|
|
1.1 root 1: \input texinfo @c -*-texinfo-*-
2: @setfilename gprof
3: @settitle gprof
4: @ifinfo
5: This file documents the gprof profiler of the GNU system.
6:
7: Copyright (C) 1988 Free Software Foundation, Inc.
8:
9: Permission is granted to make and distribute verbatim copies of
10: this manual provided the copyright notice and this permission notice
11: are preserved on all copies.
12:
13: @ignore
14: Permission is granted to process this file through Tex and print the
15: results, provided the printed document carries copying permission
16: notice identical to this one except for the removal of this paragraph
17: (this paragraph not being relevant to the printed manual).
18:
19: @end ignore
20: Permission is granted to copy and distribute modified versions of this
21: manual under the conditions for verbatim copying, provided that the entire
22: resulting derived work is distributed under the terms of a permission
23: notice identical to this one.
24:
25: Permission is granted to copy and distribute translations of this manual
26: into another language, under the above conditions for modified versions.
27: @end ifinfo
28:
29: @titlepage
30: @center @titlefont{gprof}
31: @sp 1
32: @center The GNU Profiler
33: @sp 2
34: @center Jay Fenlason and Richard Stallman
35: @sp 4
36: This manual describes the GNU profiler, @code{gprof}, and how you can use
37: it to determine which parts of a program are taking most of the execution
38: time. We assume that you know how to write, compile, and execute programs.
39: GNU @code{gprof} was written by Jay Fenlason.
40: @sp 8
41: Copyright @copyright{} 1988 Free Software Foundation, Inc.
42:
43: Permission is granted to make and distribute verbatim copies of
44: this manual provided the copyright notice and this permission notice
45: are preserved on all copies.
46:
47: @ignore
48: Permission is granted to process this file through Tex and print the
49: results, provided the printed document carries copying permission
50: notice identical to this one except for the removal of this paragraph
51: (this paragraph not being relevant to the printed manual).
52:
53: @end ignore
54: Permission is granted to copy and distribute modified versions of this
55: manual under the conditions for verbatim copying, provided that the entire
56: resulting derived work is distributed under the terms of a permission
57: notice identical to this one.
58:
59: Permission is granted to copy and distribute translations of this manual
60: into another language, under the same conditions as for modified versions.
61:
62: @end titlepage
63:
64: @ifinfo
65: @node Top, Why, Top, (dir)
66: @ichapter Profiling a Program: Where Does It Spend Its Time?
67:
68: This manual describes the GNU profiler @code{gprof}, and how you can use it
69: to determine which parts of a program are taking most of the execution
70: time. We assume that you know how to write, compile, and execute programs.
71: GNU @code{gprof} was written by Jay Fenlason.
72:
73: @menu
74: * Why:: What profiling means, and why it is useful.
75: * Compiling:: How to compile your program for profiling.
76: * Executing:: How to execute your program to generate the
77: profile data file @file{gmon.out}.
78: * Analyzing:: How to run @code{gprof}, and how to specify
79: options for it.
80:
81: * Flat Profile:: The flat profile shows how much time was spent
82: executing directly in each function.
83: * Call Graph:: The call graph shows which functions called which
84: others, and how much time each function used
85: when its subroutine calls are included.
86:
87: * Implementation:: How the profile data is recorded and written.
88: * Sampling Error:: Statistical margins of error.
89: How to accumulate data from several runs
90: to make it more accurate.
91:
92: * Assumptions:: Some of @code{gprof}'s measurements are based
93: on assumptions about your program
94: that could be very wrong.
95:
96: * Incompatibilities:: (between GNU @code{gprof} and Unix @code{gprof}.)
97: @end menu
98: @end ifinfo
99:
100: @node Why, Compiling, Top, Top
101: @chapter Why Profile
102:
103: Profiling allows you to learn where your program spent its time and which
104: functions called which other functions while it was executing. This
105: information can show you which pieces of your program are slower than you
106: expected, and might be candidates for rewriting to make your program
107: execute faster. It can also tell you which functions are being called more
108: or less often than you expected. This may help you spot bugs that had
109: otherwise been unnoticed.
110:
111: Since the profiler uses information collected during the actual execution
112: of your program, it can be used on programs that are too large or too
113: complex to analyze by reading the source. However, how your program is run
114: will affect the information that shows up in the profile data. If you
115: don't use some feature of your program while it is being profiled, no
116: profile information will be generated for that feature.
117:
118: Profiling has several steps:
119:
120: @itemize @bullet
121: @item
122: You must compile and link your program with profiling enabled.
123: @xref{Compiling}.
124:
125: @item
126: You must execute your program to generate a profile data file.
