A place for me to store various play projects and notes. The notes are below...
- Clean up warnings in HTTP and wiki code and get wiki working. Current blocker: need bcrypt wrapper for password storage.
- Finish coil library.
- Fix the testing situation. There should be a test suite for every library. All test suites should use the "lib or exe" hack. It should be trivial for anyone to run the tests and compare with previous results. Tests should be run automatically after each push. There should be a web page showing historical results.
- Multi-line strings and raw strings. Try implementing raw strings as r-parser.
https://en.wikipedia.org/wiki/Dylan_(programming_language) needs a lot of love.
Revamp the FAQ. It's so random and negative. Look at eiffel.com FAQ for examples.
It can be recovered from https://github.com/cgay/dylan-wiki-data/blob/master/pages/main/FAQ/FAQ and everything relating to GD should be removed. Just use a single page.
Questions to add:
Q: How do I change the way my object prints? A: use io; use print; use format; override print-message and print-object
- Single file libraries. Implement DEP 6. (I've since done an about face on DEP 6. I think it would be best to define the library and modules in the normal way, but in the same source file.)
- Improved scripting support, including
- Package it up for various platforms. apt, brew, whatever.
- Scala collections API (Bruce Mitchener recommendation)
- trunk/doc/opendylan design notes
What's the reasoning behind not having format-string and format-arguments slots in <error>, but rather putting them in <simple-error>? This is a wart on the language. You probably don't want to risk the format arguments not having dynamic extent anyway, most of the time. (See comment in element-range-error.) I think it's an optimization to prevent having to do the formatting unless the error is actually signalled.
Note that in many cases OD does embed collections into errors when it doesn't use element-range-error. This should be fixed and the errors messages should at least include the collection type and size.
Why is there no equivalent to Python's __repr__? (See previous item.)
How do you create Unicode string constants if they only contain characters that can be expressed in a <byte-string>? e.g., in Python u"foo". Does "<666>" create a unicode string or a byte string? Is this even useful? How about #u"..."?
For writing code with regular expressions in it, Python's r"..." syntax is very nice. What about adding something similar to Dylan? e.g., #r"..." or #/.../
Why are set operations (union, intersection) defined on <sequence> instead of <collection> or <set>?
I think the main concentration should be on the Linux port. It should be easy for anyone with the right tools installed to build OD on both Linux and Windows.
I love using registry projects because they seem to be much less broken than non-registry projects and they allow for easily hacking on OD itself. The major drawback for me is that they don't seem to support loose/development mode very well. (Do some experiments to figure out exactly what DOES happen...I've been working around it for so long I can't quite remember.)
I'd be happy with one Big Red Switch that set the default compilation mode for everything.
- All warnings should be displayed during compilation, even those of sub-projects. I don't care if this is done in the compiler or if we provide a script that will display them. (It would have to be a bit smart about which sub-projects were actually used, I guess.) This works in the IDE so I would imagine it's not a deep problem.
- It should be possible to get a Dylan-level backtrace. I'm not talking about interactive debugging here, just the ability to display a readable backtrace when an error is signalled. e.g., Python's default behavior is to display a backtrace before it exits if there's an unhandled exception.
Increase the initial size of windows to reflect today's minimum screen sizes.
Move the Debug options to a top-level tab in the Project window. I change these all the time when testing.
Combine the Build dialog and Warnings tab into a single tab called Build. They seem a natural fit and the larger screens these days can easily accomodate the change. Note that the two lines of text in the current Build progress window can easily be combined into one: <library-name>: <action>
Change win32-environment command-line parsing to use commmand-line-parser.
Use config file instead of registry on Windows. Housel says it would be nice if it were pluggable so that you could use either one. Just configs would be enough for me, but since registry is already used it probably makes sense to be Windowsy on Windows. Config files could be used for testing on Windows.
I imagine adding a --config command-line option would be enough to switch to the use of config files and ignore the registry. This would especially be useful for testing.
