source: doc/papers/concurrency/Paper.tex @ ae6b6cf

ADTarm-ehast-experimentalcleanup-dtorsenumforall-pointer-decayjacob/cs343-translationjenkins-sandboxnew-astnew-ast-unique-exprpthread-emulationqualifiedEnum
Last change on this file since ae6b6cf was a927662, checked in by Peter A. Buhr <pabuhr@…>, 6 years ago

more intro and coroutine changes

  • Property mode set to 100644
File size: 134.7 KB
Line 
1\documentclass[AMA,STIX1COL]{WileyNJD-v2}
2
3\articletype{RESEARCH ARTICLE}%
4
5\received{XXXXX}
6\revised{XXXXX}
7\accepted{XXXXX}
8
9\raggedbottom
10
11%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
12
13% Latex packages used in the document.
14
15\usepackage{epic,eepic}
16\usepackage{xspace}
17\usepackage{comment}
18\usepackage{upquote}                                            % switch curled `'" to straight
19\usepackage{listings}                                           % format program code
20\usepackage[labelformat=simple,aboveskip=0pt,farskip=0pt]{subfig}
21\renewcommand{\thesubfigure}{(\Alph{subfigure})}
22\captionsetup{justification=raggedright,singlelinecheck=false}
23\usepackage{dcolumn}                                            % align decimal points in tables
24\usepackage{capt-of}
25
26\hypersetup{breaklinks=true}
27\definecolor{OliveGreen}{cmyk}{0.64 0 0.95 0.40}
28\definecolor{Mahogany}{cmyk}{0 0.85 0.87 0.35}
29\definecolor{Plum}{cmyk}{0.50 1 0 0}
30
31\usepackage[pagewise]{lineno}
32\renewcommand{\linenumberfont}{\scriptsize\sffamily}
33
34\renewcommand{\topfraction}{0.8}                        % float must be greater than X of the page before it is forced onto its own page
35\renewcommand{\bottomfraction}{0.8}                     % float must be greater than X of the page before it is forced onto its own page
36\renewcommand{\floatpagefraction}{0.8}          % float must be greater than X of the page before it is forced onto its own page
37\renewcommand{\textfraction}{0.0}                       % the entire page maybe devoted to floats with no text on the page at all
38
39\lefthyphenmin=3                                                        % hyphen only after 4 characters
40\righthyphenmin=3
41
42%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
43
44% Names used in the document.
45
46\newcommand{\CFAIcon}{\textsf{C}\raisebox{\depth}{\rotatebox{180}{\textsf{A}}}\xspace} % Cforall symbolic name
47\newcommand{\CFA}{\protect\CFAIcon}             % safe for section/caption
48\newcommand{\CFL}{\textrm{Cforall}\xspace}      % Cforall symbolic name
49\newcommand{\Celeven}{\textrm{C11}\xspace}      % C11 symbolic name
50\newcommand{\CC}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}\xspace} % C++ symbolic name
51\newcommand{\CCeleven}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}11\xspace} % C++11 symbolic name
52\newcommand{\CCfourteen}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}14\xspace} % C++14 symbolic name
53\newcommand{\CCseventeen}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}17\xspace} % C++17 symbolic name
54\newcommand{\CCtwenty}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}20\xspace} % C++20 symbolic name
55\newcommand{\Csharp}{C\raisebox{-0.7ex}{\Large$^\sharp$}\xspace} % C# symbolic name
56
57%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
58
59\newcommand{\Textbf}[2][red]{{\color{#1}{\textbf{#2}}}}
60\newcommand{\Emph}[2][red]{{\color{#1}\textbf{\emph{#2}}}}
61\newcommand{\uC}{$\mu$\CC}
62\newcommand{\TODO}[1]{{\Textbf{#1}}}
63
64%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
65
66% Default underscore is too low and wide. Cannot use lstlisting "literate" as replacing underscore
67% removes it as a variable-name character so keywords in variables are highlighted. MUST APPEAR
68% AFTER HYPERREF.
69%\DeclareTextCommandDefault{\textunderscore}{\leavevmode\makebox[1.2ex][c]{\rule{1ex}{0.1ex}}}
70\renewcommand{\textunderscore}{\leavevmode\makebox[1.2ex][c]{\rule{1ex}{0.075ex}}}
71
72\renewcommand*{\thefootnote}{\Alph{footnote}} % hack because fnsymbol does not work
73%\renewcommand*{\thefootnote}{\fnsymbol{footnote}}
74
75\makeatletter
76% parindent is relative, i.e., toggled on/off in environments like itemize, so store the value for
77% use rather than use \parident directly.
78\newlength{\parindentlnth}
79\setlength{\parindentlnth}{\parindent}
80
81\newcommand{\LstBasicStyle}[1]{{\lst@basicstyle{\lst@basicstyle{#1}}}}
82\newcommand{\LstKeywordStyle}[1]{{\lst@basicstyle{\lst@keywordstyle{#1}}}}
83\newcommand{\LstCommentStyle}[1]{{\lst@basicstyle{\lst@commentstyle{#1}}}}
84
85\newlength{\gcolumnposn}                                        % temporary hack because lstlisting does not handle tabs correctly
86\newlength{\columnposn}
87\setlength{\gcolumnposn}{3.5in}
88\setlength{\columnposn}{\gcolumnposn}
89
90\newcommand{\C}[2][\@empty]{\ifx#1\@empty\else\global\setlength{\columnposn}{#1}\global\columnposn=\columnposn\fi\hfill\makebox[\textwidth-\columnposn][l]{\lst@basicstyle{\LstCommentStyle{#2}}}}
91\newcommand{\CRT}{\global\columnposn=\gcolumnposn}
92
93% Denote newterms in particular font and index them without particular font and in lowercase, e.g., \newterm{abc}.
94% The option parameter provides an index term different from the new term, e.g., \newterm[\texttt{abc}]{abc}
95% The star version does not lowercase the index information, e.g., \newterm*{IBM}.
