This Week on perl5-porters - 9-15 March 2008

This Week on perl5-porters - 9-15 March 2008

Take our command, strip off the package, pass the short name and the calling package to _make_fatal(), and then use magic-goto to call our subroutine at the end. It's simple, right? Yet I struggle to find any time when it's correct. -- Paul Fenwick and his Fatal attraction.

Topics of Interest

Getting blead via git

Nicholas Clark pondered the layout of files in the build directory. He thought that it would be useful to have dual-lifed modules positioned under ext/, and in git parlance they would become submodules. This would simplify managing the changes between the CPAN version and blead or maint. But he was wondered how git would keep track of the same file that is in lib/ for maint but in ext/ for blead.

Sébastien Aperghis-Tramoni wondered if Nicholas was thinking about pragma modules (like strict and constant) too.

Nicholas also wanted to know what support git provided to answer questions such as "which changes from this branch have been integrated into that branch". Rafael seemed to think it should be possible, but no people with strong git-fu responded.

Elsewhere, there was some idle chatter of converting everything to UTF-8, but no resolution.

A lexical Fatal for Perl 5.10

Now that 5.10 allows people to write lexical pragmas, Paul Fenwick set about writing a scoped version of the Fatal module (whereby warnings become fatal errors).

He thought that it would be possible to bolt the additional functionality onto a dual-lifed Fatal module, but the syntax seemed clumsy, and a better idea would be to build on Fatal to produce an all-new lethal pragma.

There was a certain amount of bikeshed discussion to suggest better names, such as deadly, autodie and sillier, but lethal gathered currency as the thread moved along.

Fatal::AUTOLOAD - Feature or Bug?

In the middle of adding lethal support to Perl, Paul stumbled upon a thing of wond'rous beauty. It turns out that you can inherit from Fatal in a package, and through the magic of AUTOLOAD, invoke Fatal behaviour by prepending an ampersand to a builtin (à la &open my $in, '<', '/no/such/file') or not, by omitting it.

Of course, this functionality is not documented anywhere: neither the POD, nor code, nor test suite make any allusion to it. As such, Rafael Garcia-Suarez and a chorus of porters called for the chain-saw to have the beast put out of its misery.

Fedora 9 and 5.10.0

Tom Callaway was happy to announce that, thanks in large part to the efforts of Andy Armstrong, Nicholas Clark and Rafael Garcia-Suarez, the upcoming Fedora 9 release will contain perl 5.10.0. This should be available in late April.

Safely supporting POSIX SA_SIGINFO

POSIX offers various ways to define C routines that may be used as signal handlers, one of which gives the routine access to additional information. This in turn offers the handler more context with which to figure out what is going on. Some years back, Jarkko Hietaniemi wrote some code to expose this richer interface to a signal handler written in Perl.

Nicholas Clark ran across Jarkko's work this week, and realised that it no longer worked in the age of safe signals, because the lag between the signal's arrival and its delivery to Perl results in the extra information being lost. He thought that there was a way to make things work, as one of the allowed system calls in a signal handler is write, and with this he could squirrel the information away somewhere until it is safe to fetch it.

His scheme was to install a shim at the signal arrival to write the information to a conveniently pre-opened pipe, and then after the current opcode has been run, see if there are any piped signals waiting. If there are, then the information is pulled out of the pipe, unpacked (since it is all intra-host communication, we'd be free to write out raw structs), prettied up, and then the safe signal handler is called.

Tim Bunce admired the deviousness, but wondered if using pipes was overkill, and asked whether a slab of memory could be set aside for this use instead. Nicholas pointed out that there is nothing from preventing a process from receiving multiple signals simultaneously (they are asynchronous after all), so when all is said and done one would probably wind up with something that resembled the pipe infrastructure anyway, only buggy.

Current flaws in the idea include the fact that on a couple of platforms the siginfo structure is larger than 512 bytes, the largest atomic write permitted to a pipe. Craig Berry reminded people that for multiple deliveries of the same signal during the execution of an opcode are thrown away. Nicholas wondered if that was a bug that needed fixing. Craig pointed back to previous thoughts he had had on the subject perl's signal implementation.

Nicholas also mentioned the hoops one would have to jump through in order to deliver the signal to the right thread in a multi-threaded environment.

Refine make regen to be more selective

Jim Cromie was annoyed by the fact that running a make regen will cause everything to be recompiled (because it generates a slew of files that tickle major Makefile target rules). So Jim added smarts to to have each file built in a holding pen, and only update the target file when it differs from previous run.

At first there was considerable debate about using checksums and hash digests to check whether the files differed. Ben Morrow pointed out that as both the old and new files were present, and read in their entirety one could do away with extra trickery and just perform a simple byte-for-byte comparison.

  shades of mv-if-diff

A question of inheritance, encoding and aliases

It started out with H.Merijn Brand tracking down the cause of utf8.t (from Test::Simple) failure on HP-UX.