There would need to be a user config file (default ~/.opendylan/opendylan.coil) and an auto-generated config for settings that are changed via the IDE (e.g., ~/.opendylan/autogen.coil). The user file would override options in the autogen file.
It would be nice to allow "--config none" to mean "don't load any config".
Make it possible to sort the tree of definitions (alphabetically, by definition type, by textual order) by right-clicking on a file in the Project window.
Improvements to the "select buffers" menu:
- Sort by most-recently-visited. There should be a drop-down menu to choose sorting method. The choice should be remembered.
- Show/hide which buffers were loaded due to Edit Methods? Or buffers which are under the project's directory (as determined by the shared file root)? Or just an option to show those that are part of any open project? Or add a filter widget...this might be the most general.
- An option to display as <dir>/<file> rather than <file> <dir>. This can make it easier to find files if you know the directory. They just line up better. Could play around with showing only the unique parts, or grouping them by directory...
Show the filename in separator lines in composite buffers.
Make M-. work for "m(t1, t2, t3, ...)" Even better, if it's not too slow, right click on a name could put all known methods under a submenu of Edit Methods, plus All.
Integrate Testworks into the IDE. Needs design. 10 seconds thought ideas:
- Add a test-library keyword to the LID file.
- Project -> Run Test Suite... (use good default locations for tests)
- Store test results in standard locations.
- Display results in a new project window tab and have a button to "make them canonical".
- Select arbirtrary tests and/or suites to re-run.
- I really, really want hierarchical suite and test names for this.
- Needs to support external resources, e.g., servers. This can be done via a --config argument.
- Integrate with the Test Specification to indicate in Deuce whether the visible methods have corresponding tests, and if so, what they are.
Export Test Suite Skeleton can generate invalid module specs. The one it generated for regular-expressions had a bad ')' in
constant <invalid-regex> :: <object>);
which is defined as
defne constant <invalid-regex> = <illegal-regex>;
Implement fast, Eclipse-style "resource search". This would use the compiler database, not be file search. (Although file search could be integrated into it as well, as in Eclipse.) It could offer to put all the found definitions into a composite buffer.
Modernize the VCS integration. As a first pass support Subversion and Git. The way it's currently done looks a bit antiquated to me although I haven't fully internalized it yet.
- There should be auto-detection of the VCS back-end based on existence of .git or .svn directories.
- The Project Window should show the VC status of files and should provide a way to commit/update/etc the entire project or selected files.
- The Editor Window should show the revision number and status of the current file, and should have options to commit/update/diff/etc.
- Give some thought to how we might support VC operations for multiple open project (via the Main Window?)
- Provide some way to collect up commit comments incrementally (from the editor, diff/merge facility, and of course from the commit UI).
- Provide integration with diff/merge tools like Meld rather than reinventing the wheel.
- Catch overflow errors when the profiling macro is called and re-run the test without profiling. Report the error.
Has float-to-string but no string-to-float. Has number-to-string but no string-to-number.
What's the reasoning behind the 'skip' parameter to some sequence functions, like 'position'? It seems like it would be less efficient to use than 'start' and 'end' parameters since the implementation would have to start from the beginning each time. I don't like it being a mandatory keyword in the 'position' generic function. (Other functions that have it: find-element, find-value, find-key.) Note that find-key(..., skip: ...) is only used twice is all of Open Dylan.
subsequence-position uses 'count', but it really means 'skip'. Need to do an exhaustive review of all the sequence functions.
find-key has a 'failure' keyword argument which it seems should be called 'default' or at least 'on-failure'. Want:
find-key(seq, fn, start: m, end: n, default: d)
Review other collection functions that might use start/end parameters.
Look at Factor as an example: http://docs.factorcode.org/content/article-sequences.html
abeaumont points out that it can be more readable to use sequence functions that copy rather than start:/end:. Agree, but it conses. Can have both by simply using slice() liberally.
Have both slice(seq, m, n) and subseq(seq, m, n) ? The former has sloppy semantics and the latter has strict?
