perltodo - Perl TO-DO List
This is a list of wishes for Perl. The tasks we think are smaller or easier are listed first. Anyone is welcome to work on any of these, but it's a good idea to first contact perl5-porters@perl.org to avoid duplication of effort, and to learn from any previous attempts. By all means contact a pumpking privately first if you prefer.
Whilst patches to make the list shorter are most welcome, ideas to add to the list are also encouraged. Check the perl5-porters archives for past ideas, and any discussion about them. One set of archives may be found at:
- http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/
What can we offer you in return? Fame, fortune, and everlasting glory? Maybe not, but if your patch is incorporated, then we'll add your name to the AUTHORS file, which ships in the official distribution. How many other programming languages offer you 1 line of immortality?
In 5.10.0 the smartmatch operator ~~
isn't working quite "right". But
before we can fix the implementation, we need to define what "right" is.
The first problem is that Robin Houston implemented the Perl 6 smart match
spec as of February 2006, when smart match was axiomatically symmetrical:
http://groups.google.com/group/perl.perl6.language/msg/bf2b486f089ad021
Since then the Perl 6 target moved, but the Perl 5 implementation did not.
So it would be useful for someone to compare the Perl 6 smartmatch table
as of February 2006 http://svn.perl.org/viewvc/perl6/doc/trunk/design/syn/S03.pod?view=markup&pathrev=7615
and the current table http://svn.perl.org/viewvc/perl6/doc/trunk/design/syn/S03.pod?revision=14556&view=markup
and tabulate the differences in Perl 6. The annotated view of changes is
http://svn.perl.org/viewvc/perl6/doc/trunk/design/syn/S03.pod?view=annotate and the diff is
svn diff -r7615:14556 http://svn.perl.org/perl6/doc/trunk/design/syn/S03.pod
-- search for =head1 Smart matching
. (In theory viewvc can generate that,
but in practice when I tried it hung forever, I assume "thinking")
With that done and published, someone (else) can then map any changed Perl 6 semantics back to Perl 5, based on how the existing semantics map to Perl 5: http://search.cpan.org/~rgarcia/perl-5.10.0/pod/perlsyn.pod#Smart_matching_in_detail
There are also some questions that need answering:
How do you negate one? (documentation issue) http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2008-01/msg00071.html
Array behaviors http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2007-12/msg00799.html
* Should smart matches be symmetrical? (Perl 6 says no)
* Other differences between Perl 5 and Perl 6 smart match?
Objects and smart match http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2007-12/msg00865.html
Schwern notes, that there's duplication of code - lots and lots of tests have
some variation on the big block of $Is_Foo
checks. We can safely put this
into a file, change it to build an %Is
hash and require it. Maybe just put
it into test.pl. Throw in the handy tainting subroutines.
Which is crazy given just how simple POD purports to be, and how simple HTML
can be. It's not actually as simple as it sounds, particularly with the
flexibility POD allows for =item
, but it would be good to improve the
visual appeal of the HTML generated, and to avoid it having any validation
errors. See also make HTML install work, as the layout of installation tree
is needed to improve the cross-linking.
The addition of Pod::Simple
and its related modules may make this task
easier to complete.
pod/checkpods.PL (and make check
in the pod/ subdirectory)
implements a very basic check for pod files, but the errors it discovers
aren't found by podchecker. Add this check to podchecker, get rid of
checkpods and have make check
use podchecker.
(This probably impacts much more than the core: also the Test::Harness and TAP::* modules on CPAN.)
All of the tests in t/ can now be run in parallel, if $ENV{TEST_JOBS}
is set. However, tests within each directory in ext and lib are still
run in series, with directories run in parallel. This is an adequate
heuristic, but it might be possible to relax it further, and get more
throughput. Specifically, it would be good to audit all of lib/*.t, and
make them use File::Temp
.
We should have tests for everything. When all the core's modules are tested, Schwern has promised to donate to $500 to TPF. We may need volunteers to hold him upside down and shake vigorously in order to actually extract the cash.
Use Devel::Cover to ascertain the core modules's test coverage, then add tests that are currently missing.
A full test suite for the B module would be nice.
Code such as this
will currently deparse as
because the tokenizer inlines the value of the constant subroutine PI
.
This allows various compile time optimisations, such as constant folding
and dead code elimination. Where these haven't happened (such as the example
above) it ought be possible to make B::Deparse work out the name of the
original constant, because just enough information survives in the symbol
table to do this. Specifically, the same scalar is used for the constant in
the optree as is used for the constant subroutine, so by iterating over all
symbol tables and generating a mapping of SV address to constant name, it
would be possible to provide B::Deparse with this functionality.
perlbench
seems impervious to any recent changes made to the perl core. It
would be useful to have a reasonable general benchmarking suite that roughly
represented what current perl programs do, and measurably reported whether
tweaks to the core improve, degrade or don't really affect performance, to
guide people attempting to optimise the guts of perl. Gisle would welcome
new tests for perlbench.