This in turn led Michael G. Schwern to discover that there was something wrong with and Encode. Rafael Garcia-Suarez realised that the cause of the problem was due to some sort of assumption of a routine name being present in the Encode namespace, something that the new method dispatch of 5.10 may have broken.

Michael G.Schwern extracted the problem and filed it as a bug. Rafael fixed that up with change #33486. H. Merijn Brand produced a similarly entertaining error message by using Encode and encoding without an encoding name. Michael thought the results were very pretty, but a better error message wouldn't go astray.

  Encode::Alias + open go boom (#51608)

After the dust settled, Dan Kogai, the Encode maintainer, release Encode 2.24. H.Merijn was, however, still having difficulties with the utf8.t with open failing to know what to do about roman8 encoding which appears to be an HP specialty. Michael G. Schwern tried, and failed, to understand what Encode::Alias was doing, and/or whether it was doing it incorrectly.

Unfortunately, at about the same time, David Cantrell discovered that his own 5.10 smoke tests were spewing black, and so Rafael and H.Merijn suggested he try upgrading to Encode 2.24. Alas, that did not solve the problem.

Fortunately, Jarkko Hietaniemi came to the rescue with his "646" patch, and this had David up and running again. Alas, Jarkko did not think that this would help with H.Merijn's roman8 problem.

Taint (PL_tainting, SvTAINTED_on, SvTAINTED_off, SvTAINT)

Bram has a large application that wasn't designed to be taint-clean. Nevertheless, in the middle of the code, he wanted to enable taint mode briefly in order to bring taint checks to bear on unsafe feeds coming into the system.

So he used Taint::Util and Taint::Runtime for his nefarious purposes, yet was surprised when reality didn't quite meet his expectations, in that a scalar, created before taint mode was enabled, reading from STDIN will not be considered to have tainted contents.

Paul Fenwick was horrified by the idea of enabling taint mode at an arbitrary point during the execution of a program (perhaps missing the point that Bram's application wasn't taint-clean in the first place, in which case some taint is better than none). He thought that the problem of scalars not honouring taint mode if created before tainting was enables was probably a performance consideration.

  say it taint so

5.8.9 for VMS

Craig Berry issued a status report for the current 5.8 snapshot on VMS. There are four main problems (three in ExtUtils, one to do with threads). Solutions exist for two problems in ExtUtils, the third may be a question of housecleaning. The threads issue is related to bug #45053. As this is failing with 5.8.8, Craig suggested that it could be left documented that way so as not to hold up 5.8.9.

  99.63% okay

Tests failed on PPC64

Sérgio Durigan Júnior reported a couple of failures in the test suite when building 5.10 on the PPC64 platform. Dominic Dunlop realised that the messages were semi-harmless, being indicative of tests making assumptions on the availability of modules that had not yet been built.

Since none of the porters run PPC64 machines, it seems likely that some new hints will be needed for the configuration process to allow building straight out of the box.


Jim Cromie regretted the lack of a Perl_ck_* routine to check all constructed ops, and thought we needed one. The routine would check that the C struct of the op had sane values.

Test failures for perl 5.10 on Solaris 10

Ken Williams took 5.10 for a spin on Solaris and discovered a number of failures. He traced this down to the fact that the source was built using Sun's make, but that some tests wind up running with a different make, and they fail.

Patches of Interest

More sv.c consting

Steve Schubiger continued with his consting crusade, creating a series of patches that Rafael applied. At the end of the week, patch number 13 was as yet unapplied.

Fix ExtUtils::Install under Cygwin

This problem continued to occupy Steve Hay, Nicholas Clark and Jan Dubois this week. The underlying issue is whether a read-only directory may not be added to, or may not be deleted. Compounding the problem was the fact that different compilers return contradictory information concerning the information returned from the operating system.

By the end of the week, things were cleaned up enough to give the green light for the release of 5.8.9.

Testing madly

Gerard Goossen set about fixing up the errors with the Perl5-to-Perl5 conversion that the MAD infrastructure provides.



Gerard also wanted to rip out a chunk of code that was no longer needed for guessing if a bareword was a subroutine name. Rafael wondered if this was wise, and Gerard explained that changes elsewhere in the MAD codebase had rendered the code unnecessary.

Misleading example in perlsyn.pod (given/when/default)

Paul Fenwick looked at given and when, and was misled by the documentation. He proposed removing the ambiguity with a slight edit, and Rafael applied it.

Rever cygwin archname hints

Reini Urban reverted the archname tweaks on the Cygwin platform, the main reason being to tidy up cpan-testers reports.

