gprof.texi revision 1.1.1.7
1\input texinfo @c -*-texinfo-*- 2@setfilename gprof.info 3@c Copyright (C) 1988-2022 Free Software Foundation, Inc. 4@settitle GNU gprof 5@setchapternewpage odd 6 7@c man begin INCLUDE 8@include bfdver.texi 9@c man end 10 11@ifnottex 12@c This is a dir.info fragment to support semi-automated addition of 13@c manuals to an info tree. zoo@cygnus.com is developing this facility. 14@dircategory Software development 15@direntry 16* gprof: (gprof). Profiling your program's execution 17@end direntry 18@end ifnottex 19 20@copying 21This file documents the gprof profiler of the GNU system. 22 23@c man begin COPYRIGHT 24Copyright @copyright{} 1988-2022 Free Software Foundation, Inc. 25 26Permission is granted to copy, distribute and/or modify this document 27under the terms of the GNU Free Documentation License, Version 1.3 28or any later version published by the Free Software Foundation; 29with no Invariant Sections, with no Front-Cover Texts, and with no 30Back-Cover Texts. A copy of the license is included in the 31section entitled ``GNU Free Documentation License''. 32 33@c man end 34@end copying 35 36@finalout 37@smallbook 38 39@titlepage 40@title GNU gprof 41@subtitle The @sc{gnu} Profiler 42@ifset VERSION_PACKAGE 43@subtitle @value{VERSION_PACKAGE} 44@end ifset 45@subtitle Version @value{VERSION} 46@author Jay Fenlason and Richard Stallman 47 48@page 49 50This manual describes the @sc{gnu} profiler, @code{gprof}, and how you 51can use it to determine which parts of a program are taking most of the 52execution time. We assume that you know how to write, compile, and 53execute programs. @sc{gnu} @code{gprof} was written by Jay Fenlason. 54Eric S. Raymond made some minor corrections and additions in 2003. 55 56@vskip 0pt plus 1filll 57Copyright @copyright{} 1988-2022 Free Software Foundation, Inc. 58 59 Permission is granted to copy, distribute and/or modify this document 60 under the terms of the GNU Free Documentation License, Version 1.3 61 or any later version published by the Free Software Foundation; 62 with no Invariant Sections, with no Front-Cover Texts, and with no 63 Back-Cover Texts. A copy of the license is included in the 64 section entitled ``GNU Free Documentation License''. 65 66@end titlepage 67@contents 68 69@ifnottex 70@node Top 71@top Profiling a Program: Where Does It Spend Its Time? 72 73This manual describes the @sc{gnu} profiler, @code{gprof}, and how you 74can use it to determine which parts of a program are taking most of the 75execution time. We assume that you know how to write, compile, and 76execute programs. @sc{gnu} @code{gprof} was written by Jay Fenlason. 77 78This manual is for @code{gprof} 79@ifset VERSION_PACKAGE 80@value{VERSION_PACKAGE} 81@end ifset 82version @value{VERSION}. 83 84This document is distributed under the terms of the GNU Free 85Documentation License version 1.3. A copy of the license is included 86in the section entitled ``GNU Free Documentation License''. 87 88@menu 89* Introduction:: What profiling means, and why it is useful. 90 91* Compiling:: How to compile your program for profiling. 92* Executing:: Executing your program to generate profile data 93* Invoking:: How to run @code{gprof}, and its options 94 95* Output:: Interpreting @code{gprof}'s output 96 97* Inaccuracy:: Potential problems you should be aware of 98* How do I?:: Answers to common questions 99* Incompatibilities:: (between @sc{gnu} @code{gprof} and Unix @code{gprof}.) 100* Details:: Details of how profiling is done 101* GNU Free Documentation License:: GNU Free Documentation License 102@end menu 103@end ifnottex 104 105@node Introduction 106@chapter Introduction to Profiling 107 108@ifset man 109@c man title gprof display call graph profile data 110 111@smallexample 112@c man begin SYNOPSIS 113gprof [ -[abcDhilLrsTvwxyz] ] [ -[ACeEfFJnNOpPqQRStZ][@var{name}] ] 114 [ -I @var{dirs} ] [ -d[@var{num}] ] [ -k @var{from/to} ] 115 [ -m @var{min-count} ] [ -R @var{map_file} ] [ -t @var{table-length} ] 116 [ --[no-]annotated-source[=@var{name}] ] 117 [ --[no-]exec-counts[=@var{name}] ] 118 [ --[no-]flat-profile[=@var{name}] ] [ --[no-]graph[=@var{name}] ] 119 [ --[no-]time=@var{name}] [ --all-lines ] [ --brief ] 120 [ --debug[=@var{level}] ] [ --function-ordering ] 121 [ --file-ordering @var{map_file} ] [ --directory-path=@var{dirs} ] 122 [ --display-unused-functions ] [ --file-format=@var{name} ] 123 [ --file-info ] [ --help ] [ --line ] [ --inline-file-names ] 124 [ --min-count=@var{n} ] [ --no-static ] [ --print-path ] 125 [ --separate-files ] [ --static-call-graph ] [ --sum ] 126 [ --table-length=@var{len} ] [ --traditional ] [ --version ] 127 [ --width=@var{n} ] [ --ignore-non-functions ] 128 [ --demangle[=@var{STYLE}] ] [ --no-demangle ] 129 [--external-symbol-table=name] 130 [ @var{image-file} ] [ @var{profile-file} @dots{} ] 131@c man end 132@end smallexample 133 134@c man begin DESCRIPTION 135@code{gprof} produces an execution profile of C, Pascal, or Fortran77 136programs. The effect of called routines is incorporated in the profile 137of each caller. The profile data is taken from the call graph profile file 138(@file{gmon.out} default) which is created by programs 139that are compiled with the @samp{-pg} option of 140@code{cc}, @code{pc}, and @code{f77}. 141The @samp{-pg} option also links in versions of the library routines 142that are compiled for profiling. @code{Gprof} reads the given object 143file (the default is @code{a.out}) and establishes the relation between 144its symbol table and the call graph profile from @file{gmon.out}. 145If more than one profile file is specified, the @code{gprof} 146output shows the sum of the profile information in the given profile files. 147 148@code{Gprof} calculates the amount of time spent in each routine. 149Next, these times are propagated along the edges of the call graph. 150Cycles are discovered, and calls into a cycle are made to share the time 151of the cycle. 152 153@c man end 154 155@c man begin BUGS 156The granularity of the sampling is shown, but remains 157statistical at best. 158We assume that the time for each execution of a function 159can be expressed by the total time for the function divided 160by the number of times the function is called. 161Thus the time propagated along the call graph arcs to the function's 162parents is directly proportional to the number of times that 163arc is traversed. 164 165Parents that are not themselves profiled will have the time of 166their profiled children propagated to them, but they will appear 167to be spontaneously invoked in the call graph listing, and will 168not have their time propagated further. 169Similarly, signal catchers, even though profiled, will appear 170to be spontaneous (although for more obscure reasons). 171Any profiled children of signal catchers should have their times 172propagated properly, unless the signal catcher was invoked during 173the execution of the profiling routine, in which case all is lost. 174 175The profiled program must call @code{exit}(2) 176or return normally for the profiling information to be saved 177in the @file{gmon.out} file. 178@c man end 179 180@c man begin FILES 181@table @code 182@item @file{a.out} 183the namelist and text space. 184@item @file{gmon.out} 185dynamic call graph and profile. 186@item @file{gmon.sum} 187summarized dynamic call graph and profile. 188@end table 189@c man end 190 191@c man begin SEEALSO 192monitor(3), profil(2), cc(1), prof(1), and the Info entry for @file{gprof}. 193 194``An Execution Profiler for Modular Programs'', 195by S. Graham, P. Kessler, M. McKusick; 196Software - Practice and Experience, 197Vol. 13, pp. 671-685, 1983. 198 199``gprof: A Call Graph Execution Profiler'', 200by S. Graham, P. Kessler, M. McKusick; 201Proceedings of the SIGPLAN '82 Symposium on Compiler Construction, 202SIGPLAN Notices, Vol. 17, No 6, pp. 120-126, June 1982. 203@c man end 204@end ifset 205 206Profiling allows you to learn where your program spent its time and which 207functions called which other functions while it was executing. This 208information can show you which pieces of your program are slower than you 209expected, and might be candidates for rewriting to make your program 210execute faster. It can also tell you which functions are being called more 211or less often than you expected. This may help you spot bugs that had 212otherwise been unnoticed. 213 214Since the profiler uses information collected during the actual execution 215of your program, it can be used on programs that are too large or too 216complex to analyze by reading the source. However, how your program is run 217will affect the information that shows up in the profile data. If you 218don't use some feature of your program while it is being profiled, no 219profile information will be generated for that feature. 220 221Profiling has several steps: 222 223@itemize @bullet 224@item 225You must compile and link your program with profiling enabled. 226@xref{Compiling, ,Compiling a Program for Profiling}. 227 228@item 229You must execute your program to generate a profile data file. 230@xref{Executing, ,Executing the Program}. 231 232@item 233You must run @code{gprof} to analyze the profile data. 234@xref{Invoking, ,@code{gprof} Command Summary}. 235@end itemize 236 237The next three chapters explain these steps in greater detail. 238 239@c man begin DESCRIPTION 240 241Several forms of output are available from the analysis. 242 243The @dfn{flat profile} shows how much time your program spent in each function, 244and how many times that function was called. If you simply want to know 245which functions burn most of the cycles, it is stated concisely here. 246@xref{Flat Profile, ,The Flat Profile}. 247 248The @dfn{call graph} shows, for each function, which functions called it, which 249other functions it called, and how many times. There is also an estimate 250of how much time was spent in the subroutines of each function. This can 251suggest places where you might try to eliminate function calls that use a 252lot of time. @xref{Call Graph, ,The Call Graph}. 253 254The @dfn{annotated source} listing is a copy of the program's 255source code, labeled with the number of times each line of the 256program was executed. @xref{Annotated Source, ,The Annotated Source 257Listing}. 258@c man end 259 260To better understand how profiling works, you may wish to read 261a description of its implementation. 262@xref{Implementation, ,Implementation of Profiling}. 263 264@node Compiling 265@chapter Compiling a Program for Profiling 266 267The first step in generating profile information for your program is 268to compile and link it with profiling enabled. 269 270To compile a source file for profiling, specify the @samp{-pg} option when 271you run the compiler. (This is in addition to the options you normally 272use.) 273 274To link the program for profiling, if you use a compiler such as @code{cc} 275to do the linking, simply specify @samp{-pg} in addition to your usual 276options. The same option, @samp{-pg}, alters either compilation or linking 277to do what is necessary for profiling. Here are examples: 278 279@example 280cc -g -c myprog.c utils.c -pg 281cc -o myprog myprog.o utils.o -pg 282@end example 283 284The @samp{-pg} option also works with a command that both compiles and links: 285 286@example 287cc -o myprog myprog.c utils.c -g -pg 288@end example 289 290Note: The @samp{-pg} option must be part of your compilation options 291as well as your link options. If it is not then no call-graph data 292will be gathered and when you run @code{gprof} you will get an error 293message like this: 294 295@example 296gprof: gmon.out file is missing call-graph data 297@end example 298 299If you add the @samp{-Q} switch to suppress the printing of the call 300graph data you will still be able to see the time samples: 301 302@example 303Flat profile: 304 305Each sample counts as 0.01 seconds. 