Presentation
authorpimjager <pim@pimjager.nl>
Thu, 14 Apr 2016 10:51:46 +0000 (12:51 +0200)
committerpimjager <pim@pimjager.nl>
Thu, 14 Apr 2016 10:51:46 +0000 (12:51 +0200)
deliverables/p2/p2.tex

index a3fca20..9f5fc65 100644 (file)
@@ -69,8 +69,11 @@ sem :: AST -> SemOutput
                Does matter for variables, not for functions.
                \pause
                \begin{itemize}
-                       \item Note that this means that \texttt{ones=1:ones} is 
+                       \item Note that this means that \CI{ones=1:ones} is 
                                        not allowed.
+                       \pause
+                       \item Functions which have not been encountered yet are temporarily
+                               typed $\alpha$.
                \end{itemize}
        \end{block}
        \pause%
@@ -120,8 +123,8 @@ flop(n, l) :: Int -> [Int] -> Bool {
 }
                \end{CleanCode}
                \pause
-               \item It is also correctly determined that \texttt{Bool} and the return 
-               type of \texttt{flop(n,l)} don't match.
+               \item It is also correctly determined that \CI{Bool} and the return 
+               type of \CI{flop(n,l)} don't match.
        \end{itemize}
 \end{frame}
 
@@ -175,7 +178,7 @@ id(x) :: a -> a {
 }
                \end{CleanCode}
                \pause
-               \item Is typed fun, but when we introduce: 
+               \item Is typed fine, but when we introduce: 
                \begin{CleanCode}
 var x = id(5);
 var y = id(True);
@@ -190,8 +193,71 @@ var y = id(True);
 \begin{frame}[fragile]
        \frametitle{But wait, there is more!}
        \framesubtitle{Trouble that is}
-       
-       \begin
+               \begin{block}{Our type inference algorithm is too greedy}
+                       It globally types a function once it is applied to a value, even
+                       if this types it more specified then needed. 
+               \end{block}
+               \pause
+               \begin{block}{Basically type inference for \CI{VarDecl} works great}
+                       All instances of VarDecl work well. Including those where a var is
+                       assigned by function application.
+               \end{block}
+               \pause
+               \begin{block}{Inference for functions is a completely different story}
+                       \begin{itemize}
+                               \item Type checking for functions works very well
+                               \item Type inference for functions works spotty at best
+                       \end{itemize}
+               \end{block}
+\end{frame}
+
+\begin{frame}[fragile]
+       \frametitle{We changed to much from the presented algorithms}
+       \begin{itemize}
+               \item Our type checker is written with the origin algorithm 
+                       \emph{in mind}
+               \pause
+               \item We were confident that that would be sufficient and we would be 
+                       able to implement type inference this way.
+               \pause
+               \item As it turns out, \emph{we couldn't}!
+       \end{itemize}
+\end{frame}
+
+\begin{frame}[fragile]
+       \frametitle{This not a problem for code generation}
+       \begin{block}{Sufficiently typed programs can be generated}
+               When all functions in the SPL program are completely typed then the
+               type inference algorithm yields a fully typed AST.\\  
+               \CI{VarDecls} \emph{types are correctly infered!}
+       \end{block}
+       \begin{block}{We can fix the type inference after code generation}
+               And this time do it right.
+       \end{block}
+\end{frame}
+
+\begin{frame}[fragile]
+       \frametitle{Doing it right}
+       \framesubtitle{How we will redo the type inference}
+       \begin{block}{Properly implement the \emph{exact} algorithm}
+               We will implement the Hindley-Miller algorithm exactly instead of 
+               \emph{``kinda''}
+       \end{block}
+       \begin{block}{Split constraint generation and solving}
+               Hide all the nasty details of constraint generation using a 
+               Reader-Writer-State-Transformer-Monad
+               \begin{description}
+                       \item[Reader] Reader environment to read fresh type variable
+                       \item[Writer] Write constraints to \CI{Constraint} environment
+                       \item[State] Gamma
+                       \item[Transformer] (Either SemError) 
+               \end{description}
+               \emph{Sadly these monads are not in the Clean library.}\\  
+               Then solve with a Relatively simple Solver-Monad
+               \begin{CleanCode}
+:: Solve a = StateT Unifier (Either TypeError) a
+               \end{CleanCode}
+       \end{block}
 \end{frame}
 
 % - Can functions that are defined later in a file call earlier defined functions?