96\newcommand{\newtermFontInline}{\emph}
97\newcommand{\newterm}{\@ifstar\@snewterm\@newterm}
98\newcommand{\@newterm}[2][\@empty]{\lowercase{\def\temp{#2}}{\newtermFontInline{#2}}\ifx#1\@empty\index{\temp}\else\index{#1@{\protect#2}}\fi}
99\newcommand{\@snewterm}[2][\@empty]{{\newtermFontInline{#2}}\ifx#1\@empty\index{#2}\else\index{#1@{\protect#2}}\fi}
100
101% Latin abbreviation
102\newcommand{\abbrevFont}{\textit}                       % set empty for no italics
103\@ifundefined{eg}{
104\newcommand{\EG}{\abbrevFont{e}\abbrevFont{g}}
105\newcommand*{\eg}{%
106        \@ifnextchar{,}{\EG}%
107                {\@ifnextchar{:}{\EG}%
108                        {\EG,\xspace}}%
109}}{}%
110\@ifundefined{ie}{
111\newcommand{\IE}{\abbrevFont{i}\abbrevFont{e}}
112\newcommand*{\ie}{%
113        \@ifnextchar{,}{\IE}%
114                {\@ifnextchar{:}{\IE}%
115                        {\IE,\xspace}}%
116}}{}%
117\@ifundefined{etc}{
118\newcommand{\ETC}{\abbrevFont{etc}}
119\newcommand*{\etc}{%
120        \@ifnextchar{.}{\ETC}%
121        {\ETC.\xspace}%
122}}{}%
123\@ifundefined{etal}{
124\newcommand{\ETAL}{\abbrevFont{et}~\abbrevFont{al}}
125\newcommand*{\etal}{%
126        \@ifnextchar{.}{\protect\ETAL}%
127                {\protect\ETAL.\xspace}%
128}}{}%
129\@ifundefined{viz}{
130\newcommand{\VIZ}{\abbrevFont{viz}}
131\newcommand*{\viz}{%
132        \@ifnextchar{.}{\VIZ}%
133                {\VIZ.\xspace}%
134}}{}%
135\makeatother
136
137\newenvironment{cquote}
138               {\list{}{\lstset{resetmargins=true,aboveskip=0pt,belowskip=0pt}\topsep=3pt\parsep=0pt\leftmargin=\parindentlnth\rightmargin\leftmargin}%
139                \item\relax}
140               {\endlist}
141
142%\newenvironment{cquote}{%
143%\list{}{\lstset{resetmargins=true,aboveskip=0pt,belowskip=0pt}\topsep=3pt\parsep=0pt\leftmargin=\parindentlnth\rightmargin\leftmargin}%
144%\item\relax%
145%}{%
146%\endlist%
147%}% cquote
148
149% CFA programming language, based on ANSI C (with some gcc additions)
150\lstdefinelanguage{CFA}[ANSI]{C}{
151        morekeywords={
152                _Alignas, _Alignof, __alignof, __alignof__, asm, __asm, __asm__, __attribute, __attribute__,
153                auto, _Bool, catch, catchResume, choose, _Complex, __complex, __complex__, __const, __const__,
154                coroutine, disable, dtype, enable, exception, __extension__, fallthrough, fallthru, finally,
155                __float80, float80, __float128, float128, forall, ftype, _Generic, _Imaginary, __imag, __imag__,
156                inline, __inline, __inline__, __int128, int128, __label__, monitor, mutex, _Noreturn, one_t, or,
157                otype, restrict, __restrict, __restrict__, __signed, __signed__, _Static_assert, thread,
158                _Thread_local, throw, throwResume, timeout, trait, try, ttype, typeof, __typeof, __typeof__,
159                virtual, __volatile, __volatile__, waitfor, when, with, zero_t},
160        moredirectives={defined,include_next}%
161}
162
163\lstset{
164language=CFA,
165columns=fullflexible,
166basicstyle=\linespread{0.9}\sf,                                                 % reduce line spacing and use sanserif font
167stringstyle=\tt,                                                                                % use typewriter font
168tabsize=5,                                                                                              % N space tabbing
169xleftmargin=\parindentlnth,                                                             % indent code to paragraph indentation
170%mathescape=true,                                                                               % LaTeX math escape in CFA code $...$
171escapechar=\$,                                                                                  % LaTeX escape in CFA code
172keepspaces=true,                                                                                %
173showstringspaces=false,                                                                 % do not show spaces with cup
174showlines=true,                                                                                 % show blank lines at end of code
175aboveskip=4pt,                                                                                  % spacing above/below code block
176belowskip=3pt,
177% replace/adjust listing characters that look bad in sanserif
178literate={-}{\makebox[1ex][c]{\raisebox{0.4ex}{\rule{0.8ex}{0.1ex}}}}1 {^}{\raisebox{0.6ex}{$\scriptstyle\land\,$}}1
179        {~}{\raisebox{0.3ex}{$\scriptstyle\sim\,$}}1 % {`}{\ttfamily\upshape\hspace*{-0.1ex}`}1
180        {<}{\textrm{\textless}}1 {>}{\textrm{\textgreater}}1
181        {<-}{$\leftarrow$}2 {=>}{$\Rightarrow$}2 {->}{\makebox[1ex][c]{\raisebox{0.5ex}{\rule{0.8ex}{0.075ex}}}\kern-0.2ex{\textrm{\textgreater}}}2,
182moredelim=**[is][\color{red}]{`}{`},
183}% lstset
184
185% uC++ programming language, based on ANSI C++
186\lstdefinelanguage{uC++}[ANSI]{C++}{
187        morekeywords={
188                _Accept, _AcceptReturn, _AcceptWait, _Actor, _At, _CatchResume, _Cormonitor, _Coroutine, _Disable,
189                _Else, _Enable, _Event, _Finally, _Monitor, _Mutex, _Nomutex, _PeriodicTask, _RealTimeTask,
190                _Resume, _Select, _SporadicTask, _Task, _Timeout, _When, _With, _Throw},
191}
192\lstdefinelanguage{Golang}{
193        morekeywords=[1]{package,import,func,type,struct,return,defer,panic,recover,select,var,const,iota,},
194        morekeywords=[2]{string,uint,uint8,uint16,uint32,uint64,int,int8,int16,int32,int64,
195                bool,float32,float64,complex64,complex128,byte,rune,uintptr, error,interface},
196        morekeywords=[3]{map,slice,make,new,nil,len,cap,copy,close,true,false,delete,append,real,imag,complex,chan,},
197        morekeywords=[4]{for,break,continue,range,goto,switch,case,fallthrough,if,else,default,},
198        morekeywords=[5]{Println,Printf,Error,},
199        sensitive=true,
200        morecomment=[l]{//},
201        morecomment=[s]{/*}{*/},
202        morestring=[b]',
203        morestring=[b]",
204        morestring=[s]{`}{`},
205}
206
207\lstnewenvironment{cfa}[1][]
208{\lstset{#1}}
209{}
210\lstnewenvironment{C++}[1][]                            % use C++ style
211{\lstset{language=C++,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
212{}
213\lstnewenvironment{uC++}[1][]
214{\lstset{#1}}
215{}
216\lstnewenvironment{Go}[1][]
217{\lstset{language=go,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
218{}
219\lstnewenvironment{python}[1][]
220{\lstset{language=python,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
221{}
222
223% inline code @...@
224\lstMakeShortInline@%
225
226\let\OLDthebibliography\thebibliography
227\renewcommand\thebibliography[1]{
228  \OLDthebibliography{#1}
229  \setlength{\parskip}{0pt}
230  \setlength{\itemsep}{4pt plus 0.3ex}
231}
232
233\title{\texorpdfstring{Advanced Control-flow and Concurrency in \protect\CFA}{Advanced Control-flow in Cforall}}
234
235\author[1]{Thierry Delisle}
236\author[1]{Peter A. Buhr*}
237\authormark{DELISLE \textsc{et al.}}
238
239\address[1]{\orgdiv{Cheriton School of Computer Science}, \orgname{University of Waterloo}, \orgaddress{\state{Waterloo, ON}, \country{Canada}}}
240
241\corres{*Peter A. Buhr, Cheriton School of Computer Science, University of Waterloo, 200 University Avenue West, Waterloo, ON, N2L 3G1, Canada. \email{pabuhr{\char`\@}uwaterloo.ca}}
242
243\fundingInfo{Natural Sciences and Engineering Research Council of Canada}
244
245\abstract[Summary]{
246\CFA is a modern, polymorphic, non-object-oriented, backwards-compatible extension of the C programming language.