127: @xref{Executing}.
128:
129: @item
130: You must run @code{gprof} to analyze the profile data.
131: @xref{Analyzing}.
132: @end itemize
133:
134: The next three chapters explain these steps in greater detail.
135:
136: The result of the analysis is a file containing two tables, the
137: @dfn{flat profile} and the @dfn{call graph} (plus blurbs which briefly
138: explain the contents of these tables).
139:
140: The flat profile shows how much time your program spent in each function,
141: and how many times that function was called. If you simply want to know
142: which functions burn most of the cycles, it is stated concisely here.
143: @xref{Flat Profile}.
144:
145: The call graph, shows, for each function, which functions called it, which
146: other functions it called, and how many times. There is also an estimate
147: of how much time was spent in the subroutines of each function. This can
148: suggest places where you might try to eliminate function calls that use a
149: lot of time. @xref{Call Graph}.
150:
151: @node Compiling, Executing, Why, Top
152: @chapter Compiling a Program for Profiling
153:
154: The first step in generating profile information for your program is
155: to compile and link it with profiling enabled.
156:
157: To compile a source file for profiling, specify the @samp{-pg} option when
158: you run the compiler. (This is in addition to the options you normally
159: use.)
160:
161: To link the program for profiling, if you use a compiler such as @code{cc}
162: to do the linking, simply specify @samp{-pg} in addition to your usual
163: options. The same option, @samp{-pg}, alters either compilation or linking
164: to do what is necessary for profiling. Here are examples:
165:
166: @example
167: cc -g myprog.c utils.c -pg
168: cc -o myprog myprog.o utils.o -pg
169: @end example
170:
171: The @samp{-pg} option also works with a command that both compiles and links:
172:
173: @example
174: cc -o myprog myprog.c utils.c -g -pg
175: @end example
176:
177: If you run the linker @code{ld} directly instead of through a compiler such
178: as @code{cc}, you must specify the profiling startup file
179: @file{/lib/gcrt0.o} as the first input file instead of the usual startup
180: file @file{/lib/crt0.o}. In addition, you would probably want to specify
181: the profiling C library, @file{/usr/lib/libc_p.a}, by writing @samp{-lc_p}
182: instead of the usual @samp{-lc}. This is not absolutely necessary, but doing
183: this gives you number-of-calls information for standard library functions such
184: as @code{read} and @code{open}. For example:
185:
186: @example
187: ld -o myprog /lib/gcrt0.o myprog.o utils.o -lc_p
188: @end example
189:
190: If you compile only some of the modules of the program with @samp{-pg}, you
191: can still profile the program, but you won't get complete information about
192: the modules that were compiled without @samp{-pg}. The only information
193: you get for the functions in those modules is the total time spent in them;
194: there is no record of how many times they were called, or from where. This
195: will not affect the flat profile (except that the @code{calls} field for
196: the functions will be blank), but will greatly reduce the usefulness of the
197: call graph.
198:
199: So far GNU @code{gprof} has been tested only with C programs, but it ought
200: to work with any language in which programs are compiled and linked to form
201: executable files. If it does not, please let us know.
202:
203: @node Executing, Analyzing, Compiling, Top
204: @chapter Executing the Program to Generate Profile Data
205:
206: Once the program is compiled for profiling, you must run it in order to
207: generate the information that @code{gprof} needs. Simply run the program
208: as usual, using the normal arguments, file names, etc. The program should
209: run normally, producing the same output as usual. It will, however, run
210: somewhat slower than normal because of the time spent collecting and the
211: writing the profile data.
212:
213: The way you run the program---the arguments and input that you give
214: it---may have a dramatic effect on what the profile information shows. The
215: profile data will describe the parts of the program that were activated for
216: the particular input you use. For example, if the first command you give
217: to your program is to quit, the profile data will show the time used in
218: initialization and in cleanup, but not much else.
219:
220: You program will write the profile data into a file called @file{gmon.out}
221: just before exiting. If there is already a file called @file{gmon.out},
222: its contents are overwritten. There is currently no way to tell the
223: program to write the profile data under a different name, but you can rename
224: the file afterward if you are concerned that it may be overwritten.
225:
226: In order to write the @file{gmon.out} file properly, your program must exit
227: normally: by returning from @code{main} or by calling @code{exit}. Calling
228: the low-level function @code{_exit} does not write the profile data, and
229: neither does abnormal termination due to an unhandled signal.
