gprof
Command Summary
gprof
's Output
gprof
Output
gprof
This manual describes the GNU profiler, gprof
, and how you
can use it to determine which parts of a program are taking most of the
execution time. We assume that you know how to write, compile, and
execute programs. GNU gprof
was written by Jay Fenlason.
This manual was edited January 1993 by Jeffrey Osier and updated September 1997 by Brent Baccala.
Copyright (C) 1988, 1992, 1997, 1998 Free Software Foundation, Inc.
Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies.
Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one.
Permission is granted to copy and distribute translations of this manual into another language, under the same conditions as for modified versions.
Profiling allows you to learn where your program spent its time and which functions called which other functions while it was executing. This information can show you which pieces of your program are slower than you expected, and might be candidates for rewriting to make your program execute faster. It can also tell you which functions are being called more or less often than you expected. This may help you spot bugs that had otherwise been unnoticed.
Since the profiler uses information collected during the actual execution of your program, it can be used on programs that are too large or too complex to analyze by reading the source. However, how your program is run will affect the information that shows up in the profile data. If you don't use some feature of your program while it is being profiled, no profile information will be generated for that feature.
Profiling has several steps:
gprof
to analyze the profile data.
See section gprof
Command Summary.
The next three chapters explain these steps in greater detail.
Several forms of output are available from the analysis.
The flat profile shows how much time your program spent in each function, and how many times that function was called. If you simply want to know which functions burn most of the cycles, it is stated concisely here. See section The Flat Profile.
The call graph shows, for each function, which functions called it, which other functions it called, and how many times. There is also an estimate of how much time was spent in the subroutines of each function. This can suggest places where you might try to eliminate function calls that use a lot of time. See section The Call Graph.
The annotated source listing is a copy of the program's source code, labeled with the number of times each line of the program was executed. See section The Annotated Source Listing.
To better understand how profiling works, you may wish to read a description of its implementation. See section Implementation of Profiling.
The first step in generating profile information for your program is to compile and link it with profiling enabled.
To compile a source file for profiling, specify the `-pg' option when you run the compiler. (This is in addition to the options you normally use.)
To link the program for profiling, if you use a compiler such as cc
to do the linking, simply specify `-pg' in addition to your usual
options. The same option, `-pg', alters either compilation or linking
to do what is necessary for profiling. Here are examples:
cc -g -c myprog.c utils.c -pg cc -o myprog myprog.o utils.o -pg
The `-pg' option also works with a command that both compiles and links:
cc -o myprog myprog.c utils.c -g -pg
If you run the linker ld
directly instead of through a compiler
such as cc
, you may have to specify a profiling startup file
`gcrt0.o' as the first input file instead of the usual startup
file `crt0.o'. In addition, you would probably want to
specify the profiling C library, `libc_p.a', by writing
`-lc_p' instead of the usual `-lc'. This is not absolutely
necessary, but doing this gives you number-of-calls information for
standard library functions such as read
and open
. For
example:
ld -o myprog /lib/gcrt0.o myprog.o utils.o -lc_p
If you compile only some of the modules of the program with `-pg', you
can still profile the program, but you won't get complete information about
the modules that were compiled without `-pg'. The only information
you get for the functions in those modules is the total time spent in them;
there is no record of how many times they were called, or from where. This
will not affect the flat profile (except that the calls
field for
the functions will be blank), but will greatly reduce the usefulness of the
call graph.
If you wish to perform line-by-line profiling, you will also need to specify the `-g' option, instructing the compiler to insert debugging symbols into the program that match program addresses to source code lines. See section Line-by-line Profiling.
In addition to the `-pg' and `-g' options,
you may also wish to specify the `-a' option when compiling.
This will instrument
the program to perform basic-block counting. As the program runs,
it will count how many times it executed each branch of each `if'
statement, each iteration of each `do' loop, etc. This will
enable gprof
to construct an annotated source code
listing showing how many times each line of code was executed.
Once the program is compiled for profiling, you must run it in order to
generate the information that gprof
needs. Simply run the program
as usual, using the normal arguments, file names, etc. The program should
run normally, producing the same output as usual. It will, however, run
somewhat slower than normal because of the time spent collecting and the
writing the profile data.
The way you run the program--the arguments and input that you give it--may have a dramatic effect on what the profile information shows. The profile data will describe the parts of the program that were activated for the particular input you use. For example, if the first command you give to your program is to quit, the profile data will show the time used in initialization and in cleanup, but not much else.
Your program will write the profile data into a file called `gmon.out' just before exiting. If there is already a file called `gmon.out', its contents are overwritten. There is currently no way to tell the program to write the profile data under a different name, but you can rename the file afterward if you are concerned that it may be overwritten.
In order to write the `gmon.out' file properly, your program must exit
normally: by returning from main
or by calling exit
. Calling
the low-level function _exit
does not write the profile data, and
neither does abnormal termination due to an unhandled signal.
The `gmon.out' file is written in the program's current working
directory at the time it exits. This means that if your program calls
chdir
, the `gmon.out' file will be left in the last directory
your program chdir
'd to. If you don't have permission to write in
this directory, the file is not written, and you will get an error message.
Older versions of the GNU profiling library may also write a file
called `bb.out'. This file, if present, contains an human-readable
listing of the basic-block execution counts. Unfortunately, the
appearance of a human-readable `bb.out' means the basic-block
counts didn't get written into `gmon.out'.
The Perl script bbconv.pl
, included with the gprof
source distribution, will convert a `bb.out' file into
a format readable by gprof
.
gprof
Command Summary
After you have a profile data file `gmon.out', you can run gprof
to interpret the information in it. The gprof
program prints a
flat profile and a call graph on standard output. Typically you would
redirect the output of gprof
into a file with `>'.
You run gprof
like this:
gprof options [executable-file [profile-data-files...]] [> outfile]
Here square-brackets indicate optional arguments.
If you omit the executable file name, the file `a.out' is used. If you give no profile data file name, the file `gmon.out' is used. If any file is not in the proper format, or if the profile data file does not appear to belong to the executable file, an error message is printed.
You can give more than one profile data file by entering all their names after the executable file name; then the statistics in all the data files are summed together.
