Known bugs, planned features, etc.
#========================================================================
#
# TODO
#
# DESCRIPTION
# TODO list for the Template Toolkit version 2.14, containing
# known bugs, limitations, planned enhancements, long term visions
# and a few whacky ideas.
#
# AUTHOR
# Andy Wardley <[email protected]>
#
#------------------------------------------------------------------------
# $Id: TODO.tt2,v 1.16 2003/04/23 12:57:38 abw Exp $
#========================================================================
#------------------------------------------------------------------------
# Miscellaneous
#------------------------------------------------------------------------
* The 'eval' filter leaks memory, as reported by Colin Johnson. The
filter subroutine created contains a reference to the context and then
gets cached in the FILTER_CACHE item of the context. Hey presto -
circular references. The reset() method should probably clear the
FILTER_CACHE. Also need to check the plugins cache for similar
problems.
* The handling of the DELIMITER parameter could be improved. At the
moments it's hardcoded and hacked to Do The Right Thing for Win32
but I'd prefer it to Do The Good Thing.
* If you use 'ttree' with a COMPILE_EXT or COMPILE_DIR option then
templates in the 'lib' directories will be compiled, but those in
the src directories will not. This is because ttree does a chdir()
to the src directory and processes files as './myfile'. TT doesn't
compile RELATIVE files by default.
* No recursion checking is performed for BLOCKs, only
Template::Document instances. This is probably the way it will stay
(unless anyone shouts loudly enough) but it should be documented
anyway. STOP PRESS: I had an idea that bare BLOCK subs should be
blessed into Template::Document class to allow $template->process()
to be called regardless. Template::Document methods would need to
test $self for CODE/HASH and Do The Right Thing. This would then
allow recursion testing for BLOCKs as well as Template::Document
objects.
* It would be nice if there was an option so that the files generated
under the COMPILE_DIR are relative to the INCLUDE_PATH and not absolute.
This could cause potential conflicts (e.g. if INCLUDE_PATH changes
between sessions and the same files in different INCLUDE_PATH dirs
maps to the samed compiled version) but is convenient for those times
when you know that's not going to be a problem.
* Richard Tietjen's patch for stash replace. Allows back references
(e.g. $1) but it would be nice to find a rock-solid way to implement
it without relying on unusual ^A delimiter character.
* Further to the above, Craig Barratt has this solution which will be
going into the next verion (2.05b) unless anyone has any further
suggestions to make before then.
It would be great if replace handled backreferences. I don't like
the ^A solution since the string could contain ^A, plus it is a
security hole. The attempt I posted only works for up to 9
backreferences and doesn't handle an escaped '\$' and uses nested
evals:
$str =~ s{$search}{
my $r = $replace;
my @d = (0, $1, $2, $3, $4, $5, $6, $7, $8, $9);
$r =~ s/\$(\d+)/$d[$1]/eg;
$r;
}eg;
I wish there was a perl predefined variable array containing all
the backreferences (is there one?). You can avoid the hard-coded
limit of 9 with extra evals, and a bit of work on the re could
handle the escaped '\$' case, so maybe that would be good enough.
* Craig also notes, in fixing the problem with NEXT not working
inside SWITCH (see Changes v2.04):
By the way, I came across another arcane bug:
NEXT FOREACH k = [1];
is legal syntax but is an infinite loop, since $_[0]->{ INFOR } in
Parser.yp is not set when the NEXT is parsed, so it generates a
plain "next;" rather than calling $factor->next(). I don't see an
easy, clean fix.
#------------------------------------------------------------------------
# Documentation
#------------------------------------------------------------------------
* Extend the FAQ.
* Document the Splash! library properly, once the interface is a little
more settled.
* Examples for libraries (HTML, Splash & PostScript) should be integrated
into the documentation.
#------------------------------------------------------------------------
# Directives
#------------------------------------------------------------------------
* A 'FOR', like 'FOREACH' but without using an iterator. You wouldn't get
the 'loop' reference to test 'first', 'last', etc., against, but it would
be faster for those cases when you didn't need it. This will likely
be implemented as a facility feature (see later).
* PRINT should be defined as a new directive, doing what the print()
method of Template::View currently does (the Right Thing).
[% PRINT node %] === [% tt.view.print(node) %]
NOTE TO SELF: this is a Very Good Idea [tm]. PRINT becomes the way to
display a data structure (e.g. hash, list, XML element, MyThingy, database
record, etc.) in an "intelligent" fashion. Implemented underneath via
the current default VIEW.
