To use this tool, you must specify
--tool=callgrind on the
Valgrind command line.
Callgrind is a profiling tool that records the call history among
functions in a program's run as a call-graph.
By default, the collected data consists of
the number of instructions executed, their relationship
to source lines, the caller/callee relationship between functions,
and the numbers of such calls.
Optionally, cache simulation and/or branch prediction (similar to Cachegrind)
can produce further information about the runtime behavior of an application.
The profile data is written out to a file at program
termination. For presentation of the data, and interactive control
of the profiling, two command line tools are provided:
This command reads in the profile data, and prints a
sorted lists of functions, optionally with source annotation.
For graphical visualization of the data, try
KCachegrind, which is a KDE/Qt based
GUI that makes it easy to navigate the large amount of data that
This command enables you to interactively observe and control
the status of a program currently running under Callgrind's control,
without stopping the program. You can get statistics information as
well as the current stack trace, and you can request zeroing of counters
or dumping of profile data.
Cachegrind collects flat profile data: event counts (data reads,
cache misses, etc.) are attributed directly to the function they
occurred in. This cost attribution mechanism is
called self or exclusive
Callgrind extends this functionality by propagating costs
across function call boundaries. If function foo calls
bar, the costs from bar are added into
foo's costs. When applied to the program as a whole,
this builds up a picture of so called inclusive
costs, that is, where the cost of each function includes the costs of
all functions it called, directly or indirectly.
As an example, the inclusive cost of
main should be almost 100 percent
of the total program cost. Because of costs arising before
main is run, such as
initialization of the run time linker and construction of global C++
objects, the inclusive cost of main
is not exactly 100 percent of the total program cost.
Together with the call graph, this allows you to find the
specific call chains starting from
main in which the majority of the
program's costs occur. Caller/callee cost attribution is also useful
for profiling functions called from multiple call sites, and where
optimization opportunities depend on changing code in the callers, in
particular by reducing the call count.
Callgrind's ability to detect function calls and returns depends
on the instruction set of the platform it is run on. It works best on
x86 and amd64, and unfortunately currently does not work so well on
PowerPC, ARM, Thumb or MIPS code. This is because there are no explicit
call or return instructions in these instruction sets, so Callgrind
has to rely on heuristics to detect calls and returns.
6.1.2. Basic Usage
As with Cachegrind, you probably want to compile with debugging info
(the -g option) and with optimization turned on.
While the simulation is running, you can observe execution with:
This will print out the current backtrace. To annotate the backtrace with
event counts, run
callgrind_control -e -b
After program termination, a profile data file named
is generated, where pid is the process ID
of the program being profiled.
The data file contains information about the calls made in the
program among the functions executed, together with
Instruction Read (Ir) event counts.
To generate a function-by-function summary from the profile
data file, use
callgrind_annotate [options] callgrind.out.<pid>
This summary is similar to the output you get from a Cachegrind
run with cg_annotate: the list
of functions is ordered by exclusive cost of functions, which also
are the ones that are shown.
Important for the additional features of Callgrind are
the following two options:
--inclusive=yes: Instead of using
exclusive cost of functions as sorting order, use and show
--tree=both: Interleave into the
top level list of functions, information on the callers and the callees
of each function. In these lines, which represents executed
calls, the cost gives the number of events spent in the call.
Indented, above each function, there is the list of callers,
and below, the list of callees. The sum of events in calls to
a given function (caller lines), as well as the sum of events in
calls from the function (callee lines) together with the self
cost, gives the total inclusive cost of the function.
Use --auto=yes to get annotated source code
for all relevant functions for which the source can be found. In
addition to source annotation as produced by
cg_annotate, you will see the
annotated call sites with call counts. For all other options,
consult the (Cachegrind) documentation for
For better call graph browsing experience, it is highly recommended
to use KCachegrind.
If your code
has a significant fraction of its cost in cycles (sets
of functions calling each other in a recursive manner), you have to
use KCachegrind, as callgrind_annotate
currently does not do any cycle detection, which is important to get correct
results in this case.
If you are additionally interested in measuring the
cache behavior of your program, use Callgrind with the option
branch prediction simulation, use --branch-sim=yes.
Expect a further slow down approximately by a factor of 2.
