\parskip .1in plus 5pt \baselineskip 12pt \lineskip 1pt
\newbox\bigstrutbox
\setbox\bigstrutbox=\hbox{\vrule height8.6pt depth3.6pt width0pt}
\def\bigstrut{\relax\ifmmode\copy\bigstrutbox\else\unhcopy\bigstrutbox\fi}

These charts are normalized. For each benchmark, the fastest time over all
of the implementations in that chart is found. For each implementation
reported, its actual time is divided by the best time, and that figure
appears in the chart.  Therefore, the best implementation in that chart
will have a $1.0$ reported for its time, and if some implementation is
twice as slow as that one, there will be a $2.0$ in the chart at that
point. Normalization is not done over more than 1 chart at a time.

The charts are broken up into 2~groups; {\bf CPU Time} and {\bf Real Time}.
{\bf CPU Time} is the best approximation to CPU time that was reported for that
implementation.
For the S-1, the LM-2, and the PERQ, the CPU time is approximated by real time; 
for InterLisp Vax~780, the CPU time is approximated by CPU time $+$ garbage
collection time; all others report CPU time directly.

{\bf Real Time} is the best approximation to real time available. For
PDP-10 MacLisp, the Symbolics 3600, the LM-2, the PERQ, and the S-1, this
is reported directly.  For the Dolphin, the Dandelion, the Dorado, PSL on
all machines, Franz on all machines, Common Lisp on the DEC machines, and
InterLisp Vax~780, this is CPU time $+$ garbage collection time.

In these charts, Franz time is reported for the Translink~$=$~T, LocalF~$=$~No
case.

\vfill\eject
\input nchart.tex
\bye