1 <?xml version="1.0"?> <!-- -*- sgml -*- -->
2 <!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
3 "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd"
4 [ <!ENTITY % vg-entities SYSTEM "../../docs/xml/vg-entities.xml"> %vg-entities; ]>
7 <chapter id="cg-manual" xreflabel="Cachegrind: a cache and branch-prediction profiler">
8 <title>Cachegrind: a cache and branch-prediction profiler</title>
10 <para>To use this tool, you must specify
11 <option>--tool=cachegrind</option> on the
12 Valgrind command line.</para>
14 <sect1 id="cg-manual.overview" xreflabel="Overview">
15 <title>Overview</title>
17 <para>Cachegrind simulates how your program interacts with a machine's cache
18 hierarchy and (optionally) branch predictor. It simulates a machine with
19 independent first level instruction and data caches (I1 and D1), backed by a
20 unified second level cache (L2). This configuration is used by almost all
21 modern machines.</para>
24 It gathers the following statistics (abbreviations used for each statistic
25 is given in parentheses):</para>
28 <para>I cache reads (<computeroutput>Ir</computeroutput>,
29 which equals the number of instructions executed),
30 I1 cache read misses (<computeroutput>I1mr</computeroutput>) and
31 L2 cache instruction read misses (<computeroutput>I1mr</computeroutput>).
35 <para>D cache reads (<computeroutput>Dr</computeroutput>, which
36 equals the number of memory reads),
37 D1 cache read misses (<computeroutput>D1mr</computeroutput>), and
38 L2 cache data read misses (<computeroutput>D2mr</computeroutput>).
42 <para>D cache writes (<computeroutput>Dw</computeroutput>, which equals
43 the number of memory writes),
44 D1 cache write misses (<computeroutput>D1mw</computeroutput>), and
45 L2 cache data write misses (<computeroutput>D2mw</computeroutput>).
49 <para>Conditional branches executed (<computeroutput>Bc</computeroutput>) and
50 conditional branches mispredicted (<computeroutput>Bcm</computeroutput>).
54 <para>Indirect branches executed (<computeroutput>Bi</computeroutput>) and
55 indirect branches mispredicted (<computeroutput>Bim</computeroutput>).
60 <para>Note that D1 total accesses is given by
61 <computeroutput>D1mr</computeroutput> +
62 <computeroutput>D1mw</computeroutput>, and that L2 total
63 accesses is given by <computeroutput>I2mr</computeroutput> +
64 <computeroutput>D2mr</computeroutput> +
65 <computeroutput>D2mw</computeroutput>.
68 <para>These statistics are presented for the entire program and for each
69 function in the program. You can also annotate each line of source code in
70 the program with the counts that were caused directly by it.</para>
72 <para>On a modern machine, an L1 miss will typically cost
73 around 10 cycles, an L2 miss can cost as much as 200
74 cycles, and a mispredicted branch costs in the region of 10
75 to 30 cycles. Detailed cache and branch profiling can be very useful
76 for understanding how your program interacts with the machine and thus how
77 to make it faster.</para>
79 <para>Also, since one instruction cache read is performed per
80 instruction executed, you can find out how many instructions are
81 executed per line, which can be useful for traditional profiling.</para>
87 <sect1 id="cg-manual.profile"
88 xreflabel="Using Cachegrind, cg_annotate and cg_merge">
89 <title>Using Cachegrind, cg_annotate and cg_merge</title>
91 <para>First off, as for normal Valgrind use, you probably want to
92 compile with debugging info (the
93 <option>-g</option> option). But by contrast with
94 normal Valgrind use, you probably do want to turn
95 optimisation on, since you should profile your program as it will
96 be normally run.</para>
98 <para>Then, you need to run Cachegrind itself to gather the profiling
99 information, and then run cg_annotate to get a detailed presentation of that
100 information. As an optional intermediate step, you can use cg_merge to sum
101 together the outputs of multiple Cachegrind runs into a single file which
102 you then use as the input for cg_annotate. Alternatively, you can use
103 cg_diff to difference the outputs of two Cachegrind runs into a signel file
104 which you then use as the input for cg_annotate.</para>
107 <sect2 id="cg-manual.running-cachegrind" xreflabel="Running Cachegrind">
108 <title>Running Cachegrind</title>
110 <para>To run Cachegrind on a program <filename>prog</filename>, run:</para>
112 valgrind --tool=cachegrind prog
115 <para>The program will execute (slowly). Upon completion,
116 summary statistics that look like this will be printed:</para>
118 <programlisting><![CDATA[
119 ==31751== I refs: 27,742,716
120 ==31751== I1 misses: 276
121 ==31751== L2i misses: 275
122 ==31751== I1 miss rate: 0.0%
123 ==31751== L2i miss rate: 0.0%
125 ==31751== D refs: 15,430,290 (10,955,517 rd + 4,474,773 wr)
126 ==31751== D1 misses: 41,185 ( 21,905 rd + 19,280 wr)
127 ==31751== L2d misses: 23,085 ( 3,987 rd + 19,098 wr)
128 ==31751== D1 miss rate: 0.2% ( 0.1% + 0.4%)
129 ==31751== L2d miss rate: 0.1% ( 0.0% + 0.4%)
131 ==31751== L2 misses: 23,360 ( 4,262 rd + 19,098 wr)
132 ==31751== L2 miss rate: 0.0% ( 0.0% + 0.4%)]]></programlisting>
134 <para>Cache accesses for instruction fetches are summarised
135 first, giving the number of fetches made (this is the number of
136 instructions executed, which can be useful to know in its own
137 right), the number of I1 misses, and the number of L2 instruction
138 (<computeroutput>L2i</computeroutput>) misses.</para>
140 <para>Cache accesses for data follow. The information is similar
141 to that of the instruction fetches, except that the values are
142 also shown split between reads and writes (note each row's
143 <computeroutput>rd</computeroutput> and
144 <computeroutput>wr</computeroutput> values add up to the row's
147 <para>Combined instruction and data figures for the L2 cache
148 follow that. Note that the L2 miss rate is computed relative to the total
149 number of memory accesses, not the number of L1 misses. I.e. it is
150 <computeroutput>(I2mr + D2mr + D2mw) / (Ir + Dr + Dw)</computeroutput>
152 <computeroutput>(I2mr + D2mr + D2mw) / (I1mr + D1mr + D1mw)</computeroutput>
155 <para>Branch prediction statistics are not collected by default.