247This paper discusses some advanced control-flow and concurrency/parallelism features in \CFA, along with the supporting runtime.
248These features are created from scratch because they do not exist in ISO C, or are low-level and/or unimplemented, so C programmers continue to rely on library features, like C pthreads.
249\CFA introduces language-level control-flow mechanisms, like coroutines, user-level threading, and monitors for mutual exclusion and synchronization.
250A unique contribution of this work is allowing multiple monitors to be safely acquired \emph{simultaneously} (deadlock free), while integrating this capability with monitor synchronization mechanisms.
251These features also integrate with the \CFA polymorphic type-system and exception handling, while respecting the expectations and style of C programmers.
252Experimental results show comparable performance of the new features with similar mechanisms in other concurrent programming-languages.
253}%
254
255\keywords{coroutines, concurrency, parallelism, threads, monitors, runtime, C, \CFA (Cforall)}
256
257
258\begin{document}
259\linenumbers                                            % comment out to turn off line numbering
260
261\maketitle
262
263
264\section{Introduction}
265
266This paper discusses the design of language-level control-flow and concurrency/parallelism extensions in \CFA and its runtime.
267\CFA is a modern, polymorphic, non-object-oriented\footnote{
268\CFA has features often associated with object-oriented programming languages, such as constructors, destructors, virtuals and simple inheritance.
269However, functions \emph{cannot} be nested in structures, so there is no lexical binding between a structure and set of functions (member/method) implemented by an implicit \lstinline@this@ (receiver) parameter.},
270backwards-compatible extension of the C programming language~\cite{Moss18}.
271Within the \CFA framework, new control-flow features are created from scratch.
272ISO \Celeven defines only a subset of the \CFA extensions, where the overlapping features are concurrency~\cite[\S~7.26]{C11}.
273However, \Celeven concurrency is largely wrappers for a subset of the pthreads library~\cite{Butenhof97,Pthreads}.
274Furthermore, \Celeven and pthreads concurrency is simple, based on thread fork/join in a function and a few locks, which is low-level and error prone;
275no high-level language concurrency features are defined.
276Interestingly, almost a decade after publication of the \Celeven standard, neither gcc-8, clang-8 nor msvc-19 (most recent versions) support the \Celeven include @threads.h@, indicating little interest in the C11 concurrency approach.
277Finally, while the \Celeven standard does not state a concurrent threading-model, the historical association with pthreads suggests implementations would adopt kernel-level threading (1:1)~\cite{ThreadModel}.
278
279In contrast, there has been a renewed interest during the past decade in user-level (M:N, green) threading in old and new programming languages.
280As multi-core hardware became available in the 1980/90s, both user and kernel threading were examined.
281Kernel threading was chosen, largely because of its simplicity and fit with the simpler operating systems and hardware architectures at the time, which gave it a performance advantage~\cite{Drepper03}.
282Libraries like pthreads were developed for C, and the Solaris operating-system switched from user (JDK 1.1~\cite{JDK1.1}) to kernel threads.
283As a result, languages like Java, Scala~\cite{Scala}, Objective-C~\cite{obj-c-book}, \CCeleven~\cite{C11}, and C\#~\cite{Csharp} adopted the 1:1 kernel-threading model, with a variety of presentation mechanisms.
284From 2000 onwards, languages like Go~\cite{Go}, Erlang~\cite{Erlang}, Haskell~\cite{Haskell}, D~\cite{D}, and \uC~\cite{uC++,uC++book} have championed the M:N user-threading model, and many user-threading libraries have appeared~\cite{Qthreads,MPC,BoostThreads}, including putting green threads back into Java~\cite{Quasar}.
285The main argument for user-level threading is that they are lighter weight than kernel threads (locking and context switching do not cross the kernel boundary), so there is less restriction on programming styles that encourage large numbers of threads performing smaller work-units to facilitate load balancing by the runtime~\cite{Verch12}.
286As well, user-threading facilitates a simpler concurrency approach using thread objects that leverage sequential patterns versus events with call-backs~\cite{vonBehren03}.
287Finally, performant user-threading implementations (both time and space) are largely competitive with direct kernel-threading implementations, while achieving the programming advantages of high concurrency levels and safety.
288
289A further effort over the past decade is the development of language memory-models to deal with the conflict between certain language features and compiler/hardware optimizations.
290This issue can be rephrased as: some language features are pervasive (language and runtime) and cannot be safely added via a library to prevent invalidation by sequential optimizations~\cite{Buhr95a,Boehm05}.
291The consequence is that a language must be cognizant of these features and provide sufficient tools to program around any safety issues.
292For example, C created the @volatile@ qualifier to provide correct execution for @setjmp@/@logjmp@ (concurrency came later).
293The common solution is to provide a handful of complex qualifiers and functions (e.g., @volatile@ and atomics) allowing programmers to write consistent/race-free programs, often in the sequentially-consistent memory-model~\cite{Boehm12}.