* ARGS. There may be a requirement for reusable template components
to define what variables they plan to use. This would allow some
optimisation and also possibly help to avoid global variable clashes.
Would also be a useful "comment" directive for human readers and maybe
also help in debugging (WARNING: expected 'title' argument).
[% ARGS title # no default
bgcol='#ffffff' # default value
%]
#------------------------------------------------------------------------
# Stash
#------------------------------------------------------------------------
* The XS Stash does not work with tied hashes (e.g. DBI.tie). Also note
that enabling the XS Stash cause all the Template::* modules to be
installed in an architecture-dependant directory. For info, see
http://www.tt2.org/pipermail/templates/2001-September/001568.html
* Jonas Liljegren reports a segfault when using the XS Stash under
Apache/mod_perl with certain undefined variables. At the time of
writing detail is scarce. Check the mailing list for further
details.
* Stas Bekman raised the issue of the Stash not being able to correctly
differentiate between scalar/list context and in particular, the
cgi.param examples not working as expected. This is fixed in v3 and
in Craig's Stash/Context.pm which does the right lookahead to allow
'scalar' and 'list' postfix operators. e.g. cgi.param.list
* Have stash, etc., add current template name/line when reporting errors.
(may be tricky under the current implementation)
#------------------------------------------------------------------------
# Parser
#------------------------------------------------------------------------
* Lists don't accept arbitrary expressions as elements. In other words
you can't do this: [% foo(bar + 1) %]. This has been fixed in the v3
parser. See http://www.tt2.org/v3/
* The parser isn't as intelligent as it could be about blocks of template
code commented out en masse. The pre-scanner find the first terminating
END_TAG after an opening tag, regardless of it being on a
commented line or not.
e.g.
[%#
#
# [% INCLUDE blah %] <- directive ends here
# foo <- this gets printed
%]
* Allow { and } as block delimiters, replacing for the ugly ';' and
big, chunky [% END %] approach.
e.g. something like:
[% FOREACH a = [ 1 2 3 ] %]
[% IF b == a %]
[% INCLUDE foo %]
[% ELSE %]
[% INCLUDE bar %]
[% END %]
[% END %]
could be written as:
[% FOREACH a = [ 1 2 3 ] {
IF b == a {
INCLUDE foo
}
ELSE {
INCLUDE bar
}
}
%]
* Ability to set different parser options for BLOCK definitions, etc.
[% BLOCK header
eval_perl = 0
pre_chomp = 1
%]
...
[% END %]
Anonymous BLOCK can then be used to set a parser scope
[% BLOCK trim=1 %]
...
[% END %]
[% BLOCK trim=0 %]
...
[% END %]
And/or set different tag styles, etc.
[% BLOCK tags='star' %]
[* INCLUDE this_is_a_directive *]
[% INCLUDE this_is_not %]
[* END *]
[% INCLUDE back_to_normal %]
This is likely to be a TT3 feature and I've already got the basic
parser for this up and running. It might get back-pactched into
version 2, otherwise you might have to wait for the first alpha
release of verion 3.
* Craig Barratt reports the following:
I looked at Parse.yp to see how hard it would be to push FILTER
evaluation down into the expr rule, so that you could put filters
inside expressions (eg: using repeat() just like "x" in
perl). More about that later.
In browsing through Parser.yp I noticed several issues:
- The operator precedence is very different to perl, C etc.
For example, these expressions evaluate differently in
TT2 versus perl, C etc:
+ "1 || 0 && 0" evaluates to 0 in TT2 and 1 in perl or C.
TT2 parses it as (1||0) && 0; in perl and C && is higher
precedence than ||.
+ "1 + !0 + 1" evaluates to 1 in TT2 and 3 in perl or C.
TT2 parses it as 1 + !(0 + 1); in perl and C ! is higher
precedence than +.
+ Many other expressions parse incorrectly, but the effect
is benign since most rules return flat text that perl
correctly re-parses. Eg, 2 * 3 + 4 is incorrectly parsed
as (2 * (3 + 4)), but happily just the string "2 * 3 + 4"
is compiled by perl, which correctly evaluates it as
(2 * 3) + 4.
- There is no unary minus and the NUMBER token is signed. So you can
write "x = -2;" but not "x = -y;". Moreover, "x = 1 -1;" is a syntax
error (since "1 -1" returns just two tokens NUMBER, NUMBER). (As a
workaround you can rewrite these as "x = 0-y;" and "x = 1 - 1".)
