------------------------------- | |
1193 | 0 0 3/6 b <cycle 1> [4] | |
1194 | 0 0 3/6 a <cycle 1> [5] | |
1195 | [6] 0.00 0 0 6 c [6] | |
1196 | ---------------------------------------- | |
1197 | @end smallexample | |
1198 | ||
1199 | The @code{self} field of the cycle's primary line is the total time | |
1200 | spent in all the functions of the cycle. It equals the sum of the | |
1201 | @code{self} fields for the individual functions in the cycle, found | |
1202 | in the entry in the subroutine lines for these functions. | |
1203 | ||
1204 | The @code{children} fields of the cycle's primary line and subroutine lines | |
1205 | count only subroutines outside the cycle. Even though @code{a} calls | |
1206 | @code{b}, the time spent in those calls to @code{b} is not counted in | |
1207 | @code{a}'s @code{children} time. Thus, we do not encounter the problem of | |
1208 | what to do when the time in those calls to @code{b} includes indirect | |
1209 | recursive calls back to @code{a}. | |
1210 | ||
1211 | The @code{children} field of a caller-line in the cycle's entry estimates | |
1212 | the amount of time spent @emph{in the whole cycle}, and its other | |
1213 | subroutines, on the times when that caller called a function in the cycle. | |
1214 | ||
1215 | The @code{calls} field in the primary line for the cycle has two numbers: | |
1216 | first, the number of times functions in the cycle were called by functions | |
1217 | outside the cycle; second, the number of times they were called by | |
1218 | functions in the cycle (including times when a function in the cycle calls | |
5af11cab | 1219 | itself). This is a generalization of the usual split into non-recursive and |
252b5132 RH |
1220 | recursive calls. |
1221 | ||
1222 | The @code{calls} field of a subroutine-line for a cycle member in the | |
1223 | cycle's entry says how many time that function was called from functions in | |
1224 | the cycle. The total of all these is the second number in the primary line's | |
1225 | @code{calls} field. | |
1226 | ||
1227 | In the individual entry for a function in a cycle, the other functions in | |
1228 | the same cycle can appear as subroutines and as callers. These lines show | |
1229 | how many times each function in the cycle called or was called from each other | |
1230 | function in the cycle. The @code{self} and @code{children} fields in these | |
1231 | lines are blank because of the difficulty of defining meanings for them | |
1232 | when recursion is going on. | |
1233 | ||
1234 | @node Line-by-line,Annotated Source,Call Graph,Output | |
1235 | @section Line-by-line Profiling | |
1236 | ||
1237 | @code{gprof}'s @samp{-l} option causes the program to perform | |
1238 | @dfn{line-by-line} profiling. In this mode, histogram | |
1239 | samples are assigned not to functions, but to individual | |
1240 | lines of source code. The program usually must be compiled | |
1241 | with a @samp{-g} option, in addition to @samp{-pg}, in order | |
1242 | to generate debugging symbols for tracking source code lines. | |
1243 | ||
1244 | The flat profile is the most useful output table | |
1245 | in line-by-line mode. | |
1246 | The call graph isn't as useful as normal, since | |
1247 | the current version of @code{gprof} does not propagate | |
1248 | call graph arcs from source code lines to the enclosing function. | |
1249 | The call graph does, however, show each line of code | |
1250 | that called each function, along with a count. | |
1251 | ||
1252 | Here is a section of @code{gprof}'s output, without line-by-line profiling. | |
1253 | Note that @code{ct_init} accounted for four histogram hits, and | |
1254 | 13327 calls to @code{init_block}. | |
1255 | ||
1256 | @smallexample | |
1257 | Flat profile: | |
1258 | ||
1259 | Each sample counts as 0.01 seconds. | |
1260 | % cumulative self self total | |
1261 | time seconds seconds calls us/call us/call name | |
1262 | 30.77 0.13 0.04 6335 6.31 6.31 ct_init | |
1263 | ||
1264 | ||
1265 | Call graph (explanation follows) | |
1266 | ||
1267 | ||
1268 | granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds | |
1269 | ||
1270 | index % time self children called name | |
1271 | ||
1272 | 0.00 0.00 1/13496 name_too_long | |
1273 | 0.00 0.00 40/13496 deflate | |
1274 | 0.00 0.00 128/13496 deflate_fast | |
1275 | 0.00 0.00 13327/13496 ct_init | |
1276 | [7] 0.0 0.00 0.00 13496 init_block | |
1277 | ||
1278 | @end smallexample | |
1279 | ||
1280 | Now let's look at some of @code{gprof}'s output from the same program run, | |
1281 | this time with line-by-line profiling enabled. Note that @code{ct_init}'s | |
1282 | four histogram hits are broken down into four lines of source code - one hit | |
5af11cab | 1283 | occurred on each of lines 349, 351, 382 and 385. In the call graph, |
252b5132 RH |
1284 | note how |
1285 | @code{ct_init}'s 13327 calls to @code{init_block} are broken down | |
1286 | into one call from line 396, 3071 calls from line 384, 3730 calls | |
1287 | from line 385, and 6525 calls from 387. | |
1288 | ||
1289 | @smallexample | |
1290 | Flat profile: | |
1291 | ||
1292 | Each sample counts as 0.01 seconds. | |
1293 | % cumulative self | |
1294 | time seconds seconds calls name | |
1295 | 7.69 0.10 0.01 ct_init (trees.c:349) | |
1296 | 7.69 0.11 0.01 ct_init (trees.c:351) | |
1297 | 7.69 0.12 0.01 ct_init (trees.c:382) | |
1298 | 7.69 0.13 0.01 ct_init (trees.c:385) | |
1299 | ||
1300 | ||
1301 | Call graph (explanation follows) | |
1302 | ||
1303 | ||
1304 | granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds | |
1305 | ||
1306 | % time self children called name | |
1307 | ||
1308 | 0.00 0.00 1/13496 name_too_long (gzip.c:1440) | |
1309 | 0.00 0.00 1/13496 deflate (deflate.c:763) | |
1310 | 0.00 0.00 1/13496 ct_init (trees.c:396) | |
1311 | 0.00 0.00 2/13496 deflate (deflate.c:727) | |
1312 | 0.00 0.00 4/13496 deflate (deflate.c:686) | |
1313 | 0.00 0.00 5/13496 deflate (deflate.c:675) | |
1314 | 0.00 0.00 12/13496 deflate (deflate.c:679) | |
1315 | 0.00 0.00 16/13496 deflate (deflate.c:730) | |
1316 | 0.00 0.00 128/13496 deflate_fast (deflate.c:654) | |
1317 | 0.00 0.00 3071/13496 ct_init (trees.c:384) | |
1318 | 0.00 0.00 3730/13496 ct_init (trees.c:385) | |
1319 | 0.00 0.00 6525/13496 ct_init (trees.c:387) | |
1320 | [6] 0.0 0.00 0.00 13496 init_block (trees.c:408) | |
1321 | ||
1322 | @end smallexample | |
1323 | ||
1324 | ||
1325 | @node Annotated Source,,Line-by-line,Output | |
1326 | @section The Annotated Source Listing | |
1327 | ||
1328 | @code{gprof}'s @samp{-A} option triggers an annotated source listing, | |
1329 | which lists the program's source code, each function labeled with the | |
1330 | number of times it was called. You may also need to specify the | |
1331 | @samp{-I} option, if @code{gprof} can't find the source code files. | |
1332 | ||
1333 | Compiling with @samp{gcc @dots{} -g -pg -a} augments your program | |
1334 | with basic-block counting code, in addition to function counting code. | |
1335 | This enables @code{gprof} to determine how many times each line | |
5af11cab | 1336 | of code was executed. |
252b5132 RH |
1337 | For example, consider the following function, taken from gzip, |
1338 | with line numbers added: | |
1339 | ||
1340 | @smallexample | |
1341 | 1 ulg updcrc(s, n) | |
1342 | 2 uch *s; | |
1343 | 3 unsigned n; | |
1344 | 4 @{ | |
1345 | 5 register ulg c; | |
1346 | 6 | |
1347 | 7 static ulg crc = (ulg)0xffffffffL; | |
1348 | 8 | |
1349 | 9 if (s == NULL) @{ | |
1350 | 10 c = 0xffffffffL; | |
1351 | 11 @} else @{ | |
1352 | 12 c = crc; | |
1353 | 13 if (n) do @{ | |
1354 | 14 c = crc_32_tab[...]; | |
1355 | 15 @} while (--n); | |
1356 | 16 @} | |
1357 | 17 crc = c; | |
1358 | 18 return c ^ 0xffffffffL; | |
1359 | 19 @} | |
1360 | ||
1361 | @end smallexample | |
1362 | ||
1363 | @code{updcrc} has at least five basic-blocks. | |
1364 | One is the function itself. The | |
1365 | @code{if} statement on line 9 generates two more basic-blocks, one | |
1366 | for each branch of the @code{if}. A fourth basic-block results from | |
1367 | the @code{if} on line 13, and the contents of the @code{do} loop form | |
1368 | the fifth basic-block. The compiler may also generate additional | |
1369 | basic-blocks to handle various special cases. | |
1370 | ||
1371 | A program augmented for basic-block counting can be analyzed with | |
5af11cab | 1372 | @samp{gprof -l -A}. I also suggest use of the @samp{-x} option, |
252b5132 RH |
1373 | which ensures that each line of code is labeled at least once. |
1374 | Here is @code{updcrc}'s | |
1375 | annotated source listing for a sample @code{gzip} run: | |
1376 | ||
1377 | @smallexample | |
1378 | ulg updcrc(s, n) | |
1379 | uch *s; | |
1380 | unsigned n; | |
1381 | 2 ->@{ | |
1382 | register ulg c; | |
1383 | ||
1384 | static ulg crc = (ulg)0xffffffffL; | |
1385 | ||
1386 | 2 -> if (s == NULL) @{ | |
1387 | 1 -> c = 0xffffffffL; | |
1388 | 1 -> @} else @{ | |
1389 | 1 -> c = crc; | |
1390 | 1 -> if (n) do @{ | |
1391 | 26312 -> c = crc_32_tab[...]; | |
1392 | 26312,1,26311 -> @} while (--n); | |
1393 | @} | |
1394 | 2 -> crc = c; | |
1395 | 2 -> return c ^ 0xffffffffL; | |
1396 | 2 ->@} | |
1397 | @end smallexample | |
1398 | ||
1399 | In this example, the function was called twice, passing once through | |
1400 | each branch of the @code{if} statement. The body of the @code{do} | |
1401 | loop was executed a total of 26312 times. Note how the @code{while} | |
1402 | statement is annotated. It began execution 26312 times, once for | |
1403 | each iteration through the loop. One of those times (the last time) | |
1404 | it exited, while it branched back to the beginning of the loop 26311 times. | |
1405 | ||
1406 | @node Inaccuracy | |
1407 | @chapter Inaccuracy of @code{gprof} Output | |
1408 | ||
1409 | @menu | |
1410 | * Sampling Error:: Statistical margins of error | |
1411 | * Assumptions:: Estimating children times | |
1412 | @end menu | |
1413 | ||
1414 | @node Sampling Error,Assumptions,,Inaccuracy | |
1415 | @section Statistical Sampling Error | |
1416 | ||
1417 | The run-time figures that @code{gprof} gives you are based on a sampling | |
1418 | process, so they are subject to statistical inaccuracy. If a function runs | |
1419 | only a small amount of time, so that on the average the sampling process | |
1420 | ought to catch that function in the act only once, there is a pretty good | |
1421 | chance it will actually find that function zero times, or twice. | |
1422 | ||
1423 | By contrast, the number-of-calls and basic-block figures | |
1424 | are derived by counting, not | |
1425 | sampling. They are completely accurate and will not vary from run to run | |
1426 | if your program is deterministic. | |
1427 | ||
1428 | The @dfn{sampling period} that is printed at the beginning of the flat | |
1429 | profile says how often samples are taken. The rule of thumb is that a | |
1430 | run-time figure is accurate if it is considerably bigger than the sampling | |
1431 | period. | |
1432 | ||
1433 | The actual amount of error can be predicted. | |
