\chapter[chap:conclusions]{Conclusions}
-At the end of this research, we have created a system called Cλash, which
-allows us to translate hardware descriptions written in Haskell to be
-translated to \VHDL, and be programmed into an FPGA.
+The product of this research is a system called Cλash, which allows hardware
+descriptions written in Haskell to be translated to \VHDL and be programmed
+into an \small{FPGA}.
In this research, we have seen that a functional language is well suited
for hardware descriptions. Function applications provide elegant notation for
component instantiation and the various choice mechanisms (pattern matching,
case expressions, if expressions) are well suited to describe conditional
-assigment in the hardware.
+assignment in the hardware.
-Useful features from the functional perspective, like polymorphism and higher
-order functions and expressions are also well-suited for translation to
-hardware.
+Useful features from the functional perspective, like polymorphism and
+higher-order functions and expressions also prove suitable to describe
+hardware and our implementation shows that they can be translated to
+\VHDL\ as well.
-Using Haskell enabled a rapid prototype supporting complex features like
-polymorphism and higher order values. Mostly typechecker would have been
-complicated otherwise.
+A prototype compiler was created in this research. For this prototype the
+Haskell language was chosen as the input language, instead of creating a new
+language from scratch. This has enabled creating the prototype rapidly,
+allowing for experimenting with various functional language features and
+interpretations in Cλash. Even supporting more complex features like
+polymorphism and higher-order values has been possible. If a new language and
+compiler would have been designed from scratch, that new language would not
+have been nearly as advanced as the current version of Cλash.
-Haskell is not the perfect language. Some of the expressiveness it offers is
-not appropriate for hardware description, but also some extra syntax sugar
-could be useful. Lack of type-safe recursion is a big downside (but also a
-hard problem!), perhaps improvements to Haskell will fix this.
+However, Haskell might not have been the best choice for describing
+hardware. Some of the expressiveness it offers is not appropriate for
+hardware description (such as infinite recursion or recursive types),
+but also some extra syntax sugar could be useful (to reduce
+boilerplate).
-Explicit state descriptions...
+The lack of real dependent typing support in Haskell has been a burden.
+Haskell provides the tools to create some type level programming
+constructs (and is improving fast), but another language might have
+support for more advanced dependent types (and even type level operations) as
+a fundamental part of the language. The need for dependent typing is
+particularly present in Cλash to be able to fix some properties (list length,
+recursion depth, etc.) at compile time. Having better support for dependent
+typing might allow the use of type-safe recursion in Cλash, though this is
+fundamentally still a hard problem.
-Transformation based system is suitable, allows for reasoning about the
-system. Easy to specify transformation in the prototype. Implmentation is stil
-lacking, especially when combined with pattern matching.
+The choice of describing state very explicitly as extra arguments and
+results is a mixed blessing. It provides very explicit specification of
+state, which allows for very clear descriptions. This also allows for
+easy modification of the description in our normalization program, since
+state can be handled just like other arguments and results.
-General design (frontend + desugaring into small language, transformation
-system, simple backend) works well and should be preserved.
+On the other hand, the explicitness of the states and in particular
+substates, mean that more complex descriptions can become cumbersome
+very quickly. One finds that dealing with unpacking, passing, receiving
+and repacking becomes tedious and even error-prone. Removing some of this
+boilerplate would make the language even easier to use.
-Usefulness of Cλash is not completely clear yet. Advanced features work
-nicely, as well as choice, but no extensive testing (actual design projects)
-have been done yet. Use in education possible due to prototype.
+On the whole, the usefulness of Cλash for describing hardware is not
+completely clear yet. Most elements of the language have proven
+suitable, and even a real world hardware circuit (a reducer circuit
+\cite[baaij09]) has been implemented. However, the language has not been
+used during a complete design process, where its rapid prototyping and
+reusability qualities could become real advantages, or perhaps the state
+boilerplate or synchronicity limitations could become real problems.
-Limited to structural, single clock domain synchronous, etc.
+It is expected that Cλash will be used as a tool in education at the
+University of Twente soon. Hopefully this will provide a better insight
+in how the system performs.
+
+The prototype compiler has a clear design. Its front-end is taken from the \GHC\
+compiler and desugars Haskell into a small, but functional and typed
+language, called \emph{Core}. Cλash adds a transformation system that reduces
+this small language to a normal form and a simple back-end that performs a
+direct translation to \VHDL. This approach has worked well and should probably
+be preserved. Especially the transformation based normalization system is
+suitable. It is easy to implement a transformation in the prototype, though it
+is not trivial to maintain enough overview to guarantee that the system is
+correct and complete. In fact, the current set of transformations is probably
+not complete yet, in particular when stateful descriptions are involved.
+However, the system can be (and has been) described in a mathematical sense,
+allowing us to reason about it and probably also prove various correctness
+properties in the future.
+
+The scope of this research has been limited to structural descriptions that
+are synchronous in a single clock domain using cycle accurate designs. Even
+though this is a broad spectrum already, it may turn out that this scope is
+too narrow for practical use of Cλash. A common response from people that hear
+about using a functional language for hardware description is that they hope
+to be able to provide a concise mathematical description of their algorithm
+and have the hardware generated for them. Since this is obviously a different
+problem altogether, we could not have hoped to even start solving it. However,
+hopefully the current Cλash system provides a solid base on top of which
+further experimentation with functional descriptions can be built.
projects / matthijs / master-project / report.git / blobdiff