From gbergman@UCBBRAHMS Mon Jul 25 16:21:43 1983 Date: 25 Jul 83 16:16:52 PDT (Mon) From: gbergman@UCBBRAHMS (George Mark Bergman) Subject: Re: editor bugs etc. Message-Id: <8307252316.AA22776@UCBBRAHMS.ARPA> Received: by UCBBRAHMS.ARPA (3.342/3.7) id AA22776; 25 Jul 83 16:16:52 PDT (Mon) Received: from UCBBRAHMS.ARPA by UCBERNIE.ARPA (3.336/3.7) id AA13678; 25 Jul 83 16:21:17 PDT (Mon) To: mckusick@UCBERNIE Status: R The following are (i) a short note from Mark Horton in reply to a note of mine saying I'd been keeping notes on editor bugs, had heard he had a new version, was interested in hearing about it and perhaps sending him notes on bugs not mentioned as corrected; (ii) a long letter from me in which I do list bugs, features I think would be desirable etc., (iii) an addendum I sent the next day, (iv) brief jottings not yet sent. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - >From mark@cbosgd.UUCP Thu Jul 21 12:31:55 1983 The new version of vi isn't very different to the user. The internals use terminfo instead of termcap, but the user interface isn't affected by this (except that it starts up faster). The major new features are set showmode will cause an indication on the status line when you are in input mode vedit is a new invocation of vi for novices more function keys now work function keys work in both command and input mode Of course, there are a few bug fixes too. There is a binary in ~mark/bin/vi on ucbarpa. It requires the /etc/term heirarchy (there is no file called /etc/terminfo) which was on ucbarpa once but might be gone now. If you want to grab them from whereever they still exist, please feel free to try them. Mark - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Sent to Mark Horton 23/7/83, about 11AM Dear Mark, Well, your note didn't say you wanted me to send my comments on editor bugs and suggestions, but you didn't say I shouldn't, so I decided to do so. I've tried to organize it into some sort of sections. I'd be interested to know which of the bugs mentioned here you have already found and corrected. (I may soon find out for myself; the person in charge of this machine says he'll try to get a copy of version 3.9 from UCBARPA if they still have it, and get it running for me. If you have any helpful information for him, he is robert@brahms, Robert Gross.) Vedit sounds like a great idea. I should mention that throughout this letter, I have avoided using actual control-characters, but faked them, e.g. used ^ and H to get the appearance of ^H, since files with real control-characters can be confusing when looked at in mail, more, etc. But this means that if you want to try any commands I refer to that use them, you won't be able to yank and source them, unless you replace my fakes with real control characters. The version I am using is 3.7. PROBLEMS WITH COUNTS Some vi operations that logically ought to be able to take counts do not, while others misbehave with counts. In this section, ``N'' will always denote a positive integer placed as a count before a vi operation. The most gross case of misbehavior is that of N^B! The effect is to redraw the screen 23N-44 lines further advanced. (Probably the numbers depend on the screen-size of the terminal; this is on a Z19, using termcap h19u.) When N=1, this does indeed move you a screenful backward, but for higher N it moves the window forward some amount! Further, whatever controls are supposed to monitor whether the command would land one at an acceptable line- number seem to have a different idea of what it is doing: If you aren't already familiar with these weird effects, try setting the cursor near the end of a file that is more than 4 screenfuls long, and hitting 3^B. (You might then try an insert at the place you get to, and a :f^] .) N/pattern/ would be useful, but is not allowed. ND would be a natural synonym for dN$, by analogy with NC for cN$, but it doesn't work that way; it just ignores the N. Finally, if N is precisely the number of lines from the current line to the end of the file, N$ will still correctly carry one to the last character of the file, but cN$, NC, dN$ and yN$ refuse to do anything! (NY does work, not being a synonym for yN$.) The failure of NC is particularly annoying; often when I am composing something, I go back to somewhere in the middle of the next-to- last line, say, and want to rewrite the rest of the sentence; 2cc would kill not only the part I want to rewrite but also the OK beginning of the line, and 2C or 2c$ won't work. I realize that I could get around this by keeping an empty line at the end of the file, but that should not be necessary. PROBLEMS REGARDING SOURCE, MACROS, MAPPINGS These are enormously useful, but seem to have all kinds of hidden restrictions. The Appendix to the Ex Reference Manual, "List of Changes from Version 3.5 to Version 3.6" says ``A bug which prevented the source command from working...from visual has been fixed''. It is true that one can now use :so from vi, but it still has a bug: When the scriptfile invoked contains a global command and some other command(s) after it, everything after the first global command is ignored. The same appears to be true of scripts in named buffers invoked from vi-bottom-line by @x. (It is, perhaps, unexpected that one can invoke scripts with multiline commands using @x from vi's bottom-line at all, since such commands will not work if typed on vi's bottom line directly. A script like s/$/a\ b invoked as @x will indeed work. But strangely, if one tries to invoke from the regular mode of vi the script :s/$/a\ b by putting it in buffer x and doing @x, only the first line will take effect.) Another serious restriction is that the command ``vi'' appears to be ignored in sourced ex-scripts, and though the command Q in macros of various flavors in vi (mapped characters, map!ed characters that contain ``...^V^[...Q...''; @x scripts) does take one into ex, any ex commands after it are ignored. I assume you are aware of whatever restrictions lead to the error-message ``Cannot yank inside global/macro'', since you must have written it, though ``inside'' seems to here have the peculiar meaning ``after a text-changing operation of the macro.'' The error-message ``Can't undo in global commands'' is more mysterious, since I get it when I have a global command after a text-changing command in an @x script (though not in a sourced file). Anyway, the fewer such restrictions these operations were subject to, the more useful they would be! Although nested source commands are allowed (and I find them useful), they leave the editor in a ``noprompt'' state. This can be gotten around by including ``se prompt'' as a line in the outermost scriptfile, but I would hope the problem causing it could be cured. When one tries to ``:unmap!'' a ``:map!'' command whose right-hand-side begins with ^H (entered as ^V^H, of course), one gets the message ``That macro wasn't mapped''. (One can get around this by using :unmap! ^V[character].) Certain termcaps apparently produce automatic mappings, which unfortunately may interfere with useful vi commands. In particular, on a tvi, ^L gets mapped to a movement command, which makes it unavailable for redrawing the screen, unless unmapped. PROBLEMS WITH DIAGNOSTICS "Hit return to continue" -- It took me a long time to realize that when I got this diagnostic there was an alternative to hitting return. I suggest it be reworded "Hit Return or :" However, the behavior of the editor when this diagnostic is given seems to be inconsistent. In particular, when the last of a serious of commands is :e otherfile and I get "Hit return to continue", then hitting : usually has no different effect from hitting return (or any other key), namely the screen is redrawn; yet I think that sometimes in this situation it has brought me directly to the bottom line as desired. Very confusing. Would it be possible to have other alternatives than : and return available, such as /pattern ? Or, more simply, when one would presently be given the diagnostic "Hit return to continue", why not just put the editor into the state it would have if one then hit :, since one would then still have the option of hitting return and getting into vi proper, but it would not require the extra keystroke : to begin a bottom-line command, nor would one go through the frequent frustrating experience of absentmindedly starting to write a bottom-line command, or a pattern-search, and then having to wait while the screen was redrawn because one had hit a key other than :. "Using open mode" Again, it took me a long time to learn that when I tried to enter vi and got this diagnostic, it meant that the system had somehow lost the termcap for the terminal I was on, and that I would have to do something to get the correct termcap into the environment. Till I realized this, I generally ended up either struggling along frustrated in open mode, or logging out and logging back in. I suggest that when someone calls for vi and the termcap is not appropriate, the editor should not be invoked in any form, but instead, a message be given such as: ``Your environment does not show a termcap entry permitting the use of the visual editor. If you are working on a terminal not supporting vi (in particular, a device with no addressable cursor), you may enter one of the other modes of the editor with the command "open filename" or "ex filename". If you are working on a terminal that should support vi, your environment entries are incorrect and should be corrected. They presently show: TERM=.... TERMCAP=.... If you know the correct name or abbreviation for your terminal- type, type it at the end of the next line; if not hit return: % setenv TERM '' If the user typed an acceptable terminal-name, the message would continue, telling how to get the appropriate termcap. If the user instead typed return, the message would ask him or her to type the name of the manufacturer shown on the terminal, not worrying about upper/lower-case distinctions, and a list of possible terminal names and abbreviations would be given... . This whole program would not be part of the editor, so there would be no problem of space within the existing crowded confines of the editor code. "No such file or directory" -- I think there should be a distinction between these two cases, because of the important distinction in the consequences when the user tries to quit the editor: If the directory exists, the file is created, but if not, the results are more complicated -- I seem to recall on one occasion simply losing what I had written on my second try at quitting; though I just now did an experiment and this time repeated ZZ's and :x's simply gave repeated error messages. "File already exists..." -- The ``List of changes from 3.5 to 3.6'' says ``If you get I/O errors, the file is considered "not edited"... .'' I presume that this correction is somehow the cause of the fact that I frequently get the above message when trying to leave the editor on a machine with version 3.7, and have to use :w! %|q to exit. But I've never seen any evidence that there were I/O errors; it mainly seems to happen when I've written some lines to another file in the process of editing. So the criteria the editor is using to decide when there have been ``I/O errors'' should be rechecked. "no such command from open/visual" -- This confused me in my first few days of using the editor, when I didn't understand that one couldn't use i and a (in either their vi or ex senses) from the bottom line of vi. A message "i -- no such command from open/visual" was perplexing because I knew that "i" was indeed a vi command. Perhaps it should say "no such command from open/visual bottom line". MISCELLANEOUS PROBLEMS In ex search and replacement patterns, \\ is supposed to represent a real \-character, but something goes wrong when this occurs at the end of a global command. E.g., though :s/^/\\ works OK (in vi or ex), the variant :.g/^/s//\\ definitely does not. In vi it turns everything off, in ex it seems to behave as though there were just a single \, and in a scriptfile, it -- does something still different, which you can discover if you don't know! The Ex Reference Manual says, ``For sanity with use from within visual mode, ex ignores a ":" preceding any command.'' But it ignores it in the wrong place! -- not at the beginning of the command line, but just before the command letter itself. I.e., it accepts 1,3:s/^/ /, but not :1,3s/^/ /. SUGGESTIONS FOR MINOR ADDED CAPABILITIES In a multiline substitute command with the "c" option, when each line is displayed one has three choices: y, n or break. There are some further options that would be useful. One would be "p" -- at present, "p" can only be included on the command