Fix the bugs revealed by running the test suite with the -t
switch (via
make test.taintwarn
).
As part of the "dists" plan, anything that doesn't belong in the smallest perl distribution needs to be dual lifed. Anything else can be too. Figure out what changes would be needed to package that module and its tests up for CPAN, and do so. Test it with older perl releases, and fix the problems you find.
To make a minimal perl distribution, it's useful to look at t/lib/commonsense.t.
For maintenance (and branch merging) reasons, it would be useful to move some architecture-independent dual-life modules from lib/ to ext/, if this has no negative impact on the build of perl itself.
Ilya observed that use POSIX; eats memory like there's no tomorrow, and at various times worked to cut it down. There is probably still fat to cut out - for example POSIX passes Exporter some very memory hungry data structures.
There is a script embed.pl that generates several header files to prefix
all of Perl's symbols in a consistent way, to provide some semblance of
namespace support in C
. Functions are declared in embed.fnc, variables
in interpvar.h. Quite a few of the functions and variables
are conditionally declared there, using #ifdef
. However, embed.pl
doesn't understand the C macros, so the rules about which symbols are present
when is duplicated in makedef.pl. Writing things twice is bad, m'kay.
It would be good to teach embed.pl
to understand the conditional
compilation, and hence remove the duplication, and the mistakes it has caused.
Currently if you write
then use strict;
isn't in force within the autoloaded subroutines. It would
be more consistent (and less surprising) to arrange for all lexical pragmas
in force at the __END__ block to be in force within each autoloaded subroutine.
There's a similar problem with SelfLoader.
The installman script is slow. All it is doing text processing, which we're told is something Perl is good at. So it would be nice to know what it is doing that is taking so much CPU, and where possible address it.
Or if you prefer, tasks that you would learn from, and broaden your skills base...
There is an installhtml
target in the Makefile. It's marked as
"experimental". It would be good to get this tested, make it work reliably, and
remove the "experimental" tag. This would include
Checking that cross linking between various parts of the documentation works. In particular that links work between the modules (files with POD in lib/) and the core documentation (files in pod/)
Work out how to split perlfunc
into chunks, preferably one per function
group, preferably with general case code that could be used elsewhere.
Challenges here are correctly identifying the groups of functions that go
together, and making the right named external cross-links point to the right
page. Things to be aware of are -X, groups such as getpwnam to
endservent, two or more =items
giving the different parameter lists, such
as
and different parameter lists having different meanings. (eg select)
Be able to install them. This would probably need a configure test to see how the system does compressed man pages (same directory/different directory? same filename/different filename), as well as tweaking the installman script to compress as necessary.
Make it easy for anyone to run Devel::Cover on the core's tests. The steps to do this manually are roughly
do a normal Configure
, but include Devel::Cover as a module to install
(see INSTALL for how to do this)
- make perl
- cd t; HARNESS_PERL_SWITCHES=-MDevel::Cover ./perl -I../lib harness
Process the resulting Devel::Cover database
This just give you the coverage of the .pms. To also get the C level coverage you need to
Additionally tell Configure
to use the appropriate C compiler flags for
gcov
- make perl.gcov
(instead of make perl
)
After running the tests run gcov
to generate all the .gcov files.
(Including down in the subdirectories of ext/
(From the top level perl directory) run gcov2perl
on all the .gcov files
to get their stats into the cover_db directory.
Then process the Devel::Cover database
It would be good to add a single switch to Configure
to specify that you
wanted to perform perl level coverage, and another to specify C level
coverage, and have Configure
and the Makefile do all the right things
automatically.
Quite often vendors ship a perl binary compiled with their (pay-for)
compilers. People install a free compiler, such as gcc. To work out how to
build extensions, Perl interrogates %Config
, so in this situation
%Config
describes compilers that aren't there, and extension building
fails. This forces people into choosing between re-compiling perl themselves
using the compiler they have, or only using modules that the vendor ships.
It would be good to find a way teach Config.pm
about the installation setup,
possibly involving probing at install time or later, so that the %Config
in
a binary distribution better describes the installed machine, when the
installed machine differs from the build machine in some significant way.