1434 | For @var{n} samples, the @emph{expected} error | |
1435 | is the square-root of @var{n}. For example, | |
1436 | if the sampling period is 0.01 seconds and @code{foo}'s run-time is 1 second, | |
1437 | @var{n} is 100 samples (1 second/0.01 seconds), sqrt(@var{n}) is 10 samples, so | |
1438 | the expected error in @code{foo}'s run-time is 0.1 seconds (10*0.01 seconds), | |
1439 | or ten percent of the observed value. | |
1440 | Again, if the sampling period is 0.01 seconds and @code{bar}'s run-time is | |
1441 | 100 seconds, @var{n} is 10000 samples, sqrt(@var{n}) is 100 samples, so | |
1442 | the expected error in @code{bar}'s run-time is 1 second, | |
1443 | or one percent of the observed value. | |
1444 | It is likely to | |
1445 | vary this much @emph{on the average} from one profiling run to the next. | |
1446 | (@emph{Sometimes} it will vary more.) | |
1447 | ||
1448 | This does not mean that a small run-time figure is devoid of information. | |
1449 | If the program's @emph{total} run-time is large, a small run-time for one | |
1450 | function does tell you that that function used an insignificant fraction of | |
1451 | the whole program's time. Usually this means it is not worth optimizing. | |
1452 | ||
1453 | One way to get more accuracy is to give your program more (but similar) | |
1454 | input data so it will take longer. Another way is to combine the data from | |
1455 | several runs, using the @samp{-s} option of @code{gprof}. Here is how: | |
1456 | ||
1457 | @enumerate | |
1458 | @item | |
1459 | Run your program once. | |
1460 | ||
1461 | @item | |
1462 | Issue the command @samp{mv gmon.out gmon.sum}. | |
1463 | ||
1464 | @item | |
1465 | Run your program again, the same as before. | |
1466 | ||
1467 | @item | |
1468 | Merge the new data in @file{gmon.out} into @file{gmon.sum} with this command: | |
1469 | ||
1470 | @example | |
1471 | gprof -s @var{executable-file} gmon.out gmon.sum | |
1472 | @end example | |
1473 | ||
1474 | @item | |
1475 | Repeat the last two steps as often as you wish. | |
1476 | ||
1477 | @item | |
1478 | Analyze the cumulative data using this command: | |
1479 | ||
1480 | @example | |
1481 | gprof @var{executable-file} gmon.sum > @var{output-file} | |
1482 | @end example | |
1483 | @end enumerate | |
1484 | ||
1485 | @node Assumptions,,Sampling Error,Inaccuracy | |
1486 | @section Estimating @code{children} Times | |
1487 | ||
1488 | Some of the figures in the call graph are estimates---for example, the | |
1489 | @code{children} time values and all the the time figures in caller and | |
1490 | subroutine lines. | |
1491 | ||
1492 | There is no direct information about these measurements in the profile | |
1493 | data itself. Instead, @code{gprof} estimates them by making an assumption | |
1494 | about your program that might or might not be true. | |
1495 | ||
1496 | The assumption made is that the average time spent in each call to any | |
1497 | function @code{foo} is not correlated with who called @code{foo}. If | |
1498 | @code{foo} used 5 seconds in all, and 2/5 of the calls to @code{foo} came | |
1499 | from @code{a}, then @code{foo} contributes 2 seconds to @code{a}'s | |
1500 | @code{children} time, by assumption. | |
1501 | ||
1502 | This assumption is usually true enough, but for some programs it is far | |
1503 | from true. Suppose that @code{foo} returns very quickly when its argument | |
1504 | is zero; suppose that @code{a} always passes zero as an argument, while | |
1505 | other callers of @code{foo} pass other arguments. In this program, all the | |
1506 | time spent in @code{foo} is in the calls from callers other than @code{a}. | |
1507 | But @code{gprof} has no way of knowing this; it will blindly and | |
1508 | incorrectly charge 2 seconds of time in @code{foo} to the children of | |
1509 | @code{a}. | |
1510 | ||
1511 | @c FIXME - has this been fixed? | |
1512 | We hope some day to put more complete data into @file{gmon.out}, so that | |
1513 | this assumption is no longer needed, if we can figure out how. For the | |
1514 | nonce, the estimated figures are usually more useful than misleading. | |
1515 | ||
1516 | @node How do I? | |
1517 | @chapter Answers to Common Questions | |
1518 | ||
1519 | @table @asis | |
1520 | @item How do I find which lines in my program were executed the most times? | |
1521 | ||
1522 | Compile your program with basic-block counting enabled, run it, then | |
1523 | use the following pipeline: | |
1524 | ||
1525 | @example | |
1526 | gprof -l -C @var{objfile} | sort -k 3 -n -r | |
1527 | @end example | |
1528 | ||
1529 | This listing will show you the lines in your code executed most often, | |
1530 | but not necessarily those that consumed the most time. | |
1531 | ||
1532 | @item How do I find which lines in my program called a particular function? | |
1533 | ||
5af11cab | 1534 | Use @samp{gprof -l} and lookup the function in the call graph. |
252b5132 RH |
1535 | The callers will be broken down by function and line number. |
1536 | ||
1537 | @item How do I analyze a program that runs for less than a second? | |
1538 | ||
1539 | Try using a shell script like this one: | |
1540 | ||
1541 | @example | |
1542 | for i in `seq 1 100`; do | |
1543 | fastprog | |
1544 | mv gmon.out gmon.out.$i | |
1545 | done | |
1546 | ||
1547 | gprof -s fastprog gmon.out.* | |
1548 | ||
1549 | gprof fastprog gmon.sum | |
1550 | @end example | |
1551 | ||
1552 | If your program is completely deterministic, all the call counts | |
1553 | will be simple multiples of 100 (i.e. a function called once in | |