156 To do so, add the option <option>--branch-sim=yes</option>.</para>
161 <sect2 id="cg-manual.outputfile" xreflabel="Output File">
162 <title>Output File</title>
164 <para>As well as printing summary information, Cachegrind also writes
165 more detailed profiling information to a file. By default this file is named
166 <filename>cachegrind.out.<pid></filename> (where
167 <filename><pid></filename> is the program's process ID), but its name
168 can be changed with the <option>--cachegrind-out-file</option> option. This
169 file is human-readable, but is intended to be interpreted by the
170 accompanying program cg_annotate, described in the next section.</para>
172 <para>The default <computeroutput>.<pid></computeroutput> suffix
173 on the output file name serves two purposes. Firstly, it means you
174 don't have to rename old log files that you don't want to overwrite.
175 Secondly, and more importantly, it allows correct profiling with the
176 <option>--trace-children=yes</option> option of
177 programs that spawn child processes.</para>
179 <para>The output file can be big, many megabytes for large applications
180 built with full debugging information.</para>
186 <sect2 id="cg-manual.running-cg_annotate" xreflabel="Running cg_annotate">
187 <title>Running cg_annotate</title>
189 <para>Before using cg_annotate,
190 it is worth widening your window to be at least 120-characters
191 wide if possible, as the output lines can be quite long.</para>
193 <para>To get a function-by-function summary, run:</para>
195 <screen>cg_annotate <filename></screen>
197 <para>on a Cachegrind output file.</para>
202 <sect2 id="cg-manual.the-output-preamble" xreflabel="The Output Preamble">
203 <title>The Output Preamble</title>
205 <para>The first part of the output looks like this:</para>
207 <programlisting><![CDATA[
208 --------------------------------------------------------------------------------
209 I1 cache: 65536 B, 64 B, 2-way associative
210 D1 cache: 65536 B, 64 B, 2-way associative
211 L2 cache: 262144 B, 64 B, 8-way associative
212 Command: concord vg_to_ucode.c
213 Events recorded: Ir I1mr I2mr Dr D1mr D2mr Dw D1mw D2mw
214 Events shown: Ir I1mr I2mr Dr D1mr D2mr Dw D1mw D2mw
215 Event sort order: Ir I1mr I2mr Dr D1mr D2mr Dw D1mw D2mw
217 Chosen for annotation:
222 <para>This is a summary of the annotation options:</para>
227 <para>I1 cache, D1 cache, L2 cache: cache configuration. So
228 you know the configuration with which these results were
233 <para>Command: the command line invocation of the program
234 under examination.</para>
238 <para>Events recorded: which events were recorded.</para>
243 <para>Events shown: the events shown, which is a subset of the events
244 gathered. This can be adjusted with the
245 <option>--show</option> option.</para>
249 <para>Event sort order: the sort order in which functions are
250 shown. For example, in this case the functions are sorted
251 from highest <computeroutput>Ir</computeroutput> counts to
252 lowest. If two functions have identical
253 <computeroutput>Ir</computeroutput> counts, they will then be
254 sorted by <computeroutput>I1mr</computeroutput> counts, and
255 so on. This order can be adjusted with the
256 <option>--sort</option> option.</para>
258 <para>Note that this dictates the order the functions appear.
259 It is <emphasis>not</emphasis> the order in which the columns
260 appear; that is dictated by the "events shown" line (and can
261 be changed with the <option>--show</option>
266 <para>Threshold: cg_annotate
267 by default omits functions that cause very low counts
268 to avoid drowning you in information. In this case,
269 cg_annotate shows summaries the functions that account for
270 99% of the <computeroutput>Ir</computeroutput> counts;
271 <computeroutput>Ir</computeroutput> is chosen as the
272 threshold event since it is the primary sort event. The
273 threshold can be adjusted with the
274 <option>--threshold</option>
279 <para>Chosen for annotation: names of files specified
280 manually for annotation; in this case none.</para>
284 <para>Auto-annotation: whether auto-annotation was requested
285 via the <option>--auto=yes</option>
286 option. In this case no.</para>
294 <sect2 id="cg-manual.the-global"
295 xreflabel="The Global and Function-level Counts">
296 <title>The Global and Function-level Counts</title>
298 <para>Then follows summary statistics for the whole
301 <programlisting><![CDATA[
302 --------------------------------------------------------------------------------
303 Ir I1mr I2mr Dr D1mr D2mr Dw D1mw D2mw
304 --------------------------------------------------------------------------------
305 27,742,716 276 275 10,955,517 21,905 3,987 4,474,773 19,280 19,098 PROGRAM TOTALS]]></programlisting>
308 These are similar to the summary provided when Cachegrind finishes running.
311 <para>Then comes function-by-function statistics:</para>
313 <programlisting><![CDATA[
314 --------------------------------------------------------------------------------
315 Ir I1mr I2mr Dr D1mr D2mr Dw D1mw D2mw file:function
316 --------------------------------------------------------------------------------
317 8,821,482 5 5 2,242,702 1,621 73 1,794,230 0 0 getc.c:_IO_getc
318 5,222,023 4 4 2,276,334 16 12 875,959 1 1 concord.c:get_word
319 2,649,248 2 2 1,344,810 7,326 1,385 . . . vg_main.c:strcmp
320 2,521,927 2 2 591,215 0 0 179,398 0 0 concord.c:hash
321 2,242,740 2 2 1,046,612 568 22 448,548 0 0 ctype.c:tolower
322 1,496,937 4 4 630,874 9,000 1,400 279,388 0 0 concord.c:insert
323 897,991 51 51 897,831 95 30 62 1 1 ???:???