294
295While having a sufficient memory-model allows sound libraries to be constructed, writing these libraries can quickly become awkward and error prone, and using these low-level libraries has the same issues.
296Essentially, using low-level explicit locks is the concurrent equivalent of assembler programming.
297Just as most assembler programming is replaced with high-level programming, explicit locks can be replaced with high-level concurrency in a programming language.
298Then the goal is for the compiler to check for correct usage and follow any complex coding conventions implicitly.
299The drawback is that language constructs may preclude certain specialized techniques, therefore introducing inefficiency or inhibiting concurrency.
300For most concurrent programs, these drawbacks are insignificant in comparison to the speed of composition, and subsequent reliability and maintainability of the high-level concurrent program.
301(The same is true for high-level programming versus assembler programming.)
302Only very rarely should it be necessary to drop down to races and/or explicit locks to apply a specialized technique to achieve maximum speed or concurrency.
303As stated, this observation applies to non-concurrent forms of complex control-flow, like exception handling and coroutines.
304
305Adapting the programming language to these features also allows matching the control-flow model with the programming-language style, versus adopting one general (sound) library/paradigm.
306For example, it is possible to provide exceptions, coroutines, monitors, and tasks as specialized types in an object-oriented language, integrating these constructs to allow leveraging the type-system (static type-checking) and all other object-oriented capabilities~\cite{uC++}.
307It is also possible to leverage call/return for blocking communication via new control structures, versus switching to alternative communication paradigms, like channels or message passing.
308As well, user threading is often a complementary feature, allowing light-weight threading to match with low-cost objects, while hiding the application/kernel boundary.
309User threading also allows layering of implicit concurrency models (no explicit thread creation), such executors, data-flow, actors, into a single language, so programmers can chose the model that best fits an algorithm.\footnote{
310All implicit concurrency models have explicit threading in their implementation, and hence, can be build from explicit threading;
311however, the reverse is seldom true, i.e., given implicit concurrency, e.g., actors, it is virtually impossible to create explicit concurrency, e.g., blocking thread objects.}
312Finally, with extended language features and user-level threading it is possible to discretely fold locking and non-blocking I/O multiplexing into the language's I/O libraries, so threading implicitly dovetails with the I/O subsystem.
313\CFA embraces language extensions and user-level threading to provide advanced control-flow (exception handling\footnote{
314\CFA exception handling will be presented in a separate paper.
315The key feature that dovetails with this paper is non-local exceptions allowing exceptions to be raised across stacks, with synchronous exceptions raised among coroutines and asynchronous exceptions raised among threads, similar to that in \uC~\cite[\S~5]{uC++}
316} and coroutines) and concurrency.
317
318Most augmented traditional (Fortran 18~\cite{Fortran18}, Cobol 14~\cite{Cobol14}, Ada 12~\cite{Ada12}, Java 11~\cite{Java11}) and new languages (Go~\cite{Go}, Rust~\cite{Rust}, and D~\cite{D}), except \CC, diverge from C with different syntax and semantics, only interoperate indirectly with C, and are not systems languages, for those with managed memory.
319As a result, there is a significant learning curve to move to these languages, and C legacy-code must be rewritten.
320While \CC, like \CFA, takes an evolutionary approach to extend C, \CC's constantly growing complex and interdependent features-set (e.g., objects, inheritance, templates, etc.) mean idiomatic \CC code is difficult to use from C, and C programmers must expend significant effort learning \CC.
321Hence, rewriting and retraining costs for these languages, even \CC, are prohibitive for companies with a large C software-base.
322\CFA with its orthogonal feature-set, its high-performance runtime, and direct access to all existing C libraries circumvents these problems.
323
324We present comparative examples so the reader can judge if the \CFA control-flow extensions are equivalent or better than those in or proposed for \Celeven, \CC and other concurrent, imperative programming languages, and perform experiments to show the \CFA runtime is competitive with other similar mechanisms.
325The detailed contributions of this work are:
326\begin{itemize}
327\item
328allowing multiple monitors to be safely acquired \emph{simultaneously} (deadlock free), while seamlessly integrating this capability with all monitor synchronization mechanisms.
329\item
330all control-flow features respect the expectations of C programmers, with statically type-safe interfaces that integrate with the \CFA polymorphic type-system and other language features.
331\item
332experimental results show comparable performance of the new features with similar mechanisms in other concurrent programming-languages.
333\end{itemize}
334
335\begin{comment}
336This paper provides a minimal concurrency \newterm{Application Program Interface} (API) that is simple, efficient and can be used to build other concurrency features.
337While the simplest concurrency system is a thread and a lock, this low-level approach is hard to master.
338An easier approach for programmers is to support higher-level constructs as the basis of concurrency.
339Indeed, for highly-productive concurrent-programming, high-level approaches are much more popular~\cite{Hochstein05}.
340Examples of high-level approaches are jobs (thread pool)~\cite{TBB}, implicit threading~\cite{OpenMP}, monitors~\cite{Java}, channels~\cite{CSP,Go}, and message passing~\cite{Erlang,MPI}.
341
342The following terminology is used.
343A \newterm{thread} is a fundamental unit of execution that runs a sequence of code and requires a stack to maintain state.
344Multiple simultaneous threads give rise to \newterm{concurrency}, which requires locking to ensure access to shared data and safe communication.
345\newterm{Locking}, and by extension \newterm{locks}, are defined as a mechanism to prevent progress of threads to provide safety.
346\newterm{Parallelism} is running multiple threads simultaneously.
347Parallelism implies \emph{actual} simultaneous execution, where concurrency only requires \emph{apparent} simultaneous execution.
348As such, parallelism only affects performance, which is observed through differences in space and/or time at runtime.
349Hence, there are two problems to be solved: concurrency and parallelism.
350While these two concepts are often combined, they are distinct, requiring different tools~\cite[\S~2]{Buhr05a}.
351Concurrency tools handle mutual exclusion and synchronization, while parallelism tools handle performance, cost, and resource utilization.
352
353The proposed concurrency API is implemented in a dialect of C, called \CFA (pronounced C-for-all).
354The paper discusses how the language features are added to the \CFA translator with respect to parsing, semantics, and type checking, and the corresponding high-performance runtime-library to implement the concurrent features.
355\end{comment}
356
357
358\begin{comment}
359\section{\CFA Overview}
360
361The following is a quick introduction to the \CFA language, specifically tailored to the features needed to support concurrency.
