Purify (tm) and Quantify (tm) are commercial software quality

assurance tools available from IBM <http://www.ibm.com/software/rational/>.

Purify is essentially a memory access

verifier and leak detector; Quantify is a C level profiler.  The rest

of this file assumes you generally know how to use Purify and

Quantify, and that you have installed valid licenses for these

products.  If you haven't installed such licenses, you can ignore the

following since it won't help you a bit!

You can easily build a Purify or Quantify instrumented version of the

Python interpreter by passing the PURIFY variable to the make command

at the top of the Python tree:

    make PURIFY=purify

This assumes that the `purify' program is on your $PATH.  Note that

you cannot both Purify and Quantify the Python interpreter (or any

program for that matter) at the same time.  If you want to build a

Quantify'd interpreter, do this:

    make PURIFY=quantify

Starting with Python 2.3, pymalloc is enabled by default.  This

will cause many supurious warnings.  Modify Objects/obmalloc.c

and enable Py_USING_MEMORY_DEBUGGER by uncommenting it.

README.valgrind has more details about why this is necessary.

See below about setting up suppressions.  Some tests may not

run well with Purify due to heavy memory or CPU usage.  These

tests may include: test_largefile, test_import, and test_long.

Please report any findings (problems or no warnings) to python-dev@python.org.

It may be useful to submit a bug report for any problems.  

When running the regression test (make test), I have found it useful

to set my PURIFYOPTIONS environment variable using the following

(bash) shell function.  Check out the Purify documentation for

details:

p() {

  chainlen='-chain-length=12'

  ignoresigs='-ignore-signals="SIGHUP,SIGINT,SIGQUIT,SIGILL,SIGTRAP,SIGAVRT,SIGEMT,SIGFPE,SIGKILL,SIGBUS,SIGSEGV,SIGPIPE,SIGTERM,SIGUSR1,SIGUSR2,SIGPOLL,SIGXCPU,SIGXFSZ,SIGFREEZE,SIGTHAW,SIGRTMIN,SIGRTMAX"'

  followchild='-follow-child-processes=yes'

  threads='-max-threads=50'

  export PURIFYOPTIONS="$chainlen $ignoresigs $followchild $threads"

  echo $PURIFYOPTIONS

}

Note that you may want to crank -chain-length up even further.  A

value of 20 should get you the entire stack up into the Python C code

in all situations.

With the regression test on a fatly configured interpreter

(i.e. including as many modules as possible in your Modules/Setup

file), you'll probably get a gabillion UMR errors, and a few MLK

errors.  I think most of these can be safely suppressed by putting the

following in your .purify file:

    suppress umr ...; "socketmodule.c"

    suppress umr ...; time_strftime

    suppress umr ...; "dbmmodule.c"

    suppress umr ...; "gdbmmodule.c"

    suppress umr ...; "grpmodule.c"

    suppress umr ...; "nismodule.c"

    suppress umr ...; "pwdmodule.c"

Note: this list is very old and may not be accurate any longer.

It's possible some of these no longer need to be suppressed.

You will also need to suppress warnings (at least umr)

from Py_ADDRESS_IN_RANGE.

This will still leave you with just a few UMR, mostly in the readline

library, which you can safely ignore.  A lot of work has gone into

Python 1.5 to plug as many leaks as possible.

Using Purify or Quantify in this way will give you coarse grained

reports on the whole Python interpreter.  You can actually get more

fine grained control over both by linking with the optional `pure'

module, which exports (most of) the Purify and Quantify C API's into

Python.  To link in this module (it must be statically linked), edit

your Modules/Setup file for your site, and rebuild the interpreter.

You might want to check out the comments in the Modules/puremodule.c

file for some idiosyncrasies.

Using this module, you can actually profile or leak test a small

section of code, instead of the whole interpreter.  Using this in

conjuction with pdb.py, dbx, or the profiler.py module really gives

you quite a bit of introspective power.

Naturally there are a couple of caveats.  This has only been tested

with Purify 4.0.1 and Quantify 2.1-beta on Solaris 2.5.  Purify 4.0.1

does not work with Solaris 2.6, but Purify 4.1 which reportedly will,

is currently in beta test.  There are funky problems when Purify'ing a 

Python interpreter build with threads.  I've had a lot of problems

getting this to work, so I generally don't build with threads when I'm 

Purify'ing.  If you get this to work, let us know!

-Barry Warsaw <bwarsaw@cnri.reston.va.us>