line, which means that one has a choice between seeing the result of every substitution or none. In practice, one would generally like to see the results of the first few cases to make sure that the command one has written does what one meant it to, and maybe a few tricky cases that come up; but not every case! Another might be "u" -- to undo the last case for which one gave a "y". Still another might be an option that would mean ``undo the "c" option -- I see that the substitute command is doing what I wanted, go ahead and finish it without me.'' In a command g/pattern/p, the pattern in question is occasionally such that it takes a while to figure out where on the line it occurs. For this purpose, an option that ``pointed out'' the instance of the pattern, in the same manner that the pattern to be replaced is pointed out in substitute command with option c, would be desirable. When g/pattern/p gives more than a screenful of lines, it would be nice to have it piped through the equivalent of ``more''. ex has the command line option "-", which ``is useful in processing editor scripts''. But if one wants to use a script in the course of an otherwise interactive editing session, it would be desirable to have a corresponding resettable option ``:se -'' (or ``:se nofb''). In strings in pattern-searches, it would be useful to have ^ and $ retain their ``magic'', so that /x[a$]/ could search for all occurrences of x before an a or a newline. (Of course, one would then have to decide whether /x[^y]/ should include the case of x followed by a newline or not.) Just as ex allows the command :vi, so I think that vi should have some bottom-line command equivalent to the regular-mode command Q. When one has done some text-changing bottom-line commands, and realizes one wants to go into ex, it can be time- consuming to hit return and then Q, and wait for the screen to be redrawn for vi before one gets the ex prompt. The option of putting several commands on one line, separated by, "|" is particularly useful in the vi bottom-line mode, because it avoids having the screen redrawn several times. It would be useful to be able sometimes do the same thing with non-bottom-line commands, e.g. in editing a troff file at a low baud rate on a dumb terminal one might like to be able to do i\fI^]3Ea\fR^] without watching the line get redrawn twice. Perhaps some key that would cause any sequence of commands to be ``held'' until some complementary key was hit would be useful. It would also be desirable to have a sequence of commands that had been given in this way available as one unit to be repeated by ``.'', if desired. The parenthesis-matching facility with % might be extended to match ` with ' and < with >. I will mention one facility that I discovered by surprize is possessed by ed but not ex -- sequences such as \1 can be used within ed search-patterns. E.g. (for the most trivial case) /\(.\)\1/ will search for doubled letters. DEBATABLE SUGGESTIONS I will mention here some possible changes which have the difficulty that they would change the meaning of existing commands, so that it could be argued that the disadvantage of users having to change their habits might outweigh the advantages. First, one might try to resolve, one way or another, the contradiction between the count arguments taken by the join commands in vi and ex: In ex, jN joins N+1 lines; in vi, NJ joins N lines (except if N=1). Second, the movement commands tx and Tx of vi (x any character) seem poorly defined. Just as fx will ignore the character on which the cursor is presently sitting, even if it is an x, and move to the next occurrence, so I would think that tx should ignore the character immediately after the cursor, and Tx the character immediately before the cursor. The point is that when one does Nfx, and finds that one had failed to count one occurrence of x and fallen short of where one wanted to go, one can hit ; and get there. Likewise, on doing Ntx and finding one has fallen short, one should be able to hit ; and get to the the next occurrence; but at present, hitting ; leaves the cursor in the same position; one must hit ``2;'' to get any further. In effect, Ntx is presently defined as Nfxh; I am suggesting that it be defined as lNfxh. The sequences cw, dw and yw are presently violations of the principle that c[movement], d[movement] and y[movement] change, delete, or yank everything from the current cursor position through the endpoint of the movement command. cw does what one would expect of ce (in fact, they seem to be synonyms), while there is no way to get the effect which cw would have if it were treated ``consistently''. (E.g., if I have a line beginning ``And if'', and I want to change it to ``If'', I cannot just put the cursor on the A and hit cwI^].) dw and yw delete up to the character immediately before the point to which ``w'' would take the cursor. I would have to agree that this behavior of dw and yw is more useful than that which a literal interpretation of the movement rule would lead to; but perhaps it would become still more useful if when applied to the last word on a line, it deleted or yanked the space immediately before the word along with the word... . On the other hand, one could argue for making a distinction between cw and ce. Though I see the motivation for the above definitions, I see no sensible reason why Y should be equivalent to yy, when C and D are equivalent to c$ and d$. I would vote for changing Y to mean y$. RADICAL SUGGESTIONS Is there any reason for maintaining the distinction between the ``:'' state of vi, and ex itself? At present, there are relative advantages that lead one to choose to go into one or the other for a given operation: From the vi-: state, it is easier to return to the regular vi state; from ex, one has a more powerful range of commands; and it is easier to give a series of commands because each carriage-return gives one a new prompt. My suggestion is that from vi, ``:'' should carry you directly to ex, and when you are in ex, carriage-return (^M) after a command should give you a new prompt, while ^] should put you into vi. Conceivably, things might be simplified even further, and carriage return rather than : could be the key that would carry one from the regular mode of vi into ex: .-------. .-------. .-------. a,i... | basic | ^M | | | vi |<------ | |----->| ex |<---. | insert| | vi | | | |^M | mode | ------>| |<-----| mode | ---' `-------' ^] | mode | ^] | | `-------' `-------' (Of course, ^M presently has a meaning in vi, but it has a synonym +.) Clearly, there would also be no need for a special "Hit return to continue" state. I have not distinguished vi and open in the above diagram. My idea here is that ^] would actually return you to either vi or open, whichever you had last been in, and that to switch to the other, you could enter ex and type vi^] or o^] respectively. (Or you could type vi^M, respectively o^M, and further ex commands, and the mode would be saved for the next time you hit a ^].) Or alternatively, these could be made settable options: se visual respectively se novisual. Having gotten used to the editor as it now exists, I admit that I feel uneasy about the above idea -- the sense of knowing that I am ``still in vi'' when I hit :, and not that ``other land'' of ex, represents a kind of of orientation that it is disconcerting to abandon. But I can't see any logical disadvantage in making such a change. Can you? Certainly, there would be things that would have to be thought out, such as what happens to bottom-line vi pattern-searches. My thought would be that ``/'' from vi should give :/ (i.e., put one in ex at the start of a pattern-search), and ^] after a pattern-search should put one into vi at the appropriate character on the line, in contrast to ^M after a pattern search, which would leave one in ex at the appropriate line. In general, I think such changes would lead to greater simplicity and learnability of the editor. I would also hope that excursions between vi and ex and back could be allowed in scriptfiles. It might also be desirable for ex to have, in addition to a concept of ``current line'', one of ``current cursor position''... . Well, on to another subject. One of the inconveniences I found very vexing when first learning to use the editor was that when in either vi insert mode, or ex/vi-bottom-line, it was very hard to edit what I was doing. Within insert mode the only ``editing'' I could do, without escaping, was with the three operations ^H, ^W and the kill character. And on a slow line with a dumb terminal, escaping to make changes could be very time-consuming because large parts of the screen would insist on being redrawn. Perhaps some other control-character could serve as a modified escape, that allowed one to edit what one had entered in the current insertion without having everything below it redrawn, and then return to it. Obviously, if carried to its logical limit this idea could lead to ridiculous nests of editing operations; but there would be no need to carry it to its logical limit. Anyway, the problem of editing ex-style commands was even worse, because there was no way to ``escape and revise''. I eventually learned enough to realize that the solution was to edit complicated commands in another file and source it. But it is sometimes very useful to have the text on which the commands are to act in front of you when composing them (e.g., you can yank and modify various pieces), which led to the variant of writing command lines within the file I was editing, and then writing those lines into another file and sourcing that, without ever leaving the current file. But this is distracting to deal with when concentrating on the editing task itself, which led me to divise a scriptfile which would handle the writing-to-another-file- and-sourcing for me. Or actually, several such files: One for single-line commands to be used essentially once; one for single-line commands that I would want to use on the same file during various editing sessions, and so would want to keep available in that file, and one for multi-line commands (to be used once). When I first got the idea, I thought one scriptfile would be enough, and I would call it ``do'', so that the command to execute a script I had written in a file I was editing would be ``:so do''. The file it would write to and source would be ``do.again'', so that if I wanted to reuse it, I could ``:so do.again''. When I realized the need for several versions, ``do'' became a directory. Here, for your amusement, are the three files. (Re the lines ``se prompt'', cf. my comment on that under PROBLEMS WITH SOURCE etc.): do/1 (for 1-time use of 1-line commands) .w! ~/do/again d so # se prompt do/1+ (like above, without deleting the command) .w! ~/do/again so # se prompt do/: (to use this, write a multi-line command script, put : at the beginning of the first line, put the cursor on the last line of the script, and then source the following:) ?