306 % cumulative self self total 307 time seconds seconds calls Ts/call Ts/call name 308 44.12 0.07 0.07 zazLoop 309 35.29 0.14 0.06 main 310 20.59 0.17 0.04 bazMillion 311@end example 312 313If you run the linker @code{ld} directly instead of through a compiler 314such as @code{cc}, you may have to specify a profiling startup file 315@file{gcrt0.o} as the first input file instead of the usual startup 316file @file{crt0.o}. In addition, you would probably want to 317specify the profiling C library, @file{libc_p.a}, by writing 318@samp{-lc_p} instead of the usual @samp{-lc}. This is not absolutely 319necessary, but doing this gives you number-of-calls information for 320standard library functions such as @code{read} and @code{open}. For 321example: 322 323@example 324ld -o myprog /lib/gcrt0.o myprog.o utils.o -lc_p 325@end example 326 327If you are running the program on a system which supports shared 328libraries you may run into problems with the profiling support code in 329a shared library being called before that library has been fully 330initialised. This is usually detected by the program encountering a 331segmentation fault as soon as it is run. The solution is to link 332against a static version of the library containing the profiling 333support code, which for @code{gcc} users can be done via the 334@samp{-static} or @samp{-static-libgcc} command-line option. For 335example: 336 337@example 338gcc -g -pg -static-libgcc myprog.c utils.c -o myprog 339@end example 340 341If you compile only some of the modules of the program with @samp{-pg}, you 342can still profile the program, but you won't get complete information about 343the modules that were compiled without @samp{-pg}. The only information 344you get for the functions in those modules is the total time spent in them; 345there is no record of how many times they were called, or from where. This 346will not affect the flat profile (except that the @code{calls} field for 347the functions will be blank), but will greatly reduce the usefulness of the 348call graph. 349 350If you wish to perform line-by-line profiling you should use the 351@code{gcov} tool instead of @code{gprof}. See that tool's manual or 352info pages for more details of how to do this. 353 354Note, older versions of @code{gcc} produce line-by-line profiling 355information that works with @code{gprof} rather than @code{gcov} so 356there is still support for displaying this kind of information in 357@code{gprof}. @xref{Line-by-line, ,Line-by-line Profiling}. 358 359It also worth noting that @code{gcc} implements a 360@samp{-finstrument-functions} command-line option which will insert 361calls to special user supplied instrumentation routines at the entry 362and exit of every function in their program. This can be used to 363implement an alternative profiling scheme. 364 365@node Executing 366@chapter Executing the Program 367 368Once the program is compiled for profiling, you must run it in order to 369generate the information that @code{gprof} needs. Simply run the program 370as usual, using the normal arguments, file names, etc. The program should 371run normally, producing the same output as usual. It will, however, run 372somewhat slower than normal because of the time spent collecting and 373writing the profile data. 374 375The way you run the program---the arguments and input that you give 376it---may have a dramatic effect on what the profile information shows. The 377profile data will describe the parts of the program that were activated for 378the particular input you use. For example, if the first command you give 379to your program is to quit, the profile data will show the time used in 380initialization and in cleanup, but not much else. 381 382Your program will write the profile data into a file called @file{gmon.out} 383just before exiting. If there is already a file called @file{gmon.out}, 384its contents are overwritten. You can rename the file afterwards if you 385are concerned that it may be overwritten. If your system libc allows you 386may be able to write the profile data under a different name. Set the 387GMON_OUT_PREFIX environment variable; this name will be appended with 388the PID of the running program. 389 390In order to write the @file{gmon.out} file properly, your program must exit 391normally: by returning from @code{main} or by calling @code{exit}. Calling 392the low-level function @code{_exit} does not write the profile data, and 393neither does abnormal termination due to an unhandled signal. 394 395The @file{gmon.out} file is written in the program's @emph{current working 396directory} at the time it exits. This means that if your program calls 397@code{chdir}, the @file{gmon.out} file will be left in the last directory 398your program @code{chdir}'d to. If you don't have permission to write in 399this directory, the file is not written, and you will get an error message. 400 401Older versions of the @sc{gnu} profiling library may also write a file 402called @file{bb.out}. This file, if present, contains an human-readable 403listing of the basic-block execution counts. Unfortunately, the 404appearance of a human-readable @file{bb.out} means the basic-block 405counts didn't get written into @file{gmon.out}. 406The Perl script @code{bbconv.pl}, included with the @code{gprof} 407source distribution, will convert a @file{bb.out} file into 408a format readable by @code{gprof}. Invoke it like this: 409 410@smallexample 411bbconv.pl < bb.out > @var{bh-data} 412@end smallexample 413 414This translates the information in @file{bb.out} into a form that 415@code{gprof} can understand. But you still need to tell @code{gprof} 416about the existence of this translated information. To do that, include 417@var{bb-data} on the @code{gprof} command line, @emph{along with 418@file{gmon.out}}, like this: 419 420@smallexample 421gprof @var{options} @var{executable-file} gmon.out @var{bb-data} [@var{yet-more-profile-data-files}@dots{}] [> @var{outfile}] 422@end smallexample 423 424@node Invoking 425@chapter @code{gprof} Command Summary 426 427After you have a profile data file @file{gmon.out}, you can run @code{gprof} 428to interpret the information in it. The @code{gprof} program prints a 429flat profile and a call graph on standard output. Typically you would 430redirect the output of @code{gprof} into a file with @samp{>}. 431 432You run @code{gprof} like this: 433 434@smallexample 435gprof @var{options} [@var{executable-file} [@var{profile-data-files}@dots{}]] [> @var{outfile}] 436@end smallexample 437 438@noindent 439Here square-brackets indicate optional arguments. 440 441If you omit the executable file name, the file @file{a.out} is used. If 442you give no profile data file name, the file @file{gmon.out} is used. If 443any file is not in the proper format, or if the profile data file does not 444appear to belong to the executable file, an error message is printed. 445 446You can give more than one profile data file by entering all their names 447after the executable file name; then the statistics in all the data files 448are summed together. 449 450The order of these options does not matter. 451 452@menu 453* Output Options:: Controlling @code{gprof}'s output style 454* Analysis Options:: Controlling how @code{gprof} analyzes its data 455* Miscellaneous Options:: 456* Deprecated Options:: Options you no longer need to use, but which 457 have been retained for compatibility 458* Symspecs:: Specifying functions to include or exclude 459@end menu 460 461@node Output Options 462@section Output Options 463 464@c man begin OPTIONS 465These options specify which of several output formats 466@code{gprof} should produce. 467 468Many of these options take an optional @dfn{symspec} to specify 469functions to be included or excluded. These options can be 470specified multiple times, with different symspecs, to include 471or exclude sets of symbols. @xref{Symspecs, ,Symspecs}. 472 473Specifying any of these options overrides the default (@samp{-p -q}), 474which prints a flat profile and call graph analysis 475for all functions. 476 477@table @code 478 479@item -A[@var{symspec}] 480@itemx --annotated-source[=@var{symspec}] 481The @samp{-A} option causes @code{gprof} to print annotated source code. 482If @var{symspec} is specified, print output only for matching symbols. 483@xref{Annotated Source, ,The Annotated Source Listing}. 484 485@item -b 486@itemx --brief 487If the @samp{-b} option is given, @code{gprof} doesn't print the 488verbose blurbs that try to explain the meaning of all of the fields in 489the tables. This is useful if you intend to print out the output, or 490are tired of seeing the blurbs. 491 492@item -C[@var{symspec}] 493@itemx --exec-counts[=@var{symspec}] 494The @samp{-C} option causes @code{gprof} to 495print a tally of functions and the number of times each was called. 496If @var{symspec} is specified, print tally only for matching symbols. 497 498If the profile data file contains basic-block count records, specifying 499the @samp{-l} option, along with @samp{-C}, will cause basic-block 500execution counts to be tallied and displayed. 501 502@item -i 503@itemx --file-info 504The @samp{-i} option causes @code{gprof} to display summary information 505about the profile data file(s) and then exit. The number of histogram, 506call graph, and basic-block count records is displayed. 507 508@item -I @var{dirs} 509@itemx --directory-path=@var{dirs} 510The @samp{-I} option specifies a list of search directories in 511which to find source files. Environment variable @var{GPROF_PATH} 512can also be used to convey this information. 513Used mostly for annotated source output. 514 515@item -J[@var{symspec}] 516@itemx --no-annotated-source[=@var{symspec}] 517The @samp{-J} option causes @code{gprof} not to 518print annotated source code. 519If @var{symspec} is specified, @code{gprof} prints annotated source, 520but excludes matching symbols. 521 522@item -L 523@itemx --print-path 524Normally, source filenames are printed with the path 525component suppressed. The @samp{-L} option causes @code{gprof} 526to print the full pathname of 527source filenames, which is determined 528from symbolic debugging information in the image file 529and is relative to the directory in which the compiler 530was invoked. 531 532@item -p[@var{symspec}] 533@itemx --flat-profile[=@var{symspec}] 534The @samp{-p} option causes @code{gprof} to print a flat profile. 535If @var{symspec} is specified, print flat profile only for matching symbols. 536@xref{Flat Profile, ,The Flat Profile}. 537 538@item -P[@var{symspec}] 539@itemx --no-flat-profile[=@var{symspec}] 540The @samp{-P} option causes @code{gprof} to suppress printing a flat profile. 541If @var{symspec} is specified, @code{gprof} prints a flat profile, 542but excludes matching symbols. 543 544@item -q[@var{symspec}] 545@itemx --graph[=@var{symspec}] 546The @samp{-q} option causes @code{gprof} to print the call graph analysis. 547If @var{symspec} is specified, print call graph only for matching symbols 548and their children. 549@xref{Call Graph, ,The Call Graph}. 550 551@item -Q[@var{symspec}] 552@itemx --no-graph[=@var{symspec}] 553The @samp{-Q} option causes @code{gprof} to suppress printing the 554call graph. 555If @var{symspec} is specified, @code{gprof} prints a call graph, 556but excludes matching symbols. 557 558@item -t 559@itemx --table-length=@var{num} 560The @samp{-t} option causes the @var{num} most active source lines in 561each source file to be listed when source annotation is enabled. The 562default is 10. 563 564@item -y 565@itemx --separate-files 566This option affects annotated source output only. 567Normally, @code{gprof} prints annotated source files 568to standard-output. If this option is specified, 569annotated source for a file named @file{path/@var{filename}} 570is generated in the file @file{@var{filename}-ann}. If the underlying 571file system would truncate @file{@var{filename}-ann} so that it 572overwrites the original @file{@var{filename}}, @code{gprof} generates 573annotated source in the file @file{@var{filename}.ann} instead (if the 574original file name has an extension, that extension is @emph{replaced} 575with @file{.ann}). 