- You cannot have expressions in lists ([..]) and function arguments.
I have modified the Parser.pm (to make NUMBER unsigned) and modified
Grammar.pm.skel and Parser.yp to fix most of these issues (improved
operator precedence, unary minus and plus), and also to allow
expressions in a few more places (eg: range). But the last item
has me stuck.
The parse rules for lists and function arguments make COMMA optional,
so you can equivalently write [1 2 3 4] or [1,,,,,2 3 4] or [1,2,3,4].
This makes it very difficult to make each term an expression, because
the resulting grammar has many ambiguities. For example, is [1 -1]
two elements [1, -1] or a single element [0]? One partial solution is
to move the bracketed expression rule '(' expr ')' to the term rule,
allowing expressions to be included via parens. But there are also
ambiguities, eg: does [foo (1+1)] have 2 elements or is it a function
call to foo?
Without allowing expressions in lists or function arguments, the unary
minus change I've made means that the NUMBER token is unsigned, so with
my changes you cannot write [-1, 2, 3]. Not a good thing.
One solution is to change the grammar so that COMMAs are required in
lists and arguments, but that would break several test cases and
probably break lots of old templates. But this might be the only
way to produce a grammar that is a lot more similar to perl.
Another solution is to ignore these issues altogether and use temporary
variables to precompute expressions that you need in lists or function
arguments, or use explicit lvalue assignments, eg:
foo(x + 2); becomes temp = x + 2;
foo(temp);
or
List = [x+1,x+2,x+4]; becomes List = [];
List.0 = x+1;
List.1 = x+2;
List.2 = x+4;
Both of these look ugly to me.
Back to the FILTER issues. Ultimately I'd like to be able to embed filters
as low precedence operators in expressions, and write:
List = [
"foo" | repeat(10),
"bar" | repeat(10)
];
but I doubt there is a non-ambiguous upward compatible grammar that
supports this.
Comments?
#------------------------------------------------------------------------
# Plugins
#------------------------------------------------------------------------
* We need a way to easily enable/disable certain plugins. This should
be addressed by facility provision. Probably something for v3.
* The Text::Autoformat module has some problems with versions of Perl
prior to 5.6.0 when using a locale which has a decimal separator
other than '.' (e.g. Swedish, which uses ','). Damian has been made
aware of the problem (and I note he now has a new version out which
I need to check). For now, the Makefile.PL issues a warning but
continues regardless.
* The Template::Plugin DBI iterator first/last() methods don't behave
the same as list first/last(). Randal also reports that get_all()
doesn't work as it should - may be a conflict in code/docs?
* PLUGINS could accept a reference to an object which is used as a
singleton factory for a plugin. (NOTE: 2.01 includes PLUGIN_FACTORY
to implement this, but currently undocumented because it's likely to
change).
* Add Leo & Leon's Page plugin, or try to find some way of implementing
it in terms of the Table plugin. I think the jury is still out on the
matter of whether it counts as duplicated functionality.
* A more general solution for XML (e.g. DOM, XPath, etc) would be for
TT to support a PerlSAX handler which generates the appropriate
callbacks to the view. This should make it possible to easily
display XML content from XML::DOM, XML::XPath, or any other SAX
compliant source.
Something like this:
# define a view
[% VIEW my_view
prefix="my/xml/dom/path/" ;
END
%]
# get some XML
[% USE dom = XML.DOM %]
[% doc = dom.parser(my.files.xmldata) %]
# ask the view to print the data
[% my_view.print(doc) %]
The view print() method will call the relevant 2SAX method on the
XML node, passing a SAX2TTView handler to make the relevant calls
back to the view to display parts of the XML data model as SAX events
are received.
#------------------------------------------------------------------------
# Views
#------------------------------------------------------------------------
The current implementation is there to get me (and anybody else who's
interested) using it and trying to identify the problems, requirements
and general issues involved. I've got a better idea now about what a
VIEW should be in notional terms, but I'm still not quite sure about
the syntax and API.
General thoughts:
* A view defines a set of templates. Things like prefix, suffix,
default, etc., can be specified to customise template selection.
In this sense, it is like a custom provider of those templates.
It implements the template() method to fetch a template according
to those rules.
* It is also a custom processor of those templates. It implements the
process() method. In this sense, it is like a custom context.
* It also implements dispatch logic to apply the right template to the
right kind of data. It does this via the print() method. It may
have all kinds of custom dispatch logic.