^:?s/:/ ,''w! ~/do/again ,''d so # se prompt (I also created another version to use in case the script had to have an internal line beginning with ``:'', so that this couldn't unambiguously mark the beginning of the script. This used a line which explicitly specified the address-range of the script. But I have never had a need for it, so I will not bother you with it.) Finally, having gotten an account on a machine with a version 3 editor recently, I have divised still another way of doing this. I have put in my EXINIT the command 'map ^A "ayy:@a^M' and now, gratifyingly, the single stroke ^A has essentially the effect of ``:so do/1+'' -- except for the restrictions to which vi ``map'' commands are subject. But I've only been using this for a couple of weeks; so I have yet to learn how chafing those restrictions will or won't be. Anyway, it might be worth thinking about whether some of these things that I've done with macros should be incorporated in some form into the editor itself; or else whether these macros might be written up in the documentation (or some tutorials) on the editor. Next subject: Complicated pattern-searches in long files can be time-consuming. I have seen the point mentioned that if a pattern-description can be begun with "^", this can speed up the search -- since the pattern-comparisons need only be begun at beginnings of lines. In some cases, this might not be possible, but the user might be aware of some other character or character-sequence in the search-pattern that will occur relatively rarely in the file. In such cases it would be desirable if the user could specify one spot from which the pattern search should start, working forward and backward from there, to minimize false starts. E.g., if for some reason one wants to delete every word containing the letter m, the script %s/[^ ]*m[^ ]*// would become much less time-consuming if one could mark the point at which to begin, say writing %s/[^ ]*\!m[^ ]*// so as to instruct the editor to search for m's, and each time one was found, to find the longest possible strings of non-space characters before and after it, and delete these. (This is a silly example, but I think the idea is clear.) Something that I've seriously felt the need for is the capability of searching for lines that satisfy more than one condition. If one just wants to locate such lines, one can of course leave the editor and do a pipeline of two or more greps; but it would be nice to be able to perform global commands on such lines. Finally, any possibility of introducing the capability of searching for patterns including embedded newlines, a la sed? Multiple windows, a la emacs? ADDENDA I logged in this morning on an adm3a at 300 baud to go over this letter once more before sending it, and ran into another bug! I had done 15^D to get a large scroll despite the low speed, and at one point I saw a line with a typo scrolling up. So I noted its line-number, 402 and without waiting for the screen to stop moving typed something like 402Gfsrd. What happened was that the change was made on line 407 rather than 402 -- presumably the cursor was sent to where 402 had been when the command was received... . Editing this letter this morning reminded me of another feature I have thought would be desirable for editing on dumb terminals at low speeds: An option that would cause lines read from a file or typed in at the bottom of the screen to appear double spaced, with @ signs @ between them, such as one gets when one deletes a line on such a @ terminal. (I have faked this effect here, though the fake will not be @ very successful if you have se nu or se list on.) The point is that @ editing operations that presently cause painfully slow screen-redrawings would simply put material in in place of these fillers -- as happens now when one is lucky enough to be adding material just above a place where material was previously deleted. - * - * - * - * - * - I hope you've found some things of interest in this hodgepodge. Yours, George (gbergman@brahms) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 23/7/83, about 8PM Maybe thinking about the editor while writing my previous ``note'' to you has made me more conscious of these things, but for some reason, I've discovered several more peculiarities of the the Version 3.7 editor today! First, the abbreviation feature has a strange bug (feature?) involving the backslash: If some string X is the abbreviation for a string Y, and if X is typed (following a space etc) immediately followed by a \ and another character c, the result is not Y\c by cY\. The case where I discovered this was a doubly complicated one: I had an abbreviation of the form :ab x x\yZ where x and y were characters and Z was a string of characters. So when I typed x, it presumably first expanded it as x\yZ, then performed the transformation I just described on the x\y turning it into yx\yZ\, and thus giving the result yx\yZ\Z. This turns out to be one of the cases that can't be unmapped without doing :una ^Vx. Further, I just tried a similar case with x replaced by a string of more than one character (namely, :ab if if\pq) and I find I can't unmap that at all. I also find that an abbreviated string containing | (which must be inserted using ^V|, of course) is difficult to unmap. Second, some peculiarities about where the cursor ends up after a yank. If the yank involved a forward movement, the cursor stays where it is, which is the beginning of the yanked segment. If the yank involves a backwards movement, the place where the cursor originally was is not the same as the beginning of the yanked segment, and there seems to be some confusion as to which principle is followed: y- or yk moves the cursor up, while yb leaves it fixed. Unfortunately, there is a snake in the grass with yb: If you hit p after it, the yanked word will not appear after the position where the cursor has remained, but after the position to which it would have gone if it had moved to the beginning of the yanked segment! Likewise if you you hit an x... . (You have no idea how much trouble I'm having with those "if"'s. Of course, I could quit the editor and come back in, and I would lose that crazy abbreviation that way.) I also notice that if the cursor is at the end of a word or between words, 2e behaves the same as e! Finally, I note that d^, when the cursor is before the first nonwhite character, is another exception to the principle that d[motion] deletes everything through the endpoint of [motion]. Similarly with c^ and y^. - *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- *- * discovered next day (not yet sent): error with yb does not just concern p and x: any command is executed as though cursor is at destination of backward in-line yank. N^B does not work consistently? (Not on a medium-length file in Kim) Can get endless-loop mapping if abbreviation forms word within abbreviated. E.g. :ab x ( x ) ``word'' must be delimited on left by space, tab, newline, start-of-insert; on right by any punctuation. Why special ``no tail recursion'' rule? Things like that ``if'' abbreviation can be undone using :una i^Vf! Mention desirability of Np copies sent to ralph@ucbarpa, lindahl@topaz 25/7/83 From gbergman@UCBCARTAN Mon Aug 1 14:19:27 1983 Date: 1 Aug 83 14:14:06 PDT (Mon) From: gbergman@UCBCARTAN (George Mark Bergman) Subject: Re: editor Message-Id: <8308012114.AA00627@UCBCARTAN.ARPA> Received: by UCBCARTAN.ARPA (3.342/3.7) id AA00627; 1 Aug 83 14:14:06 PDT (Mon) Received: from UCBCARTAN.ARPA by UCBERNIE.ARPA (3.336/3.7) id AA19324; 1 Aug 83 14:19:12 PDT (Mon) To: mckusick@UCBERNIE Status: R Here's Mark Horton's reply to my letter on bugs >From mark@cbosgd.UUCP Thu Jul 28 14:38:55 1983 Sorry this has taken so long, but your note was too long to digest in the cracks between the other stuff I do. Anyway, you've clearly put a great deal of thought into this, and I appreciate your input. I'll reply individually to your thoughts, and keep them on file for use (someday) when vi gets rewritten. Some of them are just plain bugs that ought to be fixed soon anyway. PROBLEMS WITH COUNTS The most gross case of misbehavior is that of N^B! The effect is to redraw the screen 23N-44 lines further advanced. (Probably the numbers depend on the screen-size of the terminal; this is on a Z19, using termcap h19u.) When N=1, this does indeed move you a screenful backward, but for higher N it moves the window forward some amount! Further, whatever controls are supposed to monitor whether the command would land one at an acceptable line- number seem to have a different idea of what it is doing: If you aren't already familiar with these weird effects, try setting the cursor near the end of a file that is more than 4 screenfuls long, and hitting 3^B. (You might then try an insert at the place you get to, and a :f^] .) This is a known bug, and was fixed in 3.8. The count is supposed to subtract (LINES-1)*N from the line number, but there's + that should be a -, so it goes forward instead. The check is correct, so it's possible to go off the end of the buffer. N/pattern/ would be useful, but is not allowed. N/pattern/ resets the window size (silly but true) to N. ND would be a natural synonym for dN$, by analogy with NC for cN$, but it doesn't work that way; it just ignores the N. Finally, if N is precisely the number of lines from the current line to the end of the file, N$ will still correctly carry one to the last character of the file, but cN$, NC, dN$ and yN$ refuse to do anything! (NY does work, not being a synonym for yN$.) The failure of NC is particularly annoying; often when I am composing something, I go back to somewhere in the middle of the next-to- last line, say, and want to rewrite the rest of the sentence; 2cc would kill not only the part I want to rewrite but also the OK beginning of the line, and 2C or 2c$ won't work. I realize that I could get around this by keeping an empty line at the end of the file, but that should not be necessary. While you're making valid observations here, are you aware that you can delete the current sentence with d} ? I think that's what you really want. PROBLEMS REGARDING SOURCE, MACROS, MAPPINGS These are enormously useful, but seem to have all kinds of hidden restrictions. The Appendix to the Ex Reference Manual, "List of Changes from Version 3.5 to Version 3.6" says ``A bug which prevented the source command from working...from visual has been fixed''. It is true that one can now use :so from vi, but it still has a bug: When the scriptfile invoked contains a global command and some other command(s) after it, everything after the first global command is ignored. The same appears to be true of scripts in named buffers invoked from vi-bottom-line by @x. Sounds like a bug. (It is, perhaps, unexpected that one can invoke scripts with multiline commands using @x from vi's bottom-line at all, since such commands will not work if typed on vi's bottom line directly. A script like s/$/a\ b invoked as @x will indeed work. But strangely, if one tries to invoke from the regular mode of vi the script :s/$/a\ b by putting it in buffer x and doing @x, only the first line will take effect.) In 3.7 (or 3.8, I'm not sure), you can say :s/$/a^V^Mb to get a newline on the RHS. Of course, this doesn't mean there isn't a bug in scripts. Another serious restriction is that the command ``vi'' appears to be ignored in sourced ex-scripts, and though the command Q in macros of various flavors in vi (mapped characters, map!ed characters that contain ``...^V^[...Q...''; @x scripts) does take one into ex, any ex commands after it are ignored. The internals of getting a character from the tty are completely different in ex and vi. Pushing input for one doesn't affect the other. So there isn't much hope of changing this situation. I assume you are aware of whatever restrictions lead to the error-message ``Cannot yank inside global/macro'', since you must have written it, though ``inside'' seems to here have the peculiar meaning ``after a text-changing operation of the macro.'' The error-message ``Can't undo in global commands'' is more mysterious, since I get it when I have a global command after a text-changing command in an @x script (though not in a sourced file). Anyway, the fewer such restrictions these operations were subject to, the more useful they would be! It's the way undo is done - the text for undo is saved in the "last deleted" buffer, and yank puts text there too. Couple this with the fact that globals and macros can be undone as a unit (they save their state before the first change) and you'll see that the two notions can't coexist. Of course, you can always yank into a named buffer, even inside a macro. Although nested source commands are allowed (and I find them useful), they leave the editor in a ``noprompt'' state. This can be gotten around by including ``se prompt'' as a line in the outermost scriptfile, but I would hope the problem causing it could be cured. Bug, I guess. When one tries to ``:unmap!'' a ``:map!'' command whose right-hand-side begins with ^H (entered as ^V^H, of course), one gets the message ``That macro wasn't mapped''. (One can get around this by using :unmap! ^V[character].) Bug, I guess. Certain termcaps apparently produce automatic mappings, which unfortunately may interfere with useful vi commands. In particular, on a tvi, ^L gets mapped to a movement command, which makes it unavailable for redrawing the screen, unless unmapped. Well, there's no way for vi to tell the difference between ^L that the user typed as ^L and ^L that the user typed as right arrow. However, there are a number of terminals that are upward compatible with the adm3a and use ^L for right arrow. Vi has a special case for these built in - if the terminal has insert/delete line, and ^L is right arrow, then ^R will redraw the screen. PROBLEMS WITH DIAGNOSTICS "Hit return to continue" -- It took me a long time to realize that when I got this diagnostic there was an alternative to hitting return. I suggest it be reworded "Hit Return or :" However, the behavior of the editor when this diagnostic is given seems to be inconsistent. In particular, when the last of a serious of commands is :e otherfile and I get "Hit return to continue", then hitting : usually has no different effect from hitting return (or any other key), namely the screen is redrawn; yet I think that sometimes in this situation it has brought me directly to the bottom line as desired. Very confusing. Would it be possible