576 577@item -Z[@var{symspec}] 578@itemx --no-exec-counts[=@var{symspec}] 579The @samp{-Z} option causes @code{gprof} not to 580print a tally of functions and the number of times each was called. 581If @var{symspec} is specified, print tally, but exclude matching symbols. 582 583@item -r 584@itemx --function-ordering 585The @samp{--function-ordering} option causes @code{gprof} to print a 586suggested function ordering for the program based on profiling data. 587This option suggests an ordering which may improve paging, tlb and 588cache behavior for the program on systems which support arbitrary 589ordering of functions in an executable. 590 591The exact details of how to force the linker to place functions 592in a particular order is system dependent and out of the scope of this 593manual. 594 595@item -R @var{map_file} 596@itemx --file-ordering @var{map_file} 597The @samp{--file-ordering} option causes @code{gprof} to print a 598suggested .o link line ordering for the program based on profiling data. 599This option suggests an ordering which may improve paging, tlb and 600cache behavior for the program on systems which do not support arbitrary 601ordering of functions in an executable. 602 603Use of the @samp{-a} argument is highly recommended with this option. 604 605The @var{map_file} argument is a pathname to a file which provides 606function name to object file mappings. The format of the file is similar to 607the output of the program @code{nm}. 608 609@smallexample 610@group 611c-parse.o:00000000 T yyparse 612c-parse.o:00000004 C yyerrflag 613c-lang.o:00000000 T maybe_objc_method_name 614c-lang.o:00000000 T print_lang_statistics 615c-lang.o:00000000 T recognize_objc_keyword 616c-decl.o:00000000 T print_lang_identifier 617c-decl.o:00000000 T print_lang_type 618@dots{} 619 620@end group 621@end smallexample 622 623To create a @var{map_file} with @sc{gnu} @code{nm}, type a command like 624@kbd{nm --extern-only --defined-only -v --print-file-name program-name}. 625 626@item -T 627@itemx --traditional 628The @samp{-T} option causes @code{gprof} to print its output in 629``traditional'' BSD style. 630 631@item -w @var{width} 632@itemx --width=@var{width} 633Sets width of output lines to @var{width}. 634Currently only used when printing the function index at the bottom 635of the call graph. 636 637@item -x 638@itemx --all-lines 639This option affects annotated source output only. 640By default, only the lines at the beginning of a basic-block 641are annotated. If this option is specified, every line in 642a basic-block is annotated by repeating the annotation for the 643first line. This behavior is similar to @code{tcov}'s @samp{-a}. 644 645@item --demangle[=@var{style}] 646@itemx --no-demangle 647These options control whether C++ symbol names should be demangled when 648printing output. The default is to demangle symbols. The 649@code{--no-demangle} option may be used to turn off demangling. Different 650compilers have different mangling styles. The optional demangling style 651argument can be used to choose an appropriate demangling style for your 652compiler. 653@end table 654 655@node Analysis Options 656@section Analysis Options 657 658@table @code 659 660@item -a 661@itemx --no-static 662The @samp{-a} option causes @code{gprof} to suppress the printing of 663statically declared (private) functions. (These are functions whose 664names are not listed as global, and which are not visible outside the 665file/function/block where they were defined.) Time spent in these 666functions, calls to/from them, etc., will all be attributed to the 667function that was loaded directly before it in the executable file. 668@c This is compatible with Unix @code{gprof}, but a bad idea. 669This option affects both the flat profile and the call graph. 670 671@item -c 672@itemx --static-call-graph 673The @samp{-c} option causes the call graph of the program to be 674augmented by a heuristic which examines the text space of the object 675file and identifies function calls in the binary machine code. 676Since normal call graph records are only generated when functions are 677entered, this option identifies children that could have been called, 678but never were. Calls to functions that were not compiled with 679profiling enabled are also identified, but only if symbol table 680entries are present for them. 681Calls to dynamic library routines are typically @emph{not} found 682by this option. 683Parents or children identified via this heuristic 684are indicated in the call graph with call counts of @samp{0}. 685 686@item -D 687@itemx --ignore-non-functions 688The @samp{-D} option causes @code{gprof} to ignore symbols which 689are not known to be functions. This option will give more accurate 690profile data on systems where it is supported (Solaris and HPUX for 691example). 692 693@item -k @var{from}/@var{to} 694The @samp{-k} option allows you to delete from the call graph any arcs from 695symbols matching symspec @var{from} to those matching symspec @var{to}. 696 697@item -l 698@itemx --line 699The @samp{-l} option enables line-by-line profiling, which causes 700histogram hits to be charged to individual source code lines, 701instead of functions. This feature only works with programs compiled 702by older versions of the @code{gcc} compiler. Newer versions of 703@code{gcc} are designed to work with the @code{gcov} tool instead. 704 705If the program was compiled with basic-block counting enabled, 706this option will also identify how many times each line of 707code was executed. 708While line-by-line profiling can help isolate where in a large function 709a program is spending its time, it also significantly increases 710the running time of @code{gprof}, and magnifies statistical 711inaccuracies. 712@xref{Sampling Error, ,Statistical Sampling Error}. 713 714@item --inline-file-names 715This option causes @code{gprof} to print the source file after each 716symbol in both the flat profile and the call graph. The full path to the 717file is printed if used with the @samp{-L} option. 718 719@item -m @var{num} 720@itemx --min-count=@var{num} 721This option affects execution count output only. 722Symbols that are executed less than @var{num} times are suppressed. 723 724@item -n@var{symspec} 725@itemx --time=@var{symspec} 726The @samp{-n} option causes @code{gprof}, in its call graph analysis, 727to only propagate times for symbols matching @var{symspec}. 728 729@item -N@var{symspec} 730@itemx --no-time=@var{symspec} 731The @samp{-n} option causes @code{gprof}, in its call graph analysis, 732not to propagate times for symbols matching @var{symspec}. 733 734@item -S@var{filename} 735@itemx --external-symbol-table=@var{filename} 736The @samp{-S} option causes @code{gprof} to read an external symbol table 737file, such as @file{/proc/kallsyms}, rather than read the symbol table 738from the given object file (the default is @code{a.out}). This is useful 739for profiling kernel modules. 740 741@item -z 742@itemx --display-unused-functions 743If you give the @samp{-z} option, @code{gprof} will mention all 744functions in the flat profile, even those that were never called, and 745that had no time spent in them. This is useful in conjunction with the 746@samp{-c} option for discovering which routines were never called. 747 748@end table 749 750@node Miscellaneous Options 751@section Miscellaneous Options 752 753@table @code 754 755@item -d[@var{num}] 756@itemx --debug[=@var{num}] 757The @samp{-d @var{num}} option specifies debugging options. 758If @var{num} is not specified, enable all debugging. 759@xref{Debugging, ,Debugging @code{gprof}}. 760 761@item -h 762@itemx --help 763The @samp{-h} option prints command line usage. 764 765@item -O@var{name} 766@itemx --file-format=@var{name} 767Selects the format of the profile data files. Recognized formats are 768@samp{auto} (the default), @samp{bsd}, @samp{4.4bsd}, @samp{magic}, and 769@samp{prof} (not yet supported). 770 771@item -s 772@itemx --sum 773The @samp{-s} option causes @code{gprof} to summarize the information 774in the profile data files it read in, and write out a profile data 775file called @file{gmon.sum}, which contains all the information from 776the profile data files that @code{gprof} read in. The file @file{gmon.sum} 777may be one of the specified input files; the effect of this is to 778merge the data in the other input files into @file{gmon.sum}. 779 780Eventually you can run @code{gprof} again without @samp{-s} to analyze the 781cumulative data in the file @file{gmon.sum}. 782 783@item -v 784@itemx --version 785The @samp{-v} flag causes @code{gprof} to print the current version 786number, and then exit. 787 788@end table 789 790@node Deprecated Options 791@section Deprecated Options 792 793These options have been replaced with newer versions that use symspecs. 794 795@table @code 796 797@item -e @var{function_name} 798The @samp{-e @var{function}} option tells @code{gprof} to not print 799information about the function @var{function_name} (and its 800children@dots{}) in the call graph. The function will still be listed 801as a child of any functions that call it, but its index number will be 802shown as @samp{[not printed]}. More than one @samp{-e} option may be 803given; only one @var{function_name} may be indicated with each @samp{-e} 804option. 805 806@item -E @var{function_name} 807The @code{-E @var{function}} option works like the @code{-e} option, but 808time spent in the function (and children who were not called from 809anywhere else), will not be used to compute the percentages-of-time for 810the call graph. More than one @samp{-E} option may be given; only one 811@var{function_name} may be indicated with each @samp{-E} option. 812 813@item -f @var{function_name} 814The @samp{-f @var{function}} option causes @code{gprof} to limit the 815call graph to the function @var{function_name} and its children (and 816their children@dots{}). More than one @samp{-f} option may be given; 817only one @var{function_name} may be indicated with each @samp{-f} 818option. 819 820@item -F @var{function_name} 821The @samp{-F @var{function}} option works like the @code{-f} option, but 822only time spent in the function and its children (and their 823children@dots{}) will be used to determine total-time and 824percentages-of-time for the call graph. More than one @samp{-F} option 825may be given; only one @var{function_name} may be indicated with each 826@samp{-F} option. The @samp{-F} option overrides the @samp{-E} option. 827 828@end table 829 830@c man end 831 832Note that only one function can be specified with each @code{-e}, 833@code{-E}, @code{-f} or @code{-F} option. To specify more than one 834function, use multiple options. For example, this command: 835 836@example 837gprof -e boring -f foo -f bar myprogram > gprof.output 838@end example 839 840@noindent 841lists in the call graph all functions that were reached from either 842@code{foo} or @code{bar} and were not reachable from @code{boring}. 843 844@node Symspecs 845@section Symspecs 846 847Many of the output options allow functions to be included or excluded 848using @dfn{symspecs} (symbol specifications), which observe the 849following syntax: 850 851@example 852 filename_containing_a_dot 853| funcname_not_containing_a_dot 854| linenumber 855| ( [ any_filename ] `:' ( any_funcname | linenumber ) ) 856@end example 857 858Here are some sample symspecs: 859 860@table @samp 861@item main.c 862Selects everything in file @file{main.c}---the 863dot in the string tells @code{gprof} to interpret 864the string as a filename, rather than as 865a function name. To select a file whose 866name does not contain a dot, a trailing colon 867should be specified. For example, @samp{odd:} is 868interpreted as the file named @file{odd}. 869 870@item main 871Selects all functions named @samp{main}. 