362Extended versions and explanation of the following code examples are available at the \CFA website~\cite{Cforall} or in Moss~\etal~\cite{Moss18}.
363
364\CFA is a non-object-oriented extension of ISO-C, and hence, supports all C paradigms.
365Like C, the building blocks of \CFA are structures and routines.
366Virtually all of the code generated by the \CFA translator respects C memory layouts and calling conventions.
367While \CFA is not object oriented, lacking the concept of a receiver (\eg @this@) and nominal inheritance-relationships, C has a notion of objects: ``region of data storage in the execution environment, the contents of which can represent values''~\cite[3.15]{C11}.
368While some object-oriented features appear in \CFA, they are independent capabilities, allowing \CFA to adopt them while maintaining a procedural paradigm.
369
370
371\subsection{References}
372
373\CFA provides multi-level rebindable references, as an alternative to pointers, which significantly reduces syntactic noise.
374\begin{cfa}
375int x = 1, y = 2, z = 3;
376int * p1 = &x, ** p2 = &p1,  *** p3 = &p2,      $\C{// pointers to x}$
377    `&` r1 = x,   `&&` r2 = r1,   `&&&` r3 = r2;        $\C{// references to x}$
378int * p4 = &z, `&` r4 = z;
379
380*p1 = 3; **p2 = 3; ***p3 = 3;       // change x
381r1 =  3;     r2 = 3;      r3 = 3;        // change x: implicit dereferences *r1, **r2, ***r3
382**p3 = &y; *p3 = &p4;                // change p1, p2
383`&`r3 = &y; `&&`r3 = &`&`r4;             // change r1, r2: cancel implicit dereferences (&*)**r3, (&(&*)*)*r3, &(&*)r4
384\end{cfa}
385A reference is a handle to an object, like a pointer, but is automatically dereferenced the specified number of levels.
386Referencing (address-of @&@) a reference variable cancels one of the implicit dereferences, until there are no more implicit references, after which normal expression behaviour applies.
387
388
389\subsection{\texorpdfstring{\protect\lstinline{with} Statement}{with Statement}}
390\label{s:WithStatement}
391
392Heterogeneous data is aggregated into a structure/union.
393To reduce syntactic noise, \CFA provides a @with@ statement (see Pascal~\cite[\S~4.F]{Pascal}) to elide aggregate field-qualification by opening a scope containing the field identifiers.
394\begin{cquote}
395\vspace*{-\baselineskip}%???
396\lstDeleteShortInline@%
397\begin{cfa}
398struct S { char c; int i; double d; };
399struct T { double m, n; };
400// multiple aggregate parameters
401\end{cfa}
402\begin{tabular}{@{}l@{\hspace{2\parindentlnth}}|@{\hspace{2\parindentlnth}}l@{}}
403\begin{cfa}
404void f( S & s, T & t ) {
405        `s.`c; `s.`i; `s.`d;
406        `t.`m; `t.`n;
407}
408\end{cfa}
409&
410\begin{cfa}
411void f( S & s, T & t ) `with ( s, t )` {
412        c; i; d;                // no qualification
413        m; n;
414}
415\end{cfa}
416\end{tabular}
417\lstMakeShortInline@%
418\end{cquote}
419Object-oriented programming languages only provide implicit qualification for the receiver.
420
421In detail, the @with@-statement syntax is:
422\begin{cfa}
423$\emph{with-statement}$:
424        'with' '(' $\emph{expression-list}$ ')' $\emph{compound-statement}$
425\end{cfa}
426and may appear as the body of a routine or nested within a routine body.
427Each expression in the expression-list provides a type and object.
428The type must be an aggregate type.
429(Enumerations are already opened.)
430The object is the implicit qualifier for the open structure-fields.
431All expressions in the expression list are opened in parallel within the compound statement, which is different from Pascal, which nests the openings from left to right.
432
433
434\subsection{Overloading}
435
436\CFA maximizes the ability to reuse names via overloading to aggressively address the naming problem.
437Both variables and routines may be overloaded, where selection is based on number and types of returns and arguments (as in Ada~\cite{Ada}).
438\newpage
439\vspace*{-2\baselineskip}%???
440\begin{cquote}
441\begin{cfa}
442// selection based on type
443\end{cfa}
444\lstDeleteShortInline@%
445\begin{tabular}{@{}l@{\hspace{2\parindentlnth}}|@{\hspace{2\parindentlnth}}l@{}}
446\begin{cfa}
447const short int `MIN` = -32768;
448const int `MIN` = -2147483648;
449const long int `MIN` = -9223372036854775808L;
450\end{cfa}
451&
452\begin{cfa}
453short int si = `MIN`;
454int i = `MIN`;
455long int li = `MIN`;
456\end{cfa}
457\end{tabular}
458\begin{cfa}
459// selection based on type and number of parameters
460\end{cfa}
461\begin{tabular}{@{}l@{\hspace{2.7\parindentlnth}}|@{\hspace{2\parindentlnth}}l@{}}
462\begin{cfa}
463void `f`( void );
464void `f`( char );
465void `f`( int, double );
466\end{cfa}
467&
468\begin{cfa}
469`f`();
470`f`( 'a' );
471`f`( 3, 5.2 );
472\end{cfa}
473\end{tabular}
474\begin{cfa}
475// selection based on type and number of returns
476\end{cfa}
477\begin{tabular}{@{}l@{\hspace{2\parindentlnth}}|@{\hspace{2\parindentlnth}}l@{}}
478\begin{cfa}
479char `f`( int );
480double `f`( int );
481[char, double] `f`( int );
482\end{cfa}
483&
484\begin{cfa}
485char c = `f`( 3 );
486double d = `f`( 3 );
487[d, c] = `f`( 3 );
488\end{cfa}
489\end{tabular}
490\lstMakeShortInline@%
491\end{cquote}
492Overloading is important for \CFA concurrency since the runtime system relies on creating different types to represent concurrency objects.
493Therefore, overloading eliminates long prefixes and other naming conventions to prevent name clashes.
494As seen in Section~\ref{s:Concurrency}, routine @main@ is heavily overloaded.