to have other alternatives than : and return available, such as /pattern ? Or, more simply, when one would presently be given the diagnostic "Hit return to continue", why not just put the editor into the state it would have if one then hit :, since one would then still have the option of hitting return and getting into vi proper, but it would not require the extra keystroke : to begin a bottom-line command, nor would one go through the frequent frustrating experience of absentmindedly starting to write a bottom-line command, or a pattern-search, and then having to wait while the screen was redrawn because one had hit a key other than :. There is an internal difference between "ex mode" (where it doesn't keep a screen image) and "vi mode" (where it does). Any : command that outputs more than 1 line puts you into "ex mode", requiring "hit return to continue" before clearing the screen and redrawing it with known stuff. There is no hope of this changing - the code is too spaghetti-ized already. In the worst case, the ! command scribbles on the screen and there's nothing vi can do to know what the command did. What you really want is for vi to check for typeahead and avert the refresh when it's going to have to redo it anyway. My curses does this, but I simply don't have time to rewrite vi to use it. This would also solve the other problem you mention where macros ought to redraw the screen only once. If you saw the insides of the code, you'd see it needs a rewrite to do this. Each command knows the screen things to do to fix the screen. What most of us do is hit DEL when the screen is being drawn with something we don't want to see. This aborts the update and leaves junk on the screen. Then you move the cursor where you want it, and if the screen is still garbaged, hit ^L. Ugly but effective. "Using open mode" Again, it took me a long time to learn that when I tried to enter vi and got this diagnostic, it meant that the system had somehow lost the termcap for the terminal I was on, and that I would have to do something to get the correct termcap into the environment. Till I realized this, I generally ended up either struggling along frustrated in open mode, or logging out and logging back in. I suggest that when someone calls for vi and the termcap is not appropriate, the editor should not be invoked in any form, but instead, a message be given such as: ``Your environment does not show a termcap entry permitting the use of the visual editor. If you are working on a terminal not supporting vi (in particular, a device with no addressable cursor), you may enter one of the other modes of the editor with the command "open filename" or "ex filename". If you are working on a terminal that should support vi, your environment entries are incorrect and should be corrected. They presently show: TERM=.... TERMCAP=.... If you know the correct name or abbreviation for your terminal- type, type it at the end of the next line; if not hit return: % setenv TERM '' If the user typed an acceptable terminal-name, the message would continue, telling how to get the appropriate termcap. If the user instead typed return, the message would ask him or her to type the name of the manufacturer shown on the terminal, not worrying about upper/lower-case distinctions, and a list of possible terminal names and abbreviations would be given... . This whole program would not be part of the editor, so there would be no problem of space within the existing crowded confines of the editor code. This, of course, doesn't belong in vi, but in login or tset. In fact, tset is much friendlier these days. Vi will print a better diagnostic if it knows you're on a "generic" terminal type such as "dialup" - in terminfo there's a capability to indicate this, so it can print I don't know what kind of terminal you have - all I have is "patchboard" 3.7 can't do this because it can't tell the difference between a generic terminal and a hardcopy terminal. "No such file or directory" -- I think there should be a distinction between these two cases, because of the important distinction in the consequences when the user tries to quit the editor: The kernel doesn't distinguish, so it's hard for vi to. This is just a perror string. If the directory exists, the file is created, but if not, the results are more complicated -- I seem to recall on one occasion simply losing what I had written on my second try at quitting; though I just now did an experiment and this time repeated ZZ's and :x's simply gave repeated error messages. Well, if it can't be reproduced, it doesn't stand much chance of getting fixed. (Not that it would if you could reproduce it, of course.) "File already exists..." -- The ``List of changes from 3.5 to 3.6'' says ``If you get I/O errors, the file is considered "not edited"... .'' I presume that this correction is somehow the cause of the fact that I frequently get the above message when trying to leave the editor on a machine with version 3.7, and have to use :w! %|q Ick! Most of us just type :wq! which is equivalent and much shorter. to exit. But I've never seen any evidence that there were I/O errors; it mainly seems to happen when I've written some lines to another file in the process of editing. So the criteria the editor is using to decide when there have been ``I/O errors'' should be rechecked. Actually, if you get ANY error (e.g. :foo) it resets the "buffer modified" condition. I have finally been convinced this is not right. For a while, I considered it a necessary evil in case your /tmp/Ex* file got an I/O error. "no such command from open/visual" -- This confused me in my first few days of using the editor, when I didn't understand that one couldn't use i and a (in either their vi or ex senses) from the bottom line of vi. A message "i -- no such command from open/visual" was perplexing because I knew that "i" was indeed a vi command. Perhaps it should say "no such command from open/visual bottom line". OK. MISCELLANEOUS PROBLEMS In ex search and replacement patterns, \\ is supposed to represent a real \-character, but something goes wrong when this occurs at the end of a global command. E.g., though :s/^/\\ works OK (in vi or ex), the variant :.g/^/s//\\ definitely does not. In vi it turns everything off, in ex it seems to behave as though there were just a single \, and in a scriptfile, it -- does something still different, which you can discover if you don't know! Backslash is special at the end of a line in global. You need more \\'s. The Ex Reference Manual says, ``For sanity with use from within visual mode, ex ignores a ":" preceding any command.'' But it ignores it in the wrong place! -- not at the beginning of the command line, but just before the command letter itself. I.e., it accepts 1,3:s/^/ /, but not :1,3s/^/ /. Hmm. SUGGESTIONS FOR MINOR ADDED CAPABILITIES In a multiline substitute command with the "c" option, when each line is displayed one has three choices: y, n or break. There are some further options that would be useful. One would be "p" -- at present, "p" can only be included on the command line, which means that one has a choice between seeing the result of every substitution or none. In practice, one would generally like to see the results of the first few cases to make sure that the command one has written does what one meant it to, and maybe a few tricky cases that come up; but not every case! Another might be "u" -- to undo the last case for which one gave a "y". Still another might be an option that would mean ``undo the "c" option -- I see that the substitute command is doing what I wanted, go ahead and finish it without me.'' If I were going to do something this involved (I didn't know anybody used the c option anymore - most people use "n" and "." in vi) I would do query-replace ala EMACS right. In a command g/pattern/p, the pattern in question is occasionally such that it takes a while to figure out where on the line it occurs. For this purpose, an option that ``pointed out'' the instance of the pattern, in the same manner that the pattern to be replaced is pointed out in substitute command with option c, would be desirable. When g/pattern/p gives more than a screenful of lines, it would be nice to have it piped through the equivalent of ``more''. Nice but unlikely. Unless, of course, you have page mode in your tty driver, like we do, in which case you get it for free. ex has the command line option "-", which ``is useful in processing editor scripts''. But if one wants to use a script in the course of an otherwise interactive editing session, it would be desirable to have a corresponding resettable option ``:se -'' (or ``:se nofb''). Seems like a good idea. In strings in pattern-searches, it would be useful to have ^ and $ retain their ``magic'', so that /x[a$]/ could search for all occurrences of x before an a or a newline. (Of course, one would then have to decide whether /x[^y]/ should include the case of x followed by a newline or not.) This sounds pretty hard to do. Just as ex allows the command :vi, so I think that vi should have some bottom-line command equivalent to the regular-mode command Q. When one has done some text-changing bottom-line commands, and realizes one wants to go into ex, it can be time- consuming to hit return and then Q, and wait for the screen to be redrawn for vi before one gets the ex prompt. This would be ugly, since the ex command routine would have to return an indication to the vi routine to exit back to the top level ex command. But I suppose it could be done. The option of putting several commands on one line, separated by, "|" is particularly useful in the vi bottom-line mode, because it avoids having the screen redrawn several times. It would be useful to be able sometimes do the same thing with non-bottom-line commands, e.g. in editing a troff file at a low baud rate on a dumb terminal one might like to be able to do i\fI^]3Ea\fR^] without watching the line get redrawn twice. Perhaps some key that would cause any sequence of commands to be ``held'' until some complementary key was hit would be useful. It would also be desirable to have a sequence of commands that had been given in this way available as one unit to be repeated by ``.'', if desired. See above. And get yourself a terminal with insert/delete char! The parenthesis-matching facility with % might be extended to match ` with ' and < with >. OK. I will mention one facility that I discovered by surprize is possessed by ed but not ex -- sequences such as \1 can be used within ed search-patterns. E.g. (for the most trivial case) /\(.\)\1/ will search for doubled letters. This surprises me. DEBATABLE SUGGESTIONS I will mention here some possible changes which have the difficulty that they would change the meaning of existing commands, so that it could be argued that the disadvantage of users having to change their habits might outweigh the advantages. First, one might try to resolve, one way or another, the contradiction between the count arguments taken by the join commands in vi and ex: In ex, jN joins N+1 lines; in vi, NJ joins N lines (except if N=1). Yeah, ex should be N, not N+1. Second, the movement commands tx and Tx of vi (x any character) seem poorly defined. Just as fx will ignore the character on which the cursor is presently sitting, even if it is an x, and move to the next occurrence, so I would think that tx should ignore the character immediately after the cursor, and Tx the character immediately before the cursor. The point is that when one does Nfx, and finds that one had failed to count one occurrence of x and fallen short of where one wanted to go, one can hit ; and get there. Likewise, on doing Ntx and finding one has fallen short, one should be able to hit ; and get to the the next occurrence; but at present, hitting ; leaves the cursor in the same position; one must hit ``2;'' to get any further. In effect, Ntx is presently defined as Nfxh; I am suggesting that it be defined as lNfxh. Agreed. The sequences cw, dw and yw are presently violations of the principle that c[movement], d[movement] and y[movement] change, delete, or yank everything from the current cursor position through the endpoint of the movement command. cw does what one would expect of ce (in fact, they seem to be synonyms), while there is no way to get the effect which cw would have if it were treated ``consistently''. (E.g., if I have a line beginning ``And if'', and I want to change it to ``If'', I cannot just put the cursor on the A and hit cwI^].) dw and yw delete up to the character immediately before the point to which ``w'' would take the cursor. I would have to agree that this behavior of dw and yw is more useful than that which a literal interpretation of the movement rule would lead to; but perhaps it would become still more useful if when applied to the last word on a line, it deleted or yanked the space immediately before the word along with the word... . On the other hand, one could argue for making a distinction between cw and ce. This is to make the user interface friendlier, and is a fact of life. If I wanted to change "And if" to "If", I'd type "dw~". Though I see the motivation for the above definitions, I see no sensible reason why Y should be equivalent to yy, when C and D are equivalent to c$ and d$. I would vote for changing Y to mean y$. The users wouldn't stand for such a change. Too many are used to it like it is. But you could always map it. RADICAL SUGGESTIONS Is there any reason for maintaining the distinction between the ``:'' state of vi, and ex itself? At present, there are relative advantages that lead one to choose to go into one or the other for a given operation: From the vi-: state, it is easier to return to the regular vi state; from ex, one has a more powerful range of commands; and it is easier to give a series of commands because each carriage-return gives one a new prompt. My suggestion is that from vi, ``:'' should carry you directly to ex, and when you are in ex, carriage-return (^M) after a command should give you a new prompt, while ^] should put you into vi. Conceivably, things might be simplified even further, and carriage return rather than : could be the key that would carry one from the regular mode of vi into ex: The basic problem here is that if : put you into ex mode, you'd have to redraw the screen when you hit return. The motivation for : commands is that you don't have to go through a conceptually hard mode change and wait for a screen redraw. .