1554 | each run will appear with a call count of 100). | |
1555 | ||
1556 | @end table | |
1557 | ||
1558 | @node Incompatibilities | |
1559 | @chapter Incompatibilities with Unix @code{gprof} | |
1560 | ||
1561 | @sc{gnu} @code{gprof} and Berkeley Unix @code{gprof} use the same data | |
1562 | file @file{gmon.out}, and provide essentially the same information. But | |
1563 | there are a few differences. | |
1564 | ||
1565 | @itemize @bullet | |
1566 | @item | |
1567 | @sc{gnu} @code{gprof} uses a new, generalized file format with support | |
1568 | for basic-block execution counts and non-realtime histograms. A magic | |
1569 | cookie and version number allows @code{gprof} to easily identify | |
1570 | new style files. Old BSD-style files can still be read. | |
1571 | @xref{File Format}. | |
1572 | ||
1573 | @item | |
1574 | For a recursive function, Unix @code{gprof} lists the function as a | |
1575 | parent and as a child, with a @code{calls} field that lists the number | |
1576 | of recursive calls. @sc{gnu} @code{gprof} omits these lines and puts | |
1577 | the number of recursive calls in the primary line. | |
1578 | ||
1579 | @item | |
1580 | When a function is suppressed from the call graph with @samp{-e}, @sc{gnu} | |
1581 | @code{gprof} still lists it as a subroutine of functions that call it. | |
1582 | ||
1583 | @item | |
1584 | @sc{gnu} @code{gprof} accepts the @samp{-k} with its argument | |
1585 | in the form @samp{from/to}, instead of @samp{from to}. | |
1586 | ||
1587 | @item | |
1588 | In the annotated source listing, | |
1589 | if there are multiple basic blocks on the same line, | |
5af11cab | 1590 | @sc{gnu} @code{gprof} prints all of their counts, separated by commas. |
252b5132 RH |
1591 | |
1592 | @ignore - it does this now | |
1593 | @item | |
1594 | The function names printed in @sc{gnu} @code{gprof} output do not include | |
1595 | the leading underscores that are added internally to the front of all | |
1596 | C identifiers on many operating systems. | |
1597 | @end ignore | |
1598 | ||
1599 | @item | |
1600 | The blurbs, field widths, and output formats are different. @sc{gnu} | |
1601 | @code{gprof} prints blurbs after the tables, so that you can see the | |
1602 | tables without skipping the blurbs. | |
1603 | @end itemize | |
1604 | ||
1605 | @node Details | |
1606 | @chapter Details of Profiling | |
1607 | ||
1608 | @menu | |
5af11cab | 1609 | * Implementation:: How a program collects profiling information |
252b5132 RH |
1610 | * File Format:: Format of @samp{gmon.out} files |
1611 | * Internals:: @code{gprof}'s internal operation | |
1612 | * Debugging:: Using @code{gprof}'s @samp{-d} option | |
1613 | @end menu | |
1614 | ||
1615 | @node Implementation,File Format,,Details | |
1616 | @section Implementation of Profiling | |
1617 | ||
1618 | Profiling works by changing how every function in your program is compiled | |
1619 | so that when it is called, it will stash away some information about where | |
1620 | it was called from. From this, the profiler can figure out what function | |
1621 | called it, and can count how many times it was called. This change is made | |
1622 | by the compiler when your program is compiled with the @samp{-pg} option, | |
1623 | which causes every function to call @code{mcount} | |
1624 | (or @code{_mcount}, or @code{__mcount}, depending on the OS and compiler) | |
1625 | as one of its first operations. | |
1626 | ||
1627 | The @code{mcount} routine, included in the profiling library, | |
1628 | is responsible for recording in an in-memory call graph table | |
1629 | both its parent routine (the child) and its parent's parent. This is | |
1630 | typically done by examining the stack frame to find both | |
1631 | the address of the child, and the return address in the original parent. | |
5af11cab | 1632 | Since this is a very machine-dependent operation, @code{mcount} |
252b5132 RH |
1633 | itself is typically a short assembly-language stub routine |
1634 | that extracts the required | |
1635 | information, and then calls @code{__mcount_internal} | |
1636 | (a normal C function) with two arguments - @code{frompc} and @code{selfpc}. | |
1637 | @code{__mcount_internal} is responsible for maintaining | |
1638 | the in-memory call graph, which records @code{frompc}, @code{selfpc}, | |
5af11cab | 1639 | and the number of times each of these call arcs was traversed. |
252b5132 RH |
1640 | |
1641 | GCC Version 2 provides a magical function (@code{__builtin_return_address}), | |
1642 | which allows a generic @code{mcount} function to extract the | |
1643 | required information from the stack frame. However, on some | |
1644 | architectures, most notably the SPARC, using this builtin can be | |
1645 | very computationally expensive, and an assembly language version | |
1646 | of @code{mcount} is used for performance reasons. | |
1647 | ||
1648 | Number-of-calls information for library routines is collected by using a | |
1649 | special version of the C library. The programs in it are the same as in | |
1650 | the usual C library, but they were compiled with @samp{-pg}. If you | |
1651 | link your program with @samp{gcc @dots{} -pg}, it automatically uses the | |
1652 | profiling version of the library. | |
1653 | ||
1654 | Profiling also involves watching your program as it runs, and keeping a | |
1655 | histogram of where the program counter happens to be every now and then. | |
1656 | Typically the program counter is looked at around 100 times per second of | |