The order of these options does not matter.
These options specify which of several output formats
gprof
should produce.
Many of these options take an optional symspec to specify functions to be included or excluded. These options can be specified multiple times, with different symspecs, to include or exclude sets of symbols. See section Symspecs.
Specifying any of these options overrides the default (`-p -q'), which prints a flat profile and call graph analysis for all functions.
-A[symspec]
--annotated-source[=symspec]
gprof
to print annotated source code.
If symspec is specified, print output only for matching symbols.
See section The Annotated Source Listing.
-b
--brief
gprof
doesn't print the
verbose blurbs that try to explain the meaning of all of the fields in
the tables. This is useful if you intend to print out the output, or
are tired of seeing the blurbs.
-C[symspec]
--exec-counts[=symspec]
gprof
to
print a tally of functions and the number of times each was called.
If symspec is specified, print tally only for matching symbols.
If the profile data file contains basic-block count records, specifing
the `-l' option, along with `-C', will cause basic-block
execution counts to be tallied and displayed.
-i
--file-info
gprof
to display summary information
about the profile data file(s) and then exit. The number of histogram,
call graph, and basic-block count records is displayed.
-I dirs
--directory-path=dirs
-J[symspec]
--no-annotated-source[=symspec]
gprof
not to
print annotated source code.
If symspec is specified, gprof
prints annotated source,
but excludes matching symbols.
-L
--print-path
gprof
to print the full pathname of
source filenames, which is determined
from symbolic debugging information in the image file
and is relative to the directory in which the compiler
was invoked.
-p[symspec]
--flat-profile[=symspec]
gprof
to print a flat profile.
If symspec is specified, print flat profile only for matching symbols.
See section The Flat Profile.
-P[symspec]
--no-flat-profile[=symspec]
gprof
to suppress printing a flat profile.
If symspec is specified, gprof
prints a flat profile,
but excludes matching symbols.
-q[symspec]
--graph[=symspec]
gprof
to print the call graph analysis.
If symspec is specified, print call graph only for matching symbols
and their children.
See section The Call Graph.
-Q[symspec]
--no-graph[=symspec]
gprof
to suppress printing the
call graph.
If symspec is specified, gprof
prints a call graph,
but excludes matching symbols.
-y
--separate-files
-Z[symspec]
--no-exec-counts[=symspec]
gprof
not to
print a tally of functions and the number of times each was called.
If symspec is specified, print tally, but exclude matching symbols.
--function-ordering
gprof
to print a
suggested function ordering for the program based on profiling data.
This option suggests an ordering which may improve paging, tlb and
cache behavior for the program on systems which support arbitrary
ordering of functions in an executable.
The exact details of how to force the linker to place functions
in a particular order is system dependent and out of the scope of this
manual.
--file-ordering map_file
gprof
to print a
suggested .o link line ordering for the program based on profiling data.
This option suggests an ordering which may improve paging, tlb and
cache behavior for the program on systems which do not support arbitrary
ordering of functions in an executable.
Use of the `-a' argument is highly recommended with this option.
The map_file argument is a pathname to a file which provides
function name to object file mappings. The format of the file is similar to
the output of the program nm
.
c-parse.o:00000000 T yyparse c-parse.o:00000004 C yyerrflag c-lang.o:00000000 T maybe_objc_method_name c-lang.o:00000000 T print_lang_statistics c-lang.o:00000000 T recognize_objc_keyword c-decl.o:00000000 T print_lang_identifier c-decl.o:00000000 T print_lang_type ...GNU
nm
`--extern-only' `--defined-only' `-v' `--print-file-name' can be used to create map_file.
-T
--traditional
gprof
to print its output in
"traditional" BSD style.
-w width
--width=width
-x
--all-lines
tcov
's `-a'.
--demangle
--no-demangle
--no-demangle
option may be used to turn off demangling.
-a
--no-static
gprof
to suppress the printing of
statically declared (private) functions. (These are functions whose
names are not listed as global, and which are not visible outside the
file/function/block where they were defined.) Time spent in these
functions, calls to/from them, etc, will all be attributed to the
function that was loaded directly before it in the executable file.
This option affects both the flat profile and the call graph.
-c
--static-call-graph
-D
--ignore-non-functions
gprof
to ignore symbols which
are not known to be functions. This option will give more accurate
profile data on systems where it is supported (Solaris and HPUX for
example).
-k from/to
-l
--line
gprof
, and magnifies statistical
inaccuracies.
See section Statistical Sampling Error.
-m num
--min-count=num
-n[symspec]
--time[=symspec]
gprof
, in its call graph analysis,
to only propagate times for symbols matching symspec.
-N[symspec]
--no-time[=symspec]
gprof
, in its call graph analysis,
not to propagate times for symbols matching symspec.
-z
--display-unused-functions
gprof
will mention all
functions in the flat profile, even those that were never called, and
that had no time spent in them. This is useful in conjunction with the
`-c' option for discovering which routines were never called.
-d[num]
--debug[=num]
gprof
.
-Oname
--file-format=name
-s
--sum
gprof
to summarize the information
in the profile data files it read in, and write out a profile data
file called `gmon.sum', which contains all the information from
the profile data files that gprof
read in. The file `gmon.sum'
may be one of the specified input files; the effect of this is to
merge the data in the other input files into `gmon.sum'.
Eventually you can run gprof
again without `-s' to analyze the
cumulative data in the file `gmon.sum'.
-v
--version
gprof
to print the current version
number, and then exit.
-e function_name
gprof
to not print
information about the function function_name (and its
children...) in the call graph. The function will still be listed
as a child of any functions that call it, but its index number will be
shown as `[not printed]'. More than one `-e' option may be
given; only one function_name may be indicated with each `-e'
option.
-E function_name
-E function
option works like the -e
option, but
time spent in the function (and children who were not called from
anywhere else), will not be used to compute the percentages-of-time for
the call graph. More than one `-E' option may be given; only one
function_name may be indicated with each `-E' option.
-f function_name
gprof
to limit the
call graph to the function function_name and its children (and
their children...). More than one `-f' option may be given;
only one function_name may be indicated with each `-f'
option.