-------. .-------. .-------. a,i... | basic | ^M | | | vi |<------ | |----->| ex |<---. | insert| | vi | | | |^M | mode | ------>| |<-----| mode | ---' `-------' ^] | mode | ^] | | `-------' `-------' (Of course, ^M presently has a meaning in vi, but it has a synonym +.) Clearly, there would also be no need for a special "Hit return to continue" state. I have not distinguished vi and open in the above diagram. My idea here is that ^] would actually return you to either vi or open, whichever you had last been in, and that to switch to the other, you could enter ex and type vi^] or o^] respectively. (Or you could type vi^M, respectively o^M, and further ex commands, and the mode would be saved for the next time you hit a ^].) Or alternatively, these could be made settable options: se visual respectively se novisual. Having gotten used to the editor as it now exists, I admit that I feel uneasy about the above idea -- the sense of knowing that I am ``still in vi'' when I hit :, and not that ``other land'' of ex, represents a kind of of orientation that it is disconcerting to abandon. But I can't see any logical disadvantage in making such a change. Can you? Certainly, there would be things that would have to be thought out, such as what happens to bottom-line vi pattern-searches. My thought would be that ``/'' from vi should give :/ (i.e., put one in ex at the start of a pattern-search), and ^] after a pattern-search should put one into vi at the appropriate character on the line, in contrast to ^M after a pattern search, which would leave one in ex at the appropriate line. In general, I think such changes would lead to greater simplicity and learnability of the editor. I would also hope that excursions between vi and ex and back could be allowed in scriptfiles. It might also be desirable for ex to have, in addition to a concept of ``current line'', one of ``current cursor position''... . Well, on to another subject. One of the inconveniences I found very vexing when first learning to use the editor was that when in either vi insert mode, or ex/vi-bottom-line, it was very hard to edit what I was doing. Within insert mode the only ``editing'' I could do, without escaping, was with the three operations ^H, ^W and the kill character. And on a slow line with a dumb terminal, escaping to make changes could be very time-consuming because large parts of the screen would insist on being redrawn. Perhaps some other control-character could serve as a modified escape, that allowed one to edit what one had entered in the current insertion without having everything below it redrawn, and then return to it. Obviously, if carried to its logical limit this idea could lead to ridiculous nests of editing operations; but there would be no need to carry it to its logical limit. Why not just get a terminal with insert char? You're paying in performance for having an obsolete terminal. Anyway, the problem of editing ex-style commands was even worse, because there was no way to ``escape and revise''. I eventually learned enough to realize that the solution was to edit complicated commands in another file and source it. This is a standard complaint about a moded editor. It couldn't be fixed without taking away the property of ESC ending command lines. Besides, allowing editing on the bottom line would really break a lot of code. But it is sometimes very useful to have the text on which the commands are to act in front of you when composing them (e.g., you can yank and modify various pieces), which led to the variant of writing command lines within the file I was editing, and then writing those lines into another file and sourcing that, without ever leaving the current file. But this is distracting to deal with when concentrating on the editing task itself, which led me to divise a scriptfile which would handle the writing-to-another-file- and-sourcing for me. Or actually, several such files: One for single-line commands to be used essentially once; one for single-line commands that I would want to use on the same file during various editing sessions, and so would want to keep available in that file, and one for multi-line commands (to be used once). When I first got the idea, I thought one scriptfile would be enough, and I would call it ``do'', so that the command to execute a script I had written in a file I was editing would be ``:so do''. The file it would write to and source would be ``do.again'', so that if I wanted to reuse it, I could ``:so do.again''. When I realized the need for several versions, ``do'' became a directory. Here, for your amusement, are the three files. (Re the lines ``se prompt'', cf. my comment on that under PROBLEMS WITH SOURCE etc.): do/1 (for 1-time use of 1-line commands) .w! ~/do/again d so # se prompt do/1+ (like above, without deleting the command) .w! ~/do/again so # se prompt do/: (to use this, write a multi-line command script, put : at the beginning of the first line, put the cursor on the last line of the script, and then source the following:) ?^:?s/:/ ,''w! ~/do/again ,''d so # se prompt (I also created another version to use in case the script had to have an internal line beginning with ``:'', so that this couldn't unambiguously mark the beginning of the script. This used a line which explicitly specified the address-range of the script. But I have never had a need for it, so I will not bother you with it.) Finally, having gotten an account on a machine with a version 3 editor recently, I have divised still another way of doing this. I have put in my EXINIT the command 'map ^A "ayy:@a^M' and now, gratifyingly, the single stroke ^A has essentially the effect of ``:so do/1+'' -- except for the restrictions to which vi ``map'' commands are subject. But I've only been using this for a couple of weeks; so I have yet to learn how chafing those restrictions will or won't be. Anyway, it might be worth thinking about whether some of these things that I've done with macros should be incorporated in some form into the editor itself; or else whether these macros might be written up in the documentation (or some tutorials) on the editor. Next subject: Complicated pattern-searches in long files can be time-consuming. I have seen the point mentioned that if a pattern-description can be begun with "^", this can speed up the search -- since the pattern-comparisons need only be begun at beginnings of lines. In some cases, this might not be possible, but the user might be aware of some other character or character-sequence in the search-pattern that will occur relatively rarely in the file. In such cases it would be desirable if the user could specify one spot from which the pattern search should start, working forward and backward from there, to minimize false starts. E.g., if for some reason one wants to delete every word containing the letter m, the script %s/[^ ]*m[^ ]*// would become much less time-consuming if one could mark the point at which to begin, say writing %s/[^ ]*\!m[^ ]*// so as to instruct the editor to search for m's, and each time one was found, to find the longest possible strings of non-space characters before and after it, and delete these. (This is a silly example, but I think the idea is clear.) Isn't worth doing - this is fast enough for most people in most cases. Something that I've seriously felt the need for is the capability of searching for lines that satisfy more than one condition. If one just wants to locate such lines, one can of course leave the editor and do a pipeline of two or more greps; but it would be nice to be able to perform global commands on such lines. You want the PWB "or" operator. This is hard to put in - their code is really convoluted. Finally, any possibility of introducing the capability of searching for patterns including embedded newlines, a la sed? Newlines aren't stored - the data structure is an array of lines. So this is nearly impossible. Multiple windows, a la emacs? Would be easy after a rewrite, but impossible with current code. What you really want is a window manager, anyway, most of the time. ADDENDA I logged in this morning on an adm3a at 300 baud to go over this letter once more before sending it, and ran into another bug! I had done 15^D to get a large scroll despite the low speed, and at one point I saw a line with a typo scrolling up. So I noted its line-number, 402 and without waiting for the screen to stop moving typed something like 402Gfsrd. What happened was that the change was made on line 407 rather than 402 -- presumably the cursor was sent to where 402 had been when the command was received... . Knowing the internals of vi, I'd say this is impossible. It probably just screwed up your screen from line noise (or your terminal isn't truly full duplex), or you got the wrong line number. Editing this letter this morning reminded me of another feature I have thought would be desirable for editing on dumb terminals at low speeds: An option that would cause lines read from a file or typed in at the bottom of the screen to appear double spaced, with @ signs @ between them, such as one gets when one deletes a line on such a @ terminal. (I have faked this effect here, though the fake will not be @ very successful if you have se nu or se list on.) The point is that @ editing operations that presently cause painfully slow screen-redrawings would simply put material in in place of these fillers -- as happens now when one is lucky enough to be adding material just above a place where material was previously deleted. Again, it would cost less to buy a real terminal. You can get one for $500 or so now from Falco or Liberty or Zenith. Thanks again for the input. Mark From gbergman@UCBCARTAN Fri Jul 29 16:07:10 1983 Date: 29 Jul 83 16:02:06 PDT (Fri) From: gbergman@UCBCARTAN (George Mark Bergman) Subject: editor Message-Id: <8307292302.AA09635@UCBCARTAN.ARPA> Received: by UCBCARTAN.ARPA (3.342/3.7) id AA09635; 29 Jul 83 16:02:06 PDT (Fri) Received: from UCBCARTAN.ARPA by UCBERNIE.ARPA (3.336/3.7) id AA06658; 29 Jul 83 16:07:04 PDT (Fri) To: cc-03@ucbcory, danh@kim, lindahl@topaz, mckusick@ernie, ralph@ucbarpa Status: RO 29/7/83 I got a reply from Horton: a copy of my (first) letter, annotated with comments that this should indeed be fixed, that would be impossible, etc.. I'll send a copy to anyone who'd like (or a copy of his comment on some specific bug they're interested in). Meanwhile, I'll send you my reply to him, since it discusses still more bugs and possible improvements. Dear Mark, Got your comments on my first letter! Did you get the second, shorter one? Glad to see that some things, at least, can be fixed. Robert has put version 3.9 on this machine and I'm using it. Using movement arrows within insert mode is amusing; though when there is documentation, there should be a warning to the user that these prevent ``u'' from undoing anything before those commands. But the first thing there needs to be documentation for is vedit! Is any being written? In your comment on ``Can't