File: HACKING

package info (click to toggle)
libdevel-nytprof-perl 4.04-1
  • links: PTS, VCS
  • area: main
  • in suites: squeeze
  • size: 2,124 kB
  • ctags: 1,330
  • sloc: perl: 4,392; ansic: 92; makefile: 19
file content (391 lines) | stat: -rw-r--r-- 17,544 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
# vim: ts=8 sw=2 sts=0 noexpandtab:
# $Id: HACKING 1329 2010-07-08 14:10:06Z tim.bunce@gmail.com $

HACKING Devel::NYTProf
======================

We encourage hacking Devel::NYTProf!

OBTAINING THE CURRENT RELEASE
-----------------------------
The current official release can be obtained from CPAN
http://search.cpan.org/dist/Devel-NYTProf/

OBTAINING THE LATEST DEVELOPMENT CODE
-------------------------------------
You can grab the head of the latest trunk code from the Google Code repository, see
http://code.google.com/p/perl-devel-nytprof/source/checkout

CONTRIBUTING
------------
Please work with the latest code from the repository - see above.

Small patches can be uploaded via the issue tracker at
http://code.google.com/p/perl-devel-nytprof/issues/list

For larger changes please talk to us first via the mailing list at
http://code.google.com/p/perl-devel-nytprof/source/checkout

When developing, please ensure that no new compiler warnings are output.

TESTING
-------
You MUST write test cases for your changes. All tests that are dropped into the
"t" folder will be executed. (Remember to add them to MANIFEST.)  The testing system is
customized for this module because profilers are not that easy to test.
The system still uses Test::Harness and Test::More, so it should behave just
like any other perl modules 'make test'.

Writing tests is easy!

1) Design a perl script that will trigger the new behavior/feature that you
   want to test. Name the file 't/test##-description.p'

2) Create an empty 'reference' file for the test.
   Name the file 't/test##-description.rdt'
   When the test is run you'll get an error and a diff and you'll
   find a t/test##-description.rdt.new file waiting for you.
   If, and only if, the contents of that file are correct, then rename
   it to t/test##-description.rdt and you're done!
   Of course working out if the contents are correct can be
   non-trivial, but at least you don't have to write the file :)

3) Create a corresponding CSV output file if appropriate.
   You can use the same trick of creating an empty file, but this
   time with a .x suffix: t/test##-description.x
   You still need to verify the .x.new file of course!

4) Create a test script like this:

     use strict;
     use Test::More;
     use lib qw(t/lib);
     use NYTProfTest;

     run_test_group;

   You can add additional tests as parameters to run_test_group:

     run_test_group(2 => sub {
         my ($profile, $env) = @_;
         is $profile->foo, 'bar', "...";
         is $profile->baz, 'bax', "...";
     });


Note:  While writing a test, it is helpful to be able to run it directly, 
without the test harness.  This allows you to view more output stdout and 
stderr.  Fortunately, its easy to do:

  perl -Mblib -MDevel::NYTProf t/test01.p

The output will be in the ./nytprof.out file.  You can then also run the
csv manually:

  perl -Mblib bin/nytprofcsv

The final file will be in ./nytprof/test01.p.csv

Remember, testing is VERY VERY important!  Within a day or two of releasing
code, the CPAN testers will test the release on pretty much every major platform
you can think of.  A failed test report is much easier to fix than a runtime
error like "bash: segmentation fault: core dumped"

GENERATING DISTRIBUTIONS
------------------------
Releases are generated with 'make metafile', and then fed through tar+gz.
You shouldn't ever check-in the distribution directory, any temporary files
(including Makefile.old) or change the $VERSION numbers. We'll do this for you.

RESOURCES
---------
Google Code:
http://code.google.com/p/perl-devel-nytprof/

Google Devel Group (must subscribe here):
http://groups.google.com/group/develnytprof-dev

NYTimes Open Code Blog:
http://open.nytimes.com/

TODO (unsorted, unprioritized, unconsidered, even unreasonable and daft :)
----

*** For build/test

Add (very) basic nytprofhtml test (ie it runs and produces output)

Add tests for evals in regex: s/.../ ...perl code... /e

Add tests for -block and -sub csv reports.