-F function_name
-f
option, but
only time spent in the function and its children (and their
children...) will be used to determine total-time and
percentages-of-time for the call graph. More than one `-F' option
may be given; only one function_name may be indicated with each
`-F' option. The `-F' option overrides the `-E' option.
Note that only one function can be specified with each -e
,
-E
, -f
or -F
option. To specify more than one
function, use multiple options. For example, this command:
gprof -e boring -f foo -f bar myprogram > gprof.output
lists in the call graph all functions that were reached from either
foo
or bar
and were not reachable from boring
.
Many of the output options allow functions to be included or excluded using symspecs (symbol specifications), which observe the following syntax:
filename_containing_a_dot | funcname_not_containing_a_dot | linenumber | ( [ any_filename ] `:' ( any_funcname | linenumber ) )
Here are some sample symspecs:
main.c
main
main.c:main
main.c:134
gprof
's Output
gprof
can produce several different output styles, the
most important of which are described below. The simplest output
styles (file information, execution count, and function and file ordering)
are not described here, but are documented with the respective options
that trigger them.
See section Output Options.
The flat profile shows the total amount of time your program spent executing each function. Unless the `-z' option is given, functions with no apparent time spent in them, and no apparent calls to them, are not mentioned. Note that if a function was not compiled for profiling, and didn't run long enough to show up on the program counter histogram, it will be indistinguishable from a function that was never called.
This is part of a flat profile for a small program:
Flat profile: Each sample counts as 0.01 seconds. % cumulative self self total time seconds seconds calls ms/call ms/call name 33.34 0.02 0.02 7208 0.00 0.00 open 16.67 0.03 0.01 244 0.04 0.12 offtime 16.67 0.04 0.01 8 1.25 1.25 memccpy 16.67 0.05 0.01 7 1.43 1.43 write 16.67 0.06 0.01 mcount 0.00 0.06 0.00 236 0.00 0.00 tzset 0.00 0.06 0.00 192 0.00 0.00 tolower 0.00 0.06 0.00 47 0.00 0.00 strlen 0.00 0.06 0.00 45 0.00 0.00 strchr 0.00 0.06 0.00 1 0.00 50.00 main 0.00 0.06 0.00 1 0.00 0.00 memcpy 0.00 0.06 0.00 1 0.00 10.11 print 0.00 0.06 0.00 1 0.00 0.00 profil 0.00 0.06 0.00 1 0.00 50.00 report ...
The functions are sorted by first by decreasing run-time spent in them, then by decreasing number of calls, then alphabetically by name. The functions `mcount' and `profil' are part of the profiling aparatus and appear in every flat profile; their time gives a measure of the amount of overhead due to profiling.
Just before the column headers, a statement appears indicating how much time each sample counted as. This sampling period estimates the margin of error in each of the time figures. A time figure that is not much larger than this is not reliable. In this example, each sample counted as 0.01 seconds, suggesting a 100 Hz sampling rate. The program's total execution time was 0.06 seconds, as indicated by the `cumulative seconds' field. Since each sample counted for 0.01 seconds, this means only six samples were taken during the run. Two of the samples occured while the program was in the `open' function, as indicated by the `self seconds' field. Each of the other four samples occured one each in `offtime', `memccpy', `write', and `mcount'. Since only six samples were taken, none of these values can be regarded as particularly reliable. In another run, the `self seconds' field for `mcount' might well be `0.00' or `0.02'. See section Statistical Sampling Error, for a complete discussion.
The remaining functions in the listing (those whose `self seconds' field is `0.00') didn't appear in the histogram samples at all. However, the call graph indicated that they were called, so therefore they are listed, sorted in decreasing order by the `calls' field. Clearly some time was spent executing these functions, but the paucity of histogram samples prevents any determination of how much time each took.
Here is what the fields in each line mean:
% time
cumulative seconds
self seconds
calls
self ms/call
total ms/call
name
The call graph shows how much time was spent in each function and its children. From this information, you can find functions that, while they themselves may not have used much time, called other functions that did use unusual amounts of time.
Here is a sample call from a small program. This call came from the
same gprof
run as the flat profile example in the previous
chapter.
granularity: each sample hit covers 2 byte(s) for 20.00% of 0.05 seconds index % time self children called name <spontaneous> [1] 100.0 0.00 0.05 start [1] 0.00 0.05 1/1 main [2] 0.00 0.00 1/2 on_exit [28] 0.00 0.00 1/1 exit [59] ----------------------------------------------- 0.00 0.05 1/1 start [1] [2] 100.0 0.00 0.05 1 main [2] 0.00 0.05 1/1 report [3] ----------------------------------------------- 0.00 0.05 1/1 main [2] [3] 100.0 0.00 0.05 1 report [3] 0.00 0.03 8/8 timelocal [6] 0.00 0.01 1/1 print [9] 0.00 0.01 9/9 fgets [12] 0.00 0.00 12/34 strncmp <cycle 1> [40] 0.00 0.00 8/8 lookup [20] 0.00 0.00 1/1 fopen [21] 0.00 0.00 8/8 chewtime [24] 0.00 0.00 8/16 skipspace [44] ----------------------------------------------- [4] 59.8 0.01 0.02 8+472 <cycle 2 as a whole> [4] 0.01 0.02 244+260 offtime <cycle 2> [7] 0.00 0.00 236+1 tzset <cycle 2> [26] -----------------------------------------------
The lines full of dashes divide this table into entries, one for each function. Each entry has one or more lines.
In each entry, the primary line is the one that starts with an index number in square brackets. The end of this line says which function the entry is for. The preceding lines in the entry describe the callers of this function and the following lines describe its subroutines (also called children when we speak of the call graph).
The entries are sorted by time spent in the function and its subroutines.
The internal profiling function mcount
(see section The Flat Profile)
is never mentioned in the call graph.
The primary line in a call graph entry is the line that describes the function which the entry is about and gives the overall statistics for this function.