yank inside global/macro'' you said that one can, of course, always yank to a named buffer. I checked, in the case of @a scripts, and the answer is yes and no: The yank works, but one still gets that diagnostic, and anything else in the sequence of commands is aborted. And concerning the diagnostic ``Can't undo in global commands'' -- I understand what you say about why one can't, but it still seems an unenlightening and perhaps inappropriate diagnostic to get when one has done @a and buffer a contains say x:g/foo/s//bar Here the initial ``x'', deleting the character the cursor was on when the command was given, creates the text-modified condition which, as I mention, seems to cause globals in @-scripts to give this diagnostic. I've just tested the above script -- several times, because I found it hard to believe what was happening -- and it did indeed give the diagnostic in question, but instead of replacing foo's with bar's, it replaced the first line containing a foo with a copy of the last line of the file! I leave it to you to figure that one out! (This is on version 3.9. Incidentally, I've used a true ^M in that script, so it is ready for you to try.) Further observations on some of the bugs I mentioned in my second letter: The business with yb is both more general and more complicated than I realized. The general fact seems to be that when one does a yank to a position earlier on the same line, the cursor does not move, but the editor thinks it has, so that any subsequent command, e.g. a movement, a delete, etc., has the effect that it would if it had been done with the cursor starting from the new position. The complication is that yanks to named buffers don't behave the same as simple yanks: They also leave the cursor unmoved, but what they actually yank into the buffer is the text from the cursor location to the end of the line; and whether they cause the editor to consider the cursor to be in a different location from where it really is seems to depend on the movement in question: with "ayNb, no, with "ayN|, yes (where N again stands for a positive integer). So what I called a snake in the grass seems to be a can of worms! In experimenting with this, incidentally, I've found that, while a put from a named buffer can be undone with u, if one attempts the command U, ``undo all changes made since coming onto this line'', only changes made since the last named-buffer-put are undone. I mentioned various abbreviations that were hard to unabbreviate, but could be done with the help of ^V -- but that I had found no way to undo :ab if if\pq To be precise, I had found that :una ^Vif, :una ^V^Vif, and :una ^Vif where each ^V represents two ^V's typed in, didn't work. I've finally found something that does: :una i^Vf. I find the diagnostic ``No tail recursion'' strange, since the occurrence of an abbreviation at the end of the item abbreviated shouldn't cause a recursive loop; while circumstances that do cause such a loop, namely the occurrence of the abbreviation within the item abbreviated, preceded by a space and followed by punctuation, are not censored, e.g. :ab x ( x ), or with more subtle effect, :ab x x sup 2. It seems that if one has :ab x word, then the abbreviation works when x is preceded by a beginning-of-insert, a space, a tab, or a newline, and followed by most any nonalphabetic character. For word-processing use, it would be natural to allow it to also be preceeded by (, ", `, [, perhaps even a user-specified set of characters. To my list of vi commands which I think should be able to take counts, add p and P. I often want to put in a large number of copies of one or more lines, as ``forms'' in which I will enter various kinds of additional data, and it would be convenient to be able to do compose such a line and then duplicate it the desired number of times all at once. What I currently do is produce a copy and then go yypk2yypk4yyp... till I have enough. Two more commands I think would be useful: (1) One which "puts" to the screen (in the sense of g/pattern/p, i.e. without affecting the buffer) the contents, or the first lines in cases that are longer than one line, of all nonempty buffers, named or numbered, with an indication, of course, of which each was. (Ideally one would like to be able to specify some subset of the buffers, whether one wants the first line, or everything, or some specified number of lines, etc..) (2) One which would show the script of the command stored to be used as ``.''. (And perhaps ditto with the last search pattern.) Oh, yes: it would also be nice to have a way to give commands without having them displace the one specified to be repeated by ``.'', e.g. if one wants to do a certain change time after time at various points in the file (using ``n'' and ``.'') but occasionally sees other changes to be made along the way. An alternative to the above would be a way to read the command stored to be used as ``.'' into a named buffer, so that one can give other commands and then return to that one. This would also allow one to reread the text of the command: useful if it isn't behaving as one meant it to. You might as well send any future mail to me here as gbergman@cartan -- I was using brahms while cartan's terminals were being moved but I'll probably be using cartan more from now on. But I'll check mail on both. (However, I'll be on vacation in New York State from the 3d to the 16th of August.) Best regards, George From gbergman@ucbcartan Fri Sep 23 13:57:06 1983 Date: Fri, 23 Sep 83 13:53:36 PDT From: gbergman@ucbcartan (George Mark Bergman) Message-Id: <8309232053.AA02960@ucbcartan.ARPA> Received: by ucbcartan.ARPA (4.6/3.7) id AA02960; Fri, 23 Sep 83 13:53:36 PDT Received: from ucbcartan.ARPA by UCBERNIE.ARPA (3.336/3.7) id AA20684; 23 Sep 83 13:57:01 PDT (Fri) To: cc-03@ucbcory, danh@kim, lindahl@topaz, mckusick@ucbernie, ralph@ucbarpa Status: RO Latest letter on ex~vi to Horton: Dear Mark, I hope I can assume that after my long letter to which you replied, you did get my two shorter notes (one sent the same day, the other about a week later); and that you've simply been to busy to think about them or reply. However, since I have had bad experiences with Mail, I am worried that either you may never have gotten them or that you may have replied and your replies not reached me. I hope you can send me a one-liner to relieve my doubts. Not many bugs to report this time, but I will modify or discuss a few of the suggestions I've made before. One that I want to modify is my suggestion that % should also match < with > and ` with '. All right in itself, but terrible if it would also mean that with :set match, every time one typed a > or a ' the cursor would look for the last occurrence of < or `, since <, >, and ' can occur in ways that have nothing to do with matching. So I think that making matchable pairs user-specifiable would be much more useful. Faith Fich (who send her regards -- she and Mike were our neighbors till a month ago, but have left for their new jobs at U. Seattle) suggested that user-settable matches would be particularly useful for those like herself who use distinct opening and closing delimiters with eqn. Actually, on further thought this isn't quite right, since eqn delimiters behave differently from parentheses -- if I want to use = as my delimiter to enter equation mode, and & as my delimiter to leave it, then eqn will not give any special treatment to &'s in text mode or ='s in equation mode; so perhaps a different kind of matching for that kind of use might be desirable; one would set :set dmatch =& for the delimiters suggested above; or :set dmatch $$ if one uses $ for both delimiters as in the eqn tutorials. I know checkeq is supposed to serve this function; but it would certainly be convenient to have it in the editor; and anyway, my experience was that checkeq didn't understand that one could have distinct opening and closing delimiters. (But this is a disinterested suggestion; I haven't used eqn for a long time; I put in my equations in raw troff.) The suggestion in my long original letter about an alternative to the diagnostic "using open mode" was based on a misunderstanding -- but one that might be worth making a reality. I had been assuming, since tutorials referred to the "three states" ex, vi, and open, and since I knew one could enter the editor by typing % ex filename or % vi filename, that a third way to enter the editor was by typing % open filename, and that this is what one should do when editing with a device not having an addressable cursor (which I had never done). So I supposed that the diagnostic "using open mode" would only be received if one were trying to use the wrong mode of the editor for the tty type shown in one's environment. Perhaps one should indeed have special way to enter the editor when one intended to edit in "open mode". (If "open" has other uses as a command, the command might be "vo" or "op".) Then if a user gave the command vi and his environment did not indicate a device with an addressable cursor, it would indeed be appropriate to give a diagnostic informing him that his environment was not compatible with this command. I suggest this because of the frustrating experiences I had as a beginner, of being put into open mode (because my tty type had somehow been lost) and not knowing what I could do about it. I am somewhat confused by your explanation that editor scripts that would go back and forth between ex and vi modes (using the :vi and Q commands) are impossible because "The internals of getting a character from the tty are completely different in ex and vi. Pushing input for one doesn't affect the other." A vi "script" (i.e. a mapping or a macro called by @x) can presently include bottom-line commands called with ":". I would suppose that the way such commands get characters from the terminal is about the same as the way ex does. You also write that "Any : command that outputs more than 1 line puts you into "ex mode"," -- is this a different sense of "ex mode"? If not, what happens when a vi script does this? I mentioned that the :so command from vi ignored anything after any global command in the sourced file. I also find that it will not accept the ex commands i and a; I realize that a prohibition against these is part of the way vi's bottom line differs from genuine ex. I've just done some experimenting, and I find that a script sourced from the bottom line of vi also interprets the command vi as a command to edit another file. Do we really want script files to be interpreted so differently when sourced from ex mode and from the bottom line of vi? Or if these differences are inevitable, can't the vi version at least have the positive difference of being able to go to visual mode with ^M^M or something, and continue following the script, interpreting characters as screen mode commands, just as a mapping or @x macro will if it has gone down to the bottom line with ":" and come back again? I am also still reluctant to give up the idea of a version of the editor in which there is no difference between the line-oriented mode and the bottom line of the visual mode. Suppose you took the editor as it exists now, and modified it so that in visual mode, ":" was interpreted as "Q", and in ex mode, ^[ was interpreted as ^Mvi. The one obvious disadvantage would be that it would insist on redrawing the screen after minor editing operations in ex mode. So suppose you made it remember the previous state of the screen every time it entered ex mode, and make the necessary adjustments on return to visual mode if the editing done was fairly trivial -- using the same criteria it now uses after bottom-line commands. What would be the problem? Of course, I know I am quite ignorant of what is involved, and what I am suggesting as possible may be quite impossible, or may have to wait for that far-off day when you rewrite the editor. You write that some of the restrictions on global commands are due to the fact that the state of the buffer before the command is stored in the "last deleted" buffer. But couldn't this situation be modified? Let's call the "last deleted" buffer buffer 0, since material from it seems to progress up to 1, 2, etc. with successive changes made. Suppose things were set up so that during the operation of a global or macro, deleted material, yanked material, etc. went by default into buffer 1 instead of 0...? Or alternatively, that the state before the global were saved in buffer 9, and this was then set outside of the chain of buffers into which material got pushed with successive deletions, during the operation of the global? Would this eliminate the objection to nested globals as well? Presumably, in a global within a global, the previous state would be stored in buffer 1, and new yanks put by default in buffer 2 (if the first of the above suggestions were followed). You mention in your comments to my letter that the difficulty of editing command lines "is a standard complaint about a moded editor" that would be hard to fix. But I point out in that same letter how appropriate macros or script-files can overcome that difficulty. (My files "do/1", "do/:" and the "map ^A ..." in my EXINIT.) It should be possible to introduce commands that would have the same effects as those macros. (Or at least, to describe such macros in the documentation. I have realized, by the way, that ^A was a poor choice of character for me to map, because it occurs in the output of function keys on a lot of terminals.) It would be nice to extend the kinds of "regular expressions" allowed in searches, e.g. a la egrep. Sometimes I also want to indicate something like "a substring of precisely 40 characters". This can be rendered as "........................................" at present, but things like that can easily lead to commands that exceed length limits. (Suppose one wants to look for a string of 40 successive characters other than whitespace. [^ ][^ ][^ ] ... is too long.) Have you checked out what I mentioned in my first long letter, that ed allows you to search for (e.g.) successive identical characters as \(.