872 873Note that there may be multiple instances of the same function name 874because some of the definitions may be local (i.e., static). Unless a 875function name is unique in a program, you must use the colon notation 876explained below to specify a function from a specific source file. 877 878Sometimes, function names contain dots. In such cases, it is necessary 879to add a leading colon to the name. For example, @samp{:.mul} selects 880function @samp{.mul}. 881 882In some object file formats, symbols have a leading underscore. 883@code{gprof} will normally not print these underscores. When you name a 884symbol in a symspec, you should type it exactly as @code{gprof} prints 885it in its output. For example, if the compiler produces a symbol 886@samp{_main} from your @code{main} function, @code{gprof} still prints 887it as @samp{main} in its output, so you should use @samp{main} in 888symspecs. 889 890@item main.c:main 891Selects function @samp{main} in file @file{main.c}. 892 893@item main.c:134 894Selects line 134 in file @file{main.c}. 895@end table 896 897@node Output 898@chapter Interpreting @code{gprof}'s Output 899 900@code{gprof} can produce several different output styles, the 901most important of which are described below. The simplest output 902styles (file information, execution count, and function and file ordering) 903are not described here, but are documented with the respective options 904that trigger them. 905@xref{Output Options, ,Output Options}. 906 907@menu 908* Flat Profile:: The flat profile shows how much time was spent 909 executing directly in each function. 910* Call Graph:: The call graph shows which functions called which 911 others, and how much time each function used 912 when its subroutine calls are included. 913* Line-by-line:: @code{gprof} can analyze individual source code lines 914* Annotated Source:: The annotated source listing displays source code 915 labeled with execution counts 916@end menu 917 918 919@node Flat Profile 920@section The Flat Profile 921@cindex flat profile 922 923The @dfn{flat profile} shows the total amount of time your program 924spent executing each function. Unless the @samp{-z} option is given, 925functions with no apparent time spent in them, and no apparent calls 926to them, are not mentioned. Note that if a function was not compiled 927for profiling, and didn't run long enough to show up on the program 928counter histogram, it will be indistinguishable from a function that 929was never called. 930 931This is part of a flat profile for a small program: 932 933@smallexample 934@group 935Flat profile: 936 937Each sample counts as 0.01 seconds. 938 % cumulative self self total 939 time seconds seconds calls ms/call ms/call name 940 33.34 0.02 0.02 7208 0.00 0.00 open 941 16.67 0.03 0.01 244 0.04 0.12 offtime 942 16.67 0.04 0.01 8 1.25 1.25 memccpy 943 16.67 0.05 0.01 7 1.43 1.43 write 944 16.67 0.06 0.01 mcount 945 0.00 0.06 0.00 236 0.00 0.00 tzset 946 0.00 0.06 0.00 192 0.00 0.00 tolower 947 0.00 0.06 0.00 47 0.00 0.00 strlen 948 0.00 0.06 0.00 45 0.00 0.00 strchr 949 0.00 0.06 0.00 1 0.00 50.00 main 950 0.00 0.06 0.00 1 0.00 0.00 memcpy 951 0.00 0.06 0.00 1 0.00 10.11 print 952 0.00 0.06 0.00 1 0.00 0.00 profil 953 0.00 0.06 0.00 1 0.00 50.00 report 954@dots{} 955@end group 956@end smallexample 957 958@noindent 959The functions are sorted first by decreasing run-time spent in them, 960then by decreasing number of calls, then alphabetically by name. The 961functions @samp{mcount} and @samp{profil} are part of the profiling 962apparatus and appear in every flat profile; their time gives a measure of 963the amount of overhead due to profiling. 964 965Just before the column headers, a statement appears indicating 966how much time each sample counted as. 967This @dfn{sampling period} estimates the margin of error in each of the time 968figures. A time figure that is not much larger than this is not 969reliable. In this example, each sample counted as 0.01 seconds, 970suggesting a 100 Hz sampling rate. 971The program's total execution time was 0.06 972seconds, as indicated by the @samp{cumulative seconds} field. Since 973each sample counted for 0.01 seconds, this means only six samples 974were taken during the run. Two of the samples occurred while the 975program was in the @samp{open} function, as indicated by the 976@samp{self seconds} field. Each of the other four samples 977occurred one each in @samp{offtime}, @samp{memccpy}, @samp{write}, 978and @samp{mcount}. 979Since only six samples were taken, none of these values can 980be regarded as particularly reliable. 981In another run, 982the @samp{self seconds} field for 983@samp{mcount} might well be @samp{0.00} or @samp{0.02}. 984@xref{Sampling Error, ,Statistical Sampling Error}, 985for a complete discussion. 986 987The remaining functions in the listing (those whose 988@samp{self seconds} field is @samp{0.00}) didn't appear 989in the histogram samples at all. However, the call graph 990indicated that they were called, so therefore they are listed, 991sorted in decreasing order by the @samp{calls} field. 992Clearly some time was spent executing these functions, 993but the paucity of histogram samples prevents any 994determination of how much time each took. 995 996Here is what the fields in each line mean: 997 998@table @code 999@item % time 1000This is the percentage of the total execution time your program spent 1001in this function. These should all add up to 100%. 1002 1003@item cumulative seconds 1004This is the cumulative total number of seconds the computer spent 1005executing this functions, plus the time spent in all the functions 1006above this one in this table. 1007 1008@item self seconds 1009This is the number of seconds accounted for by this function alone. 1010The flat profile listing is sorted first by this number. 1011 1012@item calls 1013This is the total number of times the function was called. If the 1014function was never called, or the number of times it was called cannot 1015be determined (probably because the function was not compiled with 1016profiling enabled), the @dfn{calls} field is blank. 1017 1018@item self ms/call 1019This represents the average number of milliseconds spent in this 1020function per call, if this function is profiled. Otherwise, this field 1021is blank for this function. 1022 1023@item total ms/call 1024This represents the average number of milliseconds spent in this 1025function and its descendants per call, if this function is profiled. 1026Otherwise, this field is blank for this function. 1027This is the only field in the flat profile that uses call graph analysis. 1028 1029@item name 1030This is the name of the function. The flat profile is sorted by this 1031field alphabetically after the @dfn{self seconds} and @dfn{calls} 1032fields are sorted. 1033@end table 1034 1035@node Call Graph 1036@section The Call Graph 1037@cindex call graph 1038 1039The @dfn{call graph} shows how much time was spent in each function 1040and its children. From this information, you can find functions that, 1041while they themselves may not have used much time, called other 1042functions that did use unusual amounts of time. 1043 1044Here is a sample call from a small program. This call came from the 1045same @code{gprof} run as the flat profile example in the previous 1046section. 1047 1048@smallexample 1049@group 1050granularity: each sample hit covers 2 byte(s) for 20.00% of 0.05 seconds 1051 1052index % time self children called name 1053 <spontaneous> 1054[1] 100.0 0.00 0.05 start [1] 1055 0.00 0.05 1/1 main [2] 1056 0.00 0.00 1/2 on_exit [28] 1057 0.00 0.00 1/1 exit [59] 1058----------------------------------------------- 1059 0.00 0.05 1/1 start [1] 1060[2] 100.0 0.00 0.05 1 main [2] 1061 0.00 0.05 1/1 report [3] 1062----------------------------------------------- 1063 0.00 0.05 1/1 main [2] 1064[3] 100.0 0.00 0.05 1 report [3] 1065 0.00 0.03 8/8 timelocal [6] 1066 0.00 0.01 1/1 print [9] 1067 0.00 0.01 9/9 fgets [12] 1068 0.00 0.00 12/34 strncmp <cycle 1> [40] 1069 0.00 0.00 8/8 lookup [20] 1070 0.00 0.00 1/1 fopen [21] 1071 0.00 0.00 8/8 chewtime [24] 1072 0.00 0.00 8/16 skipspace [44] 1073----------------------------------------------- 1074[4] 59.8 0.01 0.02 8+472 <cycle 2 as a whole> [4] 1075 0.01 0.02 244+260 offtime <cycle 2> [7] 1076 0.00 0.00 236+1 tzset <cycle 2> [26] 1077----------------------------------------------- 1078@end group 1079@end smallexample 1080 1081The lines full of dashes divide this table into @dfn{entries}, one for each 1082function. Each entry has one or more lines. 1083 1084In each entry, the primary line is the one that starts with an index number 1085in square brackets. The end of this line says which function the entry is 1086for. The preceding lines in the entry describe the callers of this 1087function and the following lines describe its subroutines (also called 1088@dfn{children} when we speak of the call graph). 1089 1090The entries are sorted by time spent in the function and its subroutines. 1091 1092The internal profiling function @code{mcount} (@pxref{Flat Profile, ,The 1093Flat Profile}) is never mentioned in the call graph. 1094 1095@menu 1096* Primary:: Details of the primary line's contents. 1097* Callers:: Details of caller-lines' contents. 1098* Subroutines:: Details of subroutine-lines' contents. 1099* Cycles:: When there are cycles of recursion, 1100 such as @code{a} calls @code{b} calls @code{a}@dots{} 1101@end menu 1102 1103@node Primary 1104@subsection The Primary Line 1105 1106The @dfn{primary line} in a call graph entry is the line that 1107describes the function which the entry is about and gives the overall 1108statistics for this function. 1109 1110For reference, we repeat the primary line from the entry for function 1111@code{report} in our main example, together with the heading line that 1112shows the names of the fields: 1113 1114@smallexample 1115@group 1116index % time self children called name 1117@dots{} 1118[3] 100.0 0.00 0.05 1 report [3] 1119@end group 1120@end smallexample 1121 1122Here is what the fields in the primary line mean: 1123 1124@table @code 1125@item index 1126Entries are numbered with consecutive integers. Each function 1127therefore has an index number, which appears at the beginning of its 1128primary line. 1129 1130Each cross-reference to a function, as a caller or subroutine of 1131another, gives its index number as well as its name. The index number 1132guides you if you wish to look for the entry for that function. 1133 1134@item % time 1135This is the percentage of the total time that was spent in this 1136function, including time spent in subroutines called from this 1137function. 1138 1139The time spent in this function is counted again for the callers of 1140this function. Therefore, adding up these percentages is meaningless. 1141 1142@item self 1143This is the total amount of time spent in this function. This 1144should be identical to the number printed in the @code{seconds} field 1145for this function in the flat profile. 1146 1147@item children 1148This is the total amount of time spent in the subroutine calls made by 1149this function. This should be equal to the sum of all the @code{self} 1150and @code{children} entries of the children listed directly below this 1151function. 1152 1153@item called 1154This is the number of times the function was called. 1155 1156If the function called itself recursively, there are two numbers, 1157separated by a @samp{+}. The first number counts non-recursive calls, 1158and the second counts recursive calls. 1159 1160In the example above, the function @code{report} was called once from 1161@code{main}. 1162 1163@item name 1164This is the name of the current function. The index number is 1165repeated after it. 1166 1167If the function is part of a cycle of recursion, the cycle number is 1168printed between the function's name and the index number 1169(@pxref{Cycles, ,How Mutually Recursive Functions Are Described}). 