zenithsiz/ftmemsim-valgrind
1
0
Fork 0
mirror of https://github.com/Zenithsiz/ftmemsim-valgrind.git synced 2026-02-03 18:13:01 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
ftmemsim-valgrind/cachegrind/docs/cg-manual.xml
Nicholas Nethercote b12703598a Fix two xmllint errors.
2023-04-21 23:00:39 +10:00

1512 lines
51 KiB
XML
Raw Blame History

<?xml version="1.0"?> <!-- -*- sgml -*- -->
<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.5//EN"
"http://www.oasis-open.org/docbook/xml/4.5/docbookx.dtd"
[ <!ENTITY % vg-entities SYSTEM "../../docs/xml/vg-entities.xml"> %vg-entities; ]>
<!-- Referenced from both the manual and manpage -->
<chapter id="&vg-cg-manual-id;" xreflabel="&vg-cg-manual-label;">
<title>Cachegrind: a high-precision tracing profiler</title>
<para>
To use this tool, specify <option>--tool=cachegrind</option> on the Valgrind
command line.
</para>
<sect1 id="cg-manual.overview" xreflabel="Overview">
<title>Overview</title>
<para>
Cachegrind is a high-precision tracing profiler. It runs slowly, but collects
precise and reproducible profiling data. It can merge and diff data from
different runs. To expand on these characteristics:
</para>
<itemizedlist>
<listitem>
<para>
<emphasis>Precise.</emphasis> Cachegrind measures the exact number of
instructions executed by your program, not an approximation. Furthermore,
it presents the gathered data at the file, function, and line level. This
is different to many other profilers that measure approximate execution
time, using sampling, and only at the function level.
</para>
</listitem>
<listitem>
<para>
<emphasis>Reproducible.</emphasis> In general, execution time is a better
metric than instruction counts because it's what users perceive. However,
execution time often has high variability. When running the exact same
program on the exact same input multiple times, execution time might vary
by several percent. Furthermore, small changes in a program can change its
memory layout and have even larger effects on runtime. In contrast,
instruction counts are highly reproducible; for some programs they are
perfectly reproducible. This means the effects of small changes in a
program can be measured with high precision.
</para>
</listitem>
</itemizedlist>
<para>
For these reasons, Cachegrind is an excellent complement to time-based profilers.
</para>
<para>
Cachegrind can annotate programs written in any language, so long as debug info
is present to map machine code back to the original source code. Cachegrind has
been used successfully on programs written in C, C++, Rust, and assembly.
</para>
<para>
Cachegrind can also simulate how your program interacts with a machine's cache
hierarchy and branch predictor. This simulation was the original motivation for
the tool, hence its name. However, the simulations are basic and unlikely to
reflect the behaviour of a modern machine. For this reason they are off by
default. If you really want cache and branch information, a profiler like
<computeroutput>perf</computeroutput> that accesses hardware counters is a
better choice.
</para>
</sect1>
<sect1 id="cg-manual.profile"
xreflabel="Using Cachegrind and cg_annotate">
<title>Using Cachegrind and cg_annotate</title>
<para>
First, as for normal Valgrind use, you should compile with debugging info (the
<option>-g</option> option in most compilers). But by contrast with normal
Valgrind use, you probably do want to turn optimisation on, since you should
profile your program as it will be normally run.
</para>
<para>
Second, run Cachegrind itself to gather the profiling data.
</para>
<para>
Third, run cg_annotate to get a detailed presentation of that data. cg_annotate
can combine the results of multiple Cachegrind output files. It can also
perform a diff between two Cachegrind output files.
</para>
<sect2 id="cg-manual.running-cachegrind" xreflabel="Running Cachegrind">
<title>Running Cachegrind</title>
<para>
To run Cachegrind on a program <filename>prog</filename>, run:
<screen><![CDATA[
valgrind --tool=cachegrind prog
]]></screen>
</para>
<para>
The program will execute (slowly). Upon completion, summary statistics that
look like this will be printed:
</para>
<programlisting><![CDATA[
==17942== I refs: 8,195,070
]]></programlisting>
<para>
The <computeroutput>I refs</computeroutput> number is short for "Instruction
cache references", which is equivalent to "instructions executed". If you
enable the cache and/or branch simulation, additional counts will be shown.
</para>
</sect2>
<sect2 id="cg-manual.outputfile" xreflabel="Output File">
<title>Output File</title>
<para>
Cachegrind also writes more detailed profiling data to a file. By default this
Cachegrind output file is named <filename>cachegrind.out.&lt;pid&gt;</filename>
(where <filename>&lt;pid&gt;</filename> is the program's process ID), but its
name can be changed with the <option>--cachegrind-out-file</option> option.
This file is human-readable, but is intended to be interpreted by the
accompanying program cg_annotate, described in the next section.
</para>
<para>
The default <computeroutput>.&lt;pid&gt;</computeroutput> suffix on the output
file name serves two purposes. First, it means existing Cachegrind output files
aren't immediately overwritten. Second, and more importantly, it allows correct
profiling with the <option>--trace-children=yes</option> option of programs
that spawn child processes.
</para>
</sect2>
<sect2 id="cg-manual.running-cg_annotate" xreflabel="Running cg_annotate">
<title>Running cg_annotate</title>
<para>
Before using cg_annotate, it is worth widening your window to be at least 120
characters wide if possible, because the output lines can be quite long.
</para>
<para>
Then run:
<screen>cg_annotate &lt;filename&gt;</screen>
on a Cachegrind output file.
