Notes

You can subscribe to an email version of this summary by sending an empty message to p5p-digest-subscribe@plover.com.

Please send corrections and additions to perl-thisweek-YYYYMM@simon-cozens.org where YYYYMM is the current year and month.

This week has been somewhat of a Unicode week; there were 369 messages this week.

Unicode and EBCDIC

The major fight, uhm, thread this week has concerned the implementation of Unicode on EBCDIC machines. As a participator in the thread, I'm obviously biased, so take the following with a pinch of salt.

At the moment, Unicode on EBCDIC just fails horribly, which is obviously a problem. Peter Prymmer had the idea of interposing Latin1-to-EBCDIC and EBCDIC-to-Latin1 conversion tables in the routines which converted to UTF8 and back. Because the first 128 characters of Latin1 correspond exactly to their UTF8 equivalents, we can make special allowances for these; this means the UTF8 conversion routines assume we're starting from Latin1. Hence, the idea was that these routines should start from Latin1, even on EBCDIC.

However, there were two problems with this approach: firstly, various parts of the core "know" that no conversion is necessary for the first 128 characters of Latin1, so they don't send those characters to the conversion routines at all. Nick Ing-Simmons pointed out that this "knowledge" didn't give us any optimization at all, so I removed it.

There was a larger problem, though, that sometimes we use UTF8 to store a number, rather than a character. What should have tipped me off to this (but didn't) was that when Peter applied the patch, Perl complained that version v5.7.0 didn't match version v5.7.0; one of these was being converted to a series of EBCDIC code points, ending up with v45.47.0! This meant that we needed to separate out the times when UTF8 was being used as characters and when as numbers.

Here Nick Ing-Simmons stepped in and wanted a clear, self-consistent view of the Unicode model as it should apply to EBCDIC, something nobody's yet been able to come up with to Nick's satisfaction. Nevertheless, even in the absence of such a model, it looks like Peter and I are close to getting Unicode working on EBCDIC.

Unicode and hash keys

Jarkko reminded us that Unicode hash keys simply don't work, because a lot of hash keys are stored as strings, (Sometimes hash keys are stored as SVs, but sometimes they aren't. Life's like that.) and strings don't inherently know if they're UTF8 encoded or not.

The suggested solution was to have a flag on the hash which signified whether or not the keys were UTF8 encoded, and if a new UTF8-encoded key was added to a non-UTF8 hash, upgrade all the keys in the hash. I beavered away on this for a while, but suddenly out of the blue came a masterly patch from INABA Hiroto, which turned the key length to negative to signify that a key was UTF8 encoded. (A long-forgotten suggestion of Sarathy) After a couple of nits involving stripping of the UTF8 bit on constants, plus fixing up shared keys and the behaviour of the exists operator, we now have fully Unicode-aware hashes.

Many thanks, Hiroto!

Unicode and PerlIO

The PerlIO line disciplines interface has finally borne its first fruit!

If you compile with -Duseperlio, you can now say:

    open $fh, ">:utf8", $filename or die $!;
    print $fh $utf8_string;
    binmode $fh, ":bytes";
    print $fh $literal_string;
    close $fh;

(And, presumably, although I've not tested it:)

    open $fh, ":utf8", $filename or die $!;
    $utf8_string = <$fh>;

This is fantastic news for anyone who needs to manipulate Unicode data - you can now directly read in UTF8 data into a variable, manipulate it and spit it back out into a UTF8-encoded file. Anyone dealing with non-ASCII data should say another big thank-you to Nick.

This more or less completes the Unicode support - once we work out how EBCDIC and other non-Latin1 platforms should behave, only regular expression polymorphism and a couple of bugs in the tr code remain.

dTHR, djSP and friends

Michael Stevens wondered what the cryptic abbreviation dTHR in the core signified; the answer was, surprisingly, "nothing at all". It's a remnant of the previous way of passing around the interpreter context, particularly the thread context, but now this is explicitly passed using aTHX_, dTHX_, and pTHX_. (See the "Background and PERL_IMPLICIT_CONTEXT" section in perlguts for how all this works.)