324 598,068 1 1 299,034 0 0 149,517 0 0 ../sysdeps/generic/lockfile.c:__flockfile
325 598,068 0 0 299,034 0 0 149,517 0 0 ../sysdeps/generic/lockfile.c:__funlockfile
326 598,024 4 4 213,580 35 16 149,506 0 0 vg_clientmalloc.c:malloc
327 446,587 1 1 215,973 2,167 430 129,948 14,057 13,957 concord.c:add_existing
328 341,760 2 2 128,160 0 0 128,160 0 0 vg_clientmalloc.c:vg_trap_here_WRAPPER
329 320,782 4 4 150,711 276 0 56,027 53 53 concord.c:init_hash_table
330 298,998 1 1 106,785 0 0 64,071 1 1 concord.c:create
331 149,518 0 0 149,516 0 0 1 0 0 ???:tolower@@GLIBC_2.0
332 149,518 0 0 149,516 0 0 1 0 0 ???:fgetc@@GLIBC_2.0
333 95,983 4 4 38,031 0 0 34,409 3,152 3,150 concord.c:new_word_node
334 85,440 0 0 42,720 0 0 21,360 0 0 vg_clientmalloc.c:vg_bogus_epilogue]]></programlisting>
338 <computeroutput>file_name:function_name</computeroutput> pair. If
339 a column contains only a dot it means the function never performs
340 that event (e.g. the third row shows that
341 <computeroutput>strcmp()</computeroutput> contains no
342 instructions that write to memory). The name
343 <computeroutput>???</computeroutput> is used if the the file name
344 and/or function name could not be determined from debugging
345 information. If most of the entries have the form
346 <computeroutput>???:???</computeroutput> the program probably
347 wasn't compiled with <option>-g</option>.</para>
349 <para>It is worth noting that functions will come both from
350 the profiled program (e.g. <filename>concord.c</filename>)
351 and from libraries (e.g. <filename>getc.c</filename>)</para>
356 <sect2 id="cg-manual.line-by-line" xreflabel="Line-by-line Counts">
357 <title>Line-by-line Counts</title>
359 <para>There are two ways to annotate source files -- by specifying them
360 manually as arguments to cg_annotate, or with the
361 <option>--auto=yes</option> option. For example, the output from running
362 <filename>cg_annotate <filename> concord.c</filename> for our example
363 produces the same output as above followed by an annotated version of
364 <filename>concord.c</filename>, a section of which looks like:</para>
366 <programlisting><![CDATA[
367 --------------------------------------------------------------------------------
368 -- User-annotated source: concord.c
369 --------------------------------------------------------------------------------
370 Ir I1mr I2mr Dr D1mr D2mr Dw D1mw D2mw
372 . . . . . . . . . void init_hash_table(char *file_name, Word_Node *table[])
374 . . . . . . . . . FILE *file_ptr;
375 . . . . . . . . . Word_Info *data;
376 1 0 0 . . . 1 1 1 int line = 1, i;
378 5 0 0 . . . 3 0 0 data = (Word_Info *) create(sizeof(Word_Info));
380 4,991 0 0 1,995 0 0 998 0 0 for (i = 0; i < TABLE_SIZE; i++)
381 3,988 1 1 1,994 0 0 997 53 52 table[i] = NULL;
383 . . . . . . . . . /* Open file, check it. */
384 6 0 0 1 0 0 4 0 0 file_ptr = fopen(file_name, "r");
385 2 0 0 1 0 0 . . . if (!(file_ptr)) {
386 . . . . . . . . . fprintf(stderr, "Couldn't open '%s'.\n", file_name);
387 1 1 1 . . . . . . exit(EXIT_FAILURE);
390 165,062 1 1 73,360 0 0 91,700 0 0 while ((line = get_word(data, line, file_ptr)) != EOF)
391 146,712 0 0 73,356 0 0 73,356 0 0 insert(data->;word, data->line, table);
393 4 0 0 1 0 0 2 0 0 free(data);
394 4 0 0 1 0 0 2 0 0 fclose(file_ptr);
395 3 0 0 2 0 0 . . . }]]></programlisting>
397 <para>(Although column widths are automatically minimised, a wide
398 terminal is clearly useful.)</para>
400 <para>Each source file is clearly marked
401 (<computeroutput>User-annotated source</computeroutput>) as
402 having been chosen manually for annotation. If the file was
403 found in one of the directories specified with the
404 <option>-I</option>/<option>--include</option> option, the directory
405 and file are both given.</para>
407 <para>Each line is annotated with its event counts. Events not
408 applicable for a line are represented by a dot. This is useful
409 for distinguishing between an event which cannot happen, and one
410 which can but did not.</para>
412 <para>Sometimes only a small section of a source file is
413 executed. To minimise uninteresting output, Cachegrind only shows
414 annotated lines and lines within a small distance of annotated
415 lines. Gaps are marked with the line numbers so you know which
416 part of a file the shown code comes from, eg:</para>
418 <programlisting><![CDATA[
419 (figures and code for line 704)
420 -- line 704 ----------------------------------------
421 -- line 878 ----------------------------------------
422 (figures and code for line 878)]]></programlisting>
424 <para>The amount of context to show around annotated lines is
425 controlled by the <option>--context</option>
428 <para>To get automatic annotation, use the <option>--auto=yes</option> option.
429 cg_annotate will automatically annotate every source file it can
430 find that is mentioned in the function-by-function summary.
431 Therefore, the files chosen for auto-annotation are affected by
432 the <option>--sort</option> and
433 <option>--threshold</option> options. Each
434 source file is clearly marked (<computeroutput>Auto-annotated
435 source</computeroutput>) as being chosen automatically. Any
436 files that could not be found are mentioned at the end of the
439 <programlisting><![CDATA[
440 ------------------------------------------------------------------
441 The following files chosen for auto-annotation could not be found:
442 ------------------------------------------------------------------
445 ../sysdeps/generic/lockfile.c]]></programlisting>
447 <para>This is quite common for library files, since libraries are
448 usually compiled with debugging information, but the source files
449 are often not present on a system. If a file is chosen for
450 annotation both manually and automatically, it
451 is marked as <computeroutput>User-annotated
452 source</computeroutput>. Use the
453 <option>-I</option>/<option>--include</option> option to tell Valgrind where
454 to look for source files if the filenames found from the debugging
455 information aren't specific enough.</para>
457 <para>Beware that cg_annotate can take some time to digest large
458 <filename>cachegrind.out.<pid></filename> files,
459 e.g. 30 seconds or more. Also beware that auto-annotation can
460 produce a lot of output if your program is large!</para>
465 <sect2 id="cg-manual.assembler" xreflabel="Annotating Assembly Code Programs">
466 <title>Annotating Assembly Code Programs</title>
468 <para>Valgrind can annotate assembly code programs too, or annotate
469 the assembly code generated for your C program. Sometimes this is
470 useful for understanding what is really happening when an
471 interesting line of C code is translated into multiple
474 <para>To do this, you just need to assemble your
475 <computeroutput>.s</computeroutput> files with assembly-level debug
476 information. You can use compile with the <option>-S</option> to compile C/C++
477 programs to assembly code, and then assemble the assembly code files with
478 <option>-g</option> to achieve this. You can then profile and annotate the
479 assembly code source files in the same way as C/C++ source files.</para>
483 <sect2 id="ms-manual.forkingprograms" xreflabel="Forking Programs">
484 <title>Forking Programs</title>
485 <para>If your program forks, the child will inherit all the profiling data that
486 has been gathered for the parent.</para>
488 <para>If the output file format string (controlled by
489 <option>--cachegrind-out-file</option>) does not contain <option>%p</option>,
490 then the outputs from the parent and child will be intermingled in a single
491 output file, which will almost certainly make it unreadable by
496 <sect2 id="cg-manual.annopts.warnings" xreflabel="cg_annotate Warnings">
497 <title>cg_annotate Warnings</title>
499 <para>There are a couple of situations in which
500 cg_annotate issues warnings.</para>
504 <para>If a source file is more recent than the
505 <filename>cachegrind.out.<pid></filename> file.