495As another example, variable overloading is useful in the parallel semantics of the @with@ statement for fields with the same name:
496\begin{cfa}
497struct S { int `i`; int j; double m; } s;
498struct T { int `i`; int k; int m; } t;
499with ( s, t ) {
500        j + k;                                                                  $\C{// unambiguous, s.j + t.k}$
501        m = 5.0;                                                                $\C{// unambiguous, s.m = 5.0}$
502        m = 1;                                                                  $\C{// unambiguous, t.m = 1}$
503        int a = m;                                                              $\C{// unambiguous, a = t.m }$
504        double b = m;                                                   $\C{// unambiguous, b = s.m}$
505        int c = `s.i` + `t.i`;                                  $\C{// unambiguous, qualification}$
506        (double)m;                                                              $\C{// unambiguous, cast s.m}$
507}
508\end{cfa}
509For parallel semantics, both @s.i@ and @t.i@ are visible with the same type, so only @i@ is ambiguous without qualification.
510
511
512\subsection{Operators}
513
514Overloading also extends to operators.
515Operator-overloading syntax creates a routine name with an operator symbol and question marks for the operands:
516\begin{cquote}
517\lstDeleteShortInline@%
518\begin{tabular}{@{}ll@{\hspace{\parindentlnth}}|@{\hspace{\parindentlnth}}l@{}}
519\begin{cfa}
520int ++?(int op);
521int ?++(int op);
522int `?+?`(int op1, int op2);
523int ?<=?(int op1, int op2);
524int ?=? (int & op1, int op2);
525int ?+=?(int & op1, int op2);
526\end{cfa}
527&
528\begin{cfa}
529// unary prefix increment
530// unary postfix increment
531// binary plus
532// binary less than
533// binary assignment
534// binary plus-assignment
535\end{cfa}
536&
537\begin{cfa}
538struct S { int i, j; };
539S `?+?`( S op1, S op2) { // add two structures
540        return (S){op1.i + op2.i, op1.j + op2.j};
541}
542S s1 = {1, 2}, s2 = {2, 3}, s3;
543s3 = s1 `+` s2;         // compute sum: s3 == {2, 5}
544\end{cfa}
545\end{tabular}
546\lstMakeShortInline@%
547\end{cquote}
548
549
550\subsection{Constructors / Destructors}
551
552Object lifetime is a challenge in non-managed programming languages.
553\CFA responds with \CC-like constructors and destructors, using a different operator-overloading syntax.
554\begin{cfa}
555struct VLA { int len, * data; };                        $\C{// variable length array of integers}$
556void ?{}( VLA & vla ) with ( vla ) { len = 10;  data = alloc( len ); }  $\C{// default constructor}$
557void ?{}( VLA & vla, int size, char fill ) with ( vla ) { len = size;  data = alloc( len, fill ); } // initialization
558void ?{}( VLA & vla, VLA other ) { vla.len = other.len;  vla.data = other.data; } $\C{// copy, shallow}$
559void ^?{}( VLA & vla ) with ( vla ) { free( data ); } $\C{// destructor}$
560{
561        VLA  x,            y = { 20, 0x01 },     z = y; $\C{// z points to y}$
562        // $\LstCommentStyle{\color{red}\ \ \ x\{\};\ \ \ \ \ \ \ \ \ y\{ 20, 0x01 \};\ \ \ \ \ \ \ \ \ \ z\{ z, y \};\ \ \ \ \ \ \ implicit calls}$
563        ^x{};                                                                   $\C{// deallocate x}$
564        x{};                                                                    $\C{// reallocate x}$
565        z{ 5, 0xff };                                                   $\C{// reallocate z, not pointing to y}$
566        ^y{};                                                                   $\C{// deallocate y}$
567        y{ x };                                                                 $\C{// reallocate y, points to x}$
568        x{};                                                                    $\C{// reallocate x, not pointing to y}$
569}       //  $\LstCommentStyle{\color{red}\^{}z\{\};\ \ \^{}y\{\};\ \ \^{}x\{\};\ \ \ implicit calls}$
570\end{cfa}
571Like \CC, construction is implicit on allocation (stack/heap) and destruction is implicit on deallocation.
572The object and all their fields are constructed/destructed.
573\CFA also provides @new@ and @delete@ as library routines, which behave like @malloc@ and @free@, in addition to constructing and destructing objects:
574\begin{cfa}
575{
576        ... struct S s = {10}; ...                              $\C{// allocation, call constructor}$
577}                                                                                       $\C{// deallocation, call destructor}$
578struct S * s = new();                                           $\C{// allocation, call constructor}$
579...
580delete( s );                                                            $\C{// deallocation, call destructor}$
581\end{cfa}
582\CFA concurrency uses object lifetime as a means of mutual exclusion and/or synchronization.
583
584
585\subsection{Parametric Polymorphism}
586\label{s:ParametricPolymorphism}
587
588The signature feature of \CFA is parametric-polymorphic routines~\cite{Cforall} with routines generalized using a @forall@ clause (giving the language its name), which allow separately compiled routines to support generic usage over multiple types.
589For example, the following sum routine works for any type that supports construction from 0 and addition:
590\begin{cfa}
591forall( otype T | { void `?{}`( T *, zero_t ); T `?+?`( T, T ); } ) // constraint type, 0 and +
592T sum( T a[$\,$], size_t size ) {
593        `T` total = { `0` };                                    $\C{// initialize by 0 constructor}$
594        for ( size_t i = 0; i < size; i += 1 )
595                total = total `+` a[i];                         $\C{// select appropriate +}$
596        return total;
597}
598S sa[5];
599int i = sum( sa, 5 );                                           $\C{// use S's 0 construction and +}$
600\end{cfa}
601Type variables can be @otype@ or @dtype@.
602@otype@ refers to a \emph{complete type}, \ie, a type with size, alignment, default constructor, copy constructor, destructor, and assignment operator.
603@dtype@ refers to an \emph{incomplete type}, \eg, void or a forward-declared type.
604The builtin types @zero_t@ and @one_t@ overload constant 0 and 1 for a new types, where both 0 and 1 have special meaning in C.
605
606\CFA provides \newterm{traits} to name a group of type assertions, where the trait name allows specifying the same set of assertions in multiple locations, preventing repetition mistakes at each routine declaration:
607\begin{cfa}
608trait `sumable`( otype T ) {
609        void `?{}`( T &, zero_t );                              $\C{// 0 literal constructor}$
610        T `?+?`( T, T );                                                $\C{// assortment of additions}$
611        T ?+=?( T &, T );
612        T ++?( T & );
613        T ?++( T & );
614};
615forall( otype T `| sumable( T )` )                      $\C{// use trait}$
616T sum( T a[$\,$], size_t size );
617\end{cfa}
618
619Using the return type for overload discrimination, it is possible to write a type-safe @alloc@ based on the C @malloc@:
620\begin{cfa}
621forall( dtype T | sized(T) ) T * alloc( void ) { return (T *)malloc( sizeof(T) ); }
622int * ip = alloc();                                                     $\C{// select type and size from left-hand side}$
623double * dp = alloc();
624struct S {...} * sp = alloc();
625\end{cfa}
626where the return type supplies the type/size of the allocation, which is impossible in most type systems.