Some platforms mandate that you provide a list of a shared library's external
symbols to the linker, so the core already has the infrastructure in place to
do this for generating shared perl libraries. My understanding is that the
GNU toolchain can accept an optional linker specification file, and restrict
visibility just to symbols declared in that file. It would be good to extend
makedef.pl to support this format, and to provide a means within
Configure
to enable it. This would allow Unix users to test that the
export list is correct, and to build a perl that does not pollute the global
namespace with private symbols.
Currently Configure
understands -Dusecrosscompile
option. This option
arranges for building miniperl
for TARGET machine, so this miniperl
is
assumed then to be copied to TARGET machine and used as a replacement of full
perl
executable.
This could be done little differently. Namely miniperl
should be built for
HOST and then full perl
with extensions should be compiled for TARGET.
This, however, might require extra trickery for %Config: we have one config
first for HOST and then another for TARGET. Tools like MakeMaker will be
mightily confused. Having around two different types of executables and
libraries (HOST and TARGET) makes life interesting for Makefiles and
shell (and Perl) scripts. There is $Config{run}, normally empty, which
can be used as an execution wrapper. Also note that in some
cross-compilation/execution environments the HOST and the TARGET do
not see the same filesystem(s), the $Config{run} may need to do some
file/directory copying back and forth.
Make pod/roffitall be updated by pod/buildtoc.
Right now, Configure probes for two commands, and sets two variables:
cc
(in cc.U)
This variable holds the name of a command to execute a C compiler which can resolve multiple global references that happen to have the same name. Usual values are cc and gcc. Fervent ANSI compilers may be called c89. AIX has xlc.
ld
(in dlsrc.U)
This variable indicates the program to be used to link
libraries for dynamic loading. On some systems, it is ld.
On ELF systems, it should be $cc
. Mostly, we'll try to respect
the hint file setting.
There is an implicit historical assumption from around Perl5.000alpha
something, that $cc
is also the correct command for linking object files
together to make an executable. This may be true on Unix, but it's not true
on other platforms, and there are a maze of work arounds in other places (such
as Makefile.SH) to cope with this.
Ideally, we should create a new variable to hold the name of the executable linker program, probe for it in Configure, and centralise all the special case logic there or in hints files.
A small bikeshed issue remains - what to call it, given that $ld
is already
taken (arguably for the wrong thing now, but on SunOS 4.1 it is the command
for creating dynamically-loadable modules) and $link
could be confused with
the Unix command line executable of the same name, which does something
completely different. Andy Dougherty makes the counter argument "In parrot, I
tried to call the command used to link object files and libraries into an
executable link, since that's what my vaguely-remembered DOS and VMS
experience suggested. I don't think any real confusion has ensued, so it's
probably a reasonable name for perl5 to use."
"Alas, I've always worried that introducing it would make things worse,
since now the module building utilities would have to look for
$Config{link}
and institute a fall-back plan if it weren't found."
Although I can see that as confusing, given that $Config{d_link}
is true
when (hard) links are available.
These tasks would need a little C knowledge, but don't need any specific background or experience with XS, or how the Perl interpreter works
The C code uses the macro PERL_UNUSED_ARG
to stop compilers warning about
unused arguments. Often the arguments can't be removed, as there is an
external constraint that determines the prototype of the function, so this
approach is valid. However, there are some cases where PERL_UNUSED_ARG
could be removed. Specifically
The prototypes of (nearly all) static functions can be changed
Unused arguments generated by short cut macros are wasteful - the short cut macro used can be changed.
The way @INC is laid out by default, one cannot upgrade core (dual-life) modules without overwriting files. This causes problems for binary package builders. One possible proposal is laid out in this message: http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2002-04/msg02380.html.
Natively 64-bit systems need neither -Duse64bitint nor -Duse64bitall. On these systems, it might be the default compilation mode, and there is currently no guarantee that passing no use64bitall option to the Configure process will build a 32bit perl. Implementing -Duse32bit* options would be nice for perl 5.12.
Currently perl from p4
/rsync
ships with a patchlevel.h file that
usually defines one local patch, of the form "MAINT12345" or "RC1". The output
of perl -v doesn't report that a perl isn't an official release, and this
information can get lost in bugs reports. Because of this, the minor version
isn't bumped up until RC time, to minimise the possibility of versions of perl
escaping that believe themselves to be newer than they actually are.
It would be useful to find an elegant way to have the "this is an interim maintenance release" or "this is a release candidate" in the terse -v output, and have it so that it's easy for the pumpking to remove this just as the release tarball is rolled up. This way the version pulled out of rsync would always say "I'm a development release" and it would be safe to bump the reported minor version as soon as a release ships, which would aid perl developers.