1170For example, if function @code{gnurr} is part of 1171cycle number one, and has index number twelve, its primary line would 1172be end like this: 1173 1174@example 1175gnurr <cycle 1> [12] 1176@end example 1177@end table 1178 1179@node Callers 1180@subsection Lines for a Function's Callers 1181 1182A function's entry has a line for each function it was called by. 1183These lines' fields correspond to the fields of the primary line, but 1184their meanings are different because of the difference in context. 1185 1186For reference, we repeat two lines from the entry for the function 1187@code{report}, the primary line and one caller-line preceding it, together 1188with the heading line that shows the names of the fields: 1189 1190@smallexample 1191index % time self children called name 1192@dots{} 1193 0.00 0.05 1/1 main [2] 1194[3] 100.0 0.00 0.05 1 report [3] 1195@end smallexample 1196 1197Here are the meanings of the fields in the caller-line for @code{report} 1198called from @code{main}: 1199 1200@table @code 1201@item self 1202An estimate of the amount of time spent in @code{report} itself when it was 1203called from @code{main}. 1204 1205@item children 1206An estimate of the amount of time spent in subroutines of @code{report} 1207when @code{report} was called from @code{main}. 1208 1209The sum of the @code{self} and @code{children} fields is an estimate 1210of the amount of time spent within calls to @code{report} from @code{main}. 1211 1212@item called 1213Two numbers: the number of times @code{report} was called from @code{main}, 1214followed by the total number of non-recursive calls to @code{report} from 1215all its callers. 1216 1217@item name and index number 1218The name of the caller of @code{report} to which this line applies, 1219followed by the caller's index number. 1220 1221Not all functions have entries in the call graph; some 1222options to @code{gprof} request the omission of certain functions. 1223When a caller has no entry of its own, it still has caller-lines 1224in the entries of the functions it calls. 1225 1226If the caller is part of a recursion cycle, the cycle number is 1227printed between the name and the index number. 1228@end table 1229 1230If the identity of the callers of a function cannot be determined, a 1231dummy caller-line is printed which has @samp{<spontaneous>} as the 1232``caller's name'' and all other fields blank. This can happen for 1233signal handlers. 1234@c What if some calls have determinable callers' names but not all? 1235@c FIXME - still relevant? 1236 1237@node Subroutines 1238@subsection Lines for a Function's Subroutines 1239 1240A function's entry has a line for each of its subroutines---in other 1241words, a line for each other function that it called. These lines' 1242fields correspond to the fields of the primary line, but their meanings 1243are different because of the difference in context. 1244 1245For reference, we repeat two lines from the entry for the function 1246@code{main}, the primary line and a line for a subroutine, together 1247with the heading line that shows the names of the fields: 1248 1249@smallexample 1250index % time self children called name 1251@dots{} 1252[2] 100.0 0.00 0.05 1 main [2] 1253 0.00 0.05 1/1 report [3] 1254@end smallexample 1255 1256Here are the meanings of the fields in the subroutine-line for @code{main} 1257calling @code{report}: 1258 1259@table @code 1260@item self 1261An estimate of the amount of time spent directly within @code{report} 1262when @code{report} was called from @code{main}. 1263 1264@item children 1265An estimate of the amount of time spent in subroutines of @code{report} 1266when @code{report} was called from @code{main}. 1267 1268The sum of the @code{self} and @code{children} fields is an estimate 1269of the total time spent in calls to @code{report} from @code{main}. 1270 1271@item called 1272Two numbers, the number of calls to @code{report} from @code{main} 1273followed by the total number of non-recursive calls to @code{report}. 1274This ratio is used to determine how much of @code{report}'s @code{self} 1275and @code{children} time gets credited to @code{main}. 1276@xref{Assumptions, ,Estimating @code{children} Times}. 1277 1278@item name 1279The name of the subroutine of @code{main} to which this line applies, 1280followed by the subroutine's index number. 1281 1282If the caller is part of a recursion cycle, the cycle number is 1283printed between the name and the index number. 1284@end table 1285 1286@node Cycles 1287@subsection How Mutually Recursive Functions Are Described 1288@cindex cycle 1289@cindex recursion cycle 1290 1291The graph may be complicated by the presence of @dfn{cycles of 1292recursion} in the call graph. A cycle exists if a function calls 1293another function that (directly or indirectly) calls (or appears to 1294call) the original function. For example: if @code{a} calls @code{b}, 1295and @code{b} calls @code{a}, then @code{a} and @code{b} form a cycle. 1296 1297Whenever there are call paths both ways between a pair of functions, they 1298belong to the same cycle. If @code{a} and @code{b} call each other and 1299@code{b} and @code{c} call each other, all three make one cycle. Note that 1300even if @code{b} only calls @code{a} if it was not called from @code{a}, 1301@code{gprof} cannot determine this, so @code{a} and @code{b} are still 1302considered a cycle. 1303 1304The cycles are numbered with consecutive integers. When a function 1305belongs to a cycle, each time the function name appears in the call graph 1306it is followed by @samp{<cycle @var{number}>}. 1307 1308The reason cycles matter is that they make the time values in the call 1309graph paradoxical. The ``time spent in children'' of @code{a} should 1310include the time spent in its subroutine @code{b} and in @code{b}'s 1311subroutines---but one of @code{b}'s subroutines is @code{a}! How much of 1312@code{a}'s time should be included in the children of @code{a}, when 1313@code{a} is indirectly recursive? 1314 1315The way @code{gprof} resolves this paradox is by creating a single entry 1316for the cycle as a whole. The primary line of this entry describes the 1317total time spent directly in the functions of the cycle. The 1318``subroutines'' of the cycle are the individual functions of the cycle, and 1319all other functions that were called directly by them. The ``callers'' of 1320the cycle are the functions, outside the cycle, that called functions in 1321the cycle. 1322 1323Here is an example portion of a call graph which shows a cycle containing 1324functions @code{a} and @code{b}. The cycle was entered by a call to 1325@code{a} from @code{main}; both @code{a} and @code{b} called @code{c}. 1326 1327@smallexample 1328index % time self children called name 1329---------------------------------------- 1330 1.77 0 1/1 main [2] 1331[3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3] 1332 1.02 0 3 b <cycle 1> [4] 1333 0.75 0 2 a <cycle 1> [5] 1334---------------------------------------- 1335 3 a <cycle 1> [5] 1336[4] 52.85 1.02 0 0 b <cycle 1> [4] 1337 2 a <cycle 1> [5] 1338 0 0 3/6 c [6] 1339---------------------------------------- 1340 1.77 0 1/1 main [2] 1341 2 b <cycle 1> [4] 1342[5] 38.86 0.75 0 1 a <cycle 1> [5] 1343 3 b <cycle 1> [4] 1344 0 0 3/6 c [6] 1345---------------------------------------- 1346@end smallexample 1347 1348@noindent 1349(The entire call graph for this program contains in addition an entry for 1350@code{main}, which calls @code{a}, and an entry for @code{c}, with callers 1351@code{a} and @code{b}.) 1352 1353@smallexample 1354index % time self children called name 1355 <spontaneous> 1356[1] 100.00 0 1.93 0 start [1] 1357 0.16 1.77 1/1 main [2] 1358---------------------------------------- 1359 0.16 1.77 1/1 start [1] 1360[2] 100.00 0.16 1.77 1 main [2] 1361 1.77 0 1/1 a <cycle 1> [5] 1362---------------------------------------- 1363 1.77 0 1/1 main [2] 1364[3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3] 1365 1.02 0 3 b <cycle 1> [4] 1366 0.75 0 2 a <cycle 1> [5] 1367 0 0 6/6 c [6] 1368---------------------------------------- 1369 3 a <cycle 1> [5] 1370[4] 52.85 1.02 0 0 b <cycle 1> [4] 1371 2 a <cycle 1> [5] 1372 0 0 3/6 c [6] 1373---------------------------------------- 1374 1.77 0 1/1 main [2] 1375 2 b <cycle 1> [4] 1376[5] 38.86 0.75 0 1 a <cycle 1> [5] 1377 3 b <cycle 1> [4] 1378 0 0 3/6 c [6] 1379---------------------------------------- 1380 0 0 3/6 b <cycle 1> [4] 1381 0 0 3/6 a <cycle 1> [5] 1382[6] 0.00 0 0 6 c [6] 1383---------------------------------------- 1384@end smallexample 1385 1386The @code{self} field of the cycle's primary line is the total time 1387spent in all the functions of the cycle. It equals the sum of the 1388@code{self} fields for the individual functions in the cycle, found 1389in the entry in the subroutine lines for these functions. 1390 1391The @code{children} fields of the cycle's primary line and subroutine lines 1392count only subroutines outside the cycle. Even though @code{a} calls 1393@code{b}, the time spent in those calls to @code{b} is not counted in 1394@code{a}'s @code{children} time. Thus, we do not encounter the problem of 1395what to do when the time in those calls to @code{b} includes indirect 1396recursive calls back to @code{a}. 1397 1398The @code{children} field of a caller-line in the cycle's entry estimates 1399the amount of time spent @emph{in the whole cycle}, and its other 1400subroutines, on the times when that caller called a function in the cycle. 1401 1402The @code{called} field in the primary line for the cycle has two numbers: 1403first, the number of times functions in the cycle were called by functions 1404outside the cycle; second, the number of times they were called by 1405functions in the cycle (including times when a function in the cycle calls 1406itself). This is a generalization of the usual split into non-recursive and 1407recursive calls. 1408 1409The @code{called} field of a subroutine-line for a cycle member in the 1410cycle's entry says how many time that function was called from functions in 1411the cycle. The total of all these is the second number in the primary line's 1412@code{called} field. 1413 1414In the individual entry for a function in a cycle, the other functions in 1415the same cycle can appear as subroutines and as callers. These lines show 1416how many times each function in the cycle called or was called from each other 1417function in the cycle. The @code{self} and @code{children} fields in these 1418lines are blank because of the difficulty of defining meanings for them 1419when recursion is going on. 1420 1421@node Line-by-line 1422@section Line-by-line Profiling 1423 1424@code{gprof}'s @samp{-l} option causes the program to perform 1425@dfn{line-by-line} profiling. In this mode, histogram 1426samples are assigned not to functions, but to individual 1427lines of source code. This only works with programs compiled with 1428older versions of the @code{gcc} compiler. Newer versions of @code{gcc} 1429use a different program - @code{gcov} - to display line-by-line 1430profiling information. 1431 1432With the older versions of @code{gcc} the program usually has to be 1433compiled with a @samp{-g} option, in addition to @samp{-pg}, in order 1434to generate debugging symbols for tracking source code lines. 1435Note, in much older versions of @code{gcc} the program had to be 1436compiled with the @samp{-a} command-line option as well. 1437 1438The flat profile is the most useful output table 1439in line-by-line mode. 1440The call graph isn't as useful as normal, since 1441the current version of @code{gprof} does not propagate 1442call graph arcs from source code lines to the enclosing function. 1443The call graph does, however, show each line of code 1444that called each function, along with a count. 1445 1446Here is a section of @code{gprof}'s output, without line-by-line profiling. 1447Note that @code{ct_init} accounted for four histogram hits, and 144813327 calls to @code{init_block}. 1449 1450@smallexample 1451Flat profile: 1452 1453Each sample counts as 0.01 seconds. 