230:
231: The @file{gmon.out} file is written in the program's @emph{current working
232: directory} at the time it exits. This means that if your program calls
233: @code{chdir}, the @file{gmon.out} file will be left in the last directory
234: your program @code{chdir}'d to. If you don't have permission to write in
235: this directory, the file is not written. You may get a confusing error
236: message if this happens. (We have not yet replaced the part of Unix
237: responsible for this; when we do, we will make the error message
238: comprehensible.)
239:
240: @node Analyzing, Flat Profile, Executing, Top
241: @chapter Analyzing the Profile Data: @code{gprof} Command Summary
242:
243: After you have a profile data file @file{gmon.out}, you can run @code{gprof}
244: to interpret the information in it. The @code{gprof} program prints a
245: flat profile and a call graph on standard output. Typically you would
246: redirect the output of @code{gprof} into a file with @samp{>}.
247:
248: You run @code{gprof} like this:
249:
250: @example
251: gprof @var{options} [@var{executable-file} [@var{profile-data-files}@dots{}]] [> @var{outfile}]
252: @end example
253:
254: @noindent
255: Here square-brackets indicate optional arguments.
256:
257: If you omit the executable file name, the file @file{a.out} is used. If
258: you give no profile data file name, the file @file{gmon.out} is used. If
259: any file is not in the proper format, or if the profile data file does not
260: appear to belong to the executable file, an error message is printed.
261:
262: You can give more than one profile data file by entering all their names
263: after the executable file name; then the statistics in all the data files
264: are summed together.
265:
266: The following options may be used to selectively include or exclude
267: functions in the output:
268:
269: @table @code
270: @item -a
271: The @code{-a} option causes @code{gprof} to ignore static (private)
272: functions. (These are functions whose names are not listed as global,
273: and which are not visible outside the file/function/block where they
274: were defined.) Time spent in these functions, calls to/from them,
275: etc, will all be attributed to the function that was loaded directly
276: before it in the executable file. This is compatible with Unix
277: @code{gprof}, but a bad idea. This option affects both the flat
278: profile and the call graph.
279:
280: @item -e @var{function_name}
281: The @code{-e @var{function}} option tells @code{gprof} to not print
282: information about the function (and its children@dots{}) in the call
283: graph. The function will still be listed as a child of any functions
284: that call it, but its index number will be shown as @samp{[not
285: printed]}.
286:
287: @item -E @var{function_name}
288: The @code{-E @var{function}} option works like the @code{-e} option,
289: but time spent in the function (and children who were not called from
290: anywhere else), will not be used to compute the percentages-of-time
291: for the call graph.
292:
293: @item -f @var{function_name}
294: The @code{-f @var{function}} option causes @code{gprof} to limit the
295: call graph to the function and its children (and their
296: children@dots{}).
297:
298: @item -F @var{function_name}
299: The @code{-F @var{function}} option works like the @code{-f} option,
300: but only time spent in the function and its children (and their
301: children@dots{}) will be used to determine total-time and
302: percentages-of-time for the call graph.
303:
304: @item -z
305: If you give the @code{-z} option, @code{gprof} will mention all
306: functions in the flat profile, even those that were never called, and
307: that had no time spent in them.
308: @end table
309:
310: The order of these options does not matter.
311:
312: Note that only one function can be specified with each @code{-e},
313: @code{-E}, @code{-f} or @code{-F} option. To specify more than one
314: function, use multiple options. For example, this command:
315:
316: @example
317: gprof -e boring -f foo -f bar myprogram > gprof.output
318: @end example
319:
320: @noindent
321: lists in the call graph all functions that were reached from either
322: @code{foo} or @code{bar} and were not reachable from @code{boring}.
323:
324: There are two other useful @code{gprof} options:
325:
326: @table @code
327: @item -b
328: If the @code{-b} option is given, @code{gprof} doesn't print the
329: verbose blurbs that try to explain the meaning of all of the fields in
330: the tables. This is useful if you intend to print out the output, or
331: are tired of seeing the blurbs.
332:
333: @item -s
334: The @code{-s} option causes @code{gprof} to summarize the information
335: in the profile data files it read in, and write out a profile data
336: file called @file{gmon.sum}, which contains all the information from
337: the profile data files that @code{gprof} read in. The file @file{gmon.sum}
338: may be one of the specified input files; the effect of this is to
339: merge the data in the other input files into @file{gmon.sum}.
340: @xref{Sampling Error}.
341:
342: Eventually you can run @code{gprof} again without @samp{-s} to analyze the
343: cumulative data in the file @file{gmon.sum}.