</para>
</sect2>
<!--
To produce the sample date, I did the following. Note that the single hypens in
the valgrind command should be double hyphens, but XML doesn't allow double
hyphens in comments.
gcc -g -O concord.c -o concord
valgrind -tool=cachegrind -cachegrind-out-file=concord.cgout ./concord ../cg_main.c
(to exit, type `q` and hit enter)
python ../cg_annotate concord.cgout > concord.cgann
concord.c is a small C program I wrote at university. It's a good size for an example.
-->
<sect2 id="cg-manual.the-metadata" xreflabel="The Metadata Section">
<title>The Metadata Section</title>
<para>
The first part of the output looks like this:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Metadata
--------------------------------------------------------------------------------
Invocation: ../cg_annotate concord.cgout
Command: ./concord ../cg_main.c
Events recorded: Ir
Events shown: Ir
Event sort order: Ir
Threshold: 0.1%
Annotation: on
]]></programlisting>
<para>
It summarizes how Cachegrind and the profiled program were run.
</para>
<itemizedlist>
<listitem>
<para>
Invocation: the command line used to produce this output.
</para>
</listitem>
<listitem>
<para>
Command: the command line used to run the profiled program.
</para>
</listitem>
<listitem>
<para>
Events recorded: which events were recorded. By default, this is
<computeroutput>Ir</computeroutput>. More events will be recorded if cache
and/or branch simulation is enabled.
</para>
</listitem>
<listitem>
<para>
Events shown: the events shown, which is a subset of the events gathered.
This can be adjusted with the <option>--show</option> option.
</para>
</listitem>
<listitem>
<para>
Event sort order: the sort order used for the subsequent sections. For
example, in this case those sections are sorted from highest
<computeroutput>Ir</computeroutput> counts to lowest. If there are multiple
events, one will be the primary sort event, and then there can be a
secondary sort event, tertiary sort event, etc., though more than one is
rarely needed. This order can be adjusted with the <option>--sort</option>
option. Note that this does <emphasis>not</emphasis> specify the order in
which the columns appear. That is specified by the "events shown" line (and
can be changed with the <option>--show</option> option).
</para>
</listitem>
<listitem>
<para>
Threshold: cg_annotate by default omits files and functions with very low
counts to keep the output size reasonable. By default cg_annotate only
shows files and functions that account for at least 0.1% of the primary
sort event. The threshold can be adjusted with the
<option>--threshold</option> option.
</para>
</listitem>
<listitem>
<para>
Annotation: whether source file annotation is enabled. Controlled with the
<option>--annotate</option> option.
</para>
</listitem>
</itemizedlist>
<para>
If cache simulation is enabled, details of the cache parameters will be shown
above the "Invocation" line.
</para>
</sect2>
<sect2 id="cg-manual.the-global"
xreflabel="Global, File, and Function-level Counts">
<title>Global, File, and Function-level Counts</title>
<para>
Next comes the summary for the whole program:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Summary
--------------------------------------------------------------------------------
Ir________________
8,195,070 (100.0%) PROGRAM TOTALS
]]></programlisting>
<para>
The <computeroutput>Ir</computeroutput> column label is suffixed with
underscores to show the bounds of the columns underneath.
</para>
<para>
Then comes file:function counts. Here is the first part of that section:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- File:function summary
--------------------------------------------------------------------------------
Ir______________________ file:function
< 3,078,746 (37.6%, 37.6%) /home/njn/grind/ws1/cachegrind/concord.c:
1,630,232 (19.9%) get_word
630,918 (7.7%) hash
461,095 (5.6%) insert
130,560 (1.6%) add_existing
91,014 (1.1%) init_hash_table
88,056 (1.1%) create
46,676 (0.6%) new_word_node
< 1,746,038 (21.3%, 58.9%) ./malloc/./malloc/malloc.c:
1,285,938 (15.7%) _int_malloc
458,225 (5.6%) malloc
< 1,107,550 (13.5%, 72.4%) ./libio/./libio/getc.c:getc
< 551,071 (6.7%, 79.1%) ./string/../sysdeps/x86_64/multiarch/strcmp-avx2.S:__strcmp_avx2
< 521,228 (6.4%, 85.5%) ./ctype/../include/ctype.h:
260,616 (3.2%) __ctype_tolower_loc
260,612 (3.2%) __ctype_b_loc
< 468,163 (5.7%, 91.2%) ???:
468,151 (5.7%) ???
< 456,071 (5.6%, 96.8%) /usr/include/ctype.h:get_word
]]></programlisting>
<para>
Each entry covers one file, and one or more functions within that file. If
there is only one significant function within a file, as in the first entry,
the file and function are shown on the same line separate by a colon. If there
are multiple significant functions within a file, as in the third entry, each
function gets its own line.
</para>
<para>
This example involves a small C program, and shows a combination of code from
the program itself (including functions like <function>get_word</function> and
<function>hash</function> in the file <filename>concord.c</filename>) as well
as code from system libraries, such as functions like
<function>malloc</function> and <function>getc</function>.
</para>
<para>
Each entry is preceded with a <computeroutput>&lt;</computeroutput>, which can
be useful when navigating through the output in an editor, or grepping through
results.
</para>
<para>
The first percentage in each column indicates the proportion of the total event
count is covered by this line. The second percentage, which only shows on the
first line of each entry, shows the cumulative percentage of all the entries up
to and including this one. The entries shown here account for 96.8% of the
instructions executed by the program.
</para>
<para>
The name <computeroutput>???</computeroutput> is used if the file name and/or
function name could not be determined from debugging information. If
<filename>???</filename> filenames dominate, the program probably wasn't
compiled with <option>-g</option>. If <function>???</function> function names
dominate, the program may have had symbols stripped.