1657 | run time, but the exact frequency may vary from system to system. | |
1658 | ||
1659 | This is done is one of two ways. Most UNIX-like operating systems | |
1660 | provide a @code{profil()} system call, which registers a memory | |
1661 | array with the kernel, along with a scale | |
1662 | factor that determines how the program's address space maps | |
1663 | into the array. | |
1664 | Typical scaling values cause every 2 to 8 bytes of address space | |
1665 | to map into a single array slot. | |
1666 | On every tick of the system clock | |
1667 | (assuming the profiled program is running), the value of the | |
1668 | program counter is examined and the corresponding slot in | |
1669 | the memory array is incremented. Since this is done in the kernel, | |
1670 | which had to interrupt the process anyway to handle the clock | |
1671 | interrupt, very little additional system overhead is required. | |
1672 | ||
1673 | However, some operating systems, most notably Linux 2.0 (and earlier), | |
1674 | do not provide a @code{profil()} system call. On such a system, | |
1675 | arrangements are made for the kernel to periodically deliver | |
1676 | a signal to the process (typically via @code{setitimer()}), | |
1677 | which then performs the same operation of examining the | |
1678 | program counter and incrementing a slot in the memory array. | |
1679 | Since this method requires a signal to be delivered to | |
1680 | user space every time a sample is taken, it uses considerably | |
1681 | more overhead than kernel-based profiling. Also, due to the | |
1682 | added delay required to deliver the signal, this method is | |
1683 | less accurate as well. | |
1684 | ||
1685 | A special startup routine allocates memory for the histogram and | |
1686 | either calls @code{profil()} or sets up | |
1687 | a clock signal handler. | |
1688 | This routine (@code{monstartup}) can be invoked in several ways. | |
1689 | On Linux systems, a special profiling startup file @code{gcrt0.o}, | |
1690 | which invokes @code{monstartup} before @code{main}, | |
1691 | is used instead of the default @code{crt0.o}. | |
1692 | Use of this special startup file is one of the effects | |
1693 | of using @samp{gcc @dots{} -pg} to link. | |
1694 | On SPARC systems, no special startup files are used. | |
1695 | Rather, the @code{mcount} routine, when it is invoked for | |
1696 | the first time (typically when @code{main} is called), | |
1697 | calls @code{monstartup}. | |
1698 | ||
1699 | If the compiler's @samp{-a} option was used, basic-block counting | |
1700 | is also enabled. Each object file is then compiled with a static array | |
1701 | of counts, initially zero. | |
1702 | In the executable code, every time a new basic-block begins | |
1703 | (i.e. when an @code{if} statement appears), an extra instruction | |
1704 | is inserted to increment the corresponding count in the array. | |
1705 | At compile time, a paired array was constructed that recorded | |
1706 | the starting address of each basic-block. Taken together, | |
1707 | the two arrays record the starting address of every basic-block, | |
1708 | along with the number of times it was executed. | |
1709 | ||
1710 | The profiling library also includes a function (@code{mcleanup}) which is | |
1711 | typically registered using @code{atexit()} to be called as the | |
1712 | program exits, and is responsible for writing the file @file{gmon.out}. | |
1713 | Profiling is turned off, various headers are output, and the histogram | |
1714 | is written, followed by the call-graph arcs and the basic-block counts. | |
1715 | ||
1716 | The output from @code{gprof} gives no indication of parts of your program that | |
1717 | are limited by I/O or swapping bandwidth. This is because samples of the | |
1718 | program counter are taken at fixed intervals of the program's run time. | |
1719 | Therefore, the | |
1720 | time measurements in @code{gprof} output say nothing about time that your | |
1721 | program was not running. For example, a part of the program that creates | |
1722 | so much data that it cannot all fit in physical memory at once may run very | |
1723 | slowly due to thrashing, but @code{gprof} will say it uses little time. On | |
1724 | the other hand, sampling by run time has the advantage that the amount of | |
1725 | load due to other users won't directly affect the output you get. | |
1726 | ||
1727 | @node File Format,Internals,Implementation,Details | |
1728 | @section Profiling Data File Format | |
1729 | ||
1730 | The old BSD-derived file format used for profile data does not contain a | |
1731 | magic cookie that allows to check whether a data file really is a | |
5af11cab | 1732 | @code{gprof} file. Furthermore, it does not provide a version number, thus |
252b5132 RH |
1733 | rendering changes to the file format almost impossible. @sc{gnu} @code{gprof} |
1734 | uses a new file format that provides these features. For backward | |
1735 | compatibility, @sc{gnu} @code{gprof} continues to support the old BSD-derived | |
1736 | format, but not all features are supported with it. For example, | |
1737 | basic-block execution counts cannot be accommodated by the old file | |
1738 | format. | |
1739 | ||
1740 | The new file format is defined in header file @file{gmon_out.h}. It | |
1741 | consists of a header containing the magic cookie and a version number, | |
1742 | as well as some spare bytes available for future extensions. All data | |
1743 | in a profile data file is in the native format of the host on which | |