1454 % cumulative self self total 1455 time seconds seconds calls us/call us/call name 1456 30.77 0.13 0.04 6335 6.31 6.31 ct_init 1457 1458 1459 Call graph (explanation follows) 1460 1461 1462granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds 1463 1464index % time self children called name 1465 1466 0.00 0.00 1/13496 name_too_long 1467 0.00 0.00 40/13496 deflate 1468 0.00 0.00 128/13496 deflate_fast 1469 0.00 0.00 13327/13496 ct_init 1470[7] 0.0 0.00 0.00 13496 init_block 1471 1472@end smallexample 1473 1474Now let's look at some of @code{gprof}'s output from the same program run, 1475this time with line-by-line profiling enabled. Note that @code{ct_init}'s 1476four histogram hits are broken down into four lines of source code---one hit 1477occurred on each of lines 349, 351, 382 and 385. In the call graph, 1478note how 1479@code{ct_init}'s 13327 calls to @code{init_block} are broken down 1480into one call from line 396, 3071 calls from line 384, 3730 calls 1481from line 385, and 6525 calls from 387. 1482 1483@smallexample 1484Flat profile: 1485 1486Each sample counts as 0.01 seconds. 1487 % cumulative self 1488 time seconds seconds calls name 1489 7.69 0.10 0.01 ct_init (trees.c:349) 1490 7.69 0.11 0.01 ct_init (trees.c:351) 1491 7.69 0.12 0.01 ct_init (trees.c:382) 1492 7.69 0.13 0.01 ct_init (trees.c:385) 1493 1494 1495 Call graph (explanation follows) 1496 1497 1498granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds 1499 1500 % time self children called name 1501 1502 0.00 0.00 1/13496 name_too_long (gzip.c:1440) 1503 0.00 0.00 1/13496 deflate (deflate.c:763) 1504 0.00 0.00 1/13496 ct_init (trees.c:396) 1505 0.00 0.00 2/13496 deflate (deflate.c:727) 1506 0.00 0.00 4/13496 deflate (deflate.c:686) 1507 0.00 0.00 5/13496 deflate (deflate.c:675) 1508 0.00 0.00 12/13496 deflate (deflate.c:679) 1509 0.00 0.00 16/13496 deflate (deflate.c:730) 1510 0.00 0.00 128/13496 deflate_fast (deflate.c:654) 1511 0.00 0.00 3071/13496 ct_init (trees.c:384) 1512 0.00 0.00 3730/13496 ct_init (trees.c:385) 1513 0.00 0.00 6525/13496 ct_init (trees.c:387) 1514[6] 0.0 0.00 0.00 13496 init_block (trees.c:408) 1515 1516@end smallexample 1517 1518 1519@node Annotated Source 1520@section The Annotated Source Listing 1521 1522@code{gprof}'s @samp{-A} option triggers an annotated source listing, 1523which lists the program's source code, each function labeled with the 1524number of times it was called. You may also need to specify the 1525@samp{-I} option, if @code{gprof} can't find the source code files. 1526 1527With older versions of @code{gcc} compiling with @samp{gcc @dots{} -g 1528-pg -a} augments your program with basic-block counting code, in 1529addition to function counting code. This enables @code{gprof} to 1530determine how many times each line of code was executed. With newer 1531versions of @code{gcc} support for displaying basic-block counts is 1532provided by the @code{gcov} program. 1533 1534For example, consider the following function, taken from gzip, 1535with line numbers added: 1536 1537@smallexample 1538 1 ulg updcrc(s, n) 1539 2 uch *s; 1540 3 unsigned n; 1541 4 @{ 1542 5 register ulg c; 1543 6 1544 7 static ulg crc = (ulg)0xffffffffL; 1545 8 1546 9 if (s == NULL) @{ 154710 c = 0xffffffffL; 154811 @} else @{ 154912 c = crc; 155013 if (n) do @{ 155114 c = crc_32_tab[...]; 155215 @} while (--n); 155316 @} 155417 crc = c; 155518 return c ^ 0xffffffffL; 155619 @} 1557 1558@end smallexample 1559 1560@code{updcrc} has at least five basic-blocks. 1561One is the function itself. The 1562@code{if} statement on line 9 generates two more basic-blocks, one 1563for each branch of the @code{if}. A fourth basic-block results from 1564the @code{if} on line 13, and the contents of the @code{do} loop form 1565the fifth basic-block. The compiler may also generate additional 1566basic-blocks to handle various special cases. 1567 1568A program augmented for basic-block counting can be analyzed with 1569@samp{gprof -l -A}. 1570The @samp{-x} option is also helpful, 1571to ensure that each line of code is labeled at least once. 1572Here is @code{updcrc}'s 1573annotated source listing for a sample @code{gzip} run: 1574 1575@smallexample 1576 ulg updcrc(s, n) 1577 uch *s; 1578 unsigned n; 1579 2 ->@{ 1580 register ulg c; 1581 1582 static ulg crc = (ulg)0xffffffffL; 1583 1584 2 -> if (s == NULL) @{ 1585 1 -> c = 0xffffffffL; 1586 1 -> @} else @{ 1587 1 -> c = crc; 1588 1 -> if (n) do @{ 1589 26312 -> c = crc_32_tab[...]; 159026312,1,26311 -> @} while (--n); 1591 @} 1592 2 -> crc = c; 1593 2 -> return c ^ 0xffffffffL; 1594 2 ->@} 1595@end smallexample 1596 1597In this example, the function was called twice, passing once through 1598each branch of the @code{if} statement. The body of the @code{do} 1599loop was executed a total of 26312 times. Note how the @code{while} 1600statement is annotated. It began execution 26312 times, once for 1601each iteration through the loop. One of those times (the last time) 1602it exited, while it branched back to the beginning of the loop 26311 times. 1603 1604@node Inaccuracy 1605@chapter Inaccuracy of @code{gprof} Output 1606 1607@menu 1608* Sampling Error:: Statistical margins of error 1609* Assumptions:: Estimating children times 1610@end menu 1611 1612@node Sampling Error 1613@section Statistical Sampling Error 1614 1615The run-time figures that @code{gprof} gives you are based on a sampling 1616process, so they are subject to statistical inaccuracy. If a function runs 1617only a small amount of time, so that on the average the sampling process 1618ought to catch that function in the act only once, there is a pretty good 1619chance it will actually find that function zero times, or twice. 1620 1621By contrast, the number-of-calls and basic-block figures are derived 1622by counting, not sampling. They are completely accurate and will not 1623vary from run to run if your program is deterministic and single 1624threaded. In multi-threaded applications, or single threaded 1625applications that link with multi-threaded libraries, the counts are 1626only deterministic if the counting function is thread-safe. (Note: 1627beware that the mcount counting function in glibc is @emph{not} 1628thread-safe). @xref{Implementation, ,Implementation of Profiling}. 1629 1630The @dfn{sampling period} that is printed at the beginning of the flat 1631profile says how often samples are taken. The rule of thumb is that a 1632run-time figure is accurate if it is considerably bigger than the sampling 1633period. 1634 1635The actual amount of error can be predicted. 1636For @var{n} samples, the @emph{expected} error 1637is the square-root of @var{n}. For example, 1638if the sampling period is 0.01 seconds and @code{foo}'s run-time is 1 second, 1639@var{n} is 100 samples (1 second/0.01 seconds), sqrt(@var{n}) is 10 samples, so 1640the expected error in @code{foo}'s run-time is 0.1 seconds (10*0.01 seconds), 1641or ten percent of the observed value. 1642Again, if the sampling period is 0.01 seconds and @code{bar}'s run-time is 1643100 seconds, @var{n} is 10000 samples, sqrt(@var{n}) is 100 samples, so 1644the expected error in @code{bar}'s run-time is 1 second, 1645or one percent of the observed value. 1646It is likely to 1647vary this much @emph{on the average} from one profiling run to the next. 1648(@emph{Sometimes} it will vary more.) 1649 1650This does not mean that a small run-time figure is devoid of information. 1651If the program's @emph{total} run-time is large, a small run-time for one 1652function does tell you that that function used an insignificant fraction of 1653the whole program's time. Usually this means it is not worth optimizing. 1654 1655One way to get more accuracy is to give your program more (but similar) 1656input data so it will take longer. Another way is to combine the data from 1657several runs, using the @samp{-s} option of @code{gprof}. Here is how: 1658 1659@enumerate 1660@item 1661Run your program once. 1662 1663@item 1664Issue the command @samp{mv gmon.out gmon.sum}. 1665 1666@item 1667Run your program again, the same as before. 1668 1669@item 1670Merge the new data in @file{gmon.out} into @file{gmon.sum} with this command: 1671 1672@example 1673gprof -s @var{executable-file} gmon.out gmon.sum 1674@end example 1675 1676@item 1677Repeat the last two steps as often as you wish. 1678 1679@item 1680Analyze the cumulative data using this command: 1681 1682@example 1683gprof @var{executable-file} gmon.sum > @var{output-file} 1684@end example 1685@end enumerate 1686 1687@node Assumptions 1688@section Estimating @code{children} Times 1689 1690Some of the figures in the call graph are estimates---for example, the 1691@code{children} time values and all the time figures in caller and 1692subroutine lines. 1693 1694There is no direct information about these measurements in the profile 1695data itself. Instead, @code{gprof} estimates them by making an assumption 1696about your program that might or might not be true. 1697 1698The assumption made is that the average time spent in each call to any 1699function @code{foo} is not correlated with who called @code{foo}. If 1700@code{foo} used 5 seconds in all, and 2/5 of the calls to @code{foo} came 1701from @code{a}, then @code{foo} contributes 2 seconds to @code{a}'s 1702@code{children} time, by assumption. 1703 1704This assumption is usually true enough, but for some programs it is far 1705from true. Suppose that @code{foo} returns very quickly when its argument 1706is zero; suppose that @code{a} always passes zero as an argument, while 1707other callers of @code{foo} pass other arguments. In this program, all the 1708time spent in @code{foo} is in the calls from callers other than @code{a}. 1709But @code{gprof} has no way of knowing this; it will blindly and 1710incorrectly charge 2 seconds of time in @code{foo} to the children of 1711@code{a}. 1712 1713@c FIXME - has this been fixed? 1714We hope some day to put more complete data into @file{gmon.out}, so that 1715this assumption is no longer needed, if we can figure out how. For the 1716novice, the estimated figures are usually more useful than misleading. 1717 1718@node How do I? 1719@chapter Answers to Common Questions 1720 1721@table @asis 1722@item How can I get more exact information about hot spots in my program? 1723 1724Looking at the per-line call counts only tells part of the story. 1725Because @code{gprof} can only report call times and counts by function, 1726the best way to get finer-grained information on where the program 1727is spending its time is to re-factor large functions into sequences 1728of calls to smaller ones. Beware however that this can introduce 1729artificial hot spots since compiling with @samp{-pg} adds a significant 1730overhead to function calls. An alternative solution is to use a 1731non-intrusive profiler, e.g.@: oprofile. 1732 1733@item How do I find which lines in my program were executed the most times? 1734 1735Use the @code{gcov} program. 1736 1737@item How do I find which lines in my program called a particular function? 1738 1739Use @samp{gprof -l} and lookup the function in the call graph. 1740The callers will be broken down by function and line number. 1741 1742@item How do I analyze a program that runs for less than a second? 1743 1744Try using a shell script like this one: 1745 1746@example 1747for i in `seq 1 100`; do 1748 fastprog 1749 mv gmon.out gmon.out.$i 1750done 1751 1752gprof -s fastprog gmon.out.* 1753 1754gprof fastprog gmon.sum 1755@end example 1756 1757If your program is completely deterministic, all the call counts 1758will be simple multiples of 100 (i.e., a function called once in 1759each run will appear with a call count of 100). 1760 1761@end table 1762 1763@node Incompatibilities 1764@chapter Incompatibilities with Unix @code{gprof} 1765 1766@sc{gnu} @code{gprof} and Berkeley Unix @code{gprof} use the same data 1767file @file{gmon.out}, and provide essentially the same information. But 1768there are a few differences. 