</para>
<para>
After that comes function:file counts. Here is the first part of that section:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Function:file summary
--------------------------------------------------------------------------------
Ir______________________ function:file
> 2,086,303 (25.5%, 25.5%) get_word:
1,630,232 (19.9%) /home/njn/grind/ws1/cachegrind/concord.c
456,071 (5.6%) /usr/include/ctype.h
> 1,285,938 (15.7%, 41.1%) _int_malloc:./malloc/./malloc/malloc.c
> 1,107,550 (13.5%, 54.7%) getc:./libio/./libio/getc.c
> 630,918 (7.7%, 62.4%) hash:/home/njn/grind/ws1/cachegrind/concord.c
> 551,071 (6.7%, 69.1%) __strcmp_avx2:./string/../sysdeps/x86_64/multiarch/strcmp-avx2.S
> 480,248 (5.9%, 74.9%) malloc:
458,225 (5.6%) ./malloc/./malloc/malloc.c
22,023 (0.3%) ./malloc/./malloc/arena.c
> 468,151 (5.7%, 80.7%) ???:???
> 461,095 (5.6%, 86.3%) insert:/home/njn/grind/ws1/cachegrind/concord.c
]]></programlisting>
<para>
This is similar to the previous section, but is grouped by functions first and
files second. Also, the entry markers are <computeroutput>&gt;</computeroutput>
instead of <computeroutput>&lt;</computeroutput>.
</para>
<para>
You might wonder why this section is needed, and how it differs from the
previous section. The answer is inlining. In this example there are two entries
demonstrating a function whose code is effectively spread across more than one
file: <function>get_word</function> and <function>malloc</function>. Here is an
example from profiling the Rust compiler, a much larger program that uses
inlining more:
</para>
<programlisting><![CDATA[
> 30,469,230 (1.3%, 11.1%) <rustc_middle::ty::context::CtxtInterners>::intern_ty:
10,269,220 (0.5%) /home/njn/.cargo/registry/src/github.com-1ecc6299db9ec823/hashbrown-0.12.3/src/raw/mod.rs
7,696,827 (0.3%) /home/njn/dev/rust0/compiler/rustc_middle/src/ty/context.rs
3,858,099 (0.2%) /home/njn/dev/rust0/library/core/src/cell.rs
]]></programlisting>
<para>
In this case the compiled function <function>intern_ty</function> includes code
from three different source files, due to inlining. These should be examined
together. Older versions of cg_annotate presented this entry as three separate
file:function entries, which would typically be intermixed with all the other
entries, making it hard to see that they are all really part of the same
function.
</para>
</sect2>
<sect2 id="cg-manual.line-by-line" xreflabel="Per-line Counts">
<title>Per-line Counts</title>
<para>
By default, a source file is annotated if it contains at least one function
that meets the significance threshold. This can be disabled with the
<option>--annotate</option> option.
</para>
<para>
To continue the previous example, here is part of the annotation of the file
<filename>concord.c</filename>:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Annotated source file: /home/njn/grind/ws1/cachegrind/docs/concord.c
--------------------------------------------------------------------------------
Ir____________
. /* Function builds the hash table from the given file. */
. void init_hash_table(char *file_name, Word_Node *table[])
8 (0.0%) {
. FILE *file_ptr;
. Word_Info *data;
2 (0.0%) int line = 1, i;
.
. /* Structure used when reading in words and line numbers. */
3 (0.0%) data = (Word_Info *) create(sizeof(Word_Info));
.
. /* Initialise entire table to NULL. */
2,993 (0.0%) for (i = 0; i < TABLE_SIZE; i++)
997 (0.0%) table[i] = NULL;
.
. /* Open file, check it. */
4 (0.0%) file_ptr = fopen(file_name, "r");
2 (0.0%) if (!(file_ptr)) {
. fprintf(stderr, "Couldn't open '%s'.\n", file_name);
. exit(EXIT_FAILURE);
. }
.
. /* 'Get' the words and lines one at a time from the file, and insert them
. ** into the table one at a time. */
55,363 (0.7%) while ((line = get_word(data, line, file_ptr)) != EOF)
31,632 (0.4%) insert(data->word, data->line, table);
.
2 (0.0%) free(data);
2 (0.0%) fclose(file_ptr);
6 (0.0%) }
]]></programlisting>
<para>
Each executed line is annotated with its event counts. Other lines are
annotated with a dot. This may be because they contain no executable code, or
they contain executable code but were never executed.
</para>
<para>
You can easily tell if a function is inlined from this output. If it is not
inlined, it will have event counts on the lines containing the opening and
closing braces. If it is inlined, it will not have event counts on those lines.
In the example above, <function>init_hash_table</function> does have counts,
so you can tell it is not inlined.
</para>
<para>
Note again that inlining can lead to surprising results. If a function
<function>f</function> is always inlined, in the file:function and
function:file sections counts will be attributed to the functions it is inlined
into, rather than itself. However, if you look at the line-by-line annotations
for <function>f</function> you'll see the counts that belong to
<function>f</function>. So it's worth looking for large counts/percentages in the
line-by-line annotations.
</para>
<para>
Sometimes only a small section of a source file is executed. To minimise
uninteresting output, Cachegrind only shows annotated lines and lines within a
small distance of annotated lines. Gaps are marked with line numbers, for
example:
</para>
<programlisting><![CDATA[
(counts and code for line 704)
-- line 375 ----------------------------------------
-- line 514 ----------------------------------------
(counts and code for line 878)
]]></programlisting>
<para>
The number of lines of context shown around annotated lines is controlled by
the <option>--context</option> option.