627\end{comment}
628
629
630\section{Coroutines: Stepping Stone}
631\label{coroutine}
632
633Coroutines are generalized routines allowing execution to be temporarily suspended and later resumed.
634Hence, unlike a normal routine, a coroutine may not terminate when it returns to its caller, allowing it to be restarted with the values and execution location present at the point of suspension.
635This capability is accomplished via the coroutine's stack, where suspend/resume context switch among stacks.
636Because threading design-challenges are present in coroutines, their design effort is relevant, and this effort can be easily exposed to programmers giving them a useful new programming paradigm because a coroutine handles the class of problems that need to retain state between calls, \eg plugins, device drivers, and finite-state machines.
637Therefore, the two fundamental features of the core \CFA coroutine-API are independent call-stacks and @suspend@/@resume@ operations.
638
639For example, a problem made easier with coroutines is unbounded generators, \eg generating an infinite sequence of Fibonacci numbers
640\begin{displaymath}
641\mathsf{fib}(n) = \left \{
642\begin{array}{ll}
6430                                       & n = 0         \\
6441                                       & n = 1         \\
645\mathsf{fib}(n-1) + \mathsf{fib}(n-2)   & n \ge 2       \\
646\end{array}
647\right.
648\end{displaymath}
649where Figure~\ref{f:C-fibonacci} shows conventional approaches for writing a Fibonacci generator in C.
650Figure~\ref{f:GlobalVariables} illustrates the following problems: unique unencapsulated global variables necessary to retain state between calls, only one Fibonacci generator, and execution state must be explicitly retained via explicit state variables.
651Figure~\ref{f:ExternalState} addresses these issues: unencapsulated program global variables become encapsulated structure variables, unique global variables are replaced by multiple Fibonacci objects, and explicit execution state is removed by precomputing the first two Fibonacci numbers and returning $\mathsf{fib}(n-2)$.
652
653\begin{figure}
654\centering
655\newbox\myboxA
656\begin{lrbox}{\myboxA}
657\begin{cfa}[aboveskip=0pt,belowskip=0pt]
658`int fn1, fn2, state = 1;`   // single global variables
659int fib() {
660        int fn;
661        `switch ( state )` {  // explicit execution state
662          case 1: fn = 0;  fn1 = fn;  state = 2;  break;
663          case 2: fn = 1;  fn2 = fn1;  fn1 = fn;  state = 3;  break;
664          case 3: fn = fn1 + fn2;  fn2 = fn1;  fn1 = fn;  break;
665        }
666        return fn;
667}
668int main() {
669
670        for ( int i = 0; i < 10; i += 1 ) {
671                printf( "%d\n", fib() );
672        }
673}
674\end{cfa}
675\end{lrbox}
676
677\newbox\myboxB
678\begin{lrbox}{\myboxB}
679\begin{cfa}[aboveskip=0pt,belowskip=0pt]
680#define FIB_INIT `{ 0, 1 }`
681typedef struct { int fn2, fn1; } Fib;
682int fib( Fib * f ) {
683
684        int ret = f->fn2;
685        int fn = f->fn1 + f->fn2;
686        f->fn2 = f->fn1; f->fn1 = fn;
687
688        return ret;
689}
690int main() {
691        Fib f1 = FIB_INIT, f2 = FIB_INIT;
692        for ( int i = 0; i < 10; i += 1 ) {
693                printf( "%d %d\n", fib( &fn1 ), fib( &f2 ) );
694        }
695}
696\end{cfa}
697\end{lrbox}
698
699\subfloat[3 states: global variables]{\label{f:GlobalVariables}\usebox\myboxA}
700\qquad
701\subfloat[1 state: encapsulated variables]{\label{f:ExternalState}\usebox\myboxB}
702\caption{C fibonacci fsm}
703\label{f:C-fibonacci}
704
705\bigskip
706
707\newbox\myboxA
708\begin{lrbox}{\myboxA}
709\begin{cfa}[aboveskip=0pt,belowskip=0pt]
710`coroutine` Fib { int fn; };
711void main( Fib & fib ) with( fib ) {
712        fn = 0;  int fn1 = fn; `suspend()`;
713        fn = 1;  int fn2 = fn1;  fn1 = fn; `suspend()`;
714        for () {
715                fn = fn1 + fn2; fn2 = fn1; fn1 = fn; `suspend()`; }
716}
717int next( Fib & fib ) with( fib ) { `resume( fib );` return fn; }
718int main() {
719        Fib f1, f2;
720        for ( 10 )
721                sout | next( f1 ) | next( f2 );
722}
723\end{cfa}
724\end{lrbox}
725\newbox\myboxB
726\begin{lrbox}{\myboxB}
727\begin{python}[aboveskip=0pt,belowskip=0pt]
728
729def Fibonacci():
730        fn = 0; fn1 = fn; `yield fn`  # suspend
731        fn = 1; fn2 = fn1; fn1 = fn; `yield fn`
732        while True:
733                fn = fn1 + fn2; fn2 = fn1; fn1 = fn; `yield fn`
734
735
736f1 = Fibonacci()
737f2 = Fibonacci()
738for i in range( 10 ):
739        print( `next( f1 )`, `next( f2 )` ) # resume
740
741\end{python}
742\end{lrbox}
743\subfloat[\CFA]{\label{f:Coroutine3States}\usebox\myboxA}
744\qquad
745\subfloat[Python]{\label{f:Coroutine1State}\usebox\myboxB}
746\caption{Fibonacci input coroutine, 3 states, internal variables}
747\label{f:cfa-fibonacci}
748\end{figure}
749
750Using a coroutine, it is possible to express the Fibonacci formula directly without any of the C problems.
751Figure~\ref{f:Coroutine3States} creates a @coroutine@ type, @`coroutine` Fib { int fn; }@, which provides communication, @fn@, for the \newterm{coroutine main}, @main@, which runs on the coroutine stack, and possibly multiple interface routines, \eg @next@.
752Like the structure in Figure~\ref{f:ExternalState}, the coroutine type allows multiple instances, where instances of this type are passed to the (overloaded) coroutine main.