If the program section you want to profile is somewhere in the
middle of the run, it is beneficial to
fast forward to this section without any
profiling, and then enable profiling. This is achieved by using
the command line option
and running, in a shell:
callgrind_control -i on just before the
interesting code section is executed. To exactly specify
the code position where profiling should start, use the client request
If you want to be able to see assembly code level annotation, specify
--dump-instr=yes. This will produce
profile data at instruction granularity. Note that the resulting profile
can only be viewed with KCachegrind. For assembly annotation, it also is
interesting to see more details of the control flow inside of functions,
i.e. (conditional) jumps. This will be collected by further specifying
6.2. Advanced Usage
6.2.1. Multiple profiling dumps from one program run
Sometimes you are not interested in characteristics of a full
program run, but only of a small part of it, for example execution of one
algorithm. If there are multiple algorithms, or one algorithm
running with different input data, it may even be useful to get different
profile information for different parts of a single program run.
Profile data files have names of the form
where pid is the PID of the running
program, part is a number incremented on each
dump (".part" is skipped for the dump at program termination), and
threadID is a thread identification
("-threadID" is only used if you request dumps of individual
threads with --separate-threads=yes).
There are different ways to generate multiple profile dumps
while a program is running under Callgrind's supervision. Nevertheless,
all methods trigger the same action, which is "dump all profile
information since the last dump or program start, and zero cost
counters afterwards". To allow for zeroing cost counters without
dumping, there is a second action "zero all cost counters now".
The different methods are:
Dump on program termination.
This method is the standard way and doesn't need any special
action on your part.
Spontaneous, interactive dumping. Use
callgrind_control -d [hint [PID/Name]]
request the dumping of profile information of the supervised
application with PID or Name. hint is an
arbitrary string you can optionally specify to later be able to
distinguish profile dumps. The control program will not terminate
before the dump is completely written. Note that the application
must be actively running for detection of the dump command. So,
for a GUI application, resize the window, or for a server, send a
If you are using KCachegrind
for browsing of profile information, you can use the toolbar
button Force dump. This will request a dump
and trigger a reload after the dump is written.
Periodic dumping after execution of a specified
number of basic blocks. For this, use the command line
You can specify these options multiple times for different
functions. Function specifications support wildcards: e.g. use
generate dumps before entering any function starting with
If you are running a multi-threaded application and specify the
command line option --separate-threads=yes,
every thread will be profiled on its own and will create its own
profile dump. Thus, the last two methods will only generate one dump
of the currently running thread. With the other methods, you will get
multiple dumps (one for each thread) on a dump request.
6.2.2. Limiting the range of collected events
By default, whenever events are happening (such as an
instruction execution or cache hit/miss), Callgrind is aggregating
them into event counters. However, you may be interested only in
what is happening within a given function or starting from a given
program phase. To this end, you can disable event aggregation for
uninteresting program parts. While attribution of events to
functions as well as producing seperate output per program phase
can be done by other means (see previous section), there are two
benefits by disabling aggregation. First, this is very
fine-granular (e.g. just for a loop within a function). Second,
disabling event aggregation for complete program phases allows to
switch off time-consuming cache simulation and allows Callgrind to
progress at much higher speed with an slowdown of around factor 2
(identical to valgrind
There are two aspects which influence whether Callgrind is
aggregating events at some point in time of program execution.
First, there is the collection state. If this
is off, no aggregation will be done. By changing the collection
state, you can control event aggregation at a very fine
granularity. However, there is not much difference in regard to
execution speed of Callgrind. By default, collection is switched
on, but can be disabled by different means (see below). Second,
there is the instrumentation mode in which
Callgrind is running. This mode either can be on or off. If
instrumentation is off, no observation of actions in the program
will be done and thus, no actions will be forwarded to the
simulator which could trigger events. In the end, no events will
be aggregated. The huge benefit is the much higher speed with
instrumentation switched off. However, this only should be used
with care and in a coarse fashion: every mode change resets the
simulator state (ie. whether a memory block is cached or not) and
flushes Valgrinds internal cache of instrumented code blocks,
resulting in latency penalty at switching time. Also, cache
simulator results directly after switching on instrumentation will
be skewed due to identified cache misses which would not happen in
reality (if you care about this warm-up effect, you should make
sure to temporarly have collection state switched off directly
after turning instrumentation mode on). However, switching
instrumentation state is very useful to skip larger program phases
such as an initialization phase. By default, instrumentation is
switched on, but as with the collection state, can be changed by
Callgrind can start with instrumentation mode switched off by
Afterwards, instrumentation can be controlled in two ways: first,
Similarly, the collection state at program start can be
by --instr-atstart=no. During
execution, it can be controlled programatically with the
Further, you can limit event collection to a specific function by
This will toggle the collection state on entering and leaving the
specified function. When this option is in effect, the default
collection state at program start is "off". Only events happening
while running inside of the given function will be
collected. Recursive calls of the given function do not trigger
any action. This option can be given multiple times to specify
different functions of interest.