Lexical and

Paul Fenwick delivered a first cut at lexically scoped fatalities. The message thread, however, will be summarised next week.


New and old bugs from RT

Remove revision bread crumbs from short description (#48453)

The perlfaq has, for the longest time, carried the blight of revision tags in the titles, causing an unsightly mess in perltoc. Rafael explained that that the Perl FAQ was maintained in a separate SVN repository, and suggested that it would make sense in the long run to bring it back in the fold under git control.

Incorrect calculations (#50072)

Vladislav Malyshkin filed a bug back in January, and at the time Abigail explained that the bug was more in Vladislav's code rather than Perl (the heart of which was the problem of operators having side effects).

Abigail and Michael G. Schwern replied that debating the merits of operators and side effects in Perl 5 was a bit of a lost cause at this stage of the game.

Problem and "solution" for building 5.10.0 with win32+mingw+dmake (#51562)

Kjetil Skotheim reported a problem (and a fix) when building perl on Win32 with MinGW. Jan Dubois wasn't sure how Kjetil encountered the error in the first place, suspecting that some other issue was coming into play.

Scalar::Util::looks_like_number vs. Optimisation in regex? (#51568)

Steffen Ullrich thought he had found a bug in looks_like_number, but he was passing $3 as a parameter. Eric Brine explained that this was a dangerous practice, as something was probably interfering with it on the way down. It would be wise to interpolate it into a string, and pass that instead.

When that is done, the problem goes away.

segmentation fault with array ties (#51636)

"blino" encountered a problem while developing Gtk2::SimpleList and reduced it to a bug involving ties within ties. Most remarkable was the fact that (s)he was able to pin-point the problem as being change #31770, which involved optimising push @ISA and propose a tentative patch to correct the problem.

Rafael thought the patch looked safe enough, but wondered if it would not be papering over a deeper bug in handling magic. Vincent Pit was able to put forward a very good explanation as to what was really happening, and produced a better patch and a regression test. Nicholas wondered if Vincent should have used heavier machinery for saving and restoring magic; Vincent wasn't sure either way.

waitpid() example in perlfunc(1) is bogus (#51642)

"vedge" suggested that the waitpid snippet in the documentation could produce an infinite loop, and proposed an alternative. Abigail thought that the alternative could also wind up as an infinite loop.

  back to the drawing board

op/alarm.t hangs on 5.11.0 (Windows Vista only) (#51674)

Sisyphus reported problems with alarm on Vista. Robert May owned up as being the probable culprit with some recent changes he had made. With a small edit, he was able to recover the correct behaviour.

@- array is incorrect with non matching grouping (#51688)

The bug itself was not a bug, but (and perhaps because of that) Paul Fenwick asked for a point of order concerning the handling of tickets on the perlbug RT queue.

utf8::valid rejects characters in \x14_FFFF - \x1F_FFFF (#51710)

Chris Hall discovered a discrepancy between utf::valid and Encode::encode('utf8', ...), with utf::valid rejecting characters incorrectly. No takers.

Perl5 Bug Summary

  290 new + 1502 open = 1792 (+10 -8)

In Brief

Yitzchak Scott-Thoennes's fix for Archive::Extract's x.lzma test file was applied.

Andy Dougherty delivered what he thought was his most trivial patch ever, a one line suppression in MANIFEST. Nicholas Clark thought Andy could do much better, since Nicholas has made changes that involve only a single character.

The perlbal issue of not building for Fedora was sorted out.

Daisuke Maki found a leak in Text::CSV_XS->getline , which isn't part of core, but since (s)he supplied a reasonable patch, H.Merijn Brand, Text::CSV_XS's maintainer, took the time to fix the problem anyway.

Elizabeth Mattijsen puzzled over the differences in builtin() versus builtin( () ) (extra parentheses in the call) and discovered that she had been bitten by prototypes.

Nicholas Clark believes he has nailed the corruptions seen in ext/threads/t/free.t.

  dup shenanigans

In the process of tidying up a bug report for CGI, Nicholas Clark wondered if the current dual life module bug/patch work-flow was optimal. The distinctly sub-optimal part is having perl bugs reported to an RT queue at, and dual-life module bugs reported to RT queues. This alone makes it difficult to bounce tickets from one queue to another.

Which begs the question, why are we even using RT in two separate domains?

Paul Fenwick asked for (and received) a point of order concerning the cleaning out of RT non-bugs.

Last week's summary

About this summary

This summary was written by David Landgren.

Weekly summaries are published on and posted on a mailing list, (subscription: The archive is at Corrections and comments are welcome.

If you found this summary useful, please consider attending a YAPC conference or contributing to the Perl Foundation to help support the development of Perl.