506 This is because the information in
507 <filename>cachegrind.out.<pid></filename> is only
508 recorded with line numbers, so if the line numbers change at
509 all in the source (e.g. lines added, deleted, swapped), any
510 annotations will be incorrect.</para>
513 <para>If information is recorded about line numbers past the
514 end of a file. This can be caused by the above problem,
515 i.e. shortening the source file while using an old
516 <filename>cachegrind.out.<pid></filename> file. If
517 this happens, the figures for the bogus lines are printed
518 anyway (clearly marked as bogus) in case they are
527 <sect2 id="cg-manual.annopts.things-to-watch-out-for"
528 xreflabel="Unusual Annotation Cases">
529 <title>Unusual Annotation Cases</title>
531 <para>Some odd things that can occur during annotation:</para>
535 <para>If annotating at the assembler level, you might see
536 something like this:</para>
537 <programlisting><![CDATA[
538 1 0 0 . . . . . . leal -12(%ebp),%eax
539 1 0 0 . . . 1 0 0 movl %eax,84(%ebx)
540 2 0 0 0 0 0 1 0 0 movl $1,-20(%ebp)
541 . . . . . . . . . .align 4,0x90
542 1 0 0 . . . . . . movl $.LnrB,%eax
543 1 0 0 . . . 1 0 0 movl %eax,-16(%ebp)]]></programlisting>
545 <para>How can the third instruction be executed twice when
546 the others are executed only once? As it turns out, it
547 isn't. Here's a dump of the executable, using
548 <computeroutput>objdump -d</computeroutput>:</para>
549 <programlisting><![CDATA[
550 8048f25: 8d 45 f4 lea 0xfffffff4(%ebp),%eax
551 8048f28: 89 43 54 mov %eax,0x54(%ebx)
552 8048f2b: c7 45 ec 01 00 00 00 movl $0x1,0xffffffec(%ebp)
553 8048f32: 89 f6 mov %esi,%esi
554 8048f34: b8 08 8b 07 08 mov $0x8078b08,%eax
555 8048f39: 89 45 f0 mov %eax,0xfffffff0(%ebp)]]></programlisting>
557 <para>Notice the extra <computeroutput>mov
558 %esi,%esi</computeroutput> instruction. Where did this come
559 from? The GNU assembler inserted it to serve as the two
560 bytes of padding needed to align the <computeroutput>movl
561 $.LnrB,%eax</computeroutput> instruction on a four-byte
562 boundary, but pretended it didn't exist when adding debug
563 information. Thus when Valgrind reads the debug info it
564 thinks that the <computeroutput>movl
565 $0x1,0xffffffec(%ebp)</computeroutput> instruction covers the
566 address range 0x8048f2b--0x804833 by itself, and attributes
567 the counts for the <computeroutput>mov
568 %esi,%esi</computeroutput> to it.</para>
572 I think this isn't true any more, not since cost centres were moved from
573 being associated with instruction addresses to being associated with
576 <para>Inlined functions can cause strange results in the
577 function-by-function summary. If a function
578 <computeroutput>inline_me()</computeroutput> is defined in
579 <filename>foo.h</filename> and inlined in the functions
580 <computeroutput>f1()</computeroutput>,
581 <computeroutput>f2()</computeroutput> and
582 <computeroutput>f3()</computeroutput> in
583 <filename>bar.c</filename>, there will not be a
584 <computeroutput>foo.h:inline_me()</computeroutput> function
585 entry. Instead, there will be separate function entries for
586 each inlining site, i.e.
587 <computeroutput>foo.h:f1()</computeroutput>,
588 <computeroutput>foo.h:f2()</computeroutput> and
589 <computeroutput>foo.h:f3()</computeroutput>. To find the
591 <computeroutput>foo.h:inline_me()</computeroutput>, add up
592 the counts from each entry.</para>
594 <para>The reason for this is that although the debug info
595 output by GCC indicates the switch from
596 <filename>bar.c</filename> to <filename>foo.h</filename>, it
597 doesn't indicate the name of the function in
598 <filename>foo.h</filename>, so Valgrind keeps using the old
604 <para>Sometimes, the same filename might be represented with
605 a relative name and with an absolute name in different parts
606 of the debug info, eg:
607 <filename>/home/user/proj/proj.h</filename> and
608 <filename>../proj.h</filename>. In this case, if you use
609 auto-annotation, the file will be annotated twice with the
610 counts split between the two.</para>
614 <para>Files with more than 65,535 lines cause difficulties
615 for the Stabs-format debug info reader. This is because the line
616 number in the <computeroutput>struct nlist</computeroutput>
617 defined in <filename>a.out.h</filename> under Linux is only a
618 16-bit value. Valgrind can handle some files with more than
619 65,535 lines correctly by making some guesses to identify
620 line number overflows. But some cases are beyond it, in
621 which case you'll get a warning message explaining that
622 annotations for the file might be incorrect.</para>
624 <para>If you are using GCC 3.1 or later, this is most likely
625 irrelevant, since GCC switched to using the more modern DWARF2
626 format by default at version 3.1. DWARF2 does not have any such
627 limitations on line numbers.</para>
631 <para>If you compile some files with
632 <option>-g</option> and some without, some
633 events that take place in a file without debug info could be
634 attributed to the last line of a file with debug info
635 (whichever one gets placed before the non-debug-info file in
636 the executable).</para>
641 <para>This list looks long, but these cases should be fairly
647 <sect2 id="cg-manual.cg_merge" xreflabel="cg_merge">
648 <title>Merging Profiles with cg_merge</title>
651 cg_merge is a simple program which
652 reads multiple profile files, as created by Cachegrind, merges them
653 together, and writes the results into another file in the same format.