1769 1770@itemize @bullet 1771@item 1772@sc{gnu} @code{gprof} uses a new, generalized file format with support 1773for basic-block execution counts and non-realtime histograms. A magic 1774cookie and version number allows @code{gprof} to easily identify 1775new style files. Old BSD-style files can still be read. 1776@xref{File Format, ,Profiling Data File Format}. 1777 1778@item 1779For a recursive function, Unix @code{gprof} lists the function as a 1780parent and as a child, with a @code{calls} field that lists the number 1781of recursive calls. @sc{gnu} @code{gprof} omits these lines and puts 1782the number of recursive calls in the primary line. 1783 1784@item 1785When a function is suppressed from the call graph with @samp{-e}, @sc{gnu} 1786@code{gprof} still lists it as a subroutine of functions that call it. 1787 1788@item 1789@sc{gnu} @code{gprof} accepts the @samp{-k} with its argument 1790in the form @samp{from/to}, instead of @samp{from to}. 1791 1792@item 1793In the annotated source listing, 1794if there are multiple basic blocks on the same line, 1795@sc{gnu} @code{gprof} prints all of their counts, separated by commas. 1796 1797@ignore - it does this now 1798@item 1799The function names printed in @sc{gnu} @code{gprof} output do not include 1800the leading underscores that are added internally to the front of all 1801C identifiers on many operating systems. 1802@end ignore 1803 1804@item 1805The blurbs, field widths, and output formats are different. @sc{gnu} 1806@code{gprof} prints blurbs after the tables, so that you can see the 1807tables without skipping the blurbs. 1808@end itemize 1809 1810@node Details 1811@chapter Details of Profiling 1812 1813@menu 1814* Implementation:: How a program collects profiling information 1815* File Format:: Format of @samp{gmon.out} files 1816* Internals:: @code{gprof}'s internal operation 1817* Debugging:: Using @code{gprof}'s @samp{-d} option 1818@end menu 1819 1820@node Implementation 1821@section Implementation of Profiling 1822 1823Profiling works by changing how every function in your program is compiled 1824so that when it is called, it will stash away some information about where 1825it was called from. From this, the profiler can figure out what function 1826called it, and can count how many times it was called. This change is made 1827by the compiler when your program is compiled with the @samp{-pg} option, 1828which causes every function to call @code{mcount} 1829(or @code{_mcount}, or @code{__mcount}, depending on the OS and compiler) 1830as one of its first operations. 1831 1832The @code{mcount} routine, included in the profiling library, 1833is responsible for recording in an in-memory call graph table 1834both its parent routine (the child) and its parent's parent. This is 1835typically done by examining the stack frame to find both 1836the address of the child, and the return address in the original parent. 1837Since this is a very machine-dependent operation, @code{mcount} 1838itself is typically a short assembly-language stub routine 1839that extracts the required 1840information, and then calls @code{__mcount_internal} 1841(a normal C function) with two arguments---@code{frompc} and @code{selfpc}. 1842@code{__mcount_internal} is responsible for maintaining 1843the in-memory call graph, which records @code{frompc}, @code{selfpc}, 1844and the number of times each of these call arcs was traversed. 1845 1846GCC Version 2 provides a magical function (@code{__builtin_return_address}), 1847which allows a generic @code{mcount} function to extract the 1848required information from the stack frame. However, on some 1849architectures, most notably the SPARC, using this builtin can be 1850very computationally expensive, and an assembly language version 1851of @code{mcount} is used for performance reasons. 1852 1853Number-of-calls information for library routines is collected by using a 1854special version of the C library. The programs in it are the same as in 1855the usual C library, but they were compiled with @samp{-pg}. If you 1856link your program with @samp{gcc @dots{} -pg}, it automatically uses the 1857profiling version of the library. 1858 1859Profiling also involves watching your program as it runs, and keeping a 1860histogram of where the program counter happens to be every now and then. 1861Typically the program counter is looked at around 100 times per second of 1862run time, but the exact frequency may vary from system to system. 1863 1864This is done is one of two ways. Most UNIX-like operating systems 1865provide a @code{profil()} system call, which registers a memory 1866array with the kernel, along with a scale 1867factor that determines how the program's address space maps 1868into the array. 1869Typical scaling values cause every 2 to 8 bytes of address space 1870to map into a single array slot. 1871On every tick of the system clock 1872(assuming the profiled program is running), the value of the 1873program counter is examined and the corresponding slot in 1874the memory array is incremented. Since this is done in the kernel, 1875which had to interrupt the process anyway to handle the clock 1876interrupt, very little additional system overhead is required. 1877 1878However, some operating systems, most notably Linux 2.0 (and earlier), 1879do not provide a @code{profil()} system call. On such a system, 1880arrangements are made for the kernel to periodically deliver 1881a signal to the process (typically via @code{setitimer()}), 1882which then performs the same operation of examining the 1883program counter and incrementing a slot in the memory array. 1884Since this method requires a signal to be delivered to 1885user space every time a sample is taken, it uses considerably 1886more overhead than kernel-based profiling. Also, due to the 1887added delay required to deliver the signal, this method is 1888less accurate as well. 1889 1890A special startup routine allocates memory for the histogram and 1891either calls @code{profil()} or sets up 1892a clock signal handler. 1893This routine (@code{monstartup}) can be invoked in several ways. 1894On Linux systems, a special profiling startup file @code{gcrt0.o}, 1895which invokes @code{monstartup} before @code{main}, 1896is used instead of the default @code{crt0.o}. 1897Use of this special startup file is one of the effects 1898of using @samp{gcc @dots{} -pg} to link. 1899On SPARC systems, no special startup files are used. 1900Rather, the @code{mcount} routine, when it is invoked for 1901the first time (typically when @code{main} is called), 1902calls @code{monstartup}. 1903 1904If the compiler's @samp{-a} option was used, basic-block counting 1905is also enabled. Each object file is then compiled with a static array 1906of counts, initially zero. 1907In the executable code, every time a new basic-block begins 1908(i.e., when an @code{if} statement appears), an extra instruction 1909is inserted to increment the corresponding count in the array. 1910At compile time, a paired array was constructed that recorded 1911the starting address of each basic-block. Taken together, 1912the two arrays record the starting address of every basic-block, 1913along with the number of times it was executed. 1914 1915The profiling library also includes a function (@code{mcleanup}) which is 1916typically registered using @code{atexit()} to be called as the 1917program exits, and is responsible for writing the file @file{gmon.out}. 1918Profiling is turned off, various headers are output, and the histogram 1919is written, followed by the call-graph arcs and the basic-block counts. 1920 1921The output from @code{gprof} gives no indication of parts of your program that 1922are limited by I/O or swapping bandwidth. This is because samples of the 1923program counter are taken at fixed intervals of the program's run time. 1924Therefore, the 1925time measurements in @code{gprof} output say nothing about time that your 1926program was not running. For example, a part of the program that creates 1927so much data that it cannot all fit in physical memory at once may run very 1928slowly due to thrashing, but @code{gprof} will say it uses little time. On 1929the other hand, sampling by run time has the advantage that the amount of 1930load due to other users won't directly affect the output you get. 1931 1932@node File Format 1933@section Profiling Data File Format 1934 1935The old BSD-derived file format used for profile data does not contain a 1936magic cookie that allows one to check whether a data file really is a 1937@code{gprof} file. Furthermore, it does not provide a version number, thus 1938rendering changes to the file format almost impossible. @sc{gnu} @code{gprof} 1939uses a new file format that provides these features. For backward 1940compatibility, @sc{gnu} @code{gprof} continues to support the old BSD-derived 1941format, but not all features are supported with it. For example, 1942basic-block execution counts cannot be accommodated by the old file 1943format. 1944 1945The new file format is defined in header file @file{gmon_out.h}. It 1946consists of a header containing the magic cookie and a version number, 1947as well as some spare bytes available for future extensions. All data 1948in a profile data file is in the native format of the target for which 1949the profile was collected. @sc{gnu} @code{gprof} adapts automatically 1950to the byte-order in use. 1951 1952In the new file format, the header is followed by a sequence of 1953records. Currently, there are three different record types: histogram 1954records, call-graph arc records, and basic-block execution count 1955records. Each file can contain any number of each record type. When 1956reading a file, @sc{gnu} @code{gprof} will ensure records of the same type are 1957compatible with each other and compute the union of all records. For 1958example, for basic-block execution counts, the union is simply the sum 1959of all execution counts for each basic-block. 1960 1961@subsection Histogram Records 1962 1963Histogram records consist of a header that is followed by an array of 1964bins. The header contains the text-segment range that the histogram 1965spans, the size of the histogram in bytes (unlike in the old BSD 1966format, this does not include the size of the header), the rate of the 1967profiling clock, and the physical dimension that the bin counts 1968represent after being scaled by the profiling clock rate. The 1969physical dimension is specified in two parts: a long name of up to 15 1970characters and a single character abbreviation. For example, a 1971histogram representing real-time would specify the long name as 1972``seconds'' and the abbreviation as ``s''. This feature is useful for 1973architectures that support performance monitor hardware (which, 1974fortunately, is becoming increasingly common). For example, under DEC 1975OSF/1, the ``uprofile'' command can be used to produce a histogram of, 1976say, instruction cache misses. In this case, the dimension in the 1977histogram header could be set to ``i-cache misses'' and the abbreviation 1978could be set to ``1'' (because it is simply a count, not a physical 1979dimension). Also, the profiling rate would have to be set to 1 in 1980this case. 1981 1982Histogram bins are 16-bit numbers and each bin represent an equal 1983amount of text-space. For example, if the text-segment is one 1984thousand bytes long and if there are ten bins in the histogram, each 1985bin represents one hundred bytes. 1986 1987 1988@subsection Call-Graph Records 1989 1990Call-graph records have a format that is identical to the one used in 1991the BSD-derived file format. It consists of an arc in the call graph 1992and a count indicating the number of times the arc was traversed 1993during program execution. Arcs are specified by a pair of addresses: 1994the first must be within caller's function and the second must be 1995within the callee's function. When performing profiling at the 1996function level, these addresses can point anywhere within the 1997respective function. However, when profiling at the line-level, it is 1998better if the addresses are as close to the call-site/entry-point as 1999possible. This will ensure that the line-level call-graph is able to 2000identify exactly which line of source code performed calls to a 2001function. 