344: @end table
345:
346: @node Flat Profile, Call Graph, Analyzing, Top
347: @chapter How to Understand the Flat Profile
348: @cindex flat profile
349:
350: The @dfn{flat profile} shows the total amount of time your program
351: spent executing each function. Unless the @samp{-z} option is given,
352: functions with no apparent time spent in them, and no apparent calls
353: to them, are not mentioned. Note that if a function was not compiled
354: for profiling, and didn't run long enough to show up on the program
355: counter histogram, it will be indistinguishable from a function that
356: was never called.
357: @c???
358:
359: Here is a sample flat profile for a small program:
360:
361: @example
362: Each sample counts as 0.01 seconds.
363:
364: % time seconds cumsec calls function
365: 79.17 0.19 0.19 6 a
366: 16.67 0.04 0.23 1 main
367: 4.17 0.01 0.24 mcount
368: 0.00 0 0.24 1 profil
369: @end example
370:
371: @noindent
372: The functions are sorted by decreasing run-time spent in them. The
373: functions @code{mcount} and @code{profil} are part of the profiling
374: aparatus and appear in every flat profile; their time gives a measure of
375: the amount of overhead due to profiling. (These internal functions are
376: omitted from the call graph.)
377:
378: The sampling period estimates the margin of error in each of the time
379: figures. A time figure that is not much larger than this is not reliable.
380: In this example, the @code{seconds} field for @code{mcount} might well be 0
381: or 0.02 in another run. @xref{Sampling Error}, for a complete discussion.
382:
383: Here is what the fields in each line mean:
384:
385: @table @code
386: @item % time
387: This is the percentage of the total execution time your program spent
388: in this function. These should all add up to 100%.
389:
390: @item seconds
391: This is the total number of seconds the computer spent executing the
392: user code of this function.
393:
394: @item cumsec
395: This is the cumulative total number of seconds the computer spent
396: executing this functions, plus the time spent in all the functions
397: above this one in this table.
398:
399: @item calls
400: This is the total number of times the function was called. If the
401: function was never called, or the number of times it was called cannot
402: be determined (probably because the function was not compiled with
403: profiling enabled), the @dfn{calls} field is blank.
404:
405: @item function
406: This is the name of the function.
407: @end table
408:
409: @node Call Graph, Implementation, Flat Profile, Top
410: @chapter How to Read the Call Graph
411:
412: @cindex call graph
413: The @dfn{call graph} shows how much time was spent in each function
414: and its children. From this information, you can find functions that,
415: while they themselves may not have used much time, called other
416: functions that did use unusual amounts of time.
417:
418: Here is a sample call from a small program. This call came from the
419: same @code{gprof} run as the flat profile example in the previous
420: chapter.
421:
422: @example
423: index % time self children called name
424: <spontaneous>
425: [1] 100.00 0 0.23 0 start [1]
426: 0.04 0.19 1/1 main [2]
427: ----------------------------------------
428: 0.04 0.19 1/1 start [1]
429: [2] 100.00 0.04 0.19 1 main [2]
430: 0.19 0 1/1 a [3]
431: ----------------------------------------
432: 0.19 0 1/1 main [2]
433: [3] 82.61 0.19 0 1+5 a [3]
434: ----------------------------------------
435: @end example
436:
437: The lines full of dashes divide this table into @dfn{entries}, one for each
438: function. Each entry has one or more lines.
439:
440: In each entry, the primary line is the one that starts with an index number
441: in square brackets. The end of this line says which function the entry is
442: for. The preceding lines in the entry describe the callers of this
443: function and the following lines describe its subroutines (also called
444: @dfn{children} when we speak of the call graph).
445:
446: The entries are sorted by time spent in the function and its subroutines.
447:
448: The internal profiling functions @code{mcount} and @code{profil}
449: (@pxref{Flat Profile}) are never mentioned in the call graph.
450:
451: @menu
452: * Primary:: Details of the primary line's contents.
453: * Callers:: Details of caller-lines' contents.
454: * Subroutines:: Details of subroutine-lines' contents.
455: * Cycles:: When there are cycles of recursion,
456: such as @code{a} calls @code{b} calls @code{a}@dots{}
457: @end menu
458:
459: @node Primary, Callers, Call Graph, Call Graph
460: @section The Primary Line
461:
462: The @dfn{primary line} in a call graph entry is the line that
463: describes the function which the entry is about and gives the overall
464: statistics for this function.
465:
466: For reference, we repeat the primary line from the entry for function
467: @code{a} in our main example, together with the heading line that shows the
468: names of the fields:
469:
470: @example
471: index % time self children called name
472: @dots{}
473: [3] 82.61 0.19 0 1+5 a [3]
474: @end example
475:
476: Here is what the fields in the primary line mean:
477:
478: @table @code
479: @item index
480: Entries are numbered with consecutive integers. Each function
481: therefore has an index number, which appears at the beginning of its
482: primary line.