Add tests with various kinds of blocks and loops (if, do, while, until, etc).

Add mechanism to specify options inside the .p file, such as
  # NYTPROF=...
though this may not be needed if t/20.runtests.t gets dropped
and the logic moved to a library for traditional t/*.t files to use.

Add mechanism to specify inside the .p file that NYTProf
should not be loaded via the command line. That's needed to test
behaviors in environments where perl is init'd first. Such as mod_perl.
Then we can test things like not having the sub line range for some subs.

*** For core only

See MemoryProfiling.pod file

Store raw NYTPROF option string in the data file. 
Include parsed version in report index page.

Add actual size and mtime of fid to data file. (Already in data file as zero,
just needs the stat() call.) Don't alter errno.

Add help option which would print a summary of the options and exit.
Could also print list of available clocks for the clock=N option
(using a set of #ifdef's)

The subroutine profiler could calculate the running variance of the samples
using this logic http://www.johndcook.com/standard_deviation.html
so the reports can display the standard deviation.

Replace DB::enable_profiling() and DB::disable_profiling() with $DB::profile = 1|0;
That a more consistent API with $DB::single etc., but more importantly it lets
users leave the code in place when NYTProf is not loaded. It'll just do nothing,
whereas currently the user will get a fatal error if NYTProf isn't loaded.
It also allows smart things like use of local() for temporary overrides.

Combine current profile_* globals into a single global int using bit fields.
That way assigning to $DB::profile can offer a finer degree of control.
Specifically to enable/disable the sub or statement profiler separately.

Add mechanism to enable control of profiling on a per-sub-name and/or
per-package-name basis. For example, specify a regex and whenever a sub is
entered (for the first time, to make it cheap) check if the sub name matches
the regex. If it does then save the current $DB::profile value and set a new one.
When the sub exits restore the previous $DB::profile value.

Work around OP_UNSTACK bug (http://rt.perl.org/rt3/Ticket/Display.html?id=60954)
  while ( foo() ) {  # all calls to foo should be from here
      ...
      ... # no calls to foo() should appear here
  }

*** For core and reports

Add @INC to data file so reports can be made more readable by removing
(possibly very long) library paths where appropriate.
Tricky thing is that @INC can change during the life of the program.
One approach might be to output it whenever we assign a new fid
but only if different to the last @INC that was ouput.

Add marker with timestamp for phases BEGIN, CHECK, INIT, END
(could combine with pid marker)
Add marker with timestamp for enable_profile and disable_profile.
Could also dump-and-zero the sub profiler data so we could report per-phase timing.
The goals here are to
a) know how long the different phases of execution took mostly for general interest, and
b) know how much time was spent with the profiler enabled to calculate accurate
percentages and also be able to spot 'leaks' in the data processing (e.g. if
the sum of the statement times don't match the time spent with the profiler
enabled, due to nested string evals for example).

*** For reports only

::Reader and its data structures need to be refactored to death.
The whole reporting framework needs a rewrite to use a single 'thin' command
line and classes for the Model (lines, files, subs), View (html, csv etc),
and Controller (composing views to form reports).
Dependent on a richer data model.

Then rework bin/ntyprof* to use the new subclasses
Ideally end up with a single nytprof command that just sets up the appropriate
classes to do the work.

Trim leading @INC portion from filename in __ANON__[/very/long/path/...]
in report output. (Keep full path in link/tooltip/title as it may be ambiguous when shortened).

Add help link in reports. Could go to docs page on search.cpan.org.

Add a 'permalink' icon (eg infinity symbol) to the right of lines that define
subs to make it easer to email/IM links to particular places in the code.

Report could track which subs it has reported caller info for
and so be able to identify subs that were called but haven't been included
in the report because we didn't know where the sub was.
They could them be included in a separate 'miscellaneous' page.
This is a more general way to view the problem of xsubs in packages
for which we don't have any perl source code.