Review split() signature. I think this would be better:
split(seq, sep, parts: n, start: s, end: e)
Move <set> into the collections library. It's Currently defined in functional-extensions, so that's where you find it with a 'grep' and that's not a library you want to use directly since it's not portable.
housel suggests that we could combine collection-extensions into collections, since collections is standard/"common dylan" while collection-extensions started as a GD-only library.
Write a DEP to introduce slices to the language. They really need to be supported by the language itself or they're not terribly useful. Thoughts so far:
- Slices inherit their (im)mutability from the target sequence.
slice(sequence, start, end)where end may be #f. Both start and end may be negative to index from the end.
reverse(slice)can sometimes just swap the start and end indices.
- For mutable slices need to look at all collection mutation functions and see what the semantics should be.
Add start/end parameters to sequence functions? Doing this could create a more consistent strings API. (Alternatively slice() could be used, but it requires allocating memory.)
- Rename to
find-subsequence. This function stands out in that it doesn't use a verb-noun naming scheme where a much better (in my opinion) name exists:
- Fix inconsistency between
find-key(..., skip: 2)and
subsequence-position(..., count: 2). Use "skip".
- Rename to
add/add! don't specify where an element is added and it differs between lists and stretchy vectors, and it's completely non-obvious what deque will do. (It adds to the front.) It would be preferable to have separate functions for adding to the beginning or end of a sequence. Perhaps push/push! and push-last/push-last!, inspired by <deque>.
I just found (what I think is) a bug in user-registry-path due to this behavior. How often does one not care where an element will be added?
- Mainly is just missing features now. First step is probably to implement the verbose: and multi-line: parameters to compile-regex.
Add a #"author-id" file property. Besides being generally useful, currently if there's an error resolving the author name we're left with nothing at all. Use this in Koala's directory listing. Also, document the new attribute and consider removing the file-properties function.
Better temp file support
Export both OS-specific and generic modules. The OS-specific modules would use the OS-specific names. e.g., getpid vs current-process-id. Then for calls that only exist on a certain platform it will be natural to use the OS-specific name instead of inventing a new one. Also if you know your code only works on Linux it's more natural to call getpid anyway.
When the exe file doesn't exist, the error is "create process failed: The system cannot find the file specified." It should say what the file was.
I would like a simpler API than this. The irregular number of return values is strange, and often one wants something as simple as:
let (exit-code, stdout, stderr) = run-program("whoami");
As for a complete API, I like the way subprocess.Popen works.
Better error messages in io/*-file-accessor.dylan:
dylan-compiler -build code/coil.lid ... Opened project coil (/Users/cgay/dylan/src/coil/code/coil.hdp) Internal error: read: Is a directory Exiting with return code -1
- Should probably be using <internet-address>es instead of <string>s, at least for the <listener> class.
- Interesting discussion about URL query values: http://lists.w3.org/Archives/Public/ietf-http-wg/2004JulSep/0009.html
- Look at the Allegro web app package more. (name?) It has some nice ideas about sessions and apps.
- create-directory(parent, name) seriously? Just pass a single pathname and figure out the parent directory, yes?
Rewrite the whole damned thing? Man, I hate this library. It seems way overly complex to me. Need to come up with a reasonable alternate design, I guess. I would probably get rid of directory locators completely, since you generally don't know whether a locator names a directory or file until you ask the file-system. It has to be sufficiently easy to use with strings wherever possible, rather than (for example) having to create useless locator objects just to merge them. I think you basically need:
<locator> <url> <file>
- <http-server> => <http-server-url>
- <ftp-server> => <ftp-server-url>
- <file-server> => <file-server-url>
Various string conversion functions in common-dylan.
Specific functions I've wanted:
. slice(sequence, bpos, epos) . slice!(...) . slice!-setter(...)
Implement multi-line strings (""") in the compiler if it's not too hairy. Probably should write a DEP first, or in parallel. See conversation on Hackers list with subject "Better string constants".
Get lisp-to-dylan (LTD) added to quicklisp.
