This Week on perl5-porters (30 September / 6 October 2002)

This Week on perl5-porters (30 September / 6 October 2002)

It was a busy week indeed, with long threads, interesting bugs, clever fixes, miscellaneous optimizations, some new ideas, a few jokes, mysterious failures, and, finally, a security hole. Read on.

Hash::Util::lock_keys inhibits bless

Andreas Koenig reported that a restricted hash (that is, a hash that has been marked as readonly), can't be blessed. In fact, no readonly variable can't be blessed, and that's a feature. As Tim Bunce said, generally hashes are locked by constructors and a constructor can bless it prior to locking it. This behavior was thus documented in the Hash::Util manpage.

Nicholas Clark then proposed to mark restricted hashes with some other flag than SVf_READONLY. This implies to find a free slot in the SV flags bitvector. Tim Bunce pointed out that this property (to use a Perl 6 word) can also be implemented as a new type of magic.

Just in time subroutine loading

Elizabeth Mattijsen requested comments on her new creation,, a module that loads modules and subroutines `just in time' (using the same autoloading mechanism as AutoLoader and SelfLoader.)

Several porters didn't like the name, for several reasons : it's too generic ; Parrot has already an unrelated JIT project, using the Java sense of the acronym, so the word may confuse people ; it doesn't begin with an uppercase letter.

Elizabeth finally decided to go for the name

In the background, Tim Bunce suggested that AutoLoader and SelfLoader could be patched so that the -c command-line option to perl triggers loading and compilation of all subroutines.

The new module's manpage from the CPAN :


The biggest thread of the week -- 65 messages -- began with a question by Christian Jaeger, about the possibility of having some sort of hash that accepts references as keys without stringifying them.

Brian Ingerson came up with a comparison to Ruby, in which any object that provides a hash() method can be used as a hash key. Semantic problems arise when arrayrefs are used as hash keys, when they are modified, and when they're used again as a hash key. Do they still refer to the same hash entry ? Do another arrayref, equivalent to the first one, refer to the same hash entry ? How to compute an hash code, or to compare for equality two arbitrary complex objects ? And, as Brian wondered, isn't that Elvis parking next to my car?!?!

The hash code for a Perl data structure could be, as a first proof-of-concept implementation, the MD5 checksum of a Storable serialization. Brian proposed an implementation of this as a new pragma, keys, that would turn on complex-key functionality for all hashing operations in that scope.

On the other hand, Tim Bunce suggested to rewrite Tie::RefHash as an XS extension.

Overriden built-in misparsing

Bradley Baetz reported that an overriden built-in function isn't always parsed like the original built-in. He provided an example with die(), which breaks the use of CGI::Carp with the Template Toolkit.

The problem is on code that looks like this :

    die MyModule->my_method();

which is incorrectly parsed as (die MyModule)->my_method() when die() is overriden. This comes from the indirect object notation, that makes perl believe that die() is a method from class MyModule. Rafael Garcia-Suarez provided a patch to the tokenizer to fix this behavior.

Bradley reported also another bug on die() (bug #17763), occurring when die() throws an exception object, for which stringification has been overloaded : this exception gets incorrectly stringified before being passed to a $SIG{__DIE__} handler. Nobody commented.

The void context

Yitzchak Scott-Thoennes proposed a patch for the our(%hash) slowness bug reported last week. Tony Bowden also ran some benchmarks on the relative speeds of my @x and my(@x). Stephen McCamant explained that the `padav' op (declaration of a lexical array) doesn't have a specific optimization for void context, and uses scalar context instead, and he provided a patch, and a list of other ops that may benefit from the same improvement.

make too slow

Slaven Rezic noticed that make'ing a large module tree is much slower with 5.8.0 than with previous versions of Perl. If I understand correctly, that's due to a limitation of command-line lengths in MakeMaker-generated Makefiles, currently hardwired at 200 chars. He proposed to add a new parameter MAX_SHELL_LENGTH to tune this, and also to look at the POSIX constant ARG_MAX on systems that define it.

Later, he also thought about making `make test' faster by using parallel processes, and provided a proof-of-concept test script. security hole

Andreas Jurenda discovered a security hole in the Safe module (bug #17744), and also suggested a fix. The Safe module provides a way to construct restricted compartments, in which it's possible to eval some perl code, while some ops have been forbidden. An opmask describes the list of disabled ops. The problem is that it's possible to modify the opmask from within the eval'd perl code; thus, if the safe compartment is used again, it's possibly no longer safe.

This bug has been fixed in the development version of Perl, and Arthur Bergman quickly uploaded Safe 2.08 (and then 2.09) to CPAN.

Memory stats interface

H.Merijn Brand proposed an interface to the sbrk(2) C function, to help tracing the memory used by a perl interpreter. This triggerred quite a bit of discussion, because sbrk(2) has peculiar semantics, is not in POSIX, and there's other equally unportable methods to stat the memory (for example reading this data from /proc). The ultimate goal is, of course, to help us producing a better perl, that consumes less memory resources while being faster.

In brief

Alain Barbet continued to work on his web-browseable smoke database.

H.Merijn Brand reported that gdbm-1.8.1 breaks GDBM_File, but with gdbm-1.8.2 the breakage is gone. So it's probably not a perl bug.

Brian Ingerson proposed that the CPAN modules could install their regression tests somewhere in the lib directories. Then a perltest command could be used to run them again. The changelogs could be installed as well, and accessible via another ad-hoc command, esp. if they're rearranged into YAML format.

While hunting down a bug, Yitzchak Scott-Thoennes discovered another one (#17718) : if(%h){...} (or other boolean constructs) fail to test correctly the emptiness of a hash if this hash is tied.

Rafael Garcia-Suarez added a new warning, Possible precedence problem on bitwise %c operator, to warn about constructs like ($x & $y == $z), which is confusingly equivalent to ($x & ($y == $z)). This uncovered a bug in a regression test for Storable.

H.Merijn Brand provided a first patch to remove 5.005-threads from Perl. (They were deprecated in perl 5.8 in favor of ithreads.)

Chris Darroch noticed that using the English module and the study() function together breaks s///g. Slaven Rezic proposed a very simple patch -- that just disables study() when $& is used in the code. As he says, whoever introduces a performance hit in his script by using $& does not need the performance improvement of study :-). Hugo disapproved.

Dan Kogai released Encode v1.77.

The regression test t/op/closure.t fails mysteriously on some configurations, apparently when using perl's malloc and without the debugging code (the one that implements the -D command-line option). Apparently this is caused by Dave Mitchell's pad.[ch] patch. Dave proposed a fix, the smoke tests have to be run again with it.

About this summary

This summary brought to you by Rafael Garcia-Suarez. It's also available via a mailing list, which subscription address is If you wonder why I switched to yet another mail archive, here's the reason : Google Groups breaks threads at subject changes, which is probably not the Right Thing, and the mail archive breaks threads at month boundaries, which is definitively not the Right Thing either.