</para>
<para>
Any significant source files that could not be found are shown like this:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Annotated source file: ./malloc/./malloc/malloc.c
--------------------------------------------------------------------------------
Unannotated because one or more of these original files are unreadable:
- ./malloc/./malloc/malloc.c
]]></programlisting>
<para>
This is common for library files, because libraries are usually compiled with
debugging information but the source files are rarely present on a system.
</para>
<para>
Cachegrind relies heavily on accurate debug info. Sometimes compilers do not
map a particular compiled instruction to line number 0, where the 0 represents
"unknown" or "none". This is annoying but does happen in practice. cg_annotate
prints these in the following way:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Annotated source file: /home/njn/dev/rust0/compiler/rustc_borrowck/src/lib.rs
--------------------------------------------------------------------------------
Ir______________
1,046,746 (0.0%) <unknown (line 0)>
]]></programlisting>
<para>
Finally, when annotation is performed, the output ends with a summary of how
many counts were annotated and unannotated, and why. For example:
</para>
<programlisting><![CDATA[
--------------------------------------------------------------------------------
-- Annotation summary
--------------------------------------------------------------------------------
Ir_______________
3,534,817 (43.1%) annotated: files known & above threshold & readable, line numbers known
0 annotated: files known & above threshold & readable, line numbers unknown
0 unannotated: files known & above threshold & two or more non-identical
4,132,126 (50.4%) unannotated: files known & above threshold & unreadable
59,950 (0.7%) unannotated: files known & below threshold
468,163 (5.7%) unannotated: files unknown
]]></programlisting>
</sect2>
<sect2 id="cg-manual.forkingprograms" xreflabel="Forking Programs">
<title>Forking Programs</title>
<para>
If your program forks, the child will inherit all the profiling data that
has been gathered for the parent.
</para>
<para>
If the output file name (controlled by <option>--cachegrind-out-file</option>)
does not contain <option>%p</option>, then the outputs from the parent and
child will be intermingled in a single output file, which will almost certainly
make it unreadable by cg_annotate.
</para>
</sect2>
<sect2 id="cg-manual.annopts.warnings" xreflabel="cg_annotate Warnings">
<title>cg_annotate Warnings</title>
<para>
There are two situations in which cg_annotate prints warnings.
</para>
<itemizedlist>
<listitem>
<para>
If a source file is more recent than the Cachegrind output file. This is
because the information in the Cachegrind output file is only recorded with
line numbers, so if the line numbers change at all in the source (e.g.
lines added, deleted, swapped), any annotations will be incorrect.
</para>
</listitem>
<listitem>
<para>
If information is recorded about line numbers past the end of a file. This
can be caused by the above problem, e.g. shortening the source file while
using an old Cachegrind output file. If this happens, the figures for the
bogus lines are printed anyway (and clearly marked as bogus) in case they
are important.
</para>
</listitem>
</itemizedlist>
</sect2>
<sect2 id="cg-manual.cg_merge" xreflabel="cg_merge">
<title>Merging Cachegrind Output Files</title>
<para>
cg_annotate can merge data from multiple Cachegrind output files in a single
run. (There is also a program called cg_merge that can merge multiple
Cachegrind output files into a single Cachegrind output file, but it is now
deprecated because cg_annotate's merging does a better job.)
</para>
<para>
Use it as follows:
</para>
<programlisting><![CDATA[
cg_annotate file1 file2 file3 ...
]]></programlisting>
<para>
cg_annotate computes the sum of these files (effectively
<filename>file1</filename> + <filename>file2</filename> +
<filename>file3</filename>), and then produces output as usual that shows the
summed counts.
</para>
<para>
The most common merging scenario is if you want to aggregate costs over
multiple runs of the same program, possibly on different inputs.
</para>
</sect2>
<sect2 id="cg-manual.cg_diff" xreflabel="cg_diff">
<title>Differencing Cachegrind output files</title>
<para>
cg_annotate can diff data from two Cachegrind output files in a single run.
(There is also a program called cg_diff that can diff two Cachegrind output
files into a single Cachegrind output file, but it is now deprecated because
cg_annotate's differencing does a better job.)
</para>
<para>
Use it as follows:
</para>
<programlisting><![CDATA[
cg_annotate --diff file1 file2
]]></programlisting>
<para>
cg_annotate computes the difference between these two files (effectively
<filename>file2</filename> - <filename>file1</filename>), and then
produces output as usual that shows the count differences. Note that many of
the counts may be negative; this indicates that the counts for the relevant
file/function/line are smaller in the second version than those in the first
version.
</para>
<para>
The simplest common scenario is comparing two Cachegrind output files that came
from the same program, but on different inputs. cg_annotate will do a good job
on this without assistance.
</para>
<para>
A more complex scenario is if you want to compare Cachegrind output files from
two slightly different versions of a program that you have sitting
side-by-side, running on the same input. For example, you might have
<filename>version1/prog.c</filename> and <filename>version2/prog.c</filename>.
A straight comparison of the two would not be useful. Because functions are
always paired with filenames, a function <function>f</function> would be listed
as <filename>version1/prog.c:f</filename> for the first version but
<filename>version2/prog.c:f</filename> for the second version.
</para>
<para>
In this case, use the <option>--mod-filename</option> option. Its argument is a
search-and-replace expression that will be applied to all the filenames in both
Cachegrind output files. It can be used to remove minor differences in
filenames. For example, the option
<option>--mod-filename='s/version[0-9]/versionN/'</option> will suffice for the
above example.
</para>
<para>
Similarly, sometimes compilers auto-generate certain functions and give them
randomized names like <function>T.1234</function> where the suffixes vary from
build to build. You can use the <option>--mod-funcname</option> option to
remove small differences like these; it works in the same way as
<option>--mod-filename</option>.