*** Other - mostly unsorted - stuff ***

Intercept all opcodes that may fork and run perl code in the child
  ie fork, open, entersub (ie xs), others?
  and fflush before executing the op (so fpurge isn't strictly required)
  and reinit_if_forked() afterwards
  add option to force reinit_if_forked check per stmt just-in-case
Alternatively it might be better to use pthread_atfork() [if available] with a
child handler. The man page says "Remember: only async-cancel-safe functions
are allowed on the child side of fork()" so it seems that the safe thing to do
is to use a volatile flag variable, and change its value in the handler to
signal to the main code.

Support profiling programs which use threads:
  - move all relevant globals into a structure
  - add lock around output to file

Set options via import so perl -d:NYTProf=... works. Very handy. May need
alternative option syntax. Also perl gives special meaning to 't' option
(threads) so we should reserve the same for eventual thread support.
Problem with this is that the import() call happens after init_profiler()
so limits the usefulness. So we'd need to limit it to certain options
(trace would certainly be useful).

Add resolution of __ANON__ sub names (eg imported 'constants') where possible.
[I can't recall what I meant by that now. I think this means where an anon sub
has been imported, if the refcnt is 1 then use the imported name instead of the
__ANON__ name.]

The appending of an @line to BEGIN subs should also be applied to END subs.

Record $AUTOLOAD when AUTOLOAD() called. Perhaps as ...::AUTOLOAD[$AUTOLOAD]
Or perhaps just use the original name if the 'resolved' one is AUTOLOAD.
Could be argued either way.

More generally, consider the problem of code where one code path is fast 
and just sets $sql = ... (for example) and another code path executes the
sql. Some $sql may be fast and others slow. The profile can't separate the
timings based on what was in $sql because the code path was the same in both
cases. (For sql DBI::Profile can be used, but the underlying issue is general.)

The sub_caller information is currently one level deep. It would be good to
make it two levels. Especially because it would allow you to "see through"
AUTOLOADs and other kinds of 'dispatch' subs.

Refactor this HACKING file!

The data file includes the information mapping a line-level line to the
corresponding block-level and sub-level lines. This should be added to the data
structure. It would enable a much richer visualization of which lines have
contributed to the 'rolled up' counts. That's especially tricky to work out
with the block level view.

Following on from that I have a totally crazy idea that the browsers css engine
could be used to highlight the corresponding rollup line when hovering over a
source line, and/or the opposite. Needs lots of thought, but it's an interesting idea.

Profile and optimize report generation

Bug or limitation?: sub calls in a continue { ... } block of a while () get
associated with the 'next;' within the loop. Fixed by perl change 33710?

Investigate style.css problem when using --outfile=some/other/dir

Sub profiler should avoid sv_setpvf(subname_sv, "%s::%s", stash_name, GvNAME(gv));
because it's expensive (Perl_sv_setpvf_nocontext accounts for 29% of pp_entersub_profiler).
Use a two level hash: HvNAME(GvSTASH(gv)) then GvNAME(gv).
Should then also be able to avoid newSV/free for subname_sv (which accounts for 50% of its time).

Class::MOP should update %DB::sub (if $^P & 0x10 set) when it creates methods.
Sub::Name should do same (extracting the file and line from the ANON[...:...])

Add refs so a string eval fid can be related to its 'siblings' (other string
eval fids from the same line in the 'parent' fid).

Profile should report _both_ the 'raw original' filename (possibly relative)
used by the application being profiled, plus an absolute filename determined
ASAP (to avoid problems with scripts that chdir).

Add (very) basic nytprofhtml test (ie it runs and produces output) so we check the VERSION has been updated.

In the called by list in html: "by $subname line $line of $file"
make the file not include the @INC portion

Monitor and report when method cache is invalidated. Watch generation number
and output a tag when it changes. Report locations of changes. Highlight those
that happen after INIT phase.

Fix testing of t/*.pm_x files which are currently being ignored.