For reference, we repeat the primary line from the entry for function
report
in our main example, together with the heading line that
shows the names of the fields:
index % time self children called name ... [3] 100.0 0.00 0.05 1 report [3]
Here is what the fields in the primary line mean:
index
% time
self
seconds
field
for this function in the flat profile.
children
self
and children
entries of the children listed directly below this
function.
called
report
was called once from
main
.
name
gnurr
is part of
cycle number one, and has index number twelve, its primary line would
be end like this:
gnurr <cycle 1> [12]
A function's entry has a line for each function it was called by. These lines' fields correspond to the fields of the primary line, but their meanings are different because of the difference in context.
For reference, we repeat two lines from the entry for the function
report
, the primary line and one caller-line preceding it, together
with the heading line that shows the names of the fields:
index % time self children called name ... 0.00 0.05 1/1 main [2] [3] 100.0 0.00 0.05 1 report [3]
Here are the meanings of the fields in the caller-line for report
called from main
:
self
report
itself when it was
called from main
.
children
report
when report
was called from main
.
The sum of the self
and children
fields is an estimate
of the amount of time spent within calls to report
from main
.
called
report
was called from main
,
followed by the total number of nonrecursive calls to report
from
all its callers.
name and index number
report
to which this line applies,
followed by the caller's index number.
Not all functions have entries in the call graph; some
options to gprof
request the omission of certain functions.
When a caller has no entry of its own, it still has caller-lines
in the entries of the functions it calls.
If the caller is part of a recursion cycle, the cycle number is
printed between the name and the index number.
If the identity of the callers of a function cannot be determined, a dummy caller-line is printed which has `<spontaneous>' as the "caller's name" and all other fields blank. This can happen for signal handlers.
A function's entry has a line for each of its subroutines--in other words, a line for each other function that it called. These lines' fields correspond to the fields of the primary line, but their meanings are different because of the difference in context.
For reference, we repeat two lines from the entry for the function
main
, the primary line and a line for a subroutine, together
with the heading line that shows the names of the fields:
index % time self children called name ... [2] 100.0 0.00 0.05 1 main [2] 0.00 0.05 1/1 report [3]
Here are the meanings of the fields in the subroutine-line for main
calling report
:
self
report
when report
was called from main
.
children
report
when report
was called from main
.
The sum of the self
and children
fields is an estimate
of the total time spent in calls to report
from main
.
called
report
from main
followed by the total number of nonrecursive calls to report
.
This ratio is used to determine how much of report
's self
and children
time gets credited to main
.
See section Estimating children
Times.
name
main
to which this line applies,
followed by the subroutine's index number.
If the caller is part of a recursion cycle, the cycle number is
printed between the name and the index number.
The graph may be complicated by the presence of cycles of
recursion in the call graph. A cycle exists if a function calls
another function that (directly or indirectly) calls (or appears to
call) the original function. For example: if a
calls b
,
and b
calls a
, then a
and b
form a cycle.
Whenever there are call paths both ways between a pair of functions, they
belong to the same cycle. If a
and b
call each other and
b
and c
call each other, all three make one cycle. Note that
even if b
only calls a
if it was not called from a
,
gprof
cannot determine this, so a
and b
are still
considered a cycle.
The cycles are numbered with consecutive integers. When a function belongs to a cycle, each time the function name appears in the call graph it is followed by `<cycle number>'.
The reason cycles matter is that they make the time values in the call
graph paradoxical. The "time spent in children" of a
should
include the time spent in its subroutine b
and in b
's
subroutines--but one of b
's subroutines is a
! How much of
a
's time should be included in the children of a
, when
a
is indirectly recursive?
The way gprof
resolves this paradox is by creating a single entry
for the cycle as a whole. The primary line of this entry describes the
total time spent directly in the functions of the cycle. The
"subroutines" of the cycle are the individual functions of the cycle, and
all other functions that were called directly by them. The "callers" of
the cycle are the functions, outside the cycle, that called functions in
the cycle.
Here is an example portion of a call graph which shows a cycle containing
functions a
and b
. The cycle was entered by a call to
a
from main
; both a
and b
called c
.
index % time self children called name ---------------------------------------- 1.77 0 1/1 main [2] [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3] 1.02 0 3 b <cycle 1> [4] 0.75 0 2 a <cycle 1> [5] ---------------------------------------- 3 a <cycle 1> [5] [4] 52.85 1.02 0 0 b <cycle 1> [4] 2 a <cycle 1> [5] 0 0 3/6 c [6] ---------------------------------------- 1.77 0 1/1 main [2] 2 b <cycle 1> [4] [5] 38.86 0.75 0 1 a <cycle 1> [5] 3 b <cycle 1> [4] 0 0 3/6 c [6] ----------------------------------------
(The entire call graph for this program contains in addition an entry for
main
, which calls a
, and an entry for c
, with callers
a
and b
.)
index % time self children called name <spontaneous> [1] 100.00 0 1.93 0 start [1] 0.16 1.77 1/1 main [2] ---------------------------------------- 0.16 1.77 1/1 start [1] [2] 100.00 0.16 1.77 1 main [2] 1.77 0 1/1 a <cycle 1> [5] ---------------------------------------- 1.77 0 1/1 main [2] [3] 91.71 1.77 0 1+5 <cycle 1 as a whole> [3] 1.02 0 3 b <cycle 1> [4] 0.75 0 2 a <cycle 1> [5] 0 0 6/6 c [6] ---------------------------------------- 3 a <cycle 1> [5] [4] 52.85 1.02 0 0 b <cycle 1> [4] 2 a <cycle 1> [5] 0 0 3/6 c [6] ---------------------------------------- 1.77 0 1/1 main [2] 2 b <cycle 1> [4] [5] 38.86 0.75 0 1 a <cycle 1> [5] 3 b <cycle 1> [4] 0 0 3/6 c [6] ---------------------------------------- 0 0 3/6 b <cycle 1> [4] 0 0 3/6 a <cycle 1> [5] [6] 0.00 0 0 6 c [6] ----------------------------------------
The self
field of the cycle's primary line is the total time
spent in all the functions of the cycle. It equals the sum of the
self
fields for the individual functions in the cycle, found
in the entry in the subroutine lines for these functions.
The children
fields of the cycle's primary line and subroutine lines
count only subroutines outside the cycle. Even though a
calls
b
, the time spent in those calls to b
is not counted in
a
's children
time. Thus, we do not encounter the problem of
what to do when the time in those calls to b
includes indirect
recursive calls back to a
.
