Hello,
Sorr
Brian Mastenbrook writes:
I've been thinking for several weeks about the issues involved with supporting syntax-aware editing in an editor, particularly Common Lisp syntax. Recently I wrote an incremental parser to handle syntax coloring for lisppaste, and so that experience has helped me to get a better handle on what the relevant issues are to supporting robust syntax-aware editing in a text editor.
Gilbert did something similar as demonstrated in the clim/foo.lisp file in the source tree.
What I'm including in the umbrella of syntax-aware editing are the following tasks:
- Coloring the source based on syntactic type
- Detecting invalid syntax and (a) informing the user and (b) skipping over detectable sections of invalid syntax when performing other commands
- Detecting sections of comment and strings, and ignoring these for other commands appropriately
- and, commands which operate on the parse tree of the source, including the C-M-* functions which operate on an entire
- s-expression
I would like to add detecting and being able to act upon bad indentation.
There are several approaches to making this work. Two of them have been done before en masse, with various degrees of success:
- Force the user to only edit valid syntax, and insert balanced pairs of parens / quotes / comment delimiters. This approach works fine for editing new source but does not work so well when reading in an existing, possibly invalid file, and also can feel much like a straightjacket. It also requires a large amount of adaptation to the environment. The best exemplar of this approach is Interlisp's S-Edit structural editor.
Yes, this has been tried. The idea seems good, but in practice it does not work very well.
- Maintain the view that the text is merely an octet stream, and locally use regexps to try to determine what syntax things are, setting character properties along the way. This approach is (as far as I understand it) the approach used by Emacs, which leads to massive confusion when editing unbalanced strings and comments. Sometimes Emacs never recovers; especially with CL-style #||# multiline comments.
Right. Regular expressions are not powerful enough to understand nested syntax.
What I would like to see is a third path: a robust editor which always knows the syntactic type of the text, but allows the user to edit code as if they were using a plain text editor (or allows a slightly smarter mode where balanced syntax is inserted by default). This is the holy grail of useful syntactic editing. It's also very complicated.
It might not be terribly complicated.
My first approach in writing such an editor revolved around viewing a buffer as a doubly-linked list of lines, which themselves were a doubly linked list composed of segments of text broken up by markers (which delineate syntactic type and also include the cursor and mark). This approach quickly got very complex: it was too difficult to write general text-manipulation routines when text was constantly being broken up by markers, and there could be an unbounded number of markers between each character.
| "a" | #<marker> | "bc" | #<syntactic-marker> | "def" |
The "markers embedded in lines" approach: too complex.
I agree. This is to hard, and makes other operations harder like text searching, etc.
Dan Barlow had a better idea: keep the text as a doubly linked list of strings representing lines, but still use markers to represent the beginning and end of sections with a particular syntactic type. In other words, the markers would be separated out from the text itself, making it possible to have views of the same text with different collections of markers. Primitive editing operations would then be responsible for making sure that whatever marker invariants were necessary were preserved, but this could be simpler when markers are disjoint and it's easy to pull out the set of markers affected by an operation. This is, I think, the best approach, as it would allow keeping the "line" abstraction that hemlock uses as-is.
I suggest a simplification of Gilbert's idea. He uses an incremental parser that attaches its state information to the beginning and the end of each line. Whenever an altered line is requested for display, he restarts the parser on the line, sets font information associated with the line, and then displays the line according to the font information. This approach is very robust, and his incremental parser is surprisingly small and simple.
The only simplification I suggest is not to mark the font information on the line, but to invoke the parser each time a line is displayed (essentially after each keystroke). This simplification makes it unnecessary to keep font information of the line up to date.
I am envisioning that markers can be used for several different purposes - extensible at the user's request. This would include: the cursor, the mark, delimiters for syntax types, even the beginning and end of the line. Allowing multiple cursor-type markers and setting one as primary would allow collaborative editing fairly easily. (On the subject of collaborative editing, the disjoint-markers idea is probably a good one here too: it means that updates to the text can be sent out without syntactic information, and each client updates the marker set to account for the new text on its own.)
While multiple marks are a good idea, I suspect it is going to be hard to use them for syntax-aware editing. I think it is going to be difficult to know whether characters you add should be associated with the syntactic element to the left of the mark or to the right of it. Also, as you pointed out, inserting a single character such as `"' might change the syntactic categories of everything that follows. Re-parsing the line seems like a much simpler and much more robust approach.
The next question of relevance is to figure out how to use these markers to implement knowledge of the file syntax. This Ain't Easy, for several reasons.
First, we need to know the raw syntactic role of a various section of text. Is it a symbol? A string? A list opening or closing? But if we view s-expression editing via the likes of C-M-t as the same problem as syntax coloring, this means we need to maintain a nested view of the current syntax, because finding the end of the current s-expression means understanding not just the raw syntactic role of an element but its level of nesting inside other syntactic types. Nesting is also necessary for robust coloring in Common Lisp: emacs famously fails to handle nested #||# comments (and even sometimes non-nested ones), leaving most of your buffer showing in the font lock comment color.
Yes, but Gilbert's approach keeps the stack of the parser around as state information so that solves the problem quite easily.
Maintaining this nested view of syntax while allowing the user to insert unbalanced syntactic elements is more than merely a SMOP, however. A simple insertion of a character might affect the syntactic type of the entire rest of the file. Inserting a closing character would revert it back then - but possibly not a clean reversion. Deleting a character might require restarting the parser at some point prior to the previous character or syntactic type change.
That is exactly what Gilbert suggests. He restarts the parser from the start of the line that was being modified, but on lines that are required to be displayed. In the worst case, this is the entire buffer, if a modification was made to the first line, and the last line is on display in some (possibly different) window.
Here are some specific examples to think about:
The cursor is on #\A in "#3A((1) (2) (3))". The user hits the delete key.
The cursor is on the first #( in "() (+ 1 2))". The user hits the #( key.
The cursor is at the beginning of the second line of the following section of text. The user hits the #" key.
(format nil Mary had a little lamb. Its fleece was as white as ~A. #|" '|#snow|) ; |
These examples demonstrate that robust syntax-aware editing is not a simple matter of local regular expressions or of a parser that can be run until the top-level syntactic role matches what it was before the edit. A single edit may actually have a deeper meaning - for instance, in the second example, inserting a paren means "insert a level of list-ness at this position in the syntactic role stack until an unmatched close paren is found on this level".
To solve the problem of editing unbalanced strings or multiline comments, Andreas Fuchs suggested that it be possible to revert the syntactic type of an entire region when the insertion of an unmatched opening element would otherwise destroy this information. I think this is a good idea. It can be implemented by "hiding" one set of markers when the unmatched opening element is inserted, and un-hiding them after. Any edits the user makes in the meantime can be rectified to both the hidden and unhidden syntactic markers, thus meaning zero reparsing when the close element is inserted. However, this sounds like just the type of SMOP that is far more difficult than it sounds.
I'm curious to know what other people think about this. It seems to me that this is a rather difficult problem when the various nooks and crannies of CL syntax are taken into account (unless you're willing to live with the possibility of reparsing the entire source file on many edits). If this doesn't make any sense at all I'd like to know that too.
Brian