source: doc/theses/andrew_beach_MMath/intro.tex @ 63b3279

ADTast-experimentalenumforall-pointer-decaypthread-emulationqualifiedEnum
Last change on this file since 63b3279 was 9cdfa5fb, checked in by Andrew Beach <ajbeach@…>, 3 years ago

Andrew MMath: Used (most of) Gregor's feedback to update the thesis. There are still a few \todo items as well as a general request for examples.

  • Property mode set to 100644
File size: 14.6 KB
Line 
1\chapter{Introduction}
2
3% The highest level overview of Cforall and EHMs. Get this done right away.
4This thesis covers the design and implementation of the exception handling
5mechanism (EHM) of
6\CFA (pronounced sea-for-all and may be written Cforall or CFA).
7\CFA is a new programming language that extends C, which maintains
8backwards-compatibility while introducing modern programming features.
9Adding exception handling to \CFA gives it new ways to handle errors and
10make large control-flow jumps.
11
12% Now take a step back and explain what exceptions are generally.
13Exception handling provides dynamic inter-function control flow.
14A language's EHM is a combination of language syntax and run-time
15components that construct, raise, propagate and handle exceptions,
16to provide all of that control flow.
17There are two forms of exception handling covered in this thesis:
18termination, which acts as a multi-level return,
19and resumption, which is a dynamic function call.
20% About other works:
21Often, when this separation is not made, termination exceptions are assumed
22as they are more common and may be the only form of handling provided in
23a language.
24
25All types of exception handling link a raise with a handler.
26Both operations are usually language primitives, although raises can be
27treated as a function that takes an exception argument.
28Handlers are more complex, as they are added to and removed from the stack
29during execution, must specify what they can handle and must give the code to
30handle the exception.
31
32Exceptions work with different execution models but for the descriptions
33that follow a simple call stack, with functions added and removed in a
34first-in-last-out order, is assumed.
35
36Termination exception handling searches the stack for the handler, then
37unwinds the stack to where the handler was found before calling it.
38The handler is run inside the function that defined it and when it finishes
39it returns control to that function.
40\begin{center}
41\input{termination}
42\end{center}
43%\todo{What does the right half of termination.fig mean?}
44
45Resumption exception handling searches the stack for a handler and then calls
46it without removing any other stack frames.
47The handler is run on top of the existing stack, often as a new function or
48closure capturing the context in which the handler was defined.
49After the handler has finished running, it returns control to the function
50that preformed the raise, usually starting after the raise.
51\begin{center}
52\input{resumption}
53\end{center}
54
55Although a powerful feature, exception handling tends to be complex to set up
56and expensive to use,
57so it is often limited to unusual or ``exceptional" cases.
58The classic example is error handling; exceptions can be used to
59remove error handling logic from the main execution path, and pay
60most of the cost only when the error actually occurs.
61
62\section{Thesis Overview}
63This work describes the design and implementation of the \CFA EHM.
64The \CFA EHM implements all of the common exception features (or an
65equivalent) found in most other EHMs and adds some features of its own.
66The design of all the features had to be adapted to \CFA's feature set, as
67some of the underlying tools used to implement and express exception handling
68in other languages are absent in \CFA.
69Still, the resulting syntax resembles that of other languages:
70\begin{cfa}
71try {
72        ...
73        T * object = malloc(request_size);
74        if (!object) {
75                throw OutOfMemory{fixed_allocation, request_size};
76        }
77        ...
78} catch (OutOfMemory * error) {
79        ...
80}
81\end{cfa}
82% A note that yes, that was a very fast overview.
83The design and implementation of all of \CFA's EHM's features are
84described in detail throughout this thesis, whether they are a common feature
85or one unique to \CFA.
86
87% The current state of the project and what it contributes.
88All of these features have been implemented in \CFA,
89covering both changes to the compiler and the run-time.
90In addition, a suite of test cases and performance benchmarks were created
91alongside the implementation.
92The implementation techniques are generally applicable in other programming
93languages and much of the design is as well.
94Some parts of the EHM use other features unique to \CFA and would be
95harder to replicate in other programming languages.
96
97The contributions of this work are:
98\begin{enumerate}
99\item Designing \CFA's exception handling mechanism, adapting designs from
100other programming languages and creating new features.
101\item Implementing stack unwinding and the \CFA EHM, including updating
102the \CFA compiler and the run-time environment.
103\item Designing and implementing a prototype virtual system.