\)\1? Let me pass on a suggestion made recently in a system message by someone named nye. He was worried about what would happen if while one person was editing a file, someone else with write permission for that file also began editing it. Presumably, whoever did a "write" last would overwrite the other person's work. He suggested that there be a diagnostic "file has been changed since last write" for such a situation. (In particular, he wondered what would happen if he got his mail by editing /usr/spool/mail/name, rather than by using Mail -- which is certainly tempting if one is more comfortable with the editor than with Mail -- and if unbeknownst to him new mail arrived while he was editing.) I have discovered the hard way that when one has done a vi -r filename and decided one was satisfied with the version that was saved, ZZ or :x are n o t the same as :wq -- the first two discard the saved file, and only the last one writes it over the previous version. (I actually haven't checked this out on version 3.9. I tried to, and found that robert hadn't brought /usr/lib/ex3.9{preserve,recover}, if that's what they're still called, from arpa along with the rest of 3.9. But he's getting them for me.) If you don't want to make ZZ and :x actually equivalent to :wq, you might at least make them give error messages forcing the user to make the explicit choice of :wq to keep the saved version, or :q to discard it. I've been having fun using recursive "map" commands from time to time. For example, a file of mailing addresses of Math Departments is being prepared, for use with a new Computer Services command "label". This requires that each address constitute one line of the file, with the intended lines of each label separated by semicolons rather than newlines, each being no longer than 34 characters. At present, the lines in the file are of the form [university-name];[city-state-zip] and I wanted to find cases where the university name the secretaries had typed in was over 34 characters long. So I did :map ^A +36^V|F;^A0 Then, when I hit ^A, the cursor went bopping down the file, passing through each line at which it could find a ";" before the 36th position, but stopping when it hit a line where it could not. When I had corrected such a line, by abbreviating a long word or whatever, I could hit ^A again and find the next case. The "0" at the end of the map command is to prevent the "No tail recursion" rule from aborting my mapping. It has no other effect, because it is never reached. Of course, the above example of a recursive mapping is guaranteed to stop eventually, if only by reaching the end of the file. But I find that mappings that can go on indefinitely are very hard to stop. E.g. :map q ax^[q0 I've tried some like that, and sometimes hitting or whatever enough times successfully stops them (often leaving a bit of a mess -- characters that ought to have come after the string of x's embedded among them), while other times the only thing I can do is to go to another terminal and kill the editing process. Maybe it is to prevent things like that that you put in the "No tail recursion" rule -- though clearly it can be gotten around. Might it not be better to somehow make the editor not quite so deaf to s during such a process? I will mention one curious experience which I have not been able to reproduce. I use :set nu. After a delete near the beginning of a file, I found myself looking at a screen with line-numbers beginning around -2! The lines with nonpositive numbers were blank, if I recall; lines 1, 2, etc. were as they should be. I mentioned this to someone who said he'd had the same thing had happened at times, but had never been able to figure out what conditions caused it. Best regards, George From gbergman@ucbcartan Thu Nov 3 22:32:01 1983 Received: from ucbcartan.ARPA by ucbernie.ARPA (4.17/4.13) id AA25163; Thu, 3 Nov 83 22:31:44 pst Date: Thu, 3 Nov 83 22:28:11 PST From: gbergman@ucbcartan (George Mark Bergman) Message-Id: <8311040628.AA06097@ucbcartan.ARPA> Received: by ucbcartan.ARPA (4.6/3.7) id AA06097; Thu, 3 Nov 83 22:28:11 PST To: cc-03@BERKELEY, danh@kim, lindahl@topaz, mckusick@ucbernie, ralph@ucbarpa Status: R Latest letter to Mark Horton re editor bugs etc.: copies sent to cc-03@ucbcory ralph@ucbarpa, lindahl@topaz, mckusick@ucbernie 25/7/83 danh@kim 26/7 Mark Horton's reply: anderson@kim 30/7/83 Dear Mark, I got your reply to my last letter, but you don't say whether you got the two preceding ones -- in particular, I'm curious as to what you'd say about the peculiar behavior of abbreviations terminated by \[character]. (E.g., suppose someone did :ab ac artistic, and then, supposing the file was for troffing, typed ac\fP. -- Try it!) And likewise, about what happens if you hit @x, when register "x contains x:g/foo/s//bar (To see this wierd effect, you have to have a file with an occurrence of ``foo'', and a distinctive l a s t line.) In an earlier letter I commented that the mapping that the editor sets up to enable the tvi's right-arrow key makes ^L unavailable for refreshing the screen. You replied that, of course, the editor could not distinguish a ^L generated by the arrow key from one typed as -L by the user. True, but some users, such as myself, are quite happy using hjkl to move around the screen (I use them completely by reflex) and so have no wish to enable special arrow keys, but do want to have ^L available to redraw the screen when a message or something messes it up. The problem with ^L occurred in using a tvi; now, using a Z29 (in h19 mode), another version of the problem arises. It has arrow keys that transmit things like ^]A, ^]B, etc., and version 3.9 not only maps these sequences, but also map!s them. What I found happening is that if in typing I made a change somewhere in the body of a line and then wanted to add something at the end of the line, I would often type a ^] to end the first change and then an A to begin the second, with less than a second between them, and this sequence would then be map!ed into ^]ka or something, landing me in a different place from the one I wanted. Aside from this major problem, there is the minor inconvenience that the map! mechanism apparently waits a second every time I type an ^] from insert mode to see whether this is going to be one of the map!ed sequences, and this makes exiting from insert mode sluggish. My temporary solution has been to write a file of the form unmap ^]A| unmap ^]B| ...|unmap! ^]A| unmap! ^]B| which I source every time I go into vi on the Z29. I guess what I should do is create simplified termcaps that leave out the arrow keys for my own use when in version 3.7. Kevin Layer tells me that when he puts 3.9 on all systems here, there will be documentation on how to create terminfo files; so I will be able to avoid these inconveniences in 3.9 as well. What would be better, for these two-or-more character sequences, though, would be if the "timeout" feature on mappings could involve a variable interval, which could be set in the millisecond range for terminal-generated sequences (I suppose it would have to depend on the baud rate), and longer for user-mapped sequences. The likelihood of the user accidentally typing in one of the special sequences so fast is negligible. Incidentally, I notice that when I type :map to look at the automatic mappings, these are labeled "up", "down" etc., though for mappings that I create the corresponding position just shows a repeat of the mapped sequence. Is there any way the user can put mnemonics in with his own mappings? Two other minor points: In your reply to my first letter, where I suggested that N/pattern should take one to the Nth occurrence of /pattern, you said that N/pattern actually resets the window size to N while carrying one to /pattern. The tutorial says the same, I believe, but nonetheless, it doesn't work! When one has pulled a command into a buffer, say "x, and invoked it with @x, if one then tries to get a copy of this command by doing "xp, it doesn't seem to work. The way I've found to make it work is to do any other yank-and-put (using, say, the last-deleted buffer). This somehow unfreezes the mechanism, and (after undoing this last put, unless one wanted it), one can then successfully do "xp. Yours, George (Message inbox:32) Date: Sun, 10 Jun 84 16:29:50 pdt From: gbergman@ucbcartan (George Mark Bergman) To: cc-03@ucbcory, danh@kim, decvax!tarsa, gbergman@ucbcartan, hplabs!intelca!omsvax!isosvax!root, ihnp4!burl!we13!ltuxa!jab, leblanc@ucbdali, lindahl@topaz, mckusick@ucbernie, priili@Ardc.ARPA, ralph@ucbarpa, reiser@ruby, romine@seismo.ARPA, unisoft!eryk, uw-beaver!ubc-vision!mprvaxa!sonnens Subject: more editor bugs & ideas Here's another letter of comments on the editor that I'm sending to Mark Horton (mark@cbosgd). If any of you to whom I'm sending this aren't interested in staying on this mailing list, just let me know. Horton replied to an earlier letter by saying he had no idea when he'd have any time to work on the editor again, so I don't expect replies from him to this and further such letters in the near future. For those who are not familiar with the subject of item "A." below, modeline is a feature that he added without publicizing it much, whereby if any line in the vicinity of the top or bottom of the file (top and bottom 10 lines? I don't remember) contains the string vi: or ex: and then another :, everything between these is interpreted as a command and executed when this file is read by the editor. There was a big squall in net.news when someone discovered it by chance (an accidental string of this sort occurred in their /etc/password; fortunately the "command" was meaningless, and evoked a diagnostic from the editor). Some serious dangers of this feature were pointed out by various contributors, one of whom described for all who were interested how to eliminate it from the source file. Dear Mark, Here's another few month's collection of comments... A. Modeline I presume that in following the net.news discussion of the ``MAJOR BUG (modeline)'' you saw my two contributions; but I'll summarize the relevant points (not in the order I made them): 1) One possible feature that would be about as convenient as the modeline, and would avoid the dangers that people have pointed out, would be `enhanced tags', in which the 3d entry of a line of the tags file could be not merely a pattern search, but an arbitrary command line. 2) I described (both in net.news and in an earlier letter to you) a mapping in my EXINIT which makes one keystroke yank the current line (or the next N lines if preceded by a count) to a named buffer and then execute that buffer. If one keeps a set of initializing commands within a file to which they are to apply, one can then easily execute them on beginning an editing session, which gives almost the convenience of the modeline feature, without the dangers, and has an enormous range of other uses. So I think the modeline feature could be dropped. Let me add to those points: 3) A modeline vi:r %: leads to an infinite recursion! Fortunately, ^C cuts it off. 4) I agree with others' comments in net.news that if the modeline feature is not dropped altogether, it should be a settable option with default ``nomodeline''. 