This task is really about thinking of an elegant way to arrange the C source such that it's trivial for the Pumpking to flag "this is an official release" when making a tarball, yet leave the default source saying "I'm not the official release".
The Perl source code is stable enough that it makes sense to profile it, identify and optimise the hotspots. It would be good to measure the performance of the Perl interpreter using free tools such as cachegrind, gprof, and dtrace, and work to reduce the bottlenecks they reveal.
As part of this, the idea of pp_hot.c is that it contains the hot ops, the ops that are most commonly used. The idea is that by grouping them, their object code will be adjacent in the executable, so they have a greater chance of already being in the CPU cache (or swapped in) due to being near another op already in use.
Except that it's not clear if these really are the most commonly used ops. So as part of exercising your skills with coverage and profiling tools you might want to determine what ops really are the most commonly used. And in turn suggest evictions and promotions to achieve a better pp_hot.c.
One piece of Perl code that might make a good testbed is installman.
Currently all new OP structures are individually malloc()ed and free()d.
All malloc
implementations have space overheads, and are now as fast as
custom allocates so it would both use less memory and less CPU to allocate
the various OP structures from arenas. The SV arena code can probably be
re-used for this.
Note that Configuring perl with -Accflags=-DPL_OP_SLAB_ALLOC
will use
Perl_Slab_alloc() to pack optrees into a contiguous block, which is
probably superior to the use of OP arenas, esp. from a cache locality
standpoint. See Profile Perl - am I hot or not?.
Currently, numerous functions look virtually, if not completely,
identical in both win32/wince.c
and win32/win32.c
files, which can't
be good.
Visual C++ 2005 (VC++ 8.x) deprecated a number of CRT functions on the basis that they were "unsafe" and introduced differently named secure versions of them as replacements, e.g. instead of writing
- FILE* f = fopen(__FILE__, "r");
one should now write
- FILE* f;
- errno_t err = fopen_s(&f, __FILE__, "r");
Currently, the warnings about these deprecations have been disabled by adding -D_CRT_SECURE_NO_DEPRECATE to the CFLAGS. It would be nice to remove that warning suppressant and actually make use of the new secure CRT functions.
There is also a similar issue with POSIX CRT function names like fileno having been deprecated in favour of ISO C++ conformant names like _fileno. These warnings are also currently suppressed by adding -D_CRT_NONSTDC_NO_DEPRECATE. It might be nice to do as Microsoft suggest here too, although, unlike the secure functions issue, there is presumably little or no benefit in this case.
These functions currently take no account of DACLs and therefore do not behave correctly in situations where access is restricted by DACLs (as opposed to the read-only attribute).
Furthermore, POSIX::access() behaves differently for directories having the read-only attribute set depending on what CRT library is being used. For example, the _access() function in the VC6 and VC7 CRTs (wrongly) claim that such directories are not writable, whereas in fact all directories are writable unless access is denied by DACLs. (In the case of directories, the read-only attribute actually only means that the directory cannot be deleted.) This CRT bug is fixed in the VC8 and VC9 CRTs (but, of course, the directory may still not actually be writable if access is indeed denied by DACLs).
For the chdir() issue, see ActiveState bug #74552: http://bugs.activestate.com/show_bug.cgi?id=74552
Therefore, DACLs should be checked both for consistency across CRTs and for the correct answer.
(Note that perl's -w operator should not be modified to check DACLs. It has been written so that it reflects the state of the read-only attribute, even for directories (whatever CRT is being used), for symmetry with chmod().)
Maybe create a utility that checks after each libperl.a creation that none of the above (nor sprintf(), vsprintf(), or *SHUDDER* gets()) ever creep back to libperl.a.
- nm libperl.a | ./miniperl -alne '$o = $F[0] if /:$/; print "$o $F[1]" if $F[0] eq "U" && $F[1] =~ /^(?:strn?c(?:at|py)|v?sprintf|gets)$/'
Note, of course, that this will only tell whether your platform is using those naughty interfaces.
Recent glibcs support -D_FORTIFY_SOURCE=2
and recent gcc
(4.1 onwards?) supports -fstack-protector
, both of which give
protection against various kinds of buffer overflow problems.
These should probably be used for compiling Perl whenever available,
Configure and/or hints files should be adjusted to probe for the
availability of these features and enable them as appropriate.
struct gp
and struct magic
are both currently allocated by malloc
.
It might be a speed or memory saving to change to using arenas. Or it might
not. It would need some suitable benchmarking first. In particular, GP
s
can probably be changed with minimal compatibility impact (probably nothing
outside of the core, or even outside of gv.c allocates them), but they
probably aren't allocated/deallocated often enough for a speed saving. Whereas
MAGIC
is allocated/deallocated more often, but in turn, is also something
more externally visible, so changing the rules here may bite external code.