The children
field of a caller-line in the cycle's entry estimates
the amount of time spent in the whole cycle, and its other
subroutines, on the times when that caller called a function in the cycle.
The calls
field in the primary line for the cycle has two numbers:
first, the number of times functions in the cycle were called by functions
outside the cycle; second, the number of times they were called by
functions in the cycle (including times when a function in the cycle calls
itself). This is a generalization of the usual split into nonrecursive and
recursive calls.
The calls
field of a subroutine-line for a cycle member in the
cycle's entry says how many time that function was called from functions in
the cycle. The total of all these is the second number in the primary line's
calls
field.
In the individual entry for a function in a cycle, the other functions in
the same cycle can appear as subroutines and as callers. These lines show
how many times each function in the cycle called or was called from each other
function in the cycle. The self
and children
fields in these
lines are blank because of the difficulty of defining meanings for them
when recursion is going on.
gprof
's `-l' option causes the program to perform
line-by-line profiling. In this mode, histogram
samples are assigned not to functions, but to individual
lines of source code. The program usually must be compiled
with a `-g' option, in addition to `-pg', in order
to generate debugging symbols for tracking source code lines.
The flat profile is the most useful output table
in line-by-line mode.
The call graph isn't as useful as normal, since
the current version of gprof
does not propagate
call graph arcs from source code lines to the enclosing function.
The call graph does, however, show each line of code
that called each function, along with a count.
Here is a section of gprof
's output, without line-by-line profiling.
Note that ct_init
accounted for four histogram hits, and
13327 calls to init_block
.
Flat profile: Each sample counts as 0.01 seconds. % cumulative self self total time seconds seconds calls us/call us/call name 30.77 0.13 0.04 6335 6.31 6.31 ct_init Call graph (explanation follows) granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds index % time self children called name 0.00 0.00 1/13496 name_too_long 0.00 0.00 40/13496 deflate 0.00 0.00 128/13496 deflate_fast 0.00 0.00 13327/13496 ct_init [7] 0.0 0.00 0.00 13496 init_block
Now let's look at some of gprof
's output from the same program run,
this time with line-by-line profiling enabled. Note that ct_init
's
four histogram hits are broken down into four lines of source code - one hit
occured on each of lines 349, 351, 382 and 385. In the call graph,
note how
ct_init
's 13327 calls to init_block
are broken down
into one call from line 396, 3071 calls from line 384, 3730 calls
from line 385, and 6525 calls from 387.
Flat profile: Each sample counts as 0.01 seconds. % cumulative self time seconds seconds calls name 7.69 0.10 0.01 ct_init (trees.c:349) 7.69 0.11 0.01 ct_init (trees.c:351) 7.69 0.12 0.01 ct_init (trees.c:382) 7.69 0.13 0.01 ct_init (trees.c:385) Call graph (explanation follows) granularity: each sample hit covers 4 byte(s) for 7.69% of 0.13 seconds % time self children called name 0.00 0.00 1/13496 name_too_long (gzip.c:1440) 0.00 0.00 1/13496 deflate (deflate.c:763) 0.00 0.00 1/13496 ct_init (trees.c:396) 0.00 0.00 2/13496 deflate (deflate.c:727) 0.00 0.00 4/13496 deflate (deflate.c:686) 0.00 0.00 5/13496 deflate (deflate.c:675) 0.00 0.00 12/13496 deflate (deflate.c:679) 0.00 0.00 16/13496 deflate (deflate.c:730) 0.00 0.00 128/13496 deflate_fast (deflate.c:654) 0.00 0.00 3071/13496 ct_init (trees.c:384) 0.00 0.00 3730/13496 ct_init (trees.c:385) 0.00 0.00 6525/13496 ct_init (trees.c:387) [6] 0.0 0.00 0.00 13496 init_block (trees.c:408)
gprof
's `-A' option triggers an annotated source listing,
which lists the program's source code, each function labeled with the
number of times it was called. You may also need to specify the
`-I' option, if gprof
can't find the source code files.
Compiling with `gcc ... -g -pg -a' augments your program
with basic-block counting code, in addition to function counting code.
This enables gprof
to determine how many times each line
of code was exeucted.
For example, consider the following function, taken from gzip,
with line numbers added:
1 ulg updcrc(s, n) 2 uch *s; 3 unsigned n; 4 { 5 register ulg c; 6 7 static ulg crc = (ulg)0xffffffffL; 8 9 if (s == NULL) { 10 c = 0xffffffffL; 11 } else { 12 c = crc; 13 if (n) do { 14 c = crc_32_tab[...]; 15 } while (--n); 16 } 17 crc = c; 18 return c ^ 0xffffffffL; 19 }
updcrc
has at least five basic-blocks.
One is the function itself. The
if
statement on line 9 generates two more basic-blocks, one
for each branch of the if
. A fourth basic-block results from
the if
on line 13, and the contents of the do
loop form
the fifth basic-block. The compiler may also generate additional
basic-blocks to handle various special cases.
A program augmented for basic-block counting can be analyzed with
gprof -l -A
. I also suggest use of the `-x' option,
which ensures that each line of code is labeled at least once.
Here is updcrc
's
annotated source listing for a sample gzip
run:
ulg updcrc(s, n) uch *s; unsigned n; 2 ->{ register ulg c; static ulg crc = (ulg)0xffffffffL; 2 -> if (s == NULL) { 1 -> c = 0xffffffffL; 1 -> } else { 1 -> c = crc; 1 -> if (n) do { 26312 -> c = crc_32_tab[...]; 26312,1,26311 -> } while (--n); } 2 -> crc = c; 2 -> return c ^ 0xffffffffL; 2 ->}
In this example, the function was called twice, passing once through
each branch of the if
statement. The body of the do
loop was executed a total of 26312 times. Note how the while
statement is annotated. It began execution 26312 times, once for
each iteration through the loop. One of those times (the last time)
it exited, while it branched back to the beginning of the loop 26311 times.
gprof
Output
The run-time figures that gprof
gives you are based on a sampling
process, so they are subject to statistical inaccuracy. If a function runs
only a small amount of time, so that on the average the sampling process
ought to catch that function in the act only once, there is a pretty good
chance it will actually find that function zero times, or twice.