654 You can then examine the merged results using
655 <computeroutput>cg_annotate <filename></computeroutput>, as
656 described above. The merging functionality might be useful if you
657 want to aggregate costs over multiple runs of the same program, or
658 from a single parallel run with multiple instances of the same
662 cg_merge is invoked as follows:
665 <programlisting><![CDATA[
666 cg_merge -o outputfile file1 file2 file3 ...]]></programlisting>
669 It reads and checks <computeroutput>file1</computeroutput>, then read
670 and checks <computeroutput>file2</computeroutput> and merges it into
671 the running totals, then the same with
672 <computeroutput>file3</computeroutput>, etc. The final results are
673 written to <computeroutput>outputfile</computeroutput>, or to standard
674 out if no output file is specified.</para>
677 Costs are summed on a per-function, per-line and per-instruction
678 basis. Because of this, the order in which the input files does not
679 matter, although you should take care to only mention each file once,
680 since any file mentioned twice will be added in twice.</para>
683 cg_merge does not attempt to check
684 that the input files come from runs of the same executable. It will
685 happily merge together profile files from completely unrelated
686 programs. It does however check that the
687 <computeroutput>Events:</computeroutput> lines of all the inputs are
688 identical, so as to ensure that the addition of costs makes sense.
689 For example, it would be nonsensical for it to add a number indicating
690 D1 read references to a number from a different file indicating L2
694 A number of other syntax and sanity checks are done whilst reading the
695 inputs. cg_merge will stop and
696 attempt to print a helpful error message if any of the input files
697 fail these checks.</para>
702 <sect2 id="cg-manual.cg_diff" xreflabel="cg_diff">
703 <title>Differencing Profiles with cg_diff</title>
706 cg_diff is a simple program which
707 reads two profile files, as created by Cachegrind, finds the difference
708 between them, and writes the results into another file in the same format.
709 You can then examine the merged results using
710 <computeroutput>cg_annotate <filename></computeroutput>, as
711 described above. This is very useful if you want to measure how a change to
712 a program affected its performance.
716 cg_diff is invoked as follows:
719 <programlisting><![CDATA[
720 cg_diff file1 file2]]></programlisting>
723 It reads and checks <computeroutput>file1</computeroutput>, then read
724 and checks <computeroutput>file2</computeroutput>, then computes the
725 difference (effectively <computeroutput>file1</computeroutput> -
726 <computeroutput>file2</computeroutput>). The final results are written to
727 standard output.</para>
730 Costs are summed on a per-function basis. Per-line costs are not summed,
731 because doing so is too difficult. For example, consider differencing two
732 profiles, one from a single-file program A, and one from the same program A
733 where a single blank line was inserted at the top of the file. Every single
734 per-line count has changed. In comparison, the per-function counts have not
735 changed. The per-function count differences are still very useful for
736 determining differences between programs. Note that because the result is
737 the difference of two profiles, many of the counts will be negative; this
738 indicates that the counts for the relevant function are fewer in the second
739 version than those in the first version.</para>
742 cg_diff does not attempt to check
743 that the input files come from runs of the same executable. It will
744 happily merge together profile files from completely unrelated
745 programs. It does however check that the
746 <computeroutput>Events:</computeroutput> lines of all the inputs are
747 identical, so as to ensure that the addition of costs makes sense.
748 For example, it would be nonsensical for it to add a number indicating
749 D1 read references to a number from a different file indicating L2
753 A number of other syntax and sanity checks are done whilst reading the
754 inputs. cg_diff will stop and
755 attempt to print a helpful error message if any of the input files
756 fail these checks.</para>
759 Sometimes you will want to compare Cachegrind profiles of two versions of a
760 program that you have sitting side-by-side. For example, you might have
761 <computeroutput>version1/prog.c</computeroutput> and
762 <computeroutput>version2/prog.c</computeroutput>, where the second is
763 slightly different to the first. A straight comparison of the two will not
764 be useful -- because functions are qualified with filenames, a function
765 <function>f</function> will be listed as
766 <computeroutput>version1/prog.c:f</computeroutput> for the first version but
767 <computeroutput>version2/prog.c:f</computeroutput> for the second
771 When this happens, you can use the <option>--mod-filename</option> option.
772 Its argument is a Perl search-and-replace expression that will be applied
773 to all the filenames in both Cachegrind output files. It can be used to
774 remove minor differences in filenames. For example, the option
775 <option>--mod-filename='s/version[0-9]/versionN/'</option> will suffice for
785 <sect1 id="cg-manual.cgopts" xreflabel="Cachegrind Command-line Options">
786 <title>Cachegrind Command-line Options</title>
788 <!-- start of xi:include in the manpage -->
789 <para>Cachegrind-specific options are:</para>
791 <variablelist id="cg.opts.list">
793 <varlistentry id="opt.I1" xreflabel="--I1">
795 <option><![CDATA[--I1=<size>,<associativity>,<line size> ]]></option>
798 <para>Specify the size, associativity and line size of the level 1
799 instruction cache. </para>
803 <varlistentry id="opt.D1" xreflabel="--D1">
805 <option><![CDATA[--D1=<size>,<associativity>,<line size> ]]></option>
808 <para>Specify the size, associativity and line size of the level 1
813 <varlistentry id="opt.L2" xreflabel="--L2">
815 <option><![CDATA[--L2=<size>,<associativity>,<line size> ]]></option>
818 <para>Specify the size, associativity and line size of the level 2
823 <varlistentry id="opt.cache-sim" xreflabel="--cache-sim">
825 <option><![CDATA[--cache-sim=no|yes [yes] ]]></option>
828 <para>Enables or disables collection of cache access and miss
833 <varlistentry id="opt.branch-sim" xreflabel="--branch-sim">
835 <option><![CDATA[--branch-sim=no|yes [no] ]]></option>
838 <para>Enables or disables collection of branch instruction and
839 misprediction counts. By default this is disabled as it
840 slows Cachegrind down by approximately 25%. Note that you
841 cannot specify <option>--cache-sim=no</option>
842 and <option>--branch-sim=no</option>
843 together, as that would leave Cachegrind with no
844 information to collect.</para>
848 <varlistentry id="opt.cachegrind-out-file" xreflabel="--cachegrind-out-file">
850 <option><![CDATA[--cachegrind-out-file=<file> ]]></option>
853 <para>Write the profile data to
854 <computeroutput>file</computeroutput> rather than to the default
856 <filename>cachegrind.out.<pid></filename>. The
857 <option>%p</option> and <option>%q</option> format specifiers
858 can be used to embed the process ID and/or the contents of an
859 environment variable in the name, as is the case for the core
860 option <option><xref linkend="opt.log-file"/></option>.