Several SV body structs are now the same size, notably PVMG and PVGV, PVAV and PVHV, and PVCV and PVFM. It should be possible to allocate and return same sized bodies from the same actual arena, rather than maintaining one arena for each. This could save 4-6K per thread, of memory no longer tied up in the not-yet-allocated part of an arena.
These tasks would need C knowledge, and roughly the level of knowledge of the perl API that comes from writing modules that use XS to interface to C.
As a hang over from the original constant
implementation, POSIX.xs
contains a function int_macro_int
which in conjunction with AUTOLOAD
is
used to wrap the C functions WEXITSTATUS
, WIFEXITED
, WIFSIGNALED
,
WIFSTOPPED
, WSTOPSIG
and WTERMSIG
. It's probably worth replacing
this complexity with 5 simple direct wrappings of those 5 functions.
However, it would be interesting if someone could measure the memory usage
before and after, both for the case of use POSIX();
and the case of
actually calling the Perl space functions.
Some years ago Jarkko supplied patches to provide support for the POSIX SA_SIGINFO feature in Perl, passing the extra data to the Perl signal handler.
Unfortunately, it only works with "unsafe" signals, because under safe signals, by the time Perl gets to run the signal handler, the extra information has been lost. Moreover, it's not easy to store it somewhere, as you can't call mutexs, or do anything else fancy, from inside a signal handler.
So it strikes me that we could provide safe SA_SIGINFO support
Provide global variables for two file descriptors
When the first request is made via sigaction
for SA_SIGINFO
, create a
pipe, store the reader in one, the writer in the other
In the "safe" signal handler (Perl_csighandler()
/S_raise_signal()
), if
the siginfo_t
pointer non-NULL
, and the writer file handle is open,
serialise signal number, struct siginfo_t
(or at least the parts we care
about) into a small auto char buff
write() that (non-blocking) to the writer fd
if it writes 100%, flag the signal in a counter of "signals on the pipe" akin to the current per-signal-number counts
if it writes 0%, assume the pipe is full. Flag the data as lost?
if it writes partially, croak a panic, as your OS is broken.
in the regular PERL_ASYNC_CHECK()
processing, if there are "signals on
the pipe", read the data out, deserialise, build the Perl structures on
the stack (code in Perl_sighandler()
, the "unsafe" handler), and call as
usual.
I think that this gets us decent SA_SIGINFO
support, without the current risk
of running Perl code inside the signal handler context. (With all the dangers
of things like malloc
corruption that that currently offers us)
For more information see the thread starting with this message: http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2008-03/msg00305.html
Make all autovivification consistent w.r.t LVALUE/RVALUE and strict/no strict;
This task is incremental - even a little bit of work on it will help.
chdir, chmod, chown, chroot, exec, glob, link, lstat, mkdir, open, opendir, qx, readdir, readlink, rename, rmdir, stat, symlink, sysopen, system, truncate, unlink, utime, -X. All these could potentially accept Unicode filenames either as input or output (and in the case of system and qx Unicode in general, as input or output to/from the shell). Whether a filesystem - an operating system pair understands Unicode in filenames varies.
Known combinations that have some level of understanding include Microsoft NTFS, Apple HFS+ (In Mac OS 9 and X) and Apple UFS (in Mac OS X), NFS v4 is rumored to be Unicode, and of course Plan 9. How to create Unicode filenames, what forms of Unicode are accepted and used (UCS-2, UTF-16, UTF-8), what (if any) is the normalization form used, and so on, varies. Finding the right level of interfacing to Perl requires some thought. Remember that an OS does not implicate a filesystem.
(The Windows -C command flag "wide API support" has been at least temporarily retired in 5.8.1, and the -C has been repurposed, see perlrun.)
Most probably the right way to do this would be this: Virtualize operating system access.
Currently the %ENV entries are always byte strings. See Virtualize operating system access.
Currently glob patterns and filenames returned from File::Glob::glob() are always byte strings. See Virtualize operating system access.
Some built-in operators (lc, uc, etc.) behave differently, based on
what the internal encoding of their argument is. That should not be the
case. Maybe add a pragma to switch behaviour.
Investigate trade offs to switch out perl's choices on memory usage. Particularly perl should be able to give memory back.
This task is incremental - even a little bit of work on it will help.
:unique
in a way that is actually thread-safeThe old implementation made bad assumptions on several levels. A good 90%
solution might be just to make :unique
work to share the string buffer
of SvPVs. That way large constant strings can be shared between ithreads,
such as the configuration information in Config.