By contrast, the number-of-calls and basic-block figures are derived by counting, not sampling. They are completely accurate and will not vary from run to run if your program is deterministic.
The sampling period that is printed at the beginning of the flat profile says how often samples are taken. The rule of thumb is that a run-time figure is accurate if it is considerably bigger than the sampling period.
The actual amount of error can be predicted.
For n samples, the expected error
is the square-root of n. For example,
if the sampling period is 0.01 seconds and foo
's run-time is 1 second,
n is 100 samples (1 second/0.01 seconds), sqrt(n) is 10 samples, so
the expected error in foo
's run-time is 0.1 seconds (10*0.01 seconds),
or ten percent of the observed value.
Again, if the sampling period is 0.01 seconds and bar
's run-time is
100 seconds, n is 10000 samples, sqrt(n) is 100 samples, so
the expected error in bar
's run-time is 1 second,
or one percent of the observed value.
It is likely to
vary this much on the average from one profiling run to the next.
(Sometimes it will vary more.)
This does not mean that a small run-time figure is devoid of information. If the program's total run-time is large, a small run-time for one function does tell you that that function used an insignificant fraction of the whole program's time. Usually this means it is not worth optimizing.
One way to get more accuracy is to give your program more (but similar)
input data so it will take longer. Another way is to combine the data from
several runs, using the `-s' option of gprof
. Here is how:
gprof -s executable-file gmon.out gmon.sum
gprof executable-file gmon.sum > output-file
children
Times
Some of the figures in the call graph are estimates--for example, the
children
time values and all the the time figures in caller and
subroutine lines.
There is no direct information about these measurements in the profile
data itself. Instead, gprof
estimates them by making an assumption
about your program that might or might not be true.
The assumption made is that the average time spent in each call to any
function foo
is not correlated with who called foo
. If
foo
used 5 seconds in all, and 2/5 of the calls to foo
came
from a
, then foo
contributes 2 seconds to a
's
children
time, by assumption.
This assumption is usually true enough, but for some programs it is far
from true. Suppose that foo
returns very quickly when its argument
is zero; suppose that a
always passes zero as an argument, while
other callers of foo
pass other arguments. In this program, all the
time spent in foo
is in the calls from callers other than a
.
But gprof
has no way of knowing this; it will blindly and
incorrectly charge 2 seconds of time in foo
to the children of
a
.
We hope some day to put more complete data into `gmon.out', so that this assumption is no longer needed, if we can figure out how. For the nonce, the estimated figures are usually more useful than misleading.
gprof -l -C objfile | sort -k 3 -n -rThis listing will show you the lines in your code executed most often, but not necessarily those that consumed the most time.
gprof -l
and lookup the function in the call graph.
The callers will be broken down by function and line number.
for i in `seq 1 100`; do fastprog mv gmon.out gmon.out.$i done gprof -s fastprog gmon.out.* gprof fastprog gmon.sumIf your program is completely deterministic, all the call counts will be simple multiples of 100 (i.e. a function called once in each run will appear with a call count of 100).
gprof
GNU gprof
and Berkeley Unix gprof
use the same data
file `gmon.out', and provide essentially the same information. But
there are a few differences.
gprof
uses a new, generalized file format with support
for basic-block execution counts and non-realtime histograms. A magic
cookie and version number allows gprof
to easily identify
new style files. Old BSD-style files can still be read.
See section Profiling Data File Format.
gprof
lists the function as a
parent and as a child, with a calls
field that lists the number
of recursive calls. GNU gprof
omits these lines and puts
the number of recursive calls in the primary line.
gprof
still lists it as a subroutine of functions that call it.
gprof
accepts the `-k' with its argument
in the form `from/to', instead of `from to'.
gprof
prints all of their counts, seperated by commas.
gprof
prints blurbs after the tables, so that you can see the
tables without skipping the blurbs.
Profiling works by changing how every function in your program is compiled
so that when it is called, it will stash away some information about where
it was called from. From this, the profiler can figure out what function
called it, and can count how many times it was called. This change is made
by the compiler when your program is compiled with the `-pg' option,
which causes every function to call mcount
(or _mcount
, or __mcount
, depending on the OS and compiler)
as one of its first operations.
The mcount
routine, included in the profiling library,
is responsible for recording in an in-memory call graph table
both its parent routine (the child) and its parent's parent. This is
typically done by examining the stack frame to find both
the address of the child, and the return address in the original parent.
Since this is a very machine-dependant operation, mcount
itself is typically a short assembly-language stub routine
that extracts the required
information, and then calls __mcount_internal
(a normal C function) with two arguments - frompc
and selfpc
.
__mcount_internal
is responsible for maintaining
the in-memory call graph, which records frompc
, selfpc
,
and the number of times each of these call arcs was transversed.
GCC Version 2 provides a magical function (__builtin_return_address
),
which allows a generic mcount
function to extract the
required information from the stack frame. However, on some
architectures, most notably the SPARC, using this builtin can be
very computationally expensive, and an assembly language version
of mcount
is used for performance reasons.
Number-of-calls information for library routines is collected by using a special version of the C library. The programs in it are the same as in the usual C library, but they were compiled with `-pg'. If you link your program with `gcc ... -pg', it automatically uses the profiling version of the library.
Profiling also involves watching your program as it runs, and keeping a histogram of where the program counter happens to be every now and then. Typically the program counter is looked at around 100 times per second of run time, but the exact frequency may vary from system to system.
This is done is one of two ways. Most UNIX-like operating systems
provide a profil()
system call, which registers a memory
array with the kernel, along with a scale
factor that determines how the program's address space maps
into the array.
Typical scaling values cause every 2 to 8 bytes of address space
to map into a single array slot.
On every tick of the system clock
(assuming the profiled program is running), the value of the
program counter is examined and the corresponding slot in
the memory array is incremented. Since this is done in the kernel,
which had to interrupt the process anyway to handle the clock
interrupt, very little additional system overhead is required.