It might be fun to try something like this for Dylan: http://dev-tricks.net/pipe-infix-syntax-for-python I imagine syntax like:
pipe(fib() => until(f1) => where(f2) => transform(f3) => sum)
Instead of iterators we would use closures in Dylan, so fib() returns a "generator": a closure taking no args and returning the next value each time it is called. Each name following the =>, such as "until", would be a function that expects a generator function as its first argument and returns the next value according to its own rules.
I have my doubts about the general applicability of this, but it might be a neat hack and possibly a blog post. dlowe brought it to my attention.
english2cron: parse english and generate cron time specs. Could be a Python hack to actually have anyone give a shit, and would make a fun blog post. :)
Convert lisppaste to Dylan with LTD. This would be a good synthesis project, like wiki is.
Convert COIL to Dylan. XML is just such a horrible configuration syntax.
This is underway in https://github.com/cgay/coil . Parser is basically working. Needs tests and printer.
Create a general-purpose "with" macro, analogous to Python's "with" statement or C#'s "using" statement. This could use a <resource-protocol[-mixin]> class and support acquire and release generic functions. I counted around 300 "with"-style macros in fundev and libraries, but of course only some of them would fit this model. The main ones that jump out at me are: network streams, files, database connections, locks, http-server. Housel mentioned <closable-object> as one possibility, but it only supports close currently, and the name isn't quite right.
Implement enums a la Java: http://download.oracle.com/javase/1.5.0/docs/guide/language/enums.html It looks as though each element is an instance of the enum's class name. In Dylan there's no need to create a subclass for each element when methods are provided, since we can use == method dispatch on the enum constant. The elements can be stored in class slots.
Fix shootout code and update the web site.
Benchmarks -- IO performance seems to be really bad. See the count-words benchmark. Even some very basic benchmarks could be useful at this stage...
- Array access speed for different array types.
- Gabriel benchmarks
DUIM on Linux
- Consider switching wholesale to current (Java?) terminology: condition -> exception, error -> bug
- Rename <serious-condition> to <exception>. Should <error> really be a subclass of <serious-condition>? Fix system libraries' subclassing of <error> where inappropriate. cpage's comments clarified things:
- Fix condition hierarchy around <invalid-index-error> and friends (see comment in collection.dylan) and export a useful set of conditions. <key-error>?
Proposal to rename <serious-condition>
Proposal: Change the language by renaming <serious-condition> to <exception>.
For reference, here are some DRM definitions:
<serious-condition>: The class of conditions that cannot be safely ignored. <error>: The class of conditions that represent something invalid about the program.
- <serious-condition> is the normal class to inherit from for classes that represent exceptional conditions and that are not errors. (Note the use of the word "exceptional" here, which seems natural in English.) The name "<serious-condition>" is too wishy-washy and long-winded to be the default condition class to subclass or catch. Chris Page also points out that it begs the question "How serious?" Should we also have a <really-serious-condition>?
- "Exception" is common terminology for this type of condition in other languages so this makes Dylan slightly more accessible to programmers coming from those communities.
This is a fairly straight-forward change. The steps would be:
- Rename <serious-condition> to <exception>.
- Update all uses of <serious-condition> to use <exception> instead.
- Update the DRM with an erratum noting the change.
I would prefer NOT to provide backward compatibility because the user base is tiny to non-existent. Having backward compatibility leaves the door open to new code using the old class name. That's just my preference though; I don't mind going the other route. Perhaps there are bootstrap or other issues that I haven't thought of here though.
Proposal to make <serious-condition> and <error> disjoint Proposal: Make <error> be a subclass of <condition> instead of being a subclass of <serious-condition>.
In many programs there are cases for which it is legitimate to use the "diaper pattern" (i.e., catch many or all exceptions). For example in an application's top-level loop it may be preferable to catch <serious-condition> and offer the choice to continue rather than to exit the application. Sometimes library code doesn't export a particular exception class so the client is left with no choice but to cast a wider net.