* A view takes responsiblity for things template related as opposed
to anything data related (stash) or application logic related
(plugins, runtime code, etc). It is the user interface facility
within the engine.
#------------------------------------------------------------------------
# Splash!
#------------------------------------------------------------------------
The current implementation is a fairly ugly hack to get something up
and running that's good enough to use. It's mainly template driven
and doesn't scale very well, particularly with global variables
clashing all over the place. My plan is that this will become a
view-based system and will no doubt be a test-bed for the
implementation of the view facility.
To include:
* Variable management for storing metadata relating to an
interface/view, protected from the rest of template space.
* Also, style management for defining different styles (e.g. plain,
fancy, icecold, whitehot) for different interface elements (e.g.
bars, borders, menu text) in different modes (e.g. selected, unselected),
or under different, possibly custom conditions (e.g. user preferences,
guest or logged in, etc.).
* Clearly define API for different elements, allowing people to write
apps to the API which run across different conformant widget sets.
* support themes which define a set of styles
* May be wise to move Splash out to a separate distribution.
Randal Schwartz highlighted some problems with non-compliant HTML
being generated. These include:
* no DOCTYPE declaration (added to html/header)
* ALT attribute missing from many <img> tags
* <font ...><table>...</table></font> is illegal (not sure where this
gets done)
* <tr> shouldn't have HEIGHT attribute
* <H3> block element inside inline element
* <A> not allowed here (not sure)
#------------------------------------------------------------------------
# Test Suite
#------------------------------------------------------------------------
* t/file.t and t/directry.t are currently disable on Win32 until I get
a chance to fix a couple of minor bugs relating to '/' vs '\' file
separators.
#------------------------------------------------------------------------
# Facilities
#------------------------------------------------------------------------
Core facilities currently implemented by Template::Context should be
moved out into separate facilities. These currently fall into the
categories of things like stash, view, plugins, filters, parser and
maybe some others. (NOTE: this might tie in very closely with Camelot
and the resource/presenter/actor breakdown, aka model/view/controller).
* 'view' would handle template fetching and processing. It is the view
that talks to a provider, possibly adding special prefixes, suffixes,
doing default templates, special dispatch logic, etc.
* 'stash' is responsible for storing variables, as it currently is.
* 'plugins' is responsible for fetching plugins.
* 'filters' is responsible for fetching filters.
* maybe both the above would get bundled into something like 'logic'?
* 'parser' would make parts of the parser accessible
* 'output' could be used to generate output
There would be some facility to install, customise and remove facilities
to modify TT behaviour as required. This would allow us to disable
certain plugins, or remove the plugin facility altogether, for example.
Or we could install a new stash facility which generated a different kind
of code (e.g. less magic, more speed). Or install a new custom facility
to do some application or domain specific task.
Facilities should be accessible via the context:
my $stash = $context->stash(); # currently works
my $view = $context->view(); # not yet
General form:
my $xyz = $context->facility('xyz');
Multiple form:
my ($a, $b, $c) = $context->facility(qw[ a b c ]);
This can then be written into generated code pretty much as the stash
currently is. The facilities would define the code generators that
currently clutter up Template::Directives. They would ensure that the
facility is scheduled to be requested from the context at the start of
the sub:
sub {
my $context = shift;
my ($stash, $view) = $context->faciity(qw( stash view ));
and then transform the various directives into appropriate callbacks
into the facility:
$output .= $stash->get('x');
$output .= $view->process('header');
The context becomes a switching centre for the Template Toolkit, with
most, if not all of its existing functionality moved out to
facilities. The context should acquire all facilities at the start of
a process lifecycle, run the template, and then release them all
again. This should all be done according to the process contract
which specifies which facilities should be installed, which can be
modified, what can and can't be be loaded, and so on. The
contract would also define things like PRE_PROCESS templates, error
handling, etc.
Thus the role of Template::Service is to undertake a contract with the
client and attempt to fulfill it. A Template::Contract defines the
terms of the contract and the Template::Context becomes an embodiment
(runtime instance) of a contract.
The current context would be available as the 'tt2' template variable,
with facilities available as object methods. Thus, the following become
possible:
[% tt2.stash.get('foo') %] # [% foo %]
[% tt2.view.process('header') %] # [% PROCESS header %]
[% tt2.parser.start_tag %] # can't do this currently
Directives would be re-written into code like that above. You can
use the 'tt2.facility.whatever' form when directive syntax would
otherwise get in the way:
[% mycode( header = tt2.view.process('my/fancy/header'),
footer = tt2.view.process('your/dull/footer') ) %]
Or to explicitly disambiguate:
[% tt2.stash.get('foo.size') %] # object method
[% tt2.stash.list.size(foo) %] # virtual method
[% foo.size %] # lucky dip :-)
Facility management would itself be a facility. Thus, to disable runtime
facility loading, you would simply unload (or not load) the facility
management facility.