</para>
<para>
When <option>--mod-filename</option> is used to compare two different versions
of the same program, cg_annotate will not annotate any file that is different
between the two versions, because the per-line counts are not reliable in such
a case. For example, imagine if <filename>version2/prog.c</filename> is the
same as <filename>version1/prog.c</filename> except with an extra blank line at
the top of the file. Every single per-line count will have changed. In
comparison, the per-file and per-function counts have not changed, and are
still very useful for determining differences between programs. You might think
that this means every interesting file will be left unannotated, but again
inlining means that files that are identical in the two versions can have
different counts on many lines.
</para>
</sect2>
<sect2 id="cg-manual.cache-branch-sim" xreflabel="cache-branch-sim">
<title>Cache and Branch Simulation</title>
<para>
Cachegrind can simulate how your program interacts with a machine's cache
hierarchy and/or branch predictor.
The cache simulation models a machine with independent first-level instruction
and data caches (I1 and D1), backed by a unified second-level cache (L2). For
these machines (in the cases where Cachegrind can auto-detect the cache
configuration) Cachegrind simulates the first-level and last-level caches.
Therefore, Cachegrind always refers to the I1, D1 and LL (last-level) caches.
</para>
<para>
When simulating the cache, with <option>--cache-sim=yes</option>, Cachegrind
gathers the following statistics:
</para>
<itemizedlist>
<listitem>
<para>
I cache reads (<computeroutput>Ir</computeroutput>, which equals the number
of instructions executed), I1 cache read misses
(<computeroutput>I1mr</computeroutput>) and LL cache instruction read
misses (<computeroutput>ILmr</computeroutput>).
</para>
</listitem>
<listitem>
<para>
D cache reads (<computeroutput>Dr</computeroutput>, which equals the number
of memory reads), D1 cache read misses
(<computeroutput>D1mr</computeroutput>), and LL cache data read misses
(<computeroutput>DLmr</computeroutput>).
</para>
</listitem>
<listitem>
<para>
D cache writes (<computeroutput>Dw</computeroutput>, which equals the
number of memory writes), D1 cache write misses
(<computeroutput>D1mw</computeroutput>), and LL cache data write misses
(<computeroutput>DLmw</computeroutput>).
</para>
</listitem>
</itemizedlist>
<para>
Note that D1 total accesses is given by <computeroutput>D1mr</computeroutput> +
<computeroutput>D1mw</computeroutput>, and that LL total accesses is given by
<computeroutput>ILmr</computeroutput> + <computeroutput>DLmr</computeroutput> +
<computeroutput>DLmw</computeroutput>.
</para>
<para>
When simulating the branch predictor, with <option>--branch-sim=yes</option>,
Cachegrind gathers the following statistics:
</para>
<itemizedlist>
<listitem>
<para>
Conditional branches executed (<computeroutput>Bc</computeroutput>) and
conditional branches mispredicted (<computeroutput>Bcm</computeroutput>).
</para>
</listitem>
<listitem>
<para>
Indirect branches executed (<computeroutput>Bi</computeroutput>) and
indirect branches mispredicted (<computeroutput>Bim</computeroutput>).
</para>
</listitem>
</itemizedlist>
<para>
When cache and/or branch simulation is enabled, cg_annotate will print multiple
counts per line of output. For example:
</para>
<programlisting><![CDATA[
Ir______________________ Bc____________________ Bcm__________________ Bi____________________ Bim______________ function:file
> 8,547 (0.1%, 99.4%) 936 (0.1%, 99.1%) 177 (0.3%, 96.7%) 59 (0.0%, 99.9%) 38 (19.4%, 66.3%) strcmp:
8,503 (0.1%) 928 (0.1%) 175 (0.3%) 59 (0.0%) 38 (19.4%) ./string/../sysdeps/x86_64/multiarch/../multiarch/strcmp-sse2.S
]]></programlisting>
</sect2>
</sect1>
<sect1 id="cg-manual.cgopts" xreflabel="Cachegrind Command-line Options">
<title>Cachegrind Command-line Options</title>
<!-- start of xi:include in the manpage -->
<para>
Cachegrind-specific options are:
</para>
<variablelist id="cg.opts.list">
<varlistentry id="opt.cachegrind-out-file" xreflabel="--cachegrind-out-file">
<term>
<option><![CDATA[--cachegrind-out-file=<file> ]]></option>
</term>
<listitem>
<para>
Write the Cachegrind output file to <filename>file</filename> rather than
to the default output file,
<filename>cachegrind.out.&lt;pid&gt;</filename>. The <option>%p</option>
and <option>%q</option> 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 option
<option><link linkend="opt.log-file">--log-file</link></option>.
</para>
</listitem>
</varlistentry>
<varlistentry id="opt.cache-sim" xreflabel="--cache-sim">
<term>
<option><![CDATA[--cache-sim=no|yes [no] ]]></option>
</term>
<listitem>
<para>
Enables or disables collection of cache access and miss counts.
</para>
</listitem>
</varlistentry>
<varlistentry id="opt.branch-sim" xreflabel="--branch-sim">
<term>
<option><![CDATA[--branch-sim=no|yes [no] ]]></option>
</term>
<listitem>
<para>
Enables or disables collection of branch instruction and
misprediction counts.
</para>
</listitem>
</varlistentry>
<varlistentry id="cg.opt.I1" xreflabel="--I1">
<term>
<option><![CDATA[--I1=<size>,<associativity>,<line size> ]]></option>
</term>
<listitem>
<para>
Specify the size, associativity and line size of the level 1 instruction
cache. Only useful with <option>--cache-sim=yes</option>.