6.2.3. Counting global bus events
For access to shared data among threads in a multithreaded
code, synchronization is required to avoid raced conditions.
Synchronization primitives are usually implemented via atomic instructions.
However, excessive use of such instructions can lead to performance
To enable analysis of this problem, Callgrind optionally can count
the number of atomic instructions executed. More precisely, for x86/x86_64,
these are instructions using a lock prefix. For architectures supporting
LL/SC, these are the number of SC instructions executed. For both, the term
"global bus events" is used.
The short name of the event type used for global bus events is "Ge".
To count global bus events, use --collect-bus=yes.
6.2.4. Avoiding cycles
Informally speaking, a cycle is a group of functions which
call each other in a recursive way.
Formally speaking, a cycle is a nonempty set S of functions,
such that for every pair of functions F and G in S, it is possible
to call from F to G (possibly via intermediate functions) and also
from G to F. Furthermore, S must be maximal -- that is, be the
largest set of functions satisfying this property. For example, if
a third function H is called from inside S and calls back into S,
then H is also part of the cycle and should be included in S.
Recursion is quite usual in programs, and therefore, cycles
sometimes appear in the call graph output of Callgrind. However,
the title of this chapter should raise two questions: What is bad
about cycles which makes you want to avoid them? And: How can
cycles be avoided without changing program code?
Cycles are not bad in itself, but tend to make performance
analysis of your code harder. This is because inclusive costs
for calls inside of a cycle are meaningless. The definition of
inclusive cost, i.e. self cost of a function plus inclusive cost
of its callees, needs a topological order among functions. For
cycles, this does not hold true: callees of a function in a cycle include
the function itself. Therefore, KCachegrind does cycle detection
and skips visualization of any inclusive cost for calls inside
of cycles. Further, all functions in a cycle are collapsed into artifical
functions called like Cycle 1.
Now, when a program exposes really big cycles (as is
true for some GUI code, or in general code using event or callback based
programming style), you lose the nice property to let you pinpoint
the bottlenecks by following call chains from
main, guided via
inclusive cost. In addition, KCachegrind loses its ability to show
interesting parts of the call graph, as it uses inclusive costs to
cut off uninteresting areas.
Despite the meaningless of inclusive costs in cycles, the big
drawback for visualization motivates the possibility to temporarily
switch off cycle detection in KCachegrind, which can lead to
misguiding visualization. However, often cycles appear because of
unlucky superposition of independent call chains in a way that
the profile result will see a cycle. Neglecting uninteresting
calls with very small measured inclusive cost would break these
cycles. In such cases, incorrect handling of cycles by not detecting
them still gives meaningful profiling visualization.
It has to be noted that currently, callgrind_annotate
does not do any cycle detection at all. For program executions with function
recursion, it e.g. can print nonsense inclusive costs way above 100%.
After describing why cycles are bad for profiling, it is worth
talking about cycle avoidance. The key insight here is that symbols in
the profile data do not have to exactly match the symbols found in the
program. Instead, the symbol name could encode additional information
from the current execution context such as recursion level of the
current function, or even some part of the call chain leading to the
function. While encoding of additional information into symbols is
quite capable of avoiding cycles, it has to be used carefully to not cause
symbol explosion. The latter imposes large memory requirement for Callgrind
with possible out-of-memory conditions, and big profile data files.
A further possibility to avoid cycles in Callgrind's profile data
output is to simply leave out given functions in the call graph. Of course, this
also skips any call information from and to an ignored function, and thus can
break a cycle. Candidates for this typically are dispatcher functions in event
driven code. The option to ignore calls to a function is
--fn-skip=function. Aside from
possibly breaking cycles, this is used in Callgrind to skip
trampoline functions in the PLT sections
for calls to functions in shared libraries. You can see the difference
if you profile with --skip-plt=no.