866 <!-- end of xi:include in the manpage -->
872 <sect1 id="cg-manual.annopts" xreflabel="cg_annotate Command-line Options">
873 <title>cg_annotate Command-line Options</title>
875 <!-- start of xi:include in the manpage -->
876 <variablelist id="cg_annotate.opts.list">
880 <option><![CDATA[-h --help ]]></option>
883 <para>Show the help message.</para>
889 <option><![CDATA[--version ]]></option>
892 <para>Show the version number.</para>
898 <option><![CDATA[--show=A,B,C [default: all, using order in
899 cachegrind.out.<pid>] ]]></option>
902 <para>Specifies which events to show (and the column
903 order). Default is to use all present in the
904 <filename>cachegrind.out.<pid></filename> file (and
905 use the order in the file). Useful if you want to concentrate on, for
906 example, I cache misses (<option>--show=I1mr,I2mr</option>), or data
907 read misses (<option>--show=D1mr,D2mr</option>), or L2 data misses
908 (<option>--show=D2mr,D2mw</option>). Best used in conjunction with
909 <option>--sort</option>.</para>
915 <option><![CDATA[--sort=A,B,C [default: order in
916 cachegrind.out.<pid>] ]]></option>
919 <para>Specifies the events upon which the sorting of the
920 function-by-function entries will be based.</para>
926 <option><![CDATA[--threshold=X [default: 0.1%] ]]></option>
929 <para>Sets the threshold for the function-by-function
930 summary. A function is shown if it accounts for more than X%
931 of the counts for the primary sort event. If auto-annotating, also
932 affects which files are annotated.</para>
934 <para>Note: thresholds can be set for more than one of the
935 events by appending any events for the
936 <option>--sort</option> option with a colon
937 and a number (no spaces, though). E.g. if you want to see
938 each function that covers more than 1% of L2 read misses or 1% of L2
939 write misses, use this option:</para>
940 <para><option>--sort=D2mr:1,D2mw:1</option></para>
946 <option><![CDATA[--auto=<no|yes> [default: no] ]]></option>
949 <para>When enabled, automatically annotates every file that
950 is mentioned in the function-by-function summary that can be
951 found. Also gives a list of those that couldn't be found.</para>
957 <option><![CDATA[--context=N [default: 8] ]]></option>
960 <para>Print N lines of context before and after each
961 annotated line. Avoids printing large sections of source
962 files that were not executed. Use a large number
963 (e.g. 100000) to show all source lines.</para>
969 <option><![CDATA[-I<dir> --include=<dir> [default: none] ]]></option>
972 <para>Adds a directory to the list in which to search for
973 files. Multiple <option>-I</option>/<option>--include</option>
974 options can be given to add multiple directories.</para>
979 <!-- end of xi:include in the manpage -->
984 <sect1 id="cg-manual.diffopts" xreflabel="cg_diff Command-line Options">
985 <title>cg_diff Command-line Options</title>
987 <!-- start of xi:include in the manpage -->
988 <variablelist id="cg_diff.opts.list">
992 <option><![CDATA[-h --help ]]></option>
995 <para>Show the help message.</para>
1001 <option><![CDATA[--version ]]></option>
1004 <para>Show the version number.</para>
1010 <option><![CDATA[--mod-filename=<expr> [default: none]]]></option>
1013 <para>Specifies a Perl search-and-replace expression that is applied
1014 to all filenames. Useful for removing minor differences in paths
1015 between two different versions of a program that are sitting in
1016 different directories.</para>
1021 <!-- end of xi:include in the manpage -->
1028 <sect1 id="cg-manual.acting-on"
1029 xreflabel="Acting on Cachegrind's Information">
1030 <title>Acting on Cachegrind's Information</title>
1032 Cachegrind gives you lots of information, but acting on that information
1033 isn't always easy. Here are some rules of thumb that we have found to be
1037 First of all, the global hit/miss counts and miss rates are not that useful.
1038 If you have multiple programs or multiple runs of a program, comparing the
1039 numbers might identify if any are outliers and worthy of closer
1040 investigation. Otherwise, they're not enough to act on.</para>
1043 The function-by-function counts are more useful to look at, as they pinpoint
1044 which functions are causing large numbers of counts. However, beware that
1045 inlining can make these counts misleading. If a function
1046 <function>f</function> is always inlined, counts will be attributed to the
1047 functions it is inlined into, rather than itself. However, if you look at
1048 the line-by-line annotations for <function>f</function> you'll see the
1049 counts that belong to <function>f</function>. (This is hard to avoid, it's
1050 how the debug info is structured.) So it's worth looking for large numbers
1051 in the line-by-line annotations.</para>
1054 The line-by-line source code annotations are much more useful. In our
1055 experience, the best place to start is by looking at the
1056 <computeroutput>Ir</computeroutput> numbers. They simply measure how many
1057 instructions were executed for each line, and don't include any cache
1058 information, but they can still be very useful for identifying
1062 After that, we have found that L2 misses are typically a much bigger source
1063 of slow-downs than L1 misses. So it's worth looking for any snippets of
1064 code with high <computeroutput>D2mr</computeroutput> or
1065 <computeroutput>D2mw</computeroutput> counts. (You can use
1067 --sort=D2mr</option> with cg_annotate to focus just on
1068 <literal>D2mr</literal> counts, for example.) If you find any, it's still
1069 not always easy to work out how to improve things. You need to have a
1070 reasonable understanding of how caches work, the principles of locality, and
1071 your program's data access patterns. Improving things may require
1072 redesigning a data structure, for example.</para>
1075 Looking at the <computeroutput>Bcm</computeroutput> and
1076 <computeroutput>Bim</computeroutput> misses can also be helpful.