Tainting would be easier to use if it didn't take documented shortcuts and allow taint to "leak" everywhere within an expression.
system() accepts a LIST syntax (and a PROGRAM LIST syntax) to avoid running a shell. readpipe() (the function behind qx//) could be similarly extended.
Change 25773 notes
- /* Need to check SvMAGICAL, as during global destruction it may be that
- AvARYLEN(av) has been freed before av, and hence the SvANY() pointer
- is now part of the linked list of SV heads, rather than pointing to
- the original body. */
- /* FIXME - audit the code for other bugs like this one. */
adding the SvMAGICAL
check to
- if (AvARYLEN(av) && SvMAGICAL(AvARYLEN(av))) {
- MAGIC *mg = mg_find (AvARYLEN(av), PERL_MAGIC_arylen);
Go through the core and look for similar assumptions that SVs have particular types, as all bets are off during global destruction.
PerlIO::Scalar doesn't know how to truncate(). Implementing this would require extending the PerlIO vtable.
Similarly the PerlIO vtable doesn't know about formats (write()), or about stat(), or chmod()/chown(), utime(), or flock().
(For PerlIO::Scalar it's hard to see what e.g. mode bits or ownership would mean.)
PerlIO doesn't do directories or symlinks, either: mkdir(), rmdir(), opendir(), closedir(), seekdir(), rewinddir(), glob(); symlink(), readlink().
See also Virtualize operating system access.
It should be possible to make -C work correctly if found on the #! line, given that all perl command line options are strict ASCII, and -C changes only the interpretation of non-ASCII characters, and not for the script file handle. To make it work needs some investigation of the ordering of function calls during startup, and (by implication) a bit of tweaking of that order.
Change 32057 changed the parameter and return value of Perl_moreswitches()
from <char *> to <const char *>. It should now be possible to propagate
const-correctness outwards to S_parse_body()
, Perl_moreswitches()
and Perl_yylex()
.
A comment in S_method_common
notes
- /* This code tries to figure out just what went wrong with
- gv_fetchmethod. It therefore needs to duplicate a lot of
- the internals of that function. We can't move it inside
- Perl_gv_fetchmethod_autoload(), however, since that would
- cause UNIVERSAL->can("NoSuchPackage::foo") to croak, and we
- don't want that.
- */
If Perl_gv_fetchmethod_autoload
gets rewritten to take (more) flag bits,
then it ought to be possible to move the logic from S_method_common
to
the "right" place. When making this change it would probably be good to also
pass in at least the method name length, if not also pre-computed hash values
when known. (I'm contemplating a plan to pre-compute hash values for common
fixed strings such as ISA
and pass them in to functions.)
Perl's diagnostics (error messages, see perldiag) could use reorganizing and formalizing so that each error message has its stable-for-all-eternity unique id, categorized by severity, type, and subsystem. (The error messages would be listed in a datafile outside of the Perl source code, and the source code would only refer to the messages by the id.) This clean-up and regularizing should apply for all croak() messages.
This would enable all sorts of things: easier translation/localization of the messages (though please do keep in mind the caveats of Locale::Maketext about too straightforward approaches to translation), filtering by severity, and instead of grepping for a particular error message one could look for a stable error id. (Of course, changing the error messages by default would break all the existing software depending on some particular error message...)
This kind of functionality is known as message catalogs. Look for inspiration for example in the catgets() system, possibly even use it if available-- but only if available, all platforms will not have catgets().
For the really pure at heart, consider extending this item to cover
also the warning messages (see perllexwarn, warnings.pl
).
These tasks would need C knowledge, and knowledge of how the interpreter works, or a willingness to learn.
Using ;
inside brackets is a syntax error, and we don't propose to change
that by giving it any meaning. However, it's not reported very helpfully:
- $ perl -e '$a = [$b; $c];'
- syntax error at -e line 1, near "$b;"
- syntax error at -e line 1, near "$c]"
- Execution of -e aborted due to compilation errors.
It should be possible to hook into the tokeniser or the lexer, so that when a
;
is parsed where it is not legal as a statement terminator (ie inside
{}
used as a hashref, []
or ()
) it issues an error something like
';' isn't legal inside an expression - if you need multiple statements use a
do {...} block. See the thread starting at
http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2008-09/msg00573.html
The handling of Unicode is unclean in many places. For example, the regexp engine matches in Unicode semantics whenever the string or the pattern is flagged as UTF-8, but that should not be dependent on an internal storage detail of the string. Likewise, case folding behaviour is dependent on the UTF8 internal flag being on or off.