1744 | the profile was collected. @sc{gnu} @code{gprof} adapts automatically to the | |
1745 | byte-order in use. | |
1746 | ||
1747 | In the new file format, the header is followed by a sequence of | |
1748 | records. Currently, there are three different record types: histogram | |
1749 | records, call-graph arc records, and basic-block execution count | |
1750 | records. Each file can contain any number of each record type. When | |
1751 | reading a file, @sc{gnu} @code{gprof} will ensure records of the same type are | |
1752 | compatible with each other and compute the union of all records. For | |
1753 | example, for basic-block execution counts, the union is simply the sum | |
1754 | of all execution counts for each basic-block. | |
1755 | ||
1756 | @subsection Histogram Records | |
1757 | ||
1758 | Histogram records consist of a header that is followed by an array of | |
1759 | bins. The header contains the text-segment range that the histogram | |
1760 | spans, the size of the histogram in bytes (unlike in the old BSD | |
1761 | format, this does not include the size of the header), the rate of the | |
1762 | profiling clock, and the physical dimension that the bin counts | |
1763 | represent after being scaled by the profiling clock rate. The | |
1764 | physical dimension is specified in two parts: a long name of up to 15 | |
1765 | characters and a single character abbreviation. For example, a | |
1766 | histogram representing real-time would specify the long name as | |
1767 | "seconds" and the abbreviation as "s". This feature is useful for | |
1768 | architectures that support performance monitor hardware (which, | |
1769 | fortunately, is becoming increasingly common). For example, under DEC | |
1770 | OSF/1, the "uprofile" command can be used to produce a histogram of, | |
1771 | say, instruction cache misses. In this case, the dimension in the | |
1772 | histogram header could be set to "i-cache misses" and the abbreviation | |
1773 | could be set to "1" (because it is simply a count, not a physical | |
1774 | dimension). Also, the profiling rate would have to be set to 1 in | |
1775 | this case. | |
1776 | ||
1777 | Histogram bins are 16-bit numbers and each bin represent an equal | |
1778 | amount of text-space. For example, if the text-segment is one | |
1779 | thousand bytes long and if there are ten bins in the histogram, each | |
1780 | bin represents one hundred bytes. | |
1781 | ||
1782 | ||
1783 | @subsection Call-Graph Records | |
1784 | ||
1785 | Call-graph records have a format that is identical to the one used in | |
1786 | the BSD-derived file format. It consists of an arc in the call graph | |
1787 | and a count indicating the number of times the arc was traversed | |
1788 | during program execution. Arcs are specified by a pair of addresses: | |
1789 | the first must be within caller's function and the second must be | |
1790 | within the callee's function. When performing profiling at the | |
1791 | function level, these addresses can point anywhere within the | |
1792 | respective function. However, when profiling at the line-level, it is | |
1793 | better if the addresses are as close to the call-site/entry-point as | |
1794 | possible. This will ensure that the line-level call-graph is able to | |
1795 | identify exactly which line of source code performed calls to a | |
1796 | function. | |
1797 | ||
1798 | @subsection Basic-Block Execution Count Records | |
1799 | ||
1800 | Basic-block execution count records consist of a header followed by a | |
1801 | sequence of address/count pairs. The header simply specifies the | |
1802 | length of the sequence. In an address/count pair, the address | |
1803 | identifies a basic-block and the count specifies the number of times | |
1804 | that basic-block was executed. Any address within the basic-address can | |
1805 | be used. | |
1806 | ||
1807 | @node Internals,Debugging,File Format,Details | |
1808 | @section @code{gprof}'s Internal Operation | |
1809 | ||
1810 | Like most programs, @code{gprof} begins by processing its options. | |
1811 | During this stage, it may building its symspec list | |
1812 | (@code{sym_ids.c:sym_id_add}), if | |
1813 | options are specified which use symspecs. | |
1814 | @code{gprof} maintains a single linked list of symspecs, | |
1815 | which will eventually get turned into 12 symbol tables, | |
1816 | organized into six include/exclude pairs - one | |
1817 | pair each for the flat profile (INCL_FLAT/EXCL_FLAT), | |
1818 | the call graph arcs (INCL_ARCS/EXCL_ARCS), | |
1819 | printing in the call graph (INCL_GRAPH/EXCL_GRAPH), | |
1820 | timing propagation in the call graph (INCL_TIME/EXCL_TIME), | |
1821 | the annotated source listing (INCL_ANNO/EXCL_ANNO), | |
1822 | and the execution count listing (INCL_EXEC/EXCL_EXEC). | |
1823 | ||
1824 | After option processing, @code{gprof} finishes | |
1825 | building the symspec list by adding all the symspecs in | |
1826 | @code{default_excluded_list} to the exclude lists | |
1827 | EXCL_TIME and EXCL_GRAPH, and if line-by-line profiling is specified, | |
1828 | EXCL_FLAT as well. | |
1829 | These default excludes are not added to EXCL_ANNO, EXCL_ARCS, and EXCL_EXEC. | |
1830 | ||
1831 | Next, the BFD library is called to open the object file, | |
1832 | verify that it is an object file, | |
1833 | and read its symbol table (@code{core.c:core_init}), | |
1834 | using @code{bfd_canonicalize_symtab} after mallocing | |
5af11cab | 1835 | an appropriately sized array of symbols. At this point, |
252b5132 RH |