</para>
</listitem>
</varlistentry>
<varlistentry id="cg.opt.D1" xreflabel="--D1">
<term>
<option><![CDATA[--D1=<size>,<associativity>,<line size> ]]></option>
</term>
<listitem>
<para>
Specify the size, associativity and line size of the level 1 data cache.
Only useful with <option>--cache-sim=yes</option>.
</para>
</listitem>
</varlistentry>
<varlistentry id="cg.opt.LL" xreflabel="--LL">
<term>
<option><![CDATA[--LL=<size>,<associativity>,<line size> ]]></option>
</term>
<listitem>
<para>
Specify the size, associativity and line size of the last-level cache.
Only useful with <option>--cache-sim=yes</option>.
</para>
</listitem>
</varlistentry>
</variablelist>
<!-- end of xi:include in the manpage -->
</sect1>
<sect1 id="cg-manual.annopts" xreflabel="cg_annotate Command-line Options">
<title>cg_annotate Command-line Options</title>
<!-- start of xi:include in the manpage -->
<variablelist id="cg_annotate.opts.list">
<varlistentry>
<term>
<option><![CDATA[-h --help ]]></option>
</term>
<listitem>
<para>Show the help message.</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--version ]]></option>
</term>
<listitem>
<para>Show the version number.</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--diff ]]></option>
</term>
<listitem>
<para>Diff two Cachegrind output files.</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--mod-filename <regex> [default: none]]]></option>
</term>
<listitem>
<para>
Specifies an <option>s/old/new/</option> search-and-replace expression
that is applied to all filenames. Useful when differencing, for removing
minor differences in paths between two different versions of a program
that are sitting in different directories. An <option>i</option> suffix
makes the regex case-insensitive, and a <option>g</option> suffix makes
it match multiple times.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--mod-funcname <regex> [default: none]]]></option>
</term>
<listitem>
<para>
Like <option>--mod-filename</option>, but for filenames. Useful for
removing minor differences in randomized names of auto-generated
functions generated by some compilers.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--show=A,B,C [default: all, using order in
the Cachegrind output file] ]]></option>
</term>
<listitem>
<para>
Specifies which events to show (and the column order). Default is to use
all present in the Cachegrind output file (and use the order in the
file). Best used in conjunction with <option>--sort</option>.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--sort=A,B,C [default: order in the Cachegrind output file] ]]></option>
</term>
<listitem>
<para>
Specifies the events upon which the sorting of the file:function and
function:file entries will be based.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--threshold=X [default: 0.1%] ]]></option>
</term>
<listitem>
<para>
Sets the significance threshold for the file:function and function:files
sections. A file or function is shown if it accounts for more than X% of
the counts for the primary sort event. If annotating source files, this
also affects which files are annotated.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--show-percs, --no-show-percs, --show-percs=<no|yes> [default: yes] ]]></option>
</term>
<listitem>
<para>
When enabled, a percentage is printed next to all event counts. This
helps gauge the relative importance of each function and line.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--annotate, --no-annotate, --auto=<no|yes> [default: yes] ]]></option>
</term>
<listitem>
<para>
Enables or disables source file annotation.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--context=N [default: 8] ]]></option>
</term>
<listitem>
<para>
The number of lines of context to show before and after each annotated
line. Use a large number (e.g. 100000) to show all source lines.
</para>
</listitem>
</varlistentry>
</variablelist>
<!-- end of xi:include in the manpage -->
</sect1>
<sect1 id="cg-manual.mergeopts" xreflabel="cg_merge Command-line Options">
<title>cg_merge Command-line Options</title>
<!-- start of xi:include in the manpage -->
<variablelist id="cg_merge.opts.list">
<varlistentry>
<term>
<option><![CDATA[-o outfile]]></option>
</term>
<listitem>
<para>
Write the output to to <computeroutput>outfile</computeroutput>
instead of standard output.
</para>
</listitem>
</varlistentry>
</variablelist>
<!-- end of xi:include in the manpage -->
</sect1>
<sect1 id="cg-manual.diffopts" xreflabel="cg_diff Command-line Options">
<title>cg_diff Command-line Options</title>
<!-- start of xi:include in the manpage -->
<variablelist id="cg_diff.opts.list">
<varlistentry>
<term>
<option><![CDATA[-h --help ]]></option>
</term>
<listitem>
<para>Show the help message.</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--version ]]></option>
</term>
<listitem>
<para>Show the version number.</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--mod-filename=<expr> [default: none]]]></option>
</term>
<listitem>
<para>
Specifies an <option>s/old/new/</option> search-and-replace expression
that is applied to all filenames.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term>
<option><![CDATA[--mod-funcname=<expr> [default: none]]]></option>
</term>
<listitem>
<para>
Like <option>--mod-filename</option>, but for filenames.
</para>
</listitem>
</varlistentry>
</variablelist>
<!-- end of xi:include in the manpage -->
</sect1>
<sect1 id="cg-manual.sim-details"
xreflabel="Simulation Details">
<title>Simulation Details</title>
<para>
This section talks about details you don't need to know about in order to
use Cachegrind, but may be of interest to some people.
</para>
<sect2 id="cache-sim" xreflabel="Cache Simulation Specifics">
<title>Cache Simulation Specifics</title>
<para>
The cache simulation approximates the hardware of an AMD Athlon CPU circa 2002.