104% I think the virtual system and per-call site default handlers are the only
105% "new" features, everything else is a matter of implementation.
106\item Creating tests to check the behaviour of the EHM.
107\item Creating benchmarks to check the performance of the EHM,
108as compared to other languages.
109\end{enumerate}
110
111The rest of this thesis is organized as follows.
112The current state of exceptions is covered in \autoref{s:background}.
113The existing state of \CFA is covered in \autoref{c:existing}.
114New EHM features are introduced in \autoref{c:features},
115covering their usage and design.
116That is followed by the implementation of these features in
117\autoref{c:implement}.
118Performance results are examined in \autoref{c:performance}.
119Possibilities to extend this project are discussed in \autoref{c:future}.
120Finally, the project is summarized in \autoref{c:conclusion}.
121
122\section{Background}
123\label{s:background}
124
125Exception handling has been examined before in programming languages,
126with papers on the subject dating back 70s.\cite{Goodenough75}
127Early exceptions were often treated as signals, which carried no information
128except their identity.
129Ada originally used this system\cite{Ada}, but now allows for a string
130message as a payload\cite{Ada12}.
131
132The modern flag-ship for termination exceptions is \Cpp,
133which added them in its first major wave of non-object-orientated features
134in 1990.\cite{CppHistory}
135Many EHMs have special exception types,
136however \Cpp has the ability to use any type as an exception.
137These were found to be not very useful and have been pushed aside for classes
138inheriting from
139\code{C++}{std::exception}.
140Although there is a special catch-all syntax (@catch(...)@), there are no
141operations that can be performed on the caught value, not even type inspection.
142Instead, the base exception-type \code{C++}{std::exception} defines common
143functionality (such as
144the ability to describe the reason the exception was raised) and all
145exceptions have this functionality.
146That trade-off, restricting usable types to gain guaranteed functionality,
147is almost universal now, as without some common functionality it is almost
148impossible to actually handle any errors.
149
150Java was the next popular language to use exceptions.\cite{Java8}
151Its exception system largely reflects that of \Cpp, except that it requires
152you throw a child type of \code{Java}{java.lang.Throwable}
153and it uses checked exceptions.
154Checked exceptions are part of a function's interface,
155the exception signature of the function.
156Every exception that could be raised from a function, either directly or
157because it is not handled from a called function, is given.
158Using this information, it is possible to statically verify if any given
159exception is handled, and guarantee that no exception will go unhandled.
160Making exception information explicit improves clarity and safety,
161but can slow down or restrict programming.
162For example, programming high-order functions becomes much more complex
163if the argument functions could raise exceptions.
164However, as odd it may seem, the worst problems are rooted in the simple
165inconvenience of writing and updating exception signatures.
166This has caused Java programmers to develop multiple programming ``hacks''
167to circumvent checked exceptions, negating their advantages.
168One particularly problematic example is the ``catch-and-ignore'' pattern,
169where an empty handler is used to handle an exception without doing any
170recovery or repair. In theory that could be good enough to properly handle
171the exception, but more often is used to ignore an exception that the       
172programmer does not feel is worth the effort of handling, for instance if
173they do not believe it will ever be raised.
174If they are incorrect, the exception will be silenced, while in a similar
175situation with unchecked exceptions the exception would at least activate   
176the language's unhandled exception code (usually, a program abort with an
177error message).
178
179%\subsection
180Resumption exceptions are less popular,
181although resumption is as old as termination; that is, few
182programming languages have implemented them.
183% http://bitsavers.informatik.uni-stuttgart.de/pdf/xerox/parc/techReports/
184%   CSL-79-3_Mesa_Language_Manual_Version_5.0.pdf
185Mesa is one programming language that did.\cite{Mesa} Experience with Mesa
186is quoted as being one of the reasons resumptions were not
187included in the \Cpp standard.
188% https://en.wikipedia.org/wiki/Exception_handling
189Since then, resumptions have been ignored in main-stream programming languages.
190However, resumption is being revisited in the context of decades of other
191developments in programming languages.
192While rejecting resumption may have been the right decision in the past,
193the situation has changed since then.
194Some developments, such as the functional programming equivalent to resumptions,
195algebraic effects\cite{Zhang19}, are enjoying success.
196A complete reexamination of resumption is beyond this thesis,
197but their reemergence is enough reason to try them in \CFA.
198% Especially considering how much easier they are to implement than
199% termination exceptions and how much Peter likes them.
