Re: Revised dohistexpand()

[Zoltan Hidvegi <hzoli@xxxxxxxxxx>]

Peter Stephenson wrote:
> One of the bits I like least about the history code (and that's a bold
> statement) is the zle history expansion.  The problem is that it
> relies on calling the lexer, and via that the history code, and those
> together build up a new editing line directly.  This is one more
> reason why the history code is hard to maintain, and is also very
> wasteful:  neither the resulting tokens nor the resulting history line
> are used.

I think that these few additional line for building the expanded line do not
make the history code much more difficult to maintain.  I agree that it is a
bit tricky, but once you understand it it is quite obvious.  And after your
history rearrangement it become a bit simpler that it used to be.  Also I
wrote long comments about this staff in hist.c to make understanding easier.
If something is still unclear ask me and hopefully I can explain it.

> This revised version eliminates the lexer and uses hgetc() to get
> history expansion directly.  As most of the work is now done specially
> in doexpandhist(), it was possible to rewrite it so that only history
> expansions in the current word are expanded, which is now in keeping
> with other forms of expansion and completion.  This alone should make
> it all worthwhile.  For example,
>   % cmd !!:1 z<TAB>
> will now do completion on the second argument, as you would expect;
> formerly it would expand the previous history reference first.

This doesn't really a problem if you use magic-space.  !!:1 is already
expanded when you try to expand z.

> There are a couple of drawbacks: single quotes don't hide this form of
> expansion (though again that's just like glob expansion: try typing
> '*' and then tab), and complex substitutions like !:s/a b/c d/ will
> sometimes be mishandled (though usually you get what you deserve,
> particularly if you've just finished typing the whole expression).  I
> don't think either of these is at all significant.  I have also made
> sure that magic-space still works.

That's my main problem.  Magic space is unusable if it expands single quoted
history references.  E.g. you couldn't type 'Hello!' (followed by a space).
And without the lexer there is no way to get it right as there are so many
cases: '!$', "'!$'", "`'!$'`" etc.  The other thing you mention that '*' is
turned to * after TAB is an other bug and it does not justify the bogous
behaviour of single-quoted bangs.

Also before beta5 (or beta6) doexpandhist() did not call the lexer.  But
before beta4 (as I remember) single quotes did not quote history references
either.  doexpandhist() was changed by Sven exactly because single quotes
disabled history expansion.

> There is still some small collusion with the history code: when
> expanding, the history line is not modified and the pair \! is left as
> it is instead of being turned into !.  These are algorithmically

That is the way it should work.  \! should remain unchanged.  Otherwise if you
type \!$<magic-space> it would turn to !$ and after RETURN it would change to
the last word of the previous line.

> simple and self-contained, however.  It may also be possible to
> simplify the qbang logic in hist.c (it's now not needed at all when
> expanding), but I don't know enough about that (yet).

This qbang is not too important in the stock beta10 as I remember, but it is
important for my releases to fix a bug when >\!foo expanded history (or
something like that).

To summarize my opinion, I do not like this patch very much.  It doesn't fixes
any bugs but it introduces one.  I normally bind magic-space to space so each
time I hit space the whole line I typed so far goes through the lexer but the
time spet there is unnoticable even on a slow 386/DX40 system (which I still
use at home).

Zoltan