If a call is ignored, its cost events will be propagated to the
If you have a recursive function, you can distinguish the first
10 recursion levels by specifying
Or for all functions with
--separate-recs=10, but this will
give you much bigger profile data files. In the profile data, you will see
the recursion levels of "func" as the different functions with names
"func", "func'2", "func'3" and so on.
If you have call chains "A > B > C" and "A > C > B"
in your program, you usually get a "false" cycle "B <> C". Use
and functions "B" and "C" will be treated as different functions
depending on the direct caller. Using the apostrophe for appending
this "context" to the function name, you get "A > B'A > C'B"
and "A > C'A > B'C", and there will be no cycle. Use
--separate-callers=2 to get a 2-caller
dependency for all functions. Note that doing this will increase
the size of profile data files.
6.2.5. Forking Programs
If your program forks, the child will inherit all the profiling
data that has been gathered for the parent. To start with empty profile
counter values in the child, the client request
can be inserted into code to be executed by the child, directly after
However, you will have to make sure that the output file format string
(controlled by --callgrind-out-file) does contain
%p (which is true by default). Otherwise, the
outputs from the parent and child will overwrite each other or will be
intermingled, which almost certainly is not what you want.
You will be able to control the new child independently from
the parent via callgrind_control.
6.3. Callgrind Command-line Options
In the following, options are grouped into classes.
Some options allow the specification of a function/symbol name, such as
--fn-skip=function. All these options
can be specified multiple times for different functions.
In addition, the function specifications actually are patterns by supporting
the use of wildcards '*' (zero or more arbitrary characters) and '?'
(exactly one arbitrary character), similar to file name globbing in the
shell. This feature is important especially for C++, as without wildcard
usage, the function would have to be specified in full extent, including
6.3.1. Dump creation options
These options influence the name and format of the profile data files.
Write the profile data to
file rather than to the default
%p and %q format specifiers
can be used to embed the process ID and/or the contents of an
environment variable in the name, as is the case for the core
When multiple dumps are made, the file name
is modified further; see below.
--dump-line=<no|yes> [default: yes]
This specifies that event counting should be performed at
source line granularity. This allows source annotation for sources
which are compiled with debug information
--dump-instr=<no|yes> [default: no]
This specifies that event counting should be performed at
This allows for assembly code
annotation. Currently the results can only be
displayed by KCachegrind.
--compress-strings=<no|yes> [default: yes]
This option influences the output format of the profile data.
It specifies whether strings (file and function names) should be
identified by numbers. This shrinks the file,
but makes it more difficult
for humans to read (which is not recommended in any case).
--compress-pos=<no|yes> [default: yes]
This option influences the output format of the profile data.
It specifies whether numerical positions are always specified as absolute
values or are allowed to be relative to previous numbers.
This shrinks the file size.
--combine-dumps=<no|yes> [default: no]
When enabled, when multiple profile data parts are to be
generated these parts are appended to the same output file.
6.3.2. Activity options
These options specify when actions relating to event counts are to
be executed. For interactive control use callgrind_control.
--dump-every-bb=<count> [default: 0, never]
Dump profile data every count basic blocks.
Whether a dump is needed is only checked when Valgrind's internal
scheduler is run. Therefore, the minimum setting useful is about 100000.
The count is a 64-bit value to make long dump periods possible.
Specify if you want Callgrind to start simulation and
profiling from the beginning of the program.
When set to no,
Callgrind will not be able
to collect any information, including calls, but it will have at
most a slowdown of around 4, which is the minimum Valgrind
overhead. Instrumentation can be interactively enabled via
callgrind_control -i on.
Note that the resulting call graph will most probably not
contain main, but will contain all the
functions executed after instrumentation was enabled.
Instrumentation can also programatically enabled/disabled. See the
Callgrind include file
callgrind.h for the macro
you have to use in your source code.
simulation, results will be less accurate when switching on
instrumentation later in the program run, as the simulator starts
with an empty cache at that moment. Switch on event collection
later to cope with this error.
--collect-atstart=<yes|no> [default: yes]
Specify whether event collection is enabled at beginning
of the profile run.