1077 In particular, <computeroutput>Bim</computeroutput> misses are often caused
1078 by <literal>switch</literal> statements, and in some cases these
1079 <literal>switch</literal> statements can be replaced with table-driven code.
1080 For example, you might replace code like this:</para>
1082 <programlisting><![CDATA[
1093 ]]></programlisting>
1095 <para>with code like this:</para>
1097 <programlisting><![CDATA[
1100 enum E table[] = { 1, 2, 3 };
1104 ]]></programlisting>
1107 This is obviously a contrived example, but the basic principle applies in a
1108 wide variety of situations.</para>
1111 In short, Cachegrind can tell you where some of the bottlenecks in your code
1112 are, but it can't tell you how to fix them. You have to work that out for
1113 yourself. But at least you have the information!
1119 <sect1 id="cg-manual.sim-details"
1120 xreflabel="Simulation Details">
1121 <title>Simulation Details</title>
1123 This section talks about details you don't need to know about in order to
1124 use Cachegrind, but may be of interest to some people.
1127 <sect2 id="cache-sim" xreflabel="Cache Simulation Specifics">
1128 <title>Cache Simulation Specifics</title>
1130 <para>Specific characteristics of the cache simulation are as
1136 <para>Write-allocate: when a write miss occurs, the block
1137 written to is brought into the D1 cache. Most modern caches
1138 have this property.</para>
1142 <para>Bit-selection hash function: the set of line(s) in the cache
1143 to which a memory block maps is chosen by the middle bits
1144 M--(M+N-1) of the byte address, where:</para>
1147 <para>line size = 2^M bytes</para>
1150 <para>(cache size / line size / associativity) = 2^N bytes</para>
1156 <para>Inclusive L2 cache: the L2 cache typically replicates all
1157 the entries of the L1 caches, because fetching into L1 involves
1158 fetching into L2 first (this does not guarantee strict inclusiveness,
1159 as lines evicted from L2 still could reside in L1). This is
1160 standard on Pentium chips, but AMD Opterons, Athlons and Durons
1161 use an exclusive L2 cache that only holds
1162 blocks evicted from L1. Ditto most modern VIA CPUs.</para>
1167 <para>The cache configuration simulated (cache size,
1168 associativity and line size) is determined automatically using
1169 the x86 CPUID instruction. If you have a machine that (a)
1170 doesn't support the CPUID instruction, or (b) supports it in an
1171 early incarnation that doesn't give any cache information, then
1172 Cachegrind will fall back to using a default configuration (that
1173 of a model 3/4 Athlon). Cachegrind will tell you if this
1174 happens. You can manually specify one, two or all three levels
1175 (I1/D1/L2) of the cache from the command line using the
1176 <option>--I1</option>,
1177 <option>--D1</option> and
1178 <option>--L2</option> options.
1179 For cache parameters to be valid for simulation, the number
1180 of sets (with associativity being the number of cache lines in
1181 each set) has to be a power of two.</para>
1183 <para>On PowerPC platforms
1184 Cachegrind cannot automatically
1185 determine the cache configuration, so you will
1186 need to specify it with the
1187 <option>--I1</option>,
1188 <option>--D1</option> and
1189 <option>--L2</option> options.</para>
1192 <para>Other noteworthy behaviour:</para>
1196 <para>References that straddle two cache lines are treated as
1200 <para>If both blocks hit --> counted as one hit</para>
1203 <para>If one block hits, the other misses --> counted
1207 <para>If both blocks miss --> counted as one miss (not
1214 <para>Instructions that modify a memory location
1215 (e.g. <computeroutput>inc</computeroutput> and
1216 <computeroutput>dec</computeroutput>) are counted as doing
1217 just a read, i.e. a single data reference. This may seem
1218 strange, but since the write can never cause a miss (the read
1219 guarantees the block is in the cache) it's not very
1222 <para>Thus it measures not the number of times the data cache
1223 is accessed, but the number of times a data cache miss could
1229 <para>If you are interested in simulating a cache with different
1230 properties, it is not particularly hard to write your own cache
1231 simulator, or to modify the existing ones in
1232 <computeroutput>cg_sim.c</computeroutput>. We'd be
1233 interested to hear from anyone who does.</para>
1238 <sect2 id="branch-sim" xreflabel="Branch Simulation Specifics">
1239 <title>Branch Simulation Specifics</title>
1241 <para>Cachegrind simulates branch predictors intended to be
1242 typical of mainstream desktop/server processors of around 2004.</para>
1244 <para>Conditional branches are predicted using an array of 16384 2-bit
1245 saturating counters. The array index used for a branch instruction is
1246 computed partly from the low-order bits of the branch instruction's
1247 address and partly using the taken/not-taken behaviour of the last few
1248 conditional branches. As a result the predictions for any specific
1249 branch depend both on its own history and the behaviour of previous
1250 branches. This is a standard technique for improving prediction
1253 <para>For indirect branches (that is, jumps to unknown destinations)
1254 Cachegrind uses a simple branch target address predictor. Targets are
1255 predicted using an array of 512 entries indexed by the low order 9
1256 bits of the branch instruction's address. Each branch is predicted to
1257 jump to the same address it did last time. Any other behaviour causes
1258 a mispredict.</para>
1260 <para>More recent processors have better branch predictors, in
1261 particular better indirect branch predictors. Cachegrind's predictor
1262 design is deliberately conservative so as to be representative of the
1263 large installed base of processors which pre-date widespread
1264 deployment of more sophisticated indirect branch predictors. In
1265 particular, late model Pentium 4s (Prescott), Pentium M, Core and Core
1266 2 have more sophisticated indirect branch predictors than modelled by
1269 <para>Cachegrind does not simulate a return stack predictor. It
1270 assumes that processors perfectly predict function return addresses,
1271 an assumption which is probably close to being true.</para>
1273 <para>See Hennessy and Patterson's classic text "Computer
1274 Architecture: A Quantitative Approach", 4th edition (2007), Section
1275 2.3 (pages 80-89) for background on modern branch predictors.</para>
1279 <sect2 id="cg-manual.annopts.accuracy" xreflabel="Accuracy">
1280 <title>Accuracy</title>
1282 <para>Valgrind's cache profiling has a number of
1283 shortcomings:</para>
1287 <para>It doesn't account for kernel activity -- the effect of system
1288 calls on the cache and branch predictor contents is ignored.</para>
1292 <para>It doesn't account for other process activity.