483:
484: Each cross-reference to a function, as a caller or subroutine of
485: another, gives its index number as well as its name. The index number
486: guides you if you wish to look for the entry for that function.
487:
488: @item % time
489: This is the percentage of the total time that was spent in this
490: function, including time spent in subroutines called from this
491: function.
492:
493: The time spent in this function is counted again for the callers of
494: this function. Therefore, adding up these percentages is meaningless.
495:
496: @item self
497: This is the total amount of time spent in this function. This
498: should be identical to the number printed in the @code{seconds} field
499: for this function in the flat profile.
500:
501: @item children
502: This is the total amount of time spent in the subroutine calls made by
503: this function. This should be equal to the sum of all the @code{self}
504: and @code{children} entries of the children listed directly below this
505: function.
506:
507: @item called
508: This is the number of times the function was called.
509:
510: If the function called itself recursively, there are two numbers,
511: separated by a @samp{+}. The first number counts non-recursive calls,
512: and the second counts recursive calls.
513:
514: In the example above, the function @code{a} called itself five times,
515: and was called once from @code{main}.
516:
517: @item name
518: This is the name of the current function. The index number is
519: repeated after it.
520:
521: If the function is part of a cycle of recursion, the cycle number is
522: printed between the function's name and the index number
523: (@pxref{Cycles}). For example, if function @code{gnurr} is part of
524: cycle number one, and has index number twelve, its primary line would
525: be end like this:
526:
527: @example
528: gnurr <cycle 1> [12]
529: @end example
530: @end table
531:
532: @node Callers, Subroutines, Primary, Call Graph
533: @section Lines for a Function's Callers
534:
535: A function's entry has a line for each function it was called by.
536: These lines' fields correspond to the fields of the primary line, but
537: their meanings are different because of the difference in context.
538:
539: For reference, we repeat two lines from the entry for the function
540: @code{a}, the primary line and one caller-line preceding it, together
541: with the heading line that shows the names of the fields:
542:
543: @example
544: index % time self children called name
545: @dots{}
546: 0.19 0 1/1 main [2]
547: [3] 82.61 0.19 0 1+5 a [3]
548: @end example
549:
550: Here are the meanings of the fields in the caller-line for @code{a}
551: called from @code{main}:
552:
553: @table @code
554: @item self
555: An estimate of the amount of time spent in @code{a} itself when it was
556: called from @code{main}.
557:
558: @item children
559: An estimate of the amount of time spent in @code{a}'s subroutines when
560: @code{a} was called from @code{main}.
561:
562: The sum of the @code{self} and @code{children} fields is an estimate
563: of the amount of time spent within calls to @code{a} from @code{main}.
564:
565: @item called
566: Two numbers: the number of times @code{a} was called from @code{main},
567: followed by the total number of nonrecursive calls to @code{a} from
568: all its callers.
569:
570: @item name and index number
571: The name of the caller of @code{a} to which this line applies,
572: followed by the caller's index number.
573:
574: Not all functions have entries in the call graph; some
575: options to @code{gprof} request the omission of certain functions.
576: When a caller has no entry of its own, it still has caller-lines
577: in the entries of the functions it calls. Since this caller
578: has no index number, the string @samp{[not printed]} is used
579: instead of one.
580:
581: If the caller is part of a recursion cycle, the cycle number is
582: printed between the name and the index number.
583: @end table
584:
585: If the identity of the callers of a function cannot be determined, a
586: dummy caller-line is printed which has @samp{<spontaneous>} as the
587: ``caller's name'' and all other fields blank. This can happen for
588: signal handlers.
589: @c What if some calls have determinable callers' names but not all?
590:
591: @node Subroutines, Cycles, Callers, Call Graph
592: @section Lines for a Function's Subroutines
593:
594: A function's entry has a line for each of its subroutines---in other
595: words, a line for each other function that it called. These lines'
596: fields correspond to the fields of the primary line, but their meanings
597: are different because of the difference in context.
598:
599: For reference, we repeat two lines from the entry for the function
600: @code{main}, the primary line and a line for a subroutine, together
601: with the heading line that shows the names of the fields:
602:
603: @example
604: index % time self children called name
605: @dots{}
606: [2] 100.00 0.04 0.19 1 main [2]
607: 0.19 0 1/1 a [3]
608: @end example
609:
610: Here are the meanings of the fields in the subroutine-line for @code{main}
611: calling @code{a}:
612:
613: @table @code
614: @item self
615: An estimate of the amount of time spent directly within @code{a}
616: when @code{a} was called from @code{main}.