[% tt2.facility.install(module='my_facility', name='foo') %]
[% tt2.foo.bar(123) %]
Note that there are some issues relating to cross-cutting facilities,
otherwise known as "aspects" (see Aspect Oriented Programming). For
example, variable localisation cross cuts views and data management
(i.e. tt2.view.include('header', title='my title') must first localise
the stash and delocalise it again afterwards). It may be appropriate
to install aspects as separate entities (e.g. 'localisation') which
can be invoked to apply cross cutting concerns in a generic way.
Perhaps aspects are implemented as their own facility?
[% tt2.aspect.localise %] ... [% tt2.aspect.delocalise %]
or
[% aspect = tt2.aspect.localise %] # cloned & specialised context
[% aspect.view.process('header', title='my title') %]
or
...
Another example would be debugging:
[% tt2dbg = tt2.aspect.debug(...params...) %]
[% tt2dbg.view.process('header') %]
or:
[% tt2 = tt2.aspect.debug(...) %]
...
[% tt2 = tt2.release %]
#------------------------------------------------------------------------
# Output
#------------------------------------------------------------------------
It should be possible to stack output buffers. In other words, you
stop writing to the current output buffer and open a new buffer and
start writing to that. The final output is simply the concatenation
of all output buffers.
The clever part of all this is that is should be possible to keep a
handle on an earlier buffer and go back and append to it at some point
in the future. A typical use is for generating tables of contents at
the start of a document when you don't know in advance what the document
contains. Simply push a new output buffer at the point of the TOC,
generate the rest of the document (keeping track of all the section
titles, etc) and at the end of the document, go back and generate the
TOC onto the end of the first buffer.
Haven't decided on any syntax yet, but it will almost certainly be
implemented as a facility. At the lowest level, something like
this perhaps:
[% INCLUDE header %] # write content to first buffer
[% toc = tt2.output.push %] # save current buffer as toc
Blah blah blah # write content to second buffer
[% INCLUDE footer %]
[% tt2.output.buffer = toc %] # re-instate first buffer
This is the TOC # append to first buffer
The output stack would look something like this:
+-----------------+
| <header> |
| This is the TOC |
+-----------------+
|
V
+-----------------+
| Blah blah blah |
| <footer> |
+-----------------+
Which to the end user, would silently be concatenated into:
<header>
This is the TOC
Blah blah blah
<footer>
Another use is to ensure dependencies on other templates get resolved.
For example, many of the PostScript library templates rely on other
templates. At present, each template sets a global variable to
say "I've been loaded" while also inspecting these variables for any
templates that it relies on, loading them via INCLUDE/PROCESS if not
set. e.g. ps/cross:
[% # this works, but it's a bit clunky
DEFAULT radius = '5 mm';
PROCESS ps/mm UNLESS global.ps.loaded.mm;
global.ps.loaded.cross = 1;
-%]
/cross {
...
}
Better would be to have each template simply activate the flags for
those templates that it relies on. The header file pushes a new output
buffer and the footer file goes back to the header buffer and INCLUDEs
the dependant templates.
[% PROCESS ps/header + ps/complex %]
10 mm 10 mm complex
[% PROCESS ps/footer %]
ps/complex:
[% global.ps.require.simple = 1 %]
/complex {
...
simple
...
}
ps/simple:
[% global.ps.require.mm = 1 %]
/simple {
...
31 mm 41 mm moveto
...
}
ps/header:
%%!PS-Adobe-3.0
...
%%EndComments
[% global.ps.header = tt2.output.push %]
ps/footer:
[% FOREACH file = ['mm','simple',...];
INCLUDE $file IF global.ps.require.$file;
END;
%]
Ultimately, the whole ps library would become a facility. The
acquire() method (called by the Template::Service to acquire the
facility at the start of processing a template) would do the 'header'
part (i.e. push the output buffer), and the release() method (called
at the tail end to release the facility) would do the dependency
checks. The dependency registrations would be facility features,
e.g.
gs/complex:
[% tt2.ps.require.mm = 1 %] # either
[% tt2.ps.require('mm', 'simple') %] # or