To only look at parts of your program, you have two
Zero event counters before entering the program part you
want to profile, and dump the event counters to a file after
leaving that program part.
Switch on/off collection state as needed to only see
event counters happening while inside of the program part you
want to profile.
The second option can be used if the program part you want to
profile is called many times. Option 1, i.e. creating a lot of
dumps is not practical here.
Collection state can be
toggled at entry and exit of a given function with the
option --toggle-collect. If you
use this option, collection
state should be disabled at the beginning. Note that the
specification of --toggle-collect
Collection state can be toggled also by inserting the client request
at the needed code positions.
Toggle collection on entry/exit of function.
--collect-jumps=<no|yes> [default: no]
This specifies whether information for (conditional) jumps
should be collected. As above, callgrind_annotate currently is not
able to show you the data. You have to use KCachegrind to get jump
arrows in the annotated code.
--collect-systime=<no|yes> [default: no]
This specifies whether information for system call times
should be collected.
--collect-bus=<no|yes> [default: no]
This specifies whether the number of global bus events executed
should be collected. The event type "Ge" is used for these events.
6.3.4. Cost entity separation options
These options specify how event counts should be attributed to execution
For example, they specify whether the recursion level or the
call chain leading to a function should be taken into account,
and whether the thread ID should be considered.
Also see Avoiding cycles.
--separate-threads=<no|yes> [default: no]
This option specifies whether profile data should be generated
separately for every thread. If yes, the file names get "-threadID"
--separate-callers=<callers> [default: 0]
Separate contexts by at most <callers> functions in the
call chain. See Avoiding cycles.
Ignore calls to/from a given function. E.g. if you have a
call chain A > B > C, and you specify function B to be
ignored, you will only see A > C.
This is very convenient to skip functions handling callback
behaviour. For example, with the signal/slot mechanism in the
Qt graphics library, you only want
to see the function emitting a signal to call the slots connected
to that signal. First, determine the real call chain to see the
functions needed to be skipped, then use this option.
6.3.5. Simulation options
--cache-sim=<yes|no> [default: no]
Specify if you want to do full cache simulation. By default,
only instruction read accesses will be counted ("Ir").
With cache simulation, further event counters are enabled:
Cache misses on instruction reads ("I1mr"/"ILmr"),
data read accesses ("Dr") and related cache misses ("D1mr"/"DLmr"),
data write accesses ("Dw") and related cache misses ("D1mw"/"DLmw").
For more information, see Cachegrind: a cache and branch-prediction profiler.
--branch-sim=<yes|no> [default: no]
Specify if you want to do branch prediction simulation.
Further event counters are enabled: Number of executed conditional
branches and related predictor misses ("Bc"/"Bcm"), executed indirect
jumps and related misses of the jump address predictor ("Bi"/"Bim").
6.3.6. Cache simulation options
--simulate-wb=<yes|no> [default: no]
Specify whether write-back behavior should be simulated, allowing
to distinguish LL caches misses with and without write backs.
The cache model of Cachegrind/Callgrind does not specify write-through
vs. write-back behavior, and this also is not relevant for the number
of generated miss counts. However, with explicit write-back simulation
it can be decided whether a miss triggers not only the loading of a new
cache line, but also if a write back of a dirty cache line had to take
place before. The new dirty miss events are ILdmr, DLdmr, and DLdmw,
for misses because of instruction read, data read, and data write,
respectively. As they produce two memory transactions, they should
account for a doubled time estimation in relation to a normal miss.
--simulate-hwpref=<yes|no> [default: no]
Specify whether simulation of a hardware prefetcher should be
added which is able to detect stream access in the second level cache
by comparing accesses to separate to each page.
As the simulation can not decide about any timing issues of prefetching,
it is assumed that any hardware prefetch triggered succeeds before a
real access is done. Thus, this gives a best-case scenario by covering
all possible stream accesses.
--cacheuse=<yes|no> [default: no]
Specify whether cache line use should be collected. For every
cache line, from loading to it being evicted, the number of accesses
as well as the number of actually used bytes is determined. This
behavior is related to the code which triggered loading of the cache
line. In contrast to miss counters, which shows the position where
the symptoms of bad cache behavior (i.e. latencies) happens, the
use counters try to pinpoint at the reason (i.e. the code with the
bad access behavior). The new counters are defined in a way such
that worse behavior results in higher cost.