617:
618: @item children
619: An estimate of the amount of time spent in subroutines of @code{a}
620: when @code{a} was called from @code{main}.
621:
622: The sum of the @code{self} and @code{children} fields is an estimate
623: of the total time spent in calls to @code{a} from @code{main}.
624:
625: @item called
626: Two numbers, the number of calls to @code{a} from @code{main}
627: followed by the total number of nonrecursive calls to @code{a}.
628:
629: @item name
630: The name of the subroutine of @code{a} to which this line applies,
631: followed by the subroutine's index number. If the subroutine is
632: a function omitted from the call graph, it has no index number,
633: so @samp{[not printed]} appears instead.
634:
635: If the caller is part of a recursion cycle, the cycle number is
636: printed between the name and the index number.
637: @end table
638:
639: @node Cycles,, Subroutines, Call Graph
640: @section How Mutually Recursive Functions Are Described
641: @cindex cycle
642: @cindex recursion cycle
643:
644: The graph may be complicated by the presence of @dfn{cycles of
645: recursion} in the call graph. A cycle exists if a function calls
646: another function that (directly or indirectly) calls (or appears to
647: call) the original function. For example: if @code{a} calls @code{b},
648: and @code{b} calls @code{a}, then @code{a} and @code{b} form a cycle.
649:
650: Whenever there are call-paths both ways between a pair of functions, they
651: belong to the same cycle. If @code{a} and @code{b} call each other and
652: @code{b} and @code{c} call each other, all three make one cycle. Note that
653: even if @code{b} only calls @code{a} if it was not called from @code{a},
654: @code{gprof} cannot determine this, so @code{a} and @code{b} are still
655: considered a cycle.
656:
657: The cycles are numbered with consecutive integers. When a function
658: belongs to a cycle, each time the function name appears in the call graph
659: it is followed by @samp{<cycle @var{number}>}.
660:
661: The reason cycles matter is that they make the time values in the call
662: graph paradoxical. The ``time spent in children'' of @code{a} should
663: include the time spent in its subroutine @code{b} and in @code{b}'s
664: subroutines---but one of @code{b}'s subroutines is @code{a}! How much of
665: @code{a}'s time should be included in the children of @code{a}, when
666: @code{a} is indirectly recursive?
667:
668: The way @code{gprof} resolves this paradox is by creating a single entry
669: for the cycle as a whole. The primary line of this entry describes the
670: total time spent directly in the functions of the cycle. The
671: ``subroutines'' of the cycle are the individual functions of the cycle, and
672: all other functions that were called directly by them. The ``callers'' of
673: the cycle are the functions, outside the cycle, that called functions in
674: the cycle.
675:
676: Here is a portion of the call graph which shows a cycle containing
677: functions @code{a} and @code{b}. The cycle was entered by a call to
678: @code{a} from @code{main}; both @code{a} and @code{b} called @code{c}.@refill
679:
680: @example
681: index % time self children called name
682: ----------------------------------------
683: 1.77 0 1/1 main [2]
684: [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3]
685: 1.02 0 3 b <cycle 1> [4]
686: 0.75 0 2 a <cycle 1> [5]
687: ----------------------------------------
688: 3 a <cycle 1> [5]
689: [4] 52.85 1.02 0 0 b <cycle 1> [4]
690: 2 a <cycle 1> [5]
691: 0 0 3/6 c [6]
692: ----------------------------------------
693: 1.77 0 1/1 main [2]
694: 2 b <cycle 1> [4]
695: [5] 38.86 0.75 0 1 a <cycle 1> [5]
696: 3 b <cycle 1> [4]
697: 0 0 3/6 c [6]
698: ----------------------------------------
699: @end example
700:
701: @noindent
702: (The entire call graph for this program contains in addition an entry for
703: @code{main}, which calls @code{a}, and an entry for @code{c}, with callers
704: @code{a} and @code{b}.)