1836 | function mappings are read (if the @samp{--file-ordering} option |
1837 | has been specified), and the core text space is read into | |
1838 | memory (if the @samp{-c} option was given). | |
1839 | ||
1840 | @code{gprof}'s own symbol table, an array of Sym structures, | |
1841 | is now built. | |
1842 | This is done in one of two ways, by one of two routines, depending | |
1843 | on whether line-by-line profiling (@samp{-l} option) has been | |
1844 | enabled. | |
1845 | For normal profiling, the BFD canonical symbol table is scanned. | |
1846 | For line-by-line profiling, every | |
1847 | text space address is examined, and a new symbol table entry | |
1848 | gets created every time the line number changes. | |
1849 | In either case, two passes are made through the symbol | |
1850 | table - one to count the size of the symbol table required, | |
1851 | and the other to actually read the symbols. In between the | |
1852 | two passes, a single array of type @code{Sym} is created of | |
5af11cab | 1853 | the appropriate length. |
252b5132 RH |
1854 | Finally, @code{symtab.c:symtab_finalize} |
1855 | is called to sort the symbol table and remove duplicate entries | |
1856 | (entries with the same memory address). | |
1857 | ||
1858 | The symbol table must be a contiguous array for two reasons. | |
1859 | First, the @code{qsort} library function (which sorts an array) | |
1860 | will be used to sort the symbol table. | |
1861 | Also, the symbol lookup routine (@code{symtab.c:sym_lookup}), | |
1862 | which finds symbols | |
1863 | based on memory address, uses a binary search algorithm | |
1864 | which requires the symbol table to be a sorted array. | |
1865 | Function symbols are indicated with an @code{is_func} flag. | |
1866 | Line number symbols have no special flags set. | |
1867 | Additionally, a symbol can have an @code{is_static} flag | |
1868 | to indicate that it is a local symbol. | |
1869 | ||
1870 | With the symbol table read, the symspecs can now be translated | |
1871 | into Syms (@code{sym_ids.c:sym_id_parse}). Remember that a single | |
1872 | symspec can match multiple symbols. | |
1873 | An array of symbol tables | |
1874 | (@code{syms}) is created, each entry of which is a symbol table | |
1875 | of Syms to be included or excluded from a particular listing. | |
1876 | The master symbol table and the symspecs are examined by nested | |
1877 | loops, and every symbol that matches a symspec is inserted | |
1878 | into the appropriate syms table. This is done twice, once to | |
1879 | count the size of each required symbol table, and again to build | |
1880 | the tables, which have been malloced between passes. | |
1881 | From now on, to determine whether a symbol is on an include | |
1882 | or exclude symspec list, @code{gprof} simply uses its | |
1883 | standard symbol lookup routine on the appropriate table | |
1884 | in the @code{syms} array. | |
1885 | ||
1886 | Now the profile data file(s) themselves are read | |
1887 | (@code{gmon_io.c:gmon_out_read}), | |
1888 | first by checking for a new-style @samp{gmon.out} header, | |
1889 | then assuming this is an old-style BSD @samp{gmon.out} | |
1890 | if the magic number test failed. | |
1891 | ||
1892 | New-style histogram records are read by @code{hist.c:hist_read_rec}. | |
1893 | For the first histogram record, allocate a memory array to hold | |
1894 | all the bins, and read them in. | |
1895 | When multiple profile data files (or files with multiple histogram | |
1896 | records) are read, the starting address, ending address, number | |
1897 | of bins and sampling rate must match between the various histograms, | |
1898 | or a fatal error will result. | |
1899 | If everything matches, just sum the additional histograms into | |
1900 | the existing in-memory array. | |
1901 | ||
1902 | As each call graph record is read (@code{call_graph.c:cg_read_rec}), | |
1903 | the parent and child addresses | |
1904 | are matched to symbol table entries, and a call graph arc is | |
1905 | created by @code{cg_arcs.c:arc_add}, unless the arc fails a symspec | |
1906 | check against INCL_ARCS/EXCL_ARCS. As each arc is added, | |
1907 | a linked list is maintained of the parent's child arcs, and of the child's | |
1908 | parent arcs. | |
1909 | Both the child's call count and the arc's call count are | |
1910 | incremented by the record's call count. | |
1911 | ||
1912 | Basic-block records are read (@code{basic_blocks.c:bb_read_rec}), | |
1913 | but only if line-by-line profiling has been selected. | |
1914 | Each basic-block address is matched to a corresponding line | |
1915 | symbol in the symbol table, and an entry made in the symbol's | |
1916 | bb_addr and bb_calls arrays. Again, if multiple basic-block | |
1917 | records are present for the same address, the call counts | |
1918 | are cumulative. | |
1919 | ||
1920 | A gmon.sum file is dumped, if requested (@code{gmon_io.c:gmon_out_write}). | |
1921 | ||
1922 | If histograms were present in the data files, assign them to symbols | |
1923 | (@code{hist.c:hist_assign_samples}) by iterating over all the sample | |
1924 | bins and assigning them to symbols. Since the symbol table | |
1925 | is sorted in order of ascending memory addresses, we can | |
1926 | simple follow along in the symbol table as we make our pass | |
1927 | over the sample bins. | |
1928 | This step includes a symspec check against INCL_FLAT/EXCL_FLAT. | |
1929 | Depending on the histogram | |
1930 | scale factor, a sample bin may span multiple symbols, | |
1931 | in which case a fraction of the sample count is allocated | |