5) It should certainly be off when ``novice'' is set! B. Tags Having mentioned these in point A(1), I will give some other comments I have: One of your update documents mentions the fixing of a bug that left ``nomagic'' set if the file named in the tags file did not exist. Very good, but one still ends up with ``nomagic'' if the file exists but the pattern-search is unsuccessful! It would also be nice if the tags file could include file addresses of the form ~user/filename, in particular ~/filename, and if the command :set tags= recognized the same. (I suppose this makes no difference to people who get their tags from ctags, but for me, the tags file is maintained as a collection of items I have been working on recently, or mean to soon, and entries are put in or removed by hand regularly.) C. Reversal of n and N I also mentioned in one of my net.news comments a peculiar behavior that often seems to occur after I've been using my yank-and-execute mapping a lot, in which, after a command-line pattern-search :/pattern (rather than simply /pattern), the screen-mode commands n and N give searches in the reverse of the expected direction, with ? and / respectively instead of vice versa. Perhaps you can figure out what causes this; if not, would it help for me to do something like make a core dump of the editor when it is happening and send it to you? (I don't know how to send a nonascii file, though... .) A few more observations on this behavior: Though I commonly discover it when I am using my yank-and-execute mapping, it has happened on at least one occasion when I hadn't use that at all, so far as I could recall. It may actually happen quite frequently, but people just don't notice it because they usually use /pattern instead of :/pattern. (My mapping makes it more convenient for me to use the latter when the pattern is complicated, or I want to store it for repeated use.) D. Update on dangerous recursions Discovering that ^C interrupted the recursive modeline led me to test it out on a recursive mapping, :ab x ( x ). It interrupts it OK. Problem solved courtesy of 4.2BSD, I guess. I will collect under the next heading my usual list of: E. Minor bugs and modest suggestions ye and yE yank one character less than they should. If the command Ne (N a count) would land one on a 1-letter word, one generally lands at the end of the next word instead (even if it is also a 1-letter word. Exception: if one starts at or after the end of the preceding word, e behaves as it should.) Note also that in the line Sentence... . Sentence if the cursor is on the second S, the command ( causes no motion at all. I suggest that the motion commands { and } should accept indented lines as paragraph-starts, or at least that there should be some way of requesting this in the ":set para=" command. After all, these motions shouldn't be useful only to people writing troff files! In general, the command NC (where N is a count) changes material from the cursor position to the end of the Nth line, leaving material before the cursor on the current line unchanged. But if the line is indented, and the cursor is on or before the first nonwhite character, the preceding white text (spaces and tabs) is lost. It should be possible to use :unmap #1 :unmap! #1 :unab #1 when function-keys have been mapped. Sometimes a noisy phone-line, termcap padding errors, etc. cause just one or two lines of the screen to be messed up, and one may only wish to refresh those lines. Could a command be introduced which would do this? Ironically, on a dumb terminal one can generally do this by moving the cursor over the line, but not on a smart terminal. Another way one can do it is ddP, but I would sometimes feel uneasy about unnecessarily modifying the text. I would suggest that the form of the command be 1^L (2^L for 2 lines, etc.). Currently, ^L apparently ignores counts. (Actually, I'm writing at the moment on a tvi, so I've verified this for ^R rather than ^L.) If one uses the source command, and the file sourced contains a command :e newfile where newfile does not already exist, the diagnostic ``No such file or directory'' aborts the sourcing process. One ought to be able to use such commands in a sourced file. In vi screen command mode, ^[ is supposed to ``cancel partially formed commands'' and generally does so without protesting, but if the partially formed command is a count (e.g., if one has typed 110 instead of 10 and wishes to start over) it feeps, which depending on one's terminal can be a minor or a major annoyance. (Also depending on whether someone is trying to sleep in the next room.) The diagnostic, ``First address exceeds second'' should not be needed with one-address commands! The case where a series of addresses before a command, of which the first may exceed the second, is most useful is when the last address is a pattern-search preceded by a ";", e.g. :$;?^\.PP?-r otherfile but let me give simpler examples; of the two commands :3,1,2ka :1,3,2ka the first correctly marks line 2, but the second is aborted by the diagnostic quoted. F. A feature that would be very desirable, and might or might not be easy to implement. In general, when one is inserting text on a smart terminal in vi, the context below the text being added is pushed downward, line by line, till none is left on the screen. I would like a settable option that kept a certain number of lines of following context on the screen during additions. The point is that one should see the material that what one is writing will have to mesh with. What would be involved in implementing this would, of course, depend on terminal capabilities. If it would be difficult to keep an arbitrary number of lines, would it at least be possible to have an option that would keep one line, using the special bottom-line feature of some terminals? G. More radical suggestions (wishlist). 1) Editing text with _u_n_d_e_r_l_i_n_i_n_g. Although one of the valuable features of vi is the explicitness with which most nonprinting characters are shown, this can be annoying when one wants to deal with text in which many characters are ``emphasized'' using the sequence _^H; e.g. nroff output. I suggest a settable option under which such sequences would be shown as with ul. I realize that this would involve working out a great number of details, e.g. would motion commands treat _^Hx as one character or as three? How would the nth column be defined? How would one place the cursor on one of the elements of the string _^Hx for editing purposes? What would be done with _^H^A or _^H_^H....? I think the best solution would be to treat _^Hx as a single character for the purposes of motion commands, definition of nth column, deletions, etc. when this option was set. In terms of placing the cursor, two possibilities occur to me. One would be to only allow the cursor to sit on the underlined character ``as a whole'', and to have changes in underlining done by special commands: perhaps ^E as a toggle to turn emphasis on and off in insert mode, _ to change underlining in screen command mode as ~ changes capitalization ("_" is at present a synonym to "^", except that it takes counts. ^ could be modified to take counts, and _ then used as suggested above), and \e in replacement patterns. The other would be to consider a cursor sitting ``on'' a sequence _^Hx to actually be on the x, and to set things up so that if the cursor is on any of the other members of this sequence, the sequence is ``expanded'' on the screen, i.e. shown as it is in the present vi. Then define a single vi command so as not to skip over the _ and ^H in such a sequence; namely ^H. (This would mean making a distinction between h and ^H in screen command mode.) This one motion would allow one to edit parts of such a sequence. 2) Editing several files at once. When I have to do work that involves more than one file, the repeated use of :w|e#, yanking text to named buffers to move it, losing marked lines when I return to a previous file, etc. becomes annoying. I think it would be desirable if one could make a group of files behave like one file during an editing session, and move around within that file as comfortably as one move within one file. I suggest that visually, each file be separated from the next by a pattern ::::::::::::::::::: as when ``more'' is applied to a group of files. For ``Go to file 3, line 20'' I suggest a screen command syntax *3*20G. *-1* and *+2* could mean ``the preceding file'' and ``the file after next'' in such commands. (The initial * could be optional if there is no preceding + or -, as in the first example.) In a command such as :w, the default address range would be all of the file to which the current line belongs (i.e., it would be a synonym for :*.*w) To write all files, I suggest :**w. :q, :x and ZZ would quit the editing session entirely, while :*1*q would remove the buffer of file 1 from the object being edited. On the other hand, relative motions such as 10j, H, etc. would work within the ``visible object'', the union of the files. %s/pattern/repl/ would apply to the file containing the current line, and would have the synonym *.*s/pattern/repl/, while *1,2,4*s/pattern/repl/ would affect the 3 indicated files, and **s/pattern/repl/ would affect all files. 3) Input and output of shell escapes. The various commands involving shell escapes that you have set up allow four possible relations between the text being edited and the input and output of the commands: none (:!command); specified line-range as input with output not affecting text (:address-range w !command); no input from text but output inserted at specified line (:address r !command); and specified input from text with output replacing input (:address-range !command). It would be nice to have more flexibility; in particular, to be able to include input from the file and place the output somewhere else in the file without destroying the input text, and to input more than one segment of text, e.g. !egrep -f[addr.range] [other.addr.range] > [place of insertion] Obviously, there would be a problem of setting up a syntax that would avoid confusion with strings that look like address-ranges in the shell command. Perhaps \[...\] could enclose address-ranges where I have used [...] above. 4) ; The ; syntax, allowing one to do a pattern-search starting from a specified line is useful, but in setting up 2-address commands, one does not necessarily want the point from which one starts the search for the second address to be the first address. If this business were being set up now, I would suggest that address;/pattern/ should simply be a way of specifying the result of doing the indicated pattern-search relative to the indicated address, so that :address1,address2;/pattern/d would delete from address1 to the location found by the pattern-search relative to address2. Since people are used to the existing syntax of ;, I suggest that some other symbol be used in the above way, e.g. ], so that :address1,address2]/pattern/d could be interpreted as described. 5) Insertions into long lines on smart terminals at low or medium baud rate (e.g. 1200). This is annoying because the material coming after the point of insertion begins to wrap around, and the cursor must jump back and forth, inserting characters at the beginning of the continuation line, then going back to the point of insertion, and so on. (At least, this is my experience on my Z29. I haven't done editing by phone connection on any other smart terminal.) It's actually nicer on a dumb terminal, where the editor just overwrites, and shows you the result after you escape. I suppose that the need for the cursor to jump back and forth is due to the deficiency of the terminals -- has anyone suggested to terminal manufacturers that along with the wraparound feature, they add a feature which ``remembers'' when a line is a continuation of the preceding line, and automatically pushes material from the preceding line into the continuation line when characters are added to the former, eliminating the need to send all these instructions in over a slow line? (Do terminal manufacturers listen to editor-software specialists?) If not, it might just be best to not show the pushed-over earlier material till the insertion is complete. 6) Filename convention This is really a suggestion for UNIX generally, but it could be implemented on the editor separately. It is that for any file, filename/.. denote the directory in which the file lies. (This does not mean that every file should be treated as a directory, or that the ls command should show filename/..; it would just be a convenient way to refer to the directory containing a given nondirectory file, consistent with the existing convention for directories.) Within the editor, the important cases would be %/.. and #/.., allowing commands such as: :1,10w %/../othername All for now! Yours, George