However, some operating systems, most notably Linux 2.0 (and earlier),
do not provide a profil()
system call. On such a system,
arrangements are made for the kernel to periodically deliver
a signal to the process (typically via setitimer()
),
which then performs the same operation of examining the
program counter and incrementing a slot in the memory array.
Since this method requires a signal to be delivered to
user space every time a sample is taken, it uses considerably
more overhead than kernel-based profiling. Also, due to the
added delay required to deliver the signal, this method is
less accurate as well.
A special startup routine allocates memory for the histogram and
either calls profil()
or sets up
a clock signal handler.
This routine (monstartup
) can be invoked in several ways.
On Linux systems, a special profiling startup file gcrt0.o
,
which invokes monstartup
before main
,
is used instead of the default crt0.o
.
Use of this special startup file is one of the effects
of using `gcc ... -pg' to link.
On SPARC systems, no special startup files are used.
Rather, the mcount
routine, when it is invoked for
the first time (typically when main
is called),
calls monstartup
.
If the compiler's `-a' option was used, basic-block counting
is also enabled. Each object file is then compiled with a static array
of counts, initially zero.
In the executable code, every time a new basic-block begins
(i.e. when an if
statement appears), an extra instruction
is inserted to increment the corresponding count in the array.
At compile time, a paired array was constructed that recorded
the starting address of each basic-block. Taken together,
the two arrays record the starting address of every basic-block,
along with the number of times it was executed.
The profiling library also includes a function (mcleanup
) which is
typically registered using atexit()
to be called as the
program exits, and is responsible for writing the file `gmon.out'.
Profiling is turned off, various headers are output, and the histogram
is written, followed by the call-graph arcs and the basic-block counts.
The output from gprof
gives no indication of parts of your program that
are limited by I/O or swapping bandwidth. This is because samples of the
program counter are taken at fixed intervals of the program's run time.
Therefore, the
time measurements in gprof
output say nothing about time that your
program was not running. For example, a part of the program that creates
so much data that it cannot all fit in physical memory at once may run very
slowly due to thrashing, but gprof
will say it uses little time. On
the other hand, sampling by run time has the advantage that the amount of
load due to other users won't directly affect the output you get.
The old BSD-derived file format used for profile data does not contain a
magic cookie that allows to check whether a data file really is a
gprof file. Furthermore, it does not provide a version number, thus
rendering changes to the file format almost impossible. GNU gprof
uses a new file format that provides these features. For backward
compatibility, GNU gprof
continues to support the old BSD-derived
format, but not all features are supported with it. For example,
basic-block execution counts cannot be accommodated by the old file
format.
The new file format is defined in header file `gmon_out.h'. It
consists of a header containing the magic cookie and a version number,
as well as some spare bytes available for future extensions. All data
in a profile data file is in the native format of the host on which
the profile was collected. GNU gprof
adapts automatically to the
byte-order in use.
In the new file format, the header is followed by a sequence of
records. Currently, there are three different record types: histogram
records, call-graph arc records, and basic-block execution count
records. Each file can contain any number of each record type. When
reading a file, GNU gprof
will ensure records of the same type are
compatible with each other and compute the union of all records. For
example, for basic-block execution counts, the union is simply the sum
of all execution counts for each basic-block.
Histogram records consist of a header that is followed by an array of bins. The header contains the text-segment range that the histogram spans, the size of the histogram in bytes (unlike in the old BSD format, this does not include the size of the header), the rate of the profiling clock, and the physical dimension that the bin counts represent after being scaled by the profiling clock rate. The physical dimension is specified in two parts: a long name of up to 15 characters and a single character abbreviation. For example, a histogram representing real-time would specify the long name as "seconds" and the abbreviation as "s". This feature is useful for architectures that support performance monitor hardware (which, fortunately, is becoming increasingly common). For example, under DEC OSF/1, the "uprofile" command can be used to produce a histogram of, say, instruction cache misses. In this case, the dimension in the histogram header could be set to "i-cache misses" and the abbreviation could be set to "1" (because it is simply a count, not a physical dimension). Also, the profiling rate would have to be set to 1 in this case.
Histogram bins are 16-bit numbers and each bin represent an equal amount of text-space. For example, if the text-segment is one thousand bytes long and if there are ten bins in the histogram, each bin represents one hundred bytes.
Call-graph records have a format that is identical to the one used in the BSD-derived file format. It consists of an arc in the call graph and a count indicating the number of times the arc was traversed during program execution. Arcs are specified by a pair of addresses: the first must be within caller's function and the second must be within the callee's function. When performing profiling at the function level, these addresses can point anywhere within the respective function. However, when profiling at the line-level, it is better if the addresses are as close to the call-site/entry-point as possible. This will ensure that the line-level call-graph is able to identify exactly which line of source code performed calls to a function.
Basic-block execution count records consist of a header followed by a sequence of address/count pairs. The header simply specifies the length of the sequence. In an address/count pair, the address identifies a basic-block and the count specifies the number of times that basic-block was executed. Any address within the basic-address can be used.
gprof
's Internal Operation
Like most programs, gprof
begins by processing its options.
During this stage, it may building its symspec list
(sym_ids.c:sym_id_add
), if
options are specified which use symspecs.
gprof
maintains a single linked list of symspecs,
which will eventually get turned into 12 symbol tables,
organized into six include/exclude pairs - one
pair each for the flat profile (INCL_FLAT/EXCL_FLAT),
the call graph arcs (INCL_ARCS/EXCL_ARCS),
printing in the call graph (INCL_GRAPH/EXCL_GRAPH),
timing propagation in the call graph (INCL_TIME/EXCL_TIME),
the annotated source listing (INCL_ANNO/EXCL_ANNO),
and the execution count listing (INCL_EXEC/EXCL_EXEC).
After option processing, gprof
finishes
building the symspec list by adding all the symspecs in
default_excluded_list
to the exclude lists
EXCL_TIME and EXCL_GRAPH, and if line-by-line profiling is specified,
EXCL_FLAT as well.
These default excludes are not added to EXCL_ANNO, EXCL_ARCS, and EXCL_EXEC.
