On Tue, Jan 6, 2015 at 12:33 PM, Johnny Billquist <bqt@update.uu.se> wrote:
On 2015-01-06 17:56, Dan Cross wrote:
I believe that Mary Ann is referring to repeatedly looking up
(presumably different) elements in the entry.  Assuming that e.g. `vi`
looks up O(n) elements, where $n$ is the number of elements, doing a
linear scan for each, you'd end up with quadratic behavior.

Assuming that you'd look up all the elements of the termcap entry at startup, and did each one from scratch, yes.

Yes.  Isn't that exactly what Mary Ann said was happening?  :-)

Hashing, or storing in some kind of balanced-tree like structure or
something, would of course help but would also necessitate doing a copy
and would entail some additional memory inefficiency.

Hashing would indeed cause some extra memory, but not necessarily any copying.

I fail to see how you can avoid copying the data out of the environment vector (unless you intend to either (a) turn the env var into a hash table, or (b) store pointers to the datum in the env var, but you'd need to encode their length somehow.  I don't think environment variables can contain embedded NULs, can they?).

But that would beg the question, why is vi doing a repeated scan of the terminal entry at startup, if not to find all the capabilities and store this somewhere? And if doing a look for all of them, why not scan the string from start to finish and store the information as it is found? At which point we move from quadratic to linear time.

I don't think she said it did things intelligently, just that that's how it did things.  :-)

But now we're getting into the innards of vi, which I never looked at anyway, and I guess is less relevant in this thread anyway.

The short of it (from what I got out of it) is that the move from termcap to terminfo was mostly motivated by attribute name changing away from fixed 2 character names.

A secondary motivation would be performance, but I don't really buy that one. Since we only moved to terminfo on systems with plenty of memory, performance of termcap could easily be on par anyway.

I tend to agree with you and I'll go one further: it seems that frequently we tend to identify a problem and then go to 11 with the "solution."  I can buy that termcap performance was an issue; I don't know that going directly to hashed terminfo files was the optimal solution.  A dbm file of termcap data and a hash table in whatever library parsed termcap would go a long way towards fixing the performance issues.  Did termcap have to be discarded just to add longer names?  I kind of tend to doubt it, but I wasn't there and don't know what the design criteria were, so my very-much-after-the-fact second guessing is just that.

Thanks for the insights.

        Johnny