diff options
Diffstat (limited to 'thesis/results.tex')
| -rw-r--r-- | thesis/results.tex | 55 |
1 files changed, 0 insertions, 55 deletions
diff --git a/thesis/results.tex b/thesis/results.tex index 76afb27..b826138 100644 --- a/thesis/results.tex +++ b/thesis/results.tex @@ -91,42 +91,6 @@ They are: A sequential version of Warshall's shortest path algorithm on a $6\times6$ matrix. \end{description} -\begin{figure*}[b!] - \centering - \begin{tikzpicture} - \begin{axis} - [ width=.5\textwidth - , xlabel={Size decrease (\%)} - , ylabel={Running time increase (\%)} - , mark options={scale=2} - , scatter/classes={% - a={mark=x,draw=red}, - b={mark=+,draw=black}} - ] - \addplot[scatter,only marks,scatter src=explicit symbolic] - table[meta=label] { - x y label - 12.8 11.9 a - 18.6 3.6 b - 9.8 5.1 a - 18.9 5.9 b - 18.9 3.7 b - 20.1 5.4 a - 22.9 0.9 a - 0.0 2.6 a - 20.0 5.2 b - 22.0 0.0 b - 12.1 8.0 a - }; - \end{axis} - \end{tikzpicture} - \caption{ - Comparison of code size decrease and running time increase. - Red crosses indicate programs that are smaller than 1000b in ARM; - black pluses correspond to larger programs. - \label{fig:results:scatter}} -\end{figure*} - \subsection{Code size} \label{sec:results:size} We compared the code size for the code generated by the ARM and Thumb backends for all programs in the test suite. @@ -155,25 +119,6 @@ The increase in running time is relatively low compared to the decrease in code For the larger programs, this is 3.7\% with a standard deviation of 2.04pp. (though this is based on only five programs). -\subsection{Discussion} -\label{sec:results:discussion} -If we compare the code size decrease with the running time increase, - we get the plot in \cref{fig:results:scatter}. -It seems that a high code size decrease often comes together with a low increase in running time, - suggesting that Thumb is more suitable for some programs than others. -However, we do not have enough data to see if this is significant. -More research is needed to see if there actually is such a relationship, - and if so, what programs are more suitable for Thumb. - -The outliers in \cref{fig:results:scatter} are - RFib $(0.0,2.6)$, - Ack $(12.8,11.9)$, - Fib $(9.8,5.1)$ and - Tak $(12.1,8.0)$, - all small programs that rely heavily on recursion. -This is explained by the relatively high number of jumps: - in Thumb we had to add several instructions, introducing extra code size and running time (see \cref{sec:storing-pc}). - \end{multicols} % vim: nowrap |