705:
706: @example
707: index % time self children called name
708: <spontaneous>
709: [1] 100.00 0 1.93 0 start [1]
710: 0.16 1.77 1/1 main [2]
711: ----------------------------------------
712: 0.16 1.77 1/1 start [1]
713: [2] 100.00 0.16 1.77 1 main [2]
714: 1.77 0 1/1 a <cycle 1> [5]
715: ----------------------------------------
716: 1.77 0 1/1 main [2]
717: [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3]
718: 1.02 0 3 b <cycle 1> [4]
719: 0.75 0 2 a <cycle 1> [5]
720: 0 0 6/6 c [6]
721: ----------------------------------------
722: 3 a <cycle 1> [5]
723: [4] 52.85 1.02 0 0 b <cycle 1> [4]
724: 2 a <cycle 1> [5]
725: 0 0 3/6 c [6]
726: ----------------------------------------
727: 1.77 0 1/1 main [2]
728: 2 b <cycle 1> [4]
729: [5] 38.86 0.75 0 1 a <cycle 1> [5]
730: 3 b <cycle 1> [4]
731: 0 0 3/6 c [6]
732: ----------------------------------------
733: 0 0 3/6 b <cycle 1> [4]
734: 0 0 3/6 a <cycle 1> [5]
735: [6] 0.00 0 0 6 c [6]
736: ----------------------------------------
737: @end example
738:
739: The @code{self} field of the cycle's primary line is the total time
740: spent in all the functions of the cycle. It equals the sum of the
741: @code{self} fields for the individual functions in the cycle, found
742: in the entry in the subroutine lines for these functions.
743:
744: The @code{children} fields of the cycle's primary line and subroutine lines
745: count only subroutines outside the cycle. Even though @code{a} calls
746: @code{b}, the time spent in those calls to @code{b} is not counted in
747: @code{a}'s @code{children} time. Thus, we do not encounter the problem of
748: what to do when the time in those calls to @code{b} includes indirect
749: recursive calls back to @code{a}.
750:
751: The @code{children} field of a caller-line in the cycle's entry estimates
752: the amount of time spent @emph{in the whole cycle}, and its other
753: subroutines, on the times when that caller called a function in the cycle.
754:
755: The @code{calls} field in the primary line for the cycle has two numbers:
756: first, the number of times functions in the cycle were called by functions
757: outside the cycle; second, the number of times they were called by
758: functions in the cycle (including times when a function in the cycle calls
759: itself). This is a generalization of the usual split into nonrecursive and
760: recursive calls.
761:
762: The @code{calls} field of a subroutine-line for a cycle member in the
763: cycle's entry says how many time that function was called from functions in
764: the cycle. The total of all these is the second number in the primary line's
765: @code{calls} field.
766:
767: In the individual entry for a function in a cycle, the other functions in
768: the same cycle can appear as subroutines and as callers. These lines show
769: how many times each function in the cycle called or was called from each other
770: function in the cycle. The @code{self} and @code{children} fields in these
771: lines are blank because of the difficulty of defining meanings for them
772: when recursion is going on.
773:
774: @node Implementation, Sampling Error, Call Graph, Top
775: @chapter Implementation of Profiling
776:
777: Profiling works by changing how every function in your program is compiled
778: so that when it is called, it will stash away some information about where
779: it was called from. From this, the profiler can figure out what function
780: called it, and can count how many times it was called. This change is made
781: by the compiler when your program is compiled with the @samp{-pg} option.
782:
783: Profiling also involves watching your program as it runs, and keeping a
784: histogram of where the program counter happens to be every now and then.
785: Typically the program counter is looked at around 100 times per second of
786: run time, but the exact frequency may vary from system to system.
787:
788: A special startup routine allocates memory for the histogram and sets up a
789: clock signal handler to make entries in it. Use of this special startup
790: routine is one of the effects of using @samp{cc -pg} to link. The startup
791: file also includes an @code{exit} function which is responsible for writing
792: the file @file{gmon.out}.
793:
794: Number-of-calls information for library routines is collected by using a
795: special version of the C library. The programs in it are the same as in
796: the usual C library, but they were compiled with @samp{-pg}. If you link
797: your program with @samp{cc -pg}, it automatically uses the profiling
798: version of the library.
799:
800: The output @code{gprof} gives no indication of parts of your program that
801: are limited by I/O or swapping bandwidth. This is because samples of the
802: program counter are taken at fixed intervals of run time. Therefore, the
803: time measurements in @code{gprof} output say nothing about time that your
804: program was not running. For example, a part of the program that creates
805: so much data that it cannot all fit in physical memory at once may run very
806: slowly due to thrashing, but @code{gprof} will say it uses little time. On
807: the other hand, sampling by run time has the advantage that the amount of
808: load due to other users won't directly affect the output you get.
809:
810: @node Sampling Error, Assumptions, Implementation, Top
811: @chapter Statistical Inaccuracy of @code{gprof} Output
812:
813: The run-time figures that @code{gprof} gives you are based on a sampling
814: process, so they are subject to statistical inaccuracy. If a function runs
815: only a small amount of time, so that on the average the sampling process
816: ought to catch that function in the act only once, there is a pretty good
817: chance it will actually find that function zero times, or twice.
818:
819: By contrast, the number-of-calls figures are derived by counting, not
820: sampling. They are completely accurate and will not vary from run to run
821: if your program is deterministic.