The autosplit handling doesn't address the naming of pseudo-fids from string evals
inside autoloaded subs, like "(eval 0)[test14.pm (autosplit into auto/test14/bar.al):17]"
The 'file name' for the eval fid needs to be edited when read in to remove the
' (autosplit...', but care should be taken to not remove the text for evals in
autosplit files for which we've not been able to alias to the parent.

Add a FID_ATTRIB tag to allow additional info about fids to be generated after
the initial fid info is output.

Use FID_ATTRIB tag to record autoload fids being aliases to a fid so that
reports can include a list of autoloaded subs.

Check if pp_leavegiven and pp_leavewhen need handling in init_profiler().

Copy the nytprof.out file into the output report dir, so a report is more
'self-contained' and can be archived and thrown around as a tarball/zip and
still used for further analysis.

To stress-test NYTProf using perl's own test suite, set env vars:
  NYTPROF='file=/tmp/nytprof.out:addpid=1:nameanonsubs=0:nameevals=0'
  PERL5OPT='-d:NYTProf'
and hack t/TEST to not delete the PERL5OPT env var.

The findcaller option doesn't notice if the caller is an opcode.
Opcodes that call subs (like subst calling utf8::SWASHGET) probably shouldn't
appear as the caller in the call-tree because, while strictly accurate, it's
probably less useful from the users perspective.
Fixing that part is easy but handling incl/excl time needs more thought.

For xsubs and opcodes that call perl subs:
In the subroutine prologue that currently lists where the sub was called from,
for xsubs & opcodes, add a list of subs that it called (typically none).
That would be handy because currently calls from xsubs & opcodes appear in the
reports at the line of the last _perl_ statement executed, and not in the fake
'stub' that we add to the end of the package source.

Use gethrtime() on platforms that support it.
http://developers.sun.com/solaris/articles/time_stamp.html
http://www.informit.com/guides/content.aspx?g=cplusplus&seqNum=332

Currently lvalue subs aren't profiled when use_db_sub is in effect.
http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2010-02/msg00824.html

Idle conjecture: perhaps loading the profile data into an SQLite database
would be a better approach to reporting.
(An nytprofimport utility could read an nytprof.out and write an nytprof.db)
Then, rather than hand-crafting inflexible data structures we could use SQL
(or, say, DBIx::Class) to return relevant data. Would also provide another
route for other data sources (languages/tools) to be included in a report.
Could also simplify: comparing profiles, merging profiles, ...
Also 'pipelines' of plugin db-to-db transformations could be developed.
Any volunteers to explore writing an nytprofimport for SQLite?

Option to add sub call and return events into the data file as they happen.
Would enable a dprofpp -T like output. See https://rt.cpan.org/Ticket/Display.html?id=50766

String evals could/should be tied into the subroutine profiler.
That would give inclusive timings which we don't have at the moment.
The evals appear in the html report as if they're calls but the timings are
statement timings of the eval fid that don't include time spent in subs defined
elsewhere but called from the string eval. The inconsistency is confusing.

Modify csv output to optionally include extra metadata lines. Use for testing.

In html report where sub callers are shown (called N times by S from F at line L)
also show in compact form links to the callers of the caller, if there aren't too many.
This makes it faster to climb the call stack. A simple mouseover will show the
filename of the caller (perhaps a tooltip could give the file and sub).
Also needed for modules that take shortcuts calling XS functions (*cough DBI*).
And/or, add a simple up arrow that'll just to the calling sub. Typically that'll
be the surrounding "sub foo { ..." (which'll be handy for big subs) but for
cases where the call to that sub wasn't recorded (eg pre 5.8.9 or the DBI's
xs calling optimization) it'll be the most recent sub entry that was recorded.

Generate extra report pages for xsubs in packages that don't have source code.
They're currently all dumped into the 'main' file.

Docs describing how the subroutine profiler works need updating.
Add 'u' key to treemap to trigger moving 'up' a level.
Add "calls N subs" to treemap mouseover box
Upgrade treemap to JIT version 2 (which has transition animations).