Next, the BFD library is called to open the object file,
verify that it is an object file,
and read its symbol table (core.c:core_init
),
using bfd_canonicalize_symtab
after mallocing
an appropiate sized array of asymbols. At this point,
function mappings are read (if the `--file-ordering' option
has been specified), and the core text space is read into
memory (if the `-c' option was given).
gprof
's own symbol table, an array of Sym structures,
is now built.
This is done in one of two ways, by one of two routines, depending
on whether line-by-line profiling (`-l' option) has been
enabled.
For normal profiling, the BFD canonical symbol table is scanned.
For line-by-line profiling, every
text space address is examined, and a new symbol table entry
gets created every time the line number changes.
In either case, two passes are made through the symbol
table - one to count the size of the symbol table required,
and the other to actually read the symbols. In between the
two passes, a single array of type Sym
is created of
the appropiate length.
Finally, symtab.c:symtab_finalize
is called to sort the symbol table and remove duplicate entries
(entries with the same memory address).
The symbol table must be a contiguous array for two reasons.
First, the qsort
library function (which sorts an array)
will be used to sort the symbol table.
Also, the symbol lookup routine (symtab.c:sym_lookup
),
which finds symbols
based on memory address, uses a binary search algorithm
which requires the symbol table to be a sorted array.
Function symbols are indicated with an is_func
flag.
Line number symbols have no special flags set.
Additionally, a symbol can have an is_static
flag
to indicate that it is a local symbol.
With the symbol table read, the symspecs can now be translated
into Syms (sym_ids.c:sym_id_parse
). Remember that a single
symspec can match multiple symbols.
An array of symbol tables
(syms
) is created, each entry of which is a symbol table
of Syms to be included or excluded from a particular listing.
The master symbol table and the symspecs are examined by nested
loops, and every symbol that matches a symspec is inserted
into the appropriate syms table. This is done twice, once to
count the size of each required symbol table, and again to build
the tables, which have been malloced between passes.
From now on, to determine whether a symbol is on an include
or exclude symspec list, gprof
simply uses its
standard symbol lookup routine on the appropriate table
in the syms
array.
Now the profile data file(s) themselves are read
(gmon_io.c:gmon_out_read
),
first by checking for a new-style `gmon.out' header,
then assuming this is an old-style BSD `gmon.out'
if the magic number test failed.
New-style histogram records are read by hist.c:hist_read_rec
.
For the first histogram record, allocate a memory array to hold
all the bins, and read them in.
When multiple profile data files (or files with multiple histogram
records) are read, the starting address, ending address, number
of bins and sampling rate must match between the various histograms,
or a fatal error will result.
If everything matches, just sum the additional histograms into
the existing in-memory array.
As each call graph record is read (call_graph.c:cg_read_rec
),
the parent and child addresses
are matched to symbol table entries, and a call graph arc is
created by cg_arcs.c:arc_add
, unless the arc fails a symspec
check against INCL_ARCS/EXCL_ARCS. As each arc is added,
a linked list is maintained of the parent's child arcs, and of the child's
parent arcs.
Both the child's call count and the arc's call count are
incremented by the record's call count.
Basic-block records are read (basic_blocks.c:bb_read_rec
),
but only if line-by-line profiling has been selected.
Each basic-block address is matched to a corresponding line
symbol in the symbol table, and an entry made in the symbol's
bb_addr and bb_calls arrays. Again, if multiple basic-block
records are present for the same address, the call counts
are cumulative.
A gmon.sum file is dumped, if requested (gmon_io.c:gmon_out_write
).
If histograms were present in the data files, assign them to symbols
(hist.c:hist_assign_samples
) by iterating over all the sample
bins and assigning them to symbols. Since the symbol table
is sorted in order of ascending memory addresses, we can
simple follow along in the symbol table as we make our pass
over the sample bins.
This step includes a symspec check against INCL_FLAT/EXCL_FLAT.
Depending on the histogram
scale factor, a sample bin may span multiple symbols,
in which case a fraction of the sample count is allocated
to each symbol, proportional to the degree of overlap.
This effect is rare for normal profiling, but overlaps
are more common during line-by-line profiling, and can
cause each of two adjacent lines to be credited with half
a hit, for example.
If call graph data is present, cg_arcs.c:cg_assemble
is called.
First, if `-c' was specified, a machine-dependant
routine (find_call
) scans through each symbol's machine code,
looking for subroutine call instructions, and adding them
to the call graph with a zero call count.
A topological sort is performed by depth-first numbering
all the symbols (cg_dfn.c:cg_dfn
), so that
children are always numbered less than their parents,
then making a array of pointers into the symbol table and sorting it into
numerical order, which is reverse topological
order (children appear before parents).
Cycles are also detected at this point, all members
of which are assigned the same topological number.
Two passes are now made through this sorted array of symbol pointers.
The first pass, from end to beginning (parents to children),
computes the fraction of child time to propogate to each parent
and a print flag.
The print flag reflects symspec handling of INCL_GRAPH/EXCL_GRAPH,
with a parent's include or exclude (print or no print) property
being propagated to its children, unless they themselves explicitly appear
in INCL_GRAPH or EXCL_GRAPH.
A second pass, from beginning to end (children to parents) actually
propogates the timings along the call graph, subject
to a check against INCL_TIME/EXCL_TIME.
With the print flag, fractions, and timings now stored in the symbol
structures, the topological sort array is now discarded, and a
new array of pointers is assembled, this time sorted by propagated time.
Finally, print the various outputs the user requested, which is now fairly
straightforward. The call graph (cg_print.c:cg_print
) and
flat profile (hist.c:hist_print
) are regurgitations of values
already computed. The annotated source listing
(basic_blocks.c:print_annotated_source
) uses basic-block
information, if present, to label each line of code with call counts,
otherwise only the function call counts are presented.
The function ordering code is marginally well documented
in the source code itself (cg_print.c
). Basically,
the functions with the most use and the most parents are
placed first, followed by other functions with the most use,
followed by lower use functions, followed by unused functions
at the end.
gprof
If gprof
was compiled with debugging enabled,
the `-d' option triggers debugging output
(to stdout) which can be helpful in understanding its operation.
The debugging number specified is interpreted as a sum of the following
options:
This document was generated on 7 November 1998 using the texi2html translator version 1.52.