The tokeniser isn't actually very UTF-8 clean. use utf8;
is a hack -
variable names are stored in stashes as raw bytes, without the utf-8 flag
set. The pad API only takes a char *
pointer, so that's all bytes too. The
tokeniser ignores the UTF-8-ness of PL_rsfp
, or any SVs returned from
source filters. All this could be fixed.
Currently this is illegal:
- state ($a, $b) = foo();
In Perl 6, state ($a) = foo();
and (state $a) = foo();
have different
semantics, which is tricky to implement in Perl 5 as currently they produce
the same opcode trees. The Perl 6 design is firm, so it would be good to
implement the necessary code in Perl 5. There are comments in
Perl_newASSIGNOP()
that show the code paths taken by various assignment
constructions involving state variables.
It would be nice to extend the syntax of the ~~
operator to also
understand numeric (and maybe alphanumeric) ranges.
Like ref(), only useful. It would call the DOES
method on objects; it
would also tell whether something can be dereferenced as an
array/hash/etc., or used as a regexp, etc.
http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2007-03/msg00481.html
There is no method on tied filehandles to allow them to be called back by formats.
The old perltodo notes "With gdb
, you can attach the debugger to a running
program if you pass the process ID. It would be good to do this with the Perl
debugger on a running Perl program, although I'm not sure how it would be
done." ssh and screen do this with named pipes in /tmp. Maybe we can too.
Defining an empty DESTROY method might be useful (notably in AUTOLOAD-enabled classes), but it's still a bit expensive to call. That could probably be optimized.
The old perltodo notes that lvalue functions don't work for list or hash slices. This would be good to fix.
The old perltodo notes that lvalue functions don't work in the debugger. This would be good to fix.
The regexp optimiser is not optional. It should configurable to be, to allow its performance to be measured, and its bugs to be easily demonstrated.
Allow to delete functions. One can already undef them, but they're still in the stash.
/w regex modifierThat flag would enable to match whole words, and also to interpolate
arrays as alternations. With it, /P/w would be roughly equivalent to:
See http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2007-01/msg00400.html for the discussion.
Make the peephole optimizer optional. Currently it performs two tasks as it walks the optree - genuine peephole optimisations, and necessary fixups of ops. It would be good to find an efficient way to switch out the optimisations whilst keeping the fixups.
Currently contexts are void, scalar and list. split has a special mechanism in place to pass in the number of return values wanted. It would be useful to have a general mechanism for this, backwards compatible and little speed hit. This would allow proposals such as short circuiting sort to be implemented as a module on CPAN.
Allow lexical aliases (maybe via the syntax my \$alias = \$foo
.
At the moment pp_entersub is huge, and has code to deal with entering both perl and XS subroutines. Subroutine implementations rarely change between perl and XS at run time, so investigate using 2 ops to enter subs (one for XS, one for perl) and swap between if a sub is redefined.
Self-ties are currently illegal because they caused too many segfaults. Maybe the causes of these could be tracked down and self-ties on all types reinstated.
The old perltodo notes "Look at the "reification" code in av.c
".
Perl 6's Synopsis 3 says:
The ... operator is the "yada, yada, yada" list operator, which is used as the body in function prototypes. It complains bitterly (by calling fail) if it is ever executed. Variant ??? calls warn, and !!! calls die.
Those would be nice to add to Perl 5. That could be done without new ops.
Implement a set of "vtables" that virtualizes operating system access (open(), mkdir(), unlink(), readdir(), getenv(), etc.) At the very least these interfaces should take SVs as "name" arguments instead of bare char pointers; probably the most flexible and extensible way would be for the Perl-facing interfaces to accept HVs. The system needs to be per-operating-system and per-file-system hookable/filterable, preferably both from XS and Perl level (Files and Filesystems in perlport is good reading at this point, in fact, all of perlport is.)
This has actually already been implemented (but only for Win32), take a look at iperlsys.h and win32/perlhost.h. While all Win32 variants go through a set of "vtables" for operating system access, non-Win32 systems currently go straight for the POSIX/UNIX-style system/library call. Similar system as for Win32 should be implemented for all platforms. The existing Win32 implementation probably does not need to survive alongside this proposed new implementation, the approaches could be merged.
What would this give us? One often-asked-for feature this would enable is using Unicode for filenames, and other "names" like %ENV, usernames, hostnames, and so forth. (See When Unicode Does Not Happen in perlunicode.)
But this kind of virtualization would also allow for things like virtual filesystems, virtual networks, and "sandboxes" (though as long as dynamic loading of random object code is allowed, not very safe sandboxes since external code of course know not of Perl's vtables). An example of a smaller "sandbox" is that this feature can be used to implement per-thread working directories: Win32 already does this.