Its specific characteristics are as follows:</para>
<itemizedlist>
<listitem>
<para>Write-allocate: when a write miss occurs, the block
written to is brought into the D1 cache. Most modern caches
have this property.</para>
</listitem>
<listitem>
<para>Bit-selection hash function: the set of line(s) in the cache
to which a memory block maps is chosen by the middle bits
M--(M+N-1) of the byte address, where:</para>
<itemizedlist>
<listitem>
<para>line size = 2^M bytes</para>
</listitem>
<listitem>
<para>(cache size / line size / associativity) = 2^N bytes</para>
</listitem>
</itemizedlist>
</listitem>
<listitem>
<para>Inclusive LL cache: the LL cache typically replicates all
the entries of the L1 caches, because fetching into L1 involves
fetching into LL first (this does not guarantee strict inclusiveness,
as lines evicted from LL still could reside in L1). This is
standard on Pentium chips, but AMD Opterons, Athlons and Durons
use an exclusive LL cache that only holds
blocks evicted from L1. Ditto most modern VIA CPUs.</para>
</listitem>
</itemizedlist>
<para>The cache configuration simulated (cache size,
associativity and line size) is determined automatically using
the x86 CPUID instruction. If you have a machine that (a)
doesn't support the CPUID instruction, or (b) supports it in an
early incarnation that doesn't give any cache information, then
Cachegrind will fall back to using a default configuration (that
of a model 3/4 Athlon). Cachegrind will tell you if this
happens. You can manually specify one, two or all three levels
(I1/D1/LL) of the cache from the command line using the
<option>--I1</option>,
<option>--D1</option> and
<option>--LL</option> options.
For cache parameters to be valid for simulation, the number
of sets (with associativity being the number of cache lines in
each set) has to be a power of two.</para>
<para>On PowerPC platforms
Cachegrind cannot automatically
determine the cache configuration, so you will
need to specify it with the
<option>--I1</option>,
<option>--D1</option> and
<option>--LL</option> options.</para>
<para>Other noteworthy behaviour:</para>
<itemizedlist>
<listitem>
<para>References that straddle two cache lines are treated as
follows:</para>
<itemizedlist>
<listitem>
<para>If both blocks hit --&gt; counted as one hit</para>
</listitem>
<listitem>
<para>If one block hits, the other misses --&gt; counted
as one miss.</para>
</listitem>
<listitem>
<para>If both blocks miss --&gt; counted as one miss (not
two)</para>
</listitem>
</itemizedlist>
</listitem>
<listitem>
<para>Instructions that modify a memory location
(e.g. <computeroutput>inc</computeroutput> and
<computeroutput>dec</computeroutput>) are counted as doing
just a read, i.e. a single data reference. This may seem
strange, but since the write can never cause a miss (the read
guarantees the block is in the cache) it's not very
interesting.</para>
<para>Thus it measures not the number of times the data cache
is accessed, but the number of times a data cache miss could
occur.</para>
</listitem>
</itemizedlist>
<para>
If you are interested in simulating a cache with different properties, it is
not particularly hard to write your own cache simulator, or to modify the
existing ones in <computeroutput>cg_sim.c</computeroutput>.
</para>
</sect2>
<sect2 id="branch-sim" xreflabel="Branch Simulation Specifics">
<title>Branch Simulation Specifics</title>
<para>Cachegrind simulates branch predictors intended to be
typical of mainstream desktop/server processors of around 2004.</para>
<para>Conditional branches are predicted using an array of 16384 2-bit
saturating counters. The array index used for a branch instruction is
computed partly from the low-order bits of the branch instruction's
address and partly using the taken/not-taken behaviour of the last few
conditional branches. As a result the predictions for any specific
branch depend both on its own history and the behaviour of previous
branches. This is a standard technique for improving prediction
accuracy.</para>
<para>For indirect branches (that is, jumps to unknown destinations)
Cachegrind uses a simple branch target address predictor. Targets are
predicted using an array of 512 entries indexed by the low order 9
bits of the branch instruction's address. Each branch is predicted to
jump to the same address it did last time. Any other behaviour causes
a mispredict.</para>
<para>More recent processors have better branch predictors, in
particular better indirect branch predictors. Cachegrind's predictor
design is deliberately conservative so as to be representative of the
large installed base of processors which pre-date widespread
deployment of more sophisticated indirect branch predictors. In
particular, late model Pentium 4s (Prescott), Pentium M, Core and Core
2 have more sophisticated indirect branch predictors than modelled by
Cachegrind. </para>
<para>Cachegrind does not simulate a return stack predictor. It
assumes that processors perfectly predict function return addresses,
an assumption which is probably close to being true.</para>
<para>See Hennessy and Patterson's classic text "Computer
Architecture: A Quantitative Approach", 4th edition (2007), Section
2.3 (pages 80-89) for background on modern branch predictors.</para>
</sect2>
<sect2 id="cg-manual.annopts.accuracy" xreflabel="Accuracy">
<title>Accuracy</title>
<para>
Cachegrind's instruction counting has one shortcoming on x86/amd64:
</para>
<itemizedlist>
<listitem>
<para>
When a <function>REP</function>-prefixed instruction executes each
iteration is counted separately. In contrast, hardware counters count each
such instruction just once, no matter how many times it iterates. It is
arguable that Cachegrind's behaviour is more useful.
</para>
</listitem>
</itemizedlist>
<para>
Cachegrind's cache profiling has a number of shortcomings:
</para>
<itemizedlist>
<listitem>
<para>
It doesn't account for kernel activity. The effect of system calls on the
cache and branch predictor contents is ignored.
</para>
</listitem>
<listitem>
<para>
It doesn't account for other process activity. This is arguably desirable
when considering a single program.
</para>
</listitem>
<listitem>
<para>It doesn't account for virtual-to-physical address
mappings. Hence the simulation is not a true
representation of what's happening in the
cache. Most caches and branch predictors are physically indexed, but
Cachegrind simulates caches using virtual addresses.</para>
</listitem>
<listitem>
<para>It doesn't account for cache misses not visible at the
instruction level, e.g. those arising from TLB misses, or
speculative execution.</para>
</listitem>
<listitem>
<para>Valgrind will schedule
threads differently from how they would be when running natively.
