diff options
Diffstat (limited to 'thesis/reg-alloc.tex')
| -rw-r--r-- | thesis/reg-alloc.tex | 109 |
1 files changed, 56 insertions, 53 deletions
diff --git a/thesis/reg-alloc.tex b/thesis/reg-alloc.tex index beb83c5..e90b36e 100644 --- a/thesis/reg-alloc.tex +++ b/thesis/reg-alloc.tex @@ -5,6 +5,7 @@ \small \begin{subfigure}{.5\linewidth} \begin{tikzpicture} + \ifcolours\else\selectcolormodel{gray}\fi \begin{axis} [ xbar , xmin=0 @@ -41,6 +42,7 @@ \end{subfigure}% \begin{subfigure}{.5\linewidth} \begin{tikzpicture} + \ifcolours\else\selectcolormodel{gray}\fi \begin{axis} [ xbar , xmin=0 @@ -153,7 +155,7 @@ In Clean programs, the ARM registers are used as follows~\parencite[\texttt{arms \begin{table*}[b!] \centering - \newcommand{\highlight}[1]{\textcolor{red}{\textbf{#1}}} + \newcommand{\highlight}[1]{\textbf{\ifcolours\textcolor{red}{#1}\else\underline{#1}\fi}} \begin{subfigure}{.5\linewidth} \centering \begin{tabular}{l >{\tt\footnotesize}c >{\tt\footnotesize}c} @@ -200,60 +202,9 @@ For small programs it may not be negligible, The counting method used here is rather simplistic: basing a register allocation on these counts alone means assuming that every instruction has a 16-bit variant, which is not the case. -A more accurate method would only count those instructions where using a high or low register actually makes a difference. +A more accurate method would be to only count those instructions where using a high or low register actually makes a difference. This is much more complicated, and for a rough estimate the simplistic method used here will already allow us to shrink down the code size. -\begin{figure*}[b] - \small - \centering - \begin{tikzpicture} - \begin{axis} - [ anchor=south east - , height=.25\linewidth - , width=.45\linewidth - , xbar - , xmin=0, xmax=6000000 - , xlabel={Bytes} - , ytick=\empty - , reverse stacked plots - , ylabel={Clean compiler} - , ymin=-1, ymax=1 - , axis lines*=left - , nodes near coords, nodes near coords align=west - , reverse legend - , every axis plot/.append style={point meta=explicit symbolic} - , legend style={at={($(1,-0.5)+(1.5em,0pt)$)},anchor=north,legend columns=-1} - ] - \addplot coordinates { (3785144,0) [80.7\%] }; - \addplot coordinates { (3827868,0) [81.6\%] }; - \addplot coordinates { (4385964,0) [93.5\%] }; - \addplot coordinates { (4692628,0) [100\%] }; - \legend{Thumb object-code (approximation), Thumb optimised, Thumb, ARM} - \end{axis} - \begin{axis} - [ anchor=south west - , height=.25\linewidth - , width=.45\linewidth - , xshift=3em - , xbar - , xmin=0, xmax=80000 - , xlabel={Bytes} - , ytick=\empty - , ymin=-1, ymax=1 - , ylabel={\texttt{frontend/scanner.icl}} - , axis lines*=left - , every axis plot/.append style={point meta=explicit symbolic} - , nodes near coords, nodes near coords align=west - ] - \addplot coordinates { (57296,0) [82.0\%] }; - \addplot coordinates { (57864,0) [82.8\%] }; - \addplot coordinates { (64588,0) [92.4\%] }; - \addplot coordinates { (69896,0) [100\%] }; - \end{axis} - \end{tikzpicture} - \caption{Code size for different backends.\label{fig:reg-alloc:results}} -\end{figure*} - \subsection{Optimisation} \label{sec:reg-alloc:optimisation} The register allocation used in the ARM backend is inefficient for Thumb-2 on several points (\cref{tab:reg-alloc:arm-and-thumb}). @@ -329,6 +280,58 @@ To measure the code size improvement introduced by this optimisation, With the new Thumb backend, with the optimised register allocation and leaving out any \ual{.align} directives to approximate object-code level efficiency. \end{itemize} +\begin{figure*}[b] + \small + \centering + \begin{tikzpicture} + \ifcolours\else\selectcolormodel{gray}\fi + \begin{axis} + [ anchor=south east + , height=.25\linewidth + , width=.45\linewidth + , xbar + , xmin=0, xmax=6000000 + , xlabel={Bytes} + , ytick=\empty + , reverse stacked plots + , ylabel={Clean compiler} + , ymin=-1, ymax=1 + , axis lines*=left + , nodes near coords, nodes near coords align=west + , reverse legend + , every axis plot/.append style={point meta=explicit symbolic} + , legend style={at={($(1,-0.5)+(1.5em,0pt)$)},anchor=north,legend columns=-1} + ] + \addplot coordinates { (3785144,0) [80.7\%] }; + \addplot coordinates { (3827868,0) [81.6\%] }; + \addplot coordinates { (4385964,0) [93.5\%] }; + \addplot coordinates { (4692628,0) [100\%] }; + \legend{Thumb object-code (approximation), Thumb optimised, Thumb, ARM} + \end{axis} + \begin{axis} + [ anchor=south west + , height=.25\linewidth + , width=.45\linewidth + , xshift=3em + , xbar + , xmin=0, xmax=80000 + , xlabel={Bytes} + , ytick=\empty + , ymin=-1, ymax=1 + , ylabel={\texttt{frontend/scanner.icl}} + , axis lines*=left + , every axis plot/.append style={point meta=explicit symbolic} + , nodes near coords, nodes near coords align=west + ] + \addplot coordinates { (57296,0) [82.0\%] }; + \addplot coordinates { (57864,0) [82.8\%] }; + \addplot coordinates { (64588,0) [92.4\%] }; + \addplot coordinates { (69896,0) [100\%] }; + \end{axis} + \end{tikzpicture} + \caption{Code size for different backends.\label{fig:reg-alloc:results}} +\end{figure*} + The object code generator has not been finished yet at the moment these measurements were done, so we can only estimate its efficiency. A modification described in \cref{sec:storing-pc:solution} required us to add \ual{.align} to jump instructions, adding a \ual{nop} instruction in approximately half the instances. |