See also Extend PerlIO and PerlIO::Scalar.
The peephole optimier converts constants used for hash key lookups to shared hash key scalars. Under ithreads, something is undoing this work. See http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/2007-09/msg00793.html
Currently we leak ops in various cases of parse failure. I suggested that we could solve this by always using the op slab allocator, and walking it to free ops. Dave comments that as some ops are already freed during optree creation one would have to mark which ops are freed, and not double free them when walking the slab. He notes that one problem with this is that for some ops you have to know which pad was current at the time of allocation, which does change. I suggested storing a pointer to the current pad in the memory allocated for the slab, and swapping to a new slab each time the pad changes. Dave thinks that this would work.
Repacking the optree after execution order is determined could allow removal of NULL ops, and optimal ordering of OPs with respect to cache-line filling. The slab allocator could be reused for this purpose. I think that the best way to do this is to make it an optional step just before the completed optree is attached to anything else, and to use the slab allocator unchanged, so that freeing ops is identical whether or not this step runs. Note that the slab allocator allocates ops downwards in memory, so one would have to actually "allocate" the ops in reverse-execution order to get them contiguous in memory in execution order.
See http://www.nntp.perl.org/group/perl.perl5.porters/2007/12/msg131975.html
Note that running this copy, and then freeing all the old location ops would cause their slabs to be freed, which would eliminate possible memory wastage if the previous suggestion is implemented, and we swap slabs more frequently.
This code
used to produce this output:
- Use of uninitialized value in numeric eq (==) at wrong.pl line 4.
- Use of uninitialized value in numeric eq (==) at wrong.pl line 4.
where the line of the second warning was misreported - it should be line 5.
Rafael fixed this - the problem arose because there was no nextstate OP
between the execution of the if
and the elsif
, hence PL_curcop
still
reports that the currently executing line is line 4. The solution was to inject
a nextstate OPs for each elsif
, although it turned out that the nextstate
OP needed to be a nulled OP, rather than a live nextstate OP, else other line
numbers became misreported. (Jenga!)
The problem is more general than elsif
(although the elsif
case is the
most common and the most confusing). Ideally this code
would produce this output
- Use of uninitialized value $undef in addition (+) at wrong.pl line 4.
- Use of uninitialized value $undef in addition (+) at wrong.pl line 7.
(rather than lines 4 and 5), but this would seem to require every OP to carry (at least) line number information.
What might work is to have an optional line number in memory just before the
BASEOP structure, with a flag bit in the op to say whether it's present.
Initially during compile every OP would carry its line number. Then add a late
pass to the optimiser (potentially combined with repack the optree) which
looks at the two ops on every edge of the graph of the execution path. If
the line number changes, flags the destination OP with this information.
Once all paths are traced, replace every op with the flag with a
nextstate-light op (that just updates PL_curcop
), which in turn then passes
control on to the true op. All ops would then be replaced by variants that
do not store the line number. (Which, logically, why it would work best in
conjunction with repack the optree, as that is already copying/reallocating
all the OPs)
(Although I should note that we're not certain that doing this for the general case is worth it)
Tail-calls present an opportunity for broadly applicable optimization;
anywhere that return foo(...)
is called, the outer return can
be replaced by a goto, and foo will return directly to the outer
caller, saving (conservatively) 25% of perl's call&return cost, which
is relatively higher than in C. The scheme language is known to do
this heavily. B::Concise provides good insight into where this
optimization is possible, ie anywhere entersub,leavesub op-sequence
occurs.
- perl -MO=Concise,-exec,a,b,-main -e 'sub a{ 1 }; sub b {a()}; b(2)'
Bottom line on this is probably a new pp_tailcall function which combines the code in pp_entersub, pp_leavesub. This should probably be done 1st in XS, and using B::Generate to patch the new OP into the optrees.
Tasks that will get your name mentioned in the description of the "Highlights of 5.12"
Generally make ithreads more robust. See also iCOW
This task is incremental - even a little bit of work on it will help, and will be greatly appreciated.
One bit would be to write the missing code in sv.c:Perl_dirp_dup.
Fix Perl_sv_dup, et al so that threads can return objects.
Sarathy and Arthur have a proposal for an improved Copy On Write which specifically will be able to COW new ithreads. If this can be implemented it would be a good thing.
Fix (or rewrite) the implementation of the /(?{...})/
closures.
This will allow the use of a regex from inside (?{ }), (??{ }) and (?(?{ })|) constructs.
Apparently these are quite useful. Anyway, Jeffery Friedl wants them.
demerphq has this on his todo list, but right at the bottom.