822:
823: The @dfn{sampling period} that is printed at the beginning of the flat
824: profile says how often samples are taken. The rule of thumb is that a
825: run-time figure is accurate if it is considerably bigger than the sampling
826: period.
827:
828: The actual amount of error is usually more than one sampling period. In
829: fact, if a value is @var{n} times the sampling period, the @emph{expected}
830: error in it is the square-root of @var{n} sampling periods. If the
831: sampling period is 0.01 seconds and @code{foo}'s run-time is 1 second, the
832: expected error in @code{foo}'s run-time is 0.1 seconds. It is likely to
833: vary this much @emph{on the average} from one profiling run to the next.
834: (@emph{Sometimes} it will vary more.)
835:
836: This does not mean that a small run-time figure is devoid of information.
837: If the program's @emph{total} run-time is large, a small run-time for one
838: function does tell you that that function used an insignificant fraction of
839: the whole program's time. Usually this means it is not worth optimizing.
840:
841: One way to get more accuracy is to give your program more (but similar)
842: input data so it will take longer. Another way is to combine the data from
843: several runs, using the @samp{-s} option of @code{gprof}. Here is how:
844:
845: @enumerate
846: @item
847: Run your program once.
848:
849: @item
850: Issue the command @samp{mv gmon.out gmon.sum}.
851:
852: @item
853: Run your program again, the same as before.
854:
855: @item
856: Merge the new data in @file{gmon.out} into @file{gmon.sum} with this command:
857:
858: @example
859: gprof -s @var{executable-file} gmon.out gmon.sum
860: @end example
861:
862: @item
863: Repeat the last two steps as often as you wish.
864:
865: @item
866: Analyze the cumulative data using this command:
867:
868: @example
869: gprof @var{executable-file} gmon.sum > @var{output-file}
870: @end example
871: @end enumerate
872:
873: @node Assumptions, Incompatibilities, Sampling Error, Top
874: @chapter Estimating @code{children} Times Uses an Assumption
875:
876: Some of the figures in the call graph are estimates---for example, the
877: @code{children} time values and all the the time figures in caller and
878: subroutine lines.
879:
880: There is no direct information about these measurements in the profile
881: data itself. Instead, @code{gprof} estimates them by making an assumption
882: about your program that might or might not be true.
883:
884: The assumption made is that the average time spent in each call to any
885: function @code{foo} is not correlated with who called @code{foo}. If
886: @code{foo} used 5 seconds in all, and 2/5 of the calls to @code{foo} came
887: from @code{a}, then @code{foo} contributes 2 seconds to @code{a}'s
888: @code{children} time, by assumption.
889:
890: This assumption is usually true enough, but for some programs it is far
891: from true. Suppose that @code{foo} returns very quickly when its argument
892: is zero; suppose that @code{a} always passes zero as an argument, while
893: other callers of @code{foo} pass other arguments. In this program, all the
894: time spent in @code{foo} is in the calls from callers other than @code{a}.
895: But @code{gprof} has no way of knowing this; it will blindly and
896: incorrectly charge 2 seconds of time in @code{foo} to the children of
897: @code{a}.
898:
899: We hope some day to put more complete data into @file{gmon.out}, so that
900: this assumption is no longer needed, if we can figure out how. For the
901: nonce, the estimated figures are usually more useful than misleading.
902:
903: @node Incompatibilities, , Assumptions, Top
904: @chapter Incompatibilities with Unix @code{gprof}
905:
906: GNU @code{gprof} and Berkeley Unix @code{gprof} use the same data file
907: @file{gmon.out}, and provide essentially the same information. But there a
908: few differences.@refill
909:
910: GNU @code{gprof} does not support the @samp{-c} option which prints a
911: static call graph based on reading the machine language of your
912: program. We think that program cross-references ought to be based on
913: the source files, which can be analyzed in a machine-independent
914: fashion.@refill
915:
916: For a recursive function, Unix @code{gprof} lists the function as a parent
917: and as a child, with a @code{calls} field that lists the number of
918: recursive calls. GNU @code{gprof} omits these lines and puts the number of
919: recursive calls in the primary line.
920:
921: When a function is suppressed from the call graph with @samp{-e}, GNU
922: @code{gprof} still lists it as a subroutine of functions that call it.
923:
924: The function names printed in GNU @code{gprof} output do not include
925: the leading underscores that are added internally to the front of all
926: C identifiers on many operating systems.
927:
928: The blurbs, field widths, and output formats are different. GNU
929: @code{gprof} prints blurbs after the tables, so that you can see the
930: tables without skipping the blurbs.
931:
932: @contents
933: @bye
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.