200
201%\subsection
202Functional languages tend to use other solutions for their primary error
203handling mechanism, but exception-like constructs still appear.
204Termination appears in the error construct, which marks the result of an
205expression as an error; then the result of any expression that tries to use
206it also results in an error, and so on until an appropriate handler is reached.
207Resumption appears in algebraic effects, where a function dispatches its
208side-effects to its caller for handling.
209
210%\subsection
211More recently exceptions, seem to be vanishing from newer programming
212languages, replaced by ``panic".
213In Rust, a panic is just a program level abort that may be implemented by
214unwinding the stack like in termination exception
215handling.\cite{RustPanicMacro}\cite{RustPanicModule}
216Go's panic through is very similar to a termination, except it only supports
217a catch-all by calling \code{Go}{recover()}, simplifying the interface at
218the cost of flexibility.\cite{Go:2021}
219
220%\subsection
221As exception handling's most common use cases are in error handling,
222here are some other ways to handle errors with comparisons with exceptions.
223\begin{itemize}
224\item\emph{Error Codes}:
225This pattern has a function return an enumeration (or just a set of fixed
226values) to indicate if an error has occurred and possibly which error it was.
227
228Error codes mix exceptional/error and normal values, enlarging the range of
229possible return values. This can be addressed with multiple return values
230(or a tuple) or a tagged union.
231However, the main issue with error codes is forgetting to check them,
232which leads to an error being quietly and implicitly ignored.
233Some new languages and tools will try to issue warnings when an error code
234is discarded to avoid this problem.
235Checking error codes also bloats the main execution path,
236especially if the error is not handled immediately and has to be passed
237through multiple functions before it is addressed.
238
239\item\emph{Special Return with Global Store}:
240Similar to the error codes pattern but the function itself only returns
241that there was an error,
242and stores the reason for the error in a fixed global location.
243For example, many routines in the C standard library will only return some
244error value (such as -1 or a null pointer) and the error code is written into
245the standard variable @errno@.
246
247This approach avoids the multiple results issue encountered with straight
248error codes but otherwise has the same disadvantages and more.
249Every function that reads or writes to the global store must agree on all
250possible errors and managing it becomes more complex with concurrency.
251
252\item\emph{Return Union}:
253This pattern replaces error codes with a tagged union.
254Success is one tag and the errors are another.
255It is also possible to make each possible error its own tag and carry its own
256additional information, but the two-branch format is easy to make generic
257so that one type can be used everywhere in error handling code.
258
259This pattern is very popular in any functional or semi-functional language
260with primitive support for tagged unions (or algebraic data types).
261% We need listing Rust/rust to format code snippets from it.
262% Rust's \code{rust}{Result<T, E>}
263The main advantage is that an arbitrary object can be used to represent an
264error, so it can include a lot more information than a simple error code.
265The disadvantages include that the it does have to be checked along the main
266execution, and if there aren't primitive tagged unions proper, usage can be
267hard to enforce.
268
269\item\emph{Handler Functions}:
270This pattern associates errors with functions.
271On error, the function that produced the error calls another function to
272handle it.
273The handler function can be provided locally (passed in as an argument,
274either directly as as a field of a structure/object) or globally (a global
275variable).
276C++ uses this approach as its fallback system if exception handling fails,
277such as \snake{std::terminate} and, for a time,
278\snake{std::unexpected}.\footnote{\snake{std::unexpected} was part of the
279Dynamic Exception Specification, which has been removed from the standard
280as of C++20.\cite{CppExceptSpec}}
281
282Handler functions work a lot like resumption exceptions,
283but without the dynamic search for a handler.
284Since setting up the handler can be more complex/expensive,
285especially when the handler has to be passed through multiple layers of
286function calls, but cheaper (constant time) to call,
287they are more suited to more frequent (less exceptional) situations.
288\end{itemize}
289
290%\subsection
291Because of their cost, exceptions are rarely used for hot paths of execution.
292Hence, there is an element of self-fulfilling prophecy as implementation
293techniques have been focused on making them cheap to set-up,
294happily making them expensive to use in exchange.
295This difference is less important in higher-level scripting languages,
296where using exceptions for other tasks is more common.
297An iconic example is Python's
298\code{Python}{StopIteration}\cite{PythonExceptions} exception, that
299is thrown by an iterator to indicate that it is exhausted.
300When paired with Python's iterator-based for-loop, this will be thrown every
301time the end of the loop is reached.\cite{PythonForLoop}
Note: See TracBrowser for help on using the repository browser.