Michael then produced a patch which removes dTHR, only to stumble across another such macro, djSP. Now, dSP declares a local copy of the stack pointer, but it also called dTHR as well; the idea of djSP was that you had already called dTHR, and you just wanted the stack pointer. (The 'j' was for 'just'.) Now dTHR was gone, there was no difference between the two, so djSP went as well.

He also documented STRLEN which is a type used to hold the length of a string.

Module housekeeping, dependencies, etc.

Mr. Schwern got the ball rolling with what he called "the most mundane patch in the history of Perl". It added $VERSIONs to most of the core modules, which stops the CPAN module doing really evil things. Many modules had strict added to them, which is always a Good Thing. A few other tests were put in place, including one which simply tries to load modules and makes sure they compile, and various other fixes happened. Schwern is Mr. Kwalitee Control for Perl 6, so this is the sort of thing he'll be doing over there; quote of the week goes to him, with "He who dies with the most code to maintain wins!"

A few people filled in the missing gaps - the rest of the $VERSIONs, and a few more tests.

Russ Allbery came back from the Sysadmin's BOFH, uh, BoF at LISA, the USENIX sysadmin's conference, requesting a feature to the CPAN module which would scan a program for module dependencies and install required modules to suit. I decided to cause a little bit of mischief by solving the problem with some undocumented behaviour:

    use CPAN (); push @INC, sub { CPAN::install($_[1]) && eval "use
    $_[1]" }

People, predictably, claimed this was bizarre and undocumented, but nobody documented it, although Jeff Pinyan offered to. Similarly, Russ noted that many module authors don't accurately specify their modules dependencies with PREREQ_PM in their Makemaker scripts, so CPAN auto-dependency-resolving didn't work. Fix it, people!

DESTROY order

It's been a well-known problem that when objects are destroyed, the order in which they call DESTROY is non-deterministic.

After a disturbingly simple patch from Ilya, (it spikes a reference count test) this order is now predictable. If you have:

    sub x::DESTROY {print shift->[0]}
    { my $a1 = bless [1],"x";
      my $a2 = bless [2],"x";
      { my $a3 = bless [3],"x";
        my $a4 = bless [4],"x";
        567;
      }
    }
    print "outside block";
    my $a5 = bless [5],"x";
    my $a6 = bless [6],"x";
    567;

the objects will be now destroyed in the following order: 4, 3, 2, 1, (and here "outside block" is printed) 6, 5. That is, from inner block to outer block, newest objects first. Initially, the approach did not scale, but now after a few patches to the scoping code, it looks like the behaviour is global. Wow.

CGI.pm thinks Darwin is Windows

Running for the title of joke bug of the month, Wilfredo Sanchez of Apple noticed that 'Apparently, because "Darwin" has "win" in it, CGI.pm thinks we're an MS product. Ick.' The offending code, which Wilfredo fixed, was:

    if ($OS=~/Win/i)

Oops. Chris Nandor used this to argue that comparisons against $^O should be string comparisons, and never regular expression matches, noting that "MachTen" was sometimes confused with "Mac" for the same reason. Module authors take heed!

A tip of the hat to Wilfredo and the guys at Apple who've been absolutely exemplary in their help to get Perl working on Darwin.

Various

In other news, Ben Tilly's new Carp module seems to be done; Nick Clark's ARM optimizations are in place - they let you say $a = 3; $b = 5; $c = $a + $b without $c being turned into an NV, which is great for things which don't have real floating point. Jeff Pinyan noted that using multiple side-effect operators caused weird things to happen. Well, gosh. Yitzchak and others made a few fixes to the Test::Harness module. Merijn fixes library ordering on AIX, Dominic noticed a little problem in the stdio configuration. Mike Fisher found what appeared to be a problem with file handles and forking, but turned out to be a problem with exit; Nick had fixed it in perlio. Andy patched lots of things, including SVR4 support.

Non-bug of the week, month, year and decade: $a and $b are special variables. use strict ignores them. This is not a bug. The perl5-porters hellhound pack will be despatched to the next person to submit a bug report related to them.

Until next week I remain, your humble and obedient servant,


Simon Cozens