2002 2003@subsection Basic-Block Execution Count Records 2004 2005Basic-block execution count records consist of a header followed by a 2006sequence of address/count pairs. The header simply specifies the 2007length of the sequence. In an address/count pair, the address 2008identifies a basic-block and the count specifies the number of times 2009that basic-block was executed. Any address within the basic-address can 2010be used. 2011 2012@node Internals 2013@section @code{gprof}'s Internal Operation 2014 2015Like most programs, @code{gprof} begins by processing its options. 2016During this stage, it may building its symspec list 2017(@code{sym_ids.c:@-sym_id_add}), if 2018options are specified which use symspecs. 2019@code{gprof} maintains a single linked list of symspecs, 2020which will eventually get turned into 12 symbol tables, 2021organized into six include/exclude pairs---one 2022pair each for the flat profile (INCL_FLAT/EXCL_FLAT), 2023the call graph arcs (INCL_ARCS/EXCL_ARCS), 2024printing in the call graph (INCL_GRAPH/EXCL_GRAPH), 2025timing propagation in the call graph (INCL_TIME/EXCL_TIME), 2026the annotated source listing (INCL_ANNO/EXCL_ANNO), 2027and the execution count listing (INCL_EXEC/EXCL_EXEC). 2028 2029After option processing, @code{gprof} finishes 2030building the symspec list by adding all the symspecs in 2031@code{default_excluded_list} to the exclude lists 2032EXCL_TIME and EXCL_GRAPH, and if line-by-line profiling is specified, 2033EXCL_FLAT as well. 2034These default excludes are not added to EXCL_ANNO, EXCL_ARCS, and EXCL_EXEC. 2035 2036Next, the BFD library is called to open the object file, 2037verify that it is an object file, 2038and read its symbol table (@code{core.c:@-core_init}), 2039using @code{bfd_canonicalize_symtab} after mallocing 2040an appropriately sized array of symbols. At this point, 2041function mappings are read (if the @samp{--file-ordering} option 2042has been specified), and the core text space is read into 2043memory (if the @samp{-c} option was given). 2044 2045@code{gprof}'s own symbol table, an array of Sym structures, 2046is now built. 2047This is done in one of two ways, by one of two routines, depending 2048on whether line-by-line profiling (@samp{-l} option) has been 2049enabled. 2050For normal profiling, the BFD canonical symbol table is scanned. 2051For line-by-line profiling, every 2052text space address is examined, and a new symbol table entry 2053gets created every time the line number changes. 2054In either case, two passes are made through the symbol 2055table---one to count the size of the symbol table required, 2056and the other to actually read the symbols. In between the 2057two passes, a single array of type @code{Sym} is created of 2058the appropriate length. 2059Finally, @code{symtab.c:@-symtab_finalize} 2060is called to sort the symbol table and remove duplicate entries 2061(entries with the same memory address). 2062 2063The symbol table must be a contiguous array for two reasons. 2064First, the @code{qsort} library function (which sorts an array) 2065will be used to sort the symbol table. 2066Also, the symbol lookup routine (@code{symtab.c:@-sym_lookup}), 2067which finds symbols 2068based on memory address, uses a binary search algorithm 2069which requires the symbol table to be a sorted array. 2070Function symbols are indicated with an @code{is_func} flag. 2071Line number symbols have no special flags set. 2072Additionally, a symbol can have an @code{is_static} flag 2073to indicate that it is a local symbol. 2074 2075With the symbol table read, the symspecs can now be translated 2076into Syms (@code{sym_ids.c:@-sym_id_parse}). Remember that a single 2077symspec can match multiple symbols. 2078An array of symbol tables 2079(@code{syms}) is created, each entry of which is a symbol table 2080of Syms to be included or excluded from a particular listing. 2081The master symbol table and the symspecs are examined by nested 2082loops, and every symbol that matches a symspec is inserted 2083into the appropriate syms table. This is done twice, once to 2084count the size of each required symbol table, and again to build 2085the tables, which have been malloced between passes. 2086From now on, to determine whether a symbol is on an include 2087or exclude symspec list, @code{gprof} simply uses its 2088standard symbol lookup routine on the appropriate table 2089in the @code{syms} array. 2090 2091Now the profile data file(s) themselves are read 2092(@code{gmon_io.c:@-gmon_out_read}), 2093first by checking for a new-style @samp{gmon.out} header, 2094then assuming this is an old-style BSD @samp{gmon.out} 2095if the magic number test failed. 2096 2097New-style histogram records are read by @code{hist.c:@-hist_read_rec}. 2098For the first histogram record, allocate a memory array to hold 2099all the bins, and read them in. 2100When multiple profile data files (or files with multiple histogram 2101records) are read, the memory ranges of each pair of histogram records 2102must be either equal, or non-overlapping. For each pair of histogram 2103records, the resolution (memory region size divided by the number of 2104bins) must be the same. The time unit must be the same for all 2105histogram records. If the above containts are met, all histograms 2106for the same memory range are merged. 2107 2108As each call graph record is read (@code{call_graph.c:@-cg_read_rec}), 2109the parent and child addresses 2110are matched to symbol table entries, and a call graph arc is 2111created by @code{cg_arcs.c:@-arc_add}, unless the arc fails a symspec 2112check against INCL_ARCS/EXCL_ARCS. As each arc is added, 2113a linked list is maintained of the parent's child arcs, and of the child's 2114parent arcs. 2115Both the child's call count and the arc's call count are 2116incremented by the record's call count. 2117 2118Basic-block records are read (@code{basic_blocks.c:@-bb_read_rec}), 2119but only if line-by-line profiling has been selected. 2120Each basic-block address is matched to a corresponding line 2121symbol in the symbol table, and an entry made in the symbol's 2122bb_addr and bb_calls arrays. Again, if multiple basic-block 2123records are present for the same address, the call counts 2124are cumulative. 2125 2126A gmon.sum file is dumped, if requested (@code{gmon_io.c:@-gmon_out_write}). 2127 2128If histograms were present in the data files, assign them to symbols 2129(@code{hist.c:@-hist_assign_samples}) by iterating over all the sample 2130bins and assigning them to symbols. Since the symbol table 2131is sorted in order of ascending memory addresses, we can 2132simple follow along in the symbol table as we make our pass 2133over the sample bins. 2134This step includes a symspec check against INCL_FLAT/EXCL_FLAT. 2135Depending on the histogram 2136scale factor, a sample bin may span multiple symbols, 2137in which case a fraction of the sample count is allocated 2138to each symbol, proportional to the degree of overlap. 2139This effect is rare for normal profiling, but overlaps 2140are more common during line-by-line profiling, and can 2141cause each of two adjacent lines to be credited with half 2142a hit, for example. 2143 2144If call graph data is present, @code{cg_arcs.c:@-cg_assemble} is called. 2145First, if @samp{-c} was specified, a machine-dependent 2146routine (@code{find_call}) scans through each symbol's machine code, 2147looking for subroutine call instructions, and adding them 2148to the call graph with a zero call count. 2149A topological sort is performed by depth-first numbering 2150all the symbols (@code{cg_dfn.c:@-cg_dfn}), so that 2151children are always numbered less than their parents, 2152then making a array of pointers into the symbol table and sorting it into 2153numerical order, which is reverse topological 2154order (children appear before parents). 2155Cycles are also detected at this point, all members 2156of which are assigned the same topological number. 2157Two passes are now made through this sorted array of symbol pointers. 2158The first pass, from end to beginning (parents to children), 2159computes the fraction of child time to propagate to each parent 2160and a print flag. 2161The print flag reflects symspec handling of INCL_GRAPH/EXCL_GRAPH, 2162with a parent's include or exclude (print or no print) property 2163being propagated to its children, unless they themselves explicitly appear 2164in INCL_GRAPH or EXCL_GRAPH. 2165A second pass, from beginning to end (children to parents) actually 2166propagates the timings along the call graph, subject 2167to a check against INCL_TIME/EXCL_TIME. 2168With the print flag, fractions, and timings now stored in the symbol 2169structures, the topological sort array is now discarded, and a 2170new array of pointers is assembled, this time sorted by propagated time. 2171 2172Finally, print the various outputs the user requested, which is now fairly 2173straightforward. The call graph (@code{cg_print.c:@-cg_print}) and 2174flat profile (@code{hist.c:@-hist_print}) are regurgitations of values 2175already computed. The annotated source listing 2176(@code{basic_blocks.c:@-print_annotated_source}) uses basic-block 2177information, if present, to label each line of code with call counts, 2178otherwise only the function call counts are presented. 2179 2180The function ordering code is marginally well documented 2181in the source code itself (@code{cg_print.c}). Basically, 2182the functions with the most use and the most parents are 2183placed first, followed by other functions with the most use, 2184followed by lower use functions, followed by unused functions 2185at the end. 2186 2187@node Debugging 2188@section Debugging @code{gprof} 2189 2190If @code{gprof} was compiled with debugging enabled, 2191the @samp{-d} option triggers debugging output 2192(to stdout) which can be helpful in understanding its operation. 2193The debugging number specified is interpreted as a sum of the following 2194options: 2195 2196@table @asis 2197@item 2 - Topological sort 2198Monitor depth-first numbering of symbols during call graph analysis 2199@item 4 - Cycles 2200Shows symbols as they are identified as cycle heads 2201@item 16 - Tallying 2202As the call graph arcs are read, show each arc and how 2203the total calls to each function are tallied 2204@item 32 - Call graph arc sorting 2205Details sorting individual parents/children within each call graph entry 2206@item 64 - Reading histogram and call graph records 2207Shows address ranges of histograms as they are read, and each 2208call graph arc 2209@item 128 - Symbol table 2210Reading, classifying, and sorting the symbol table from the object file. 2211For line-by-line profiling (@samp{-l} option), also shows line numbers 2212being assigned to memory addresses. 2213@item 256 - Static call graph 2214Trace operation of @samp{-c} option 2215@item 512 - Symbol table and arc table lookups 2216Detail operation of lookup routines 2217@item 1024 - Call graph propagation 2218Shows how function times are propagated along the call graph 2219@item 2048 - Basic-blocks 2220Shows basic-block records as they are read from profile data 2221(only meaningful with @samp{-l} option) 2222@item 4096 - Symspecs 2223Shows symspec-to-symbol pattern matching operation 2224@item 8192 - Annotate source 2225Tracks operation of @samp{-A} option 2226@end table 2227 2228@node GNU Free Documentation License 2229@appendix GNU Free Documentation License 2230@include fdl.texi 2231 2232@bye 2233 2234NEEDS AN INDEX 2235 2236-T - "traditional BSD style": How is it different? Should the 2237differences be documented? 2238 2239example flat file adds up to 100.01%... 2240 2241note: time estimates now only go out to one decimal place (0.0), where 2242they used to extend two (78.67). 2243