This could warp the results for threaded programs.</para>
</listitem>
<listitem>
<para>
The x86/amd64 instructions <computeroutput>bts</computeroutput>,
<computeroutput>btr</computeroutput> and
<computeroutput>btc</computeroutput> will incorrectly be counted as doing a
data read if both the arguments are registers, e.g.:
<programlisting><![CDATA[
btsl %eax, %edx]]></programlisting>
This should only happen rarely.
</para>
</listitem>
<listitem>
<para>x86/amd64 FPU instructions with data sizes of 28 and 108 bytes
(e.g. <computeroutput>fsave</computeroutput>) are treated as
though they only access 16 bytes. These instructions seem to
be rare so hopefully this won't affect accuracy much.</para>
</listitem>
</itemizedlist>
<para>Another thing worth noting is that results are very sensitive.
Changing the size of the executable being profiled, or the sizes
of any of the shared libraries it uses, or even the length of their
file names, can perturb the results. Variations will be small, but
don't expect perfectly repeatable results if your program changes at
all.</para>
<para>
Many Linux distributions perform address space layout randomisation (ASLR), in
which identical runs of the same program have their shared libraries loaded at
different locations, as a security measure. This also perturbs the
results.
</para>
</sect2>
</sect1>
<sect1 id="cg-manual.impl-details"
xreflabel="Implementation Details">
<title>Implementation Details</title>
<para>
This section talks about details you don't need to know about in order to
use Cachegrind, but may be of interest to some people.
</para>
<sect2 id="cg-manual.impl-details.how-cg-works"
xreflabel="How Cachegrind Works">
<title>How Cachegrind Works</title>
<para>The best reference for understanding how Cachegrind works is chapter 3 of
"Dynamic Binary Analysis and Instrumentation", by Nicholas Nethercote. It
is available on the <ulink url="&vg-pubs-url;">Valgrind publications
page</ulink>.</para>
</sect2>
<sect2 id="cg-manual.impl-details.file-format"
xreflabel="Cachegrind Output File Format">
<title>Cachegrind Output File Format</title>
<para>The file format is fairly straightforward, basically giving the
cost centre for every line, grouped by files and
functions. It's also totally generic and self-describing, in the sense that
it can be used for any events that can be counted on a line-by-line basis,
not just cache and branch predictor events. For example, earlier versions
of Cachegrind didn't have a branch predictor simulation. When this was
added, the file format didn't need to change at all. So the format (and
consequently, cg_annotate) could be used by other tools.</para>
<para>The file format:</para>
<programlisting><![CDATA[
file ::= desc_line* cmd_line events_line data_line+ summary_line
desc_line ::= "desc:" ws? non_nl_string
cmd_line ::= "cmd:" ws? cmd
events_line ::= "events:" ws? (event ws)+
data_line ::= file_line | fn_line | count_line
file_line ::= "fl=" filename
fn_line ::= "fn=" fn_name
count_line ::= line_num (ws+ count)* ws*
summary_line ::= "summary:" ws? count (ws+ count)+ ws*
count ::= num]]></programlisting>
<para>Where:</para>
<itemizedlist>
<listitem>
<para><computeroutput>non_nl_string</computeroutput> is any
string not containing a newline.</para>
</listitem>
<listitem>
<para><computeroutput>cmd</computeroutput> is a string holding the
command line of the profiled program.</para>
</listitem>
<listitem>
<para><computeroutput>event</computeroutput> is a string containing
no whitespace.</para>
</listitem>
<listitem>
<para><computeroutput>filename</computeroutput> and
<computeroutput>fn_name</computeroutput> are strings.</para>
</listitem>
<listitem>
<para><computeroutput>num</computeroutput> and
<computeroutput>line_num</computeroutput> are decimal
numbers.</para>
</listitem>
<listitem>
<para><computeroutput>ws</computeroutput> is whitespace.</para>
</listitem>
</itemizedlist>
<para>The contents of the "desc:" lines are printed out at the top
of the summary. This is a generic way of providing simulation
specific information, e.g. for giving the cache configuration for
cache simulation.</para>
<para>More than one line of info can be present for each file/fn/line number.
In such cases, the counts for the named events will be accumulated.</para>
<para>The number of counts in each
<computeroutput>line</computeroutput> and the
<computeroutput>summary_line</computeroutput> should not exceed
the number of events in the
<computeroutput>event_line</computeroutput>. If the number in
each <computeroutput>line</computeroutput> is less, cg_annotate
treats those missing as though they were a "0" entry. This can reduce
file size.
</para>
<para>A <computeroutput>file_line</computeroutput> changes the
current file name. A <computeroutput>fn_line</computeroutput>
changes the current function name. A
<computeroutput>count_line</computeroutput> contains counts that
pertain to the current filename/fn_name. A "fn="
<computeroutput>file_line</computeroutput> and a
<computeroutput>fn_line</computeroutput> must appear before any
<computeroutput>count_line</computeroutput>s to give the context
of the first <computeroutput>count_line</computeroutput>s.</para>
<para>Similarly, each <computeroutput>file_line</computeroutput> must be
immediately followed by a <computeroutput>fn_line</computeroutput>.
</para>
<para>The summary line is redundant, because it just holds the total counts
for each event. But this serves as a useful sanity check of the data; if
the totals for each event don't match the summary line, something has gone
wrong.</para>
</sect2>
</sect1>
</chapter>
Reference in New Issue View Git Blame Copy Permalink