1932 | to each symbol, proportional to the degree of overlap. | |
1933 | This effect is rare for normal profiling, but overlaps | |
1934 | are more common during line-by-line profiling, and can | |
1935 | cause each of two adjacent lines to be credited with half | |
1936 | a hit, for example. | |
1937 | ||
1938 | If call graph data is present, @code{cg_arcs.c:cg_assemble} is called. | |
5af11cab | 1939 | First, if @samp{-c} was specified, a machine-dependent |
252b5132 RH |
1940 | routine (@code{find_call}) scans through each symbol's machine code, |
1941 | looking for subroutine call instructions, and adding them | |
1942 | to the call graph with a zero call count. | |
1943 | A topological sort is performed by depth-first numbering | |
1944 | all the symbols (@code{cg_dfn.c:cg_dfn}), so that | |
1945 | children are always numbered less than their parents, | |
1946 | then making a array of pointers into the symbol table and sorting it into | |
1947 | numerical order, which is reverse topological | |
1948 | order (children appear before parents). | |
1949 | Cycles are also detected at this point, all members | |
1950 | of which are assigned the same topological number. | |
1951 | Two passes are now made through this sorted array of symbol pointers. | |
1952 | The first pass, from end to beginning (parents to children), | |
5af11cab | 1953 | computes the fraction of child time to propagate to each parent |
252b5132 RH |
1954 | and a print flag. |
1955 | The print flag reflects symspec handling of INCL_GRAPH/EXCL_GRAPH, | |
1956 | with a parent's include or exclude (print or no print) property | |
1957 | being propagated to its children, unless they themselves explicitly appear | |
1958 | in INCL_GRAPH or EXCL_GRAPH. | |
1959 | A second pass, from beginning to end (children to parents) actually | |
5af11cab | 1960 | propagates the timings along the call graph, subject |
252b5132 RH |
1961 | to a check against INCL_TIME/EXCL_TIME. |
1962 | With the print flag, fractions, and timings now stored in the symbol | |
1963 | structures, the topological sort array is now discarded, and a | |
1964 | new array of pointers is assembled, this time sorted by propagated time. | |
1965 | ||
1966 | Finally, print the various outputs the user requested, which is now fairly | |
1967 | straightforward. The call graph (@code{cg_print.c:cg_print}) and | |
1968 | flat profile (@code{hist.c:hist_print}) are regurgitations of values | |
1969 | already computed. The annotated source listing | |
1970 | (@code{basic_blocks.c:print_annotated_source}) uses basic-block | |
1971 | information, if present, to label each line of code with call counts, | |
1972 | otherwise only the function call counts are presented. | |
1973 | ||
1974 | The function ordering code is marginally well documented | |
1975 | in the source code itself (@code{cg_print.c}). Basically, | |
1976 | the functions with the most use and the most parents are | |
1977 | placed first, followed by other functions with the most use, | |
1978 | followed by lower use functions, followed by unused functions | |
1979 | at the end. | |
1980 | ||
1981 | @node Debugging,,Internals,Details | |
1982 | @subsection Debugging @code{gprof} | |
1983 | ||
1984 | If @code{gprof} was compiled with debugging enabled, | |
1985 | the @samp{-d} option triggers debugging output | |
1986 | (to stdout) which can be helpful in understanding its operation. | |
1987 | The debugging number specified is interpreted as a sum of the following | |
1988 | options: | |
1989 | ||
1990 | @table @asis | |
1991 | @item 2 - Topological sort | |
1992 | Monitor depth-first numbering of symbols during call graph analysis | |
1993 | @item 4 - Cycles | |
1994 | Shows symbols as they are identified as cycle heads | |
1995 | @item 16 - Tallying | |
1996 | As the call graph arcs are read, show each arc and how | |
1997 | the total calls to each function are tallied | |
1998 | @item 32 - Call graph arc sorting | |
1999 | Details sorting individual parents/children within each call graph entry | |
2000 | @item 64 - Reading histogram and call graph records | |
2001 | Shows address ranges of histograms as they are read, and each | |
2002 | call graph arc | |
2003 | @item 128 - Symbol table | |
2004 | Reading, classifying, and sorting the symbol table from the object file. | |
2005 | For line-by-line profiling (@samp{-l} option), also shows line numbers | |
2006 | being assigned to memory addresses. | |
2007 | @item 256 - Static call graph | |
2008 | Trace operation of @samp{-c} option | |
2009 | @item 512 - Symbol table and arc table lookups | |
2010 | Detail operation of lookup routines | |
2011 | @item 1024 - Call graph propagation | |
2012 | Shows how function times are propagated along the call graph | |
2013 | @item 2048 - Basic-blocks | |
2014 | Shows basic-block records as they are read from profile data | |
2015 | (only meaningful with @samp{-l} option) | |
2016 | @item 4096 - Symspecs | |
2017 | Shows symspec-to-symbol pattern matching operation | |
2018 | @item 8192 - Annotate source | |
2019 | Tracks operation of @samp{-A} option | |
2020 | @end table | |
2021 | ||
2022 | @contents | |
2023 | @bye | |
2024 | ||
2025 | NEEDS AN INDEX | |
2026 | ||
2027 | -T - "traditional BSD style": How is it different? Should the | |
2028 | differences be documented? | |
2029 | ||
2030 | example flat file adds up to 100.01%... | |
2031 | ||
2032 | note: time estimates now only go out to one decimal place (0.0), where | |
2033 | they used to extend two (78.67). |