TITLE

Using optimizer to do Method Munging. (as done in Log::Log4perl::AutoCategorize)

Munging (from C++)

A Term borrowed from C++; it disambiguates same-named functions in different classes, or with different argument or return types. It allowed a C++ front-end to use a C compiler back-end.

 Example:

 public class A {
	int foo();
	int foo(int);
 }
 public class B : public A {
	String foo();
	String foo(String);
 }

yielding something like: (its internal to compile/link so doesnt matter)

foo_A_void;
foo_A_int;
foo_B_void;
foo_B_String;

The reasons we do it are different, but the operation is similar; it 'alters' the symbol table.

Reason to Munge

At compile time, we want to convert all instances of:

    FROM: {
	Logger->debug ("msg A");
	Logger->debug ("msg B");
    }
    TO:	{
	Logger->debug_001 ("msg A");
	Logger->debug_002 ("msg B");
    }

Each invocation gets its own UNIQUE function-name (an arbitrary one)

Why?

So we can create custom handlers for each.. (hold this thought)

Motivating Application: logging with Log::Log4perl

Log::Log4perl provides a powerful, flexible way to redirect and filter logging via config-file, and the 'category' that the programmer writes into his classes.

Its problem is that using the power requires that the programmer be careful and diligent with his categorization used in his logging.

Log::Log4perl powerful example

    # load a config-string (or file)
    Log::Log4perl->init(\ qq{
           log4perl.category.Candy.Twix = DEBUG, Screen
           log4perl.category.Candy.Snickers = INFO, Screen
           });

    # now use that logger
    package Candy::Twix;

    sub new {
	my $logger = Log::Log4perl->new("Candy::Twix::new");
	$logger->debug("Creating a new Twix bar");
	bless {}, shift;
    }

Notes about above example

new() has its own category, implying that each method would also have its own. This represents a coding burden on the client. The category must be manually added to each method, even though its known to perl, via caller().

You're wise to keep category consistent with (ie the same as) the package and method.

The config-string has a logging category for Bar::Twix only, not for each of its methods. This example doesnt do detailed control of logging, but its possible later (assuming a config-file) because methods use distinct categories.

Log::Log4perl qw(:easy);

In easy mode, Log4perl sets $category = __PACKAGE__

No work for the programmer, but gives a lot less control over what gets logged.

Sorry Virginia, there is no __SUBROUTINE__.

You can have both ease and power, with Log::Log4perl::AutoCategorize

AutoCategorize is a wrapper on Log::Log4perl, and uses its log-config and filtering mechanisms. However, it extends the $category, giving the filters more detailed info to act upon..

In Log4perl ':easy':

$category = "$package";

In Log::Log4perl::AutoCategorize:

$category = "$package.$subroutine.$loggingLevel.$lineNumber";

Now filtering, as specified in your log-config, can tease out every invocation individually.

OK, wheres optimizer come into it ?

In Log4perl, every time $logger->debug() is called, it must decide: what to do with the message?

I made it worse; now $category includes info that only caller() can provide. This adds significant runtime cost (esp for an application-wide logger).

However.. by differentiating each caller with a unique method name, the name can be associated with the log / dont-log choice, and 'remembered' by a custom subroutine, which is added to the symbol table.

The next time that munged method (ex: debug_00001) is called, AutoCategorize::AUTOLOAD is uninvolved; the recently vivified method is dispatched directly.

This actually executes faster than Log4perl, and up-front munging cost is quickly amortized, though at cost of more memory for code.

Youre still using caller() at runtime. Why ?

To compute the $category, caller() must be looking at the real call-stack. At compile-time, the only thing on the stack (besides the optimizer callback) is an __ANON__ block.

Since we cant determine a category at runtime, the munge is arbitrary, and _00001++ is as good as any.

Shameless plug

Since Log::Log4perl::AutoCategorize knows all logging calls, and knows which ones were actually invoked, it gives you a detailed coverage report for free.

[jimc@harpo t]$ mroe out.cover.tconf_file 
(19:44:06.204) Log.Log4perl.AutoCategorize.END.info.308: Seen Log Events:, {
 'Log.Log4perl.AutoCategorize.END.info.308' => 1,
 'log4perl.category.A.truck.debug.63' => 20,
 'log4perl.category.A.truck.debug.65' => 20,
 'log4perl.category.A.truck.debug.66' => 20,
 'log4perl.category.A.truck.warn.62' => 20,
 'log4perl.category.main.car.warn.46' => 10,
 'log4perl.category.main.main.info.28' => 5,
 'log4perl.category.main.main.warn.27' => 5,
 'log4perl.category.main.suv.warn.51' => '-10',
 'log4perl.category.main.suv.warn.52' => '-10'
}
(19:44:06.207) Log.Log4perl.AutoCategorize.END.info.308: un-Seen Log Events:, {
 'info_00011' => 'main,tconf_file.pl,36',
 'warn_00008' => 'B::C,tconf_file.pl,76'
}
(19:44:06.210) Log.Log4perl.AutoCategorize.END.info.308: cat2data:, {
 'A.truck.debug.63' => 'debug_00005',
 'A.truck.debug.65' => 'debug_00006',
 'A.truck.debug.66' => 'debug_1_00007',
 'A.truck.warn.62' => 'warn_00004',
 'main.car.warn.46' => 'warn_00001',
 'main.main.info.28' => 'info_00010',
 'main.main.warn.27' => 'warn_00009',
 'main.suv.warn.51' => 'warn_00002',
 'main.suv.warn.52' => 'warn_00003'
}

As a convenience, it also dumps ref-data args.

Notes on Coverage Report

%Seen counts number of times the call was made, positive numbers mean the logging was enabled, negative means suppressed.

%UnSeen calls were never reached, and may represent testing inadequacies.

I eat my own dog-food; the same logging mechanisms issue a report from an END block, and are re-directable using any Log4perl means.

Conclusions

optimizer is BADDASS, this is a fairly trivial usage.

I had a couple surprises, particularly with B::NULL ops. These are probably remnants of optimization done before I get called.

My optimzer callback recognizes and alters this opcode chain:

perl -MO=Concise,-exec -e 'Logger->debug([{1=>"a"}])'

Arthur Bergman is talking tomorrow on optimizer.pm

All flavors of munging require AUTOLOAD to handle the munged functions. Otherwize, you'd not bother munging, and just code it directly.

This is only useful in a situation where you need info about your caller; ie where you have to make a decision that needs it as input.

Munging worked here because:

• we're too lazy to provide our own categories

• the caller-info was an acceptable approximation of what we would do if we werent so lazy

• laziness is a virtue :-)

cpanm Devel::Size

perl -MCPAN -e shell
install Devel::Size