1293 This is probably desirable when considering a single
1298 <para>It doesn't account for virtual-to-physical address
1299 mappings. Hence the simulation is not a true
1300 representation of what's happening in the
1301 cache. Most caches and branch predictors are physically indexed, but
1302 Cachegrind simulates caches using virtual addresses.</para>
1306 <para>It doesn't account for cache misses not visible at the
1307 instruction level, e.g. those arising from TLB misses, or
1308 speculative execution.</para>
1312 <para>Valgrind will schedule
1313 threads differently from how they would be when running natively.
1314 This could warp the results for threaded programs.</para>
1318 <para>The x86/amd64 instructions <computeroutput>bts</computeroutput>,
1319 <computeroutput>btr</computeroutput> and
1320 <computeroutput>btc</computeroutput> will incorrectly be
1321 counted as doing a data read if both the arguments are
1322 registers, eg:</para>
1323 <programlisting><![CDATA[
1324 btsl %eax, %edx]]></programlisting>
1326 <para>This should only happen rarely.</para>
1330 <para>x86/amd64 FPU instructions with data sizes of 28 and 108 bytes
1331 (e.g. <computeroutput>fsave</computeroutput>) are treated as
1332 though they only access 16 bytes. These instructions seem to
1333 be rare so hopefully this won't affect accuracy much.</para>
1338 <para>Another thing worth noting is that results are very sensitive.
1339 Changing the size of the the executable being profiled, or the sizes
1340 of any of the shared libraries it uses, or even the length of their
1341 file names, can perturb the results. Variations will be small, but
1342 don't expect perfectly repeatable results if your program changes at
1345 <para>More recent GNU/Linux distributions do address space
1346 randomisation, in which identical runs of the same program have their
1347 shared libraries loaded at different locations, as a security measure.
1348 This also perturbs the results.</para>
1350 <para>While these factors mean you shouldn't trust the results to
1351 be super-accurate, they should be close enough to be useful.</para>
1359 <sect1 id="cg-manual.impl-details"
1360 xreflabel="Implementation Details">
1361 <title>Implementation Details</title>
1363 This section talks about details you don't need to know about in order to
1364 use Cachegrind, but may be of interest to some people.
1367 <sect2 id="cg-manual.impl-details.how-cg-works"
1368 xreflabel="How Cachegrind Works">
1369 <title>How Cachegrind Works</title>
1370 <para>The best reference for understanding how Cachegrind works is chapter 3 of
1371 "Dynamic Binary Analysis and Instrumentation", by Nicholas Nethercote. It
1372 is available on the <ulink url="&vg-pubs-url;">Valgrind publications
1373 page</ulink>.</para>
1376 <sect2 id="cg-manual.impl-details.file-format"
1377 xreflabel="Cachegrind Output File Format">
1378 <title>Cachegrind Output File Format</title>
1379 <para>The file format is fairly straightforward, basically giving the
1380 cost centre for every line, grouped by files and
1381 functions. It's also totally generic and self-describing, in the sense that
1382 it can be used for any events that can be counted on a line-by-line basis,
1383 not just cache and branch predictor events. For example, earlier versions
1384 of Cachegrind didn't have a branch predictor simulation. When this was
1385 added, the file format didn't need to change at all. So the format (and
1386 consequently, cg_annotate) could be used by other tools.</para>
1388 <para>The file format:</para>
1389 <programlisting><![CDATA[
1390 file ::= desc_line* cmd_line events_line data_line+ summary_line
1391 desc_line ::= "desc:" ws? non_nl_string
1392 cmd_line ::= "cmd:" ws? cmd
1393 events_line ::= "events:" ws? (event ws)+
1394 data_line ::= file_line | fn_line | count_line
1395 file_line ::= "fl=" filename
1396 fn_line ::= "fn=" fn_name
1397 count_line ::= line_num ws? (count ws)+
1398 summary_line ::= "summary:" ws? (count ws)+
1399 count ::= num | "."]]></programlisting>
1404 <para><computeroutput>non_nl_string</computeroutput> is any
1405 string not containing a newline.</para>
1408 <para><computeroutput>cmd</computeroutput> is a string holding the
1409 command line of the profiled program.</para>
1412 <para><computeroutput>event</computeroutput> is a string containing
1413 no whitespace.</para>
1416 <para><computeroutput>filename</computeroutput> and
1417 <computeroutput>fn_name</computeroutput> are strings.</para>
1420 <para><computeroutput>num</computeroutput> and
1421 <computeroutput>line_num</computeroutput> are decimal
1425 <para><computeroutput>ws</computeroutput> is whitespace.</para>
1429 <para>The contents of the "desc:" lines are printed out at the top
1430 of the summary. This is a generic way of providing simulation
1431 specific information, e.g. for giving the cache configuration for
1432 cache simulation.</para>
1434 <para>More than one line of info can be presented for each file/fn/line number.
1435 In such cases, the counts for the named events will be accumulated.</para>
1437 <para>Counts can be "." to represent zero. This makes the files easier for
1438 humans to read.</para>
1440 <para>The number of counts in each
1441 <computeroutput>line</computeroutput> and the
1442 <computeroutput>summary_line</computeroutput> should not exceed
1443 the number of events in the
1444 <computeroutput>event_line</computeroutput>. If the number in
1445 each <computeroutput>line</computeroutput> is less, cg_annotate
1446 treats those missing as though they were a "." entry. This saves space.
1449 <para>A <computeroutput>file_line</computeroutput> changes the
1450 current file name. A <computeroutput>fn_line</computeroutput>
1451 changes the current function name. A
1452 <computeroutput>count_line</computeroutput> contains counts that
1453 pertain to the current filename/fn_name. A "fn="
1454 <computeroutput>file_line</computeroutput> and a
1455 <computeroutput>fn_line</computeroutput> must appear before any
1456 <computeroutput>count_line</computeroutput>s to give the context
1457 of the first <computeroutput>count_line</computeroutput>s.</para>
1459 <para>Each <computeroutput>file_line</computeroutput> will normally be
1460 immediately followed by a <computeroutput>fn_line</computeroutput>. But it
1461 doesn't have to be.</para>
1463 <para>The summary line is redundant, because it just holds the total counts
1464 for each event. But this serves as a useful sanity check of the data; if
1465 the totals for each event don't match the summary line, something has gone