AcCost1 and AcCost2 are counters showing bad temporal locality
for L1 and LL caches, respectively. This is done by summing up
reciprocal values of the numbers of accesses of each cache line,
multiplied by 1000 (as only integer costs are allowed). E.g. for
a given source line with 5 read accesses, a value of 5000 AcCost
means that for every access, a new cache line was loaded and directly
evicted afterwards without further accesses. Similarly, SpLoss1/2
shows bad spatial locality for L1 and LL caches, respectively. It
gives the spatial loss count of bytes which
were loaded into cache but never accessed. It pinpoints at code
accessing data in a way such that cache space is wasted. This hints
at bad layout of data structures in memory. Assuming a cache line
size of 64 bytes and 100 L1 misses for a given source line, the
loading of 6400 bytes into L1 was triggered. If SpLoss1 shows a
value of 3200 for this line, this means that half of the loaded data was
never used, or using a better data layout, only half of the cache
space would have been needed.
Please note that for cache line use counters, it currently is
not possible to provide meaningful inclusive costs. Therefore,
inclusive cost of these counters should be ignored.
Specify the size, associativity and line size of the level 1
Specify the size, associativity and line size of the level 1
Specify the size, associativity and line size of the last-level
Start full Callgrind instrumentation if not already enabled.
When cache simulation is done, this will flush the simulated cache
and lead to an artifical cache warmup phase afterwards with
cache misses which would not have happened in reality. See also
Stop full Callgrind instrumentation if not already disabled.
This flushes Valgrinds translation cache, and does no additional
instrumentation afterwards: it effectivly will run at the same
speed as Nulgrind, i.e. at minimal slowdown. Use this to
speed up the Callgrind run for uninteresting code parts. Use
enable instrumentation again. See also option
6.6. callgrind_annotate Command-line Options
Show summary of options.
Show version of callgrind_annotate.
--show=A,B,C [default: all]
Only show figures for events A,B,C.
Sort columns by events A,B,C [event column order].
--threshold=<0--100> [default: 99%]
Percentage of counts (of primary sort event) we are
--auto=<yes|no> [default: no]
Annotate all source files containing functions that helped
reach the event count threshold.
--context=N [default: 8]
Print N lines of context before and after annotated
--inclusive=<yes|no> [default: no]
Add subroutine costs to functions calls.
--tree=<none|caller|calling|both> [default: none]
Print for each function their callers, the called functions
Add dir to the list of directories to search
for source files.
6.7. callgrind_control Command-line Options
By default, callgrind_control acts on all programs run by the
current user under Callgrind. It is possible to limit the actions to
specified Callgrind runs by providing a list of pids or program names as
argument. The default action is to give some brief information about the
applications being run under Callgrind.
Show a short description, usage, and summary of options.
Show version of callgrind_control.
Show also the working directory, in addition to the brief
information given by default.
Show statistics information about active Callgrind runs.
Show stack/back traces of each thread in active Callgrind runs. For
each active function in the stack trace, also the number of invocations
since program start (or last dump) is shown. This option can be
combined with -e to show inclusive cost of active functions.
-e [A,B,...] (default: all)
Show the current per-thread, exclusive cost values of event
counters. If no explicit event names are given, figures for all event
types which are collected in the given Callgrind run are
shown. Otherwise, only figures for event types A, B, ... are shown. If
this option is combined with -b, inclusive cost for the functions of
each active stack frame is provided, too.
--dump[=<desc>] (default: no description)
Request the dumping of profile information. Optionally, a
description can be specified which is written into the dump as part of
the information giving the reason which triggered the dump action. This
can be used to distinguish multiple dumps.
Zero all event counters.
Force a Callgrind run to be terminated.
Switch instrumentation mode on or off. If a Callgrind run has
instrumentation disabled, no simulation is done and no events are
counted. This is useful to skip uninteresting program parts, as there
is much less slowdown (same as with the Valgrind tool "none"). See also
the Callgrind option --instr-atstart.
Specify the vgdb prefix to use by callgrind_control.
callgrind_control internally uses vgdb to find and control the active
Callgrind runs. If the --vgdb-prefix option was used
for launching valgrind, then the same option must be given to