Greg Smith wrote:
> On Thu, 6 Mar 2008, Heikki Linnakangas wrote:
>
>> At the most conservative end, we could fall back to the current
>> method on the first escape, quote or backslash character.
>
> I would just count the number of escaped/quote characters on each
> line, and then at the end of the line switch modes between the current
> code on the new version based on what the previous line looked like.
> That way the only additional overhead is a small bit only when escapes
> show up often, plus a touch more just once per line. Barely noticable
> in the case where nothing is escaped, very small regression for
> escape-heavy stuff but certainly better than the drop you reported in
> the last rev of this patch.
>
> Rev two of that design would keep a weighted moving average of the
> total number of escaped characters per line (say
> wma=(7*wma+current)/8) and switch modes based on that instead of the
> previous one. There's enough play in the transition between where the
> two approaches work better at that this should be easy enough to get a
> decent transition between. Based on your data I would put the
> transition at wma>4, which should keep the old code in play even if
> only half the lines have the bad regression that shows up with >8
> escapes per line.
>
>
I'd be inclined just to look at the first buffer of data we read in, and
make a one-off decision there, if we can get away with it. Then the cost
of testing is fixed rather than per line.
cheers
andrew