753The coroutine main's stack holds the state for the next generation, @f1@ and @f2@, and the code represents the three states in the Fibonacci formula via the three suspend points, to context switch back to the caller's @resume@.
754The interface routine @next@, takes a Fibonacci instance and context switches to it using @resume@;
755on restart, the Fibonacci field, @fn@, contains the next value in the sequence, which is returned.
756The first @resume@ is special because it allocates the coroutine stack and cocalls its coroutine main on that stack;
757when the coroutine main returns, its stack is deallocated.
758Hence, @Fib@ is an object at creation, transitions to a coroutine on its first resume, and transitions back to an object when the coroutine main finishes.
759Figure~\ref{f:Coroutine1State} shows the coroutine version of the C version in Figure~\ref{f:ExternalState}.
760Coroutine generators are called \newterm{output coroutines} because values are only returned.
761
762Figure~\ref{f:CFAFmt} shows an \newterm{input coroutine}, @Format@, for restructuring text into groups of characters of fixed-size blocks.
763For example, the input of the left is reformatted into the output on the right.
764\begin{quote}
765\tt
766\begin{tabular}{@{}l|l@{}}
767\multicolumn{1}{c|}{\textbf{\textrm{input}}} & \multicolumn{1}{c}{\textbf{\textrm{output}}} \\
768abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyz
769&
770\begin{tabular}[t]{@{}lllll@{}}
771abcd    & efgh  & ijkl  & mnop  & qrst  \\
772uvwx    & yzab  & cdef  & ghij  & klmn  \\
773opqr    & stuv  & wxyz  &               &
774\end{tabular}
775\end{tabular}
776\end{quote}
777The example takes advantage of resuming a coroutine in the constructor to prime the loops so the first character sent for formatting appears inside the nested loops.
778The destructor provides a newline, if formatted text ends with a full line.
779Figure~\ref{f:CFmt} shows the C equivalent formatter, where the loops of the coroutine are flattened (linearized) and rechecked on each call because execution location is not retained between calls.
780(Linearized code is the bane of device drivers.)
781
782\begin{figure}
783\centering
784\newbox\myboxA
785\begin{lrbox}{\myboxA}
786\begin{cfa}[aboveskip=0pt,belowskip=0pt]
787`coroutine` Fmt {
788        char ch;   // communication variables
789        int g, b;   // needed in destructor
790};
791void main( Fmt & fmt ) with( fmt ) {
792        for () {
793                for ( g = 0; g < 5; g += 1 ) { // groups
794                        for ( b = 0; b < 4; b += 1 ) { // blocks
795                                `suspend();`
796                                sout | ch; } // print character
797                        sout | "  "; } // block separator
798                sout | nl; }  // group separator
799}
800void ?{}( Fmt & fmt ) { `resume( fmt );` } // prime
801void ^?{}( Fmt & fmt ) with( fmt ) { // destructor
802        if ( g != 0 || b != 0 ) // special case
803                sout | nl; }
804void send( Fmt & fmt, char c ) { fmt.ch = c; `resume( fmt )`; }
805int main() {
806        Fmt fmt;
807        sout | nlOff;   // turn off auto newline
808        for ( 41 )
809                send( fmt, 'a' );
810}
811\end{cfa}
812\end{lrbox}
813
814\newbox\myboxB
815\begin{lrbox}{\myboxB}
816\begin{python}[aboveskip=0pt,belowskip=0pt]
817
818
819
820def Fmt():
821        try:
822                while True:
823                        for g in range( 5 ):
824                                for b in range( 4 ):
825
826                                        print( `(yield)`, end='' )
827                                print( '  ', end='' )
828                        print()
829
830
831        except GeneratorExit:
832                if g != 0 | b != 0:
833                        print()
834
835
836fmt = Fmt()
837`next( fmt )`                    # prime
838for i in range( 41 ):
839        `fmt.send( 'a' );`      # send to yield
840
841\end{python}
842\end{lrbox}
843\subfloat[\CFA]{\label{f:CFAFmt}\usebox\myboxA}
844\qquad
845\subfloat[Python]{\label{f:CFmt}\usebox\myboxB}
846\caption{Output formatting text}
847\label{f:fmt-line}
848\end{figure}
849
850The previous examples are \newterm{asymmetric (semi) coroutine}s because one coroutine always calls a resuming routine for another coroutine, and the resumed coroutine always suspends back to its last resumer, similar to call/return for normal routines.
851However, @resume@ and @suspend@ context switch among existing stack-frames, rather than create new ones so there is no stack growth.
852\newterm{Symmetric (full) coroutine}s have a coroutine call to a resuming routine for another coroutine, and its coroutine main calls another resuming routine, which eventually forms a resuming-call cycle.
853(The trivial cycle is a coroutine resuming itself.)
854This control flow is similar to recursion for normal routines, but again there is no stack growth from the context switch.
855
856\begin{figure}
857\centering
858\lstset{language=CFA,escapechar={},moredelim=**[is][\protect\color{red}]{`}{`}}% allow $
859\begin{tabular}{@{}l@{\hspace{2\parindentlnth}}l@{}}
860\begin{cfa}
861`coroutine` Prod {
862        Cons & c;
863        int N, money, receipt;
864};
865void main( Prod & prod ) with( prod ) {
866        // 1st resume starts here
867        for ( i; N ) {
868                int p1 = random( 100 ), p2 = random( 100 );
869                sout | p1 | " " | p2;
870                int status = delivery( c, p1, p2 );
871                sout | " $" | money | nl | status;
872                receipt += 1;
873        }
874        stop( c );
875        sout | "prod stops";
876}
877int payment( Prod & prod, int money ) {
878        prod.money = money;
879        `resume( prod );`
880        return prod.receipt;
881}
882void start( Prod & prod, int N, Cons &c ) {
883        &prod.c = &c; // reassignable reference
884        prod.[N, receipt] = [N, 0];
885        `resume( prod );`
886}
887int main() {
888        Prod prod;
889        Cons cons = { prod };
890        start( prod, 5, cons );
891}
892\end{cfa}
893&
894\begin{cfa}
895`coroutine` Cons {
896        Prod & p;
897        int p1, p2, status;
898        bool done;
899};
900void ?{}( Cons & cons, Prod & p ) {
901        &cons.p = &p; // reassignable reference
902        cons.[status, done ] = [0, false];
903}
904void ^?{}( Cons & cons ) {}
905void main( Cons & cons ) with( cons ) {
906        // 1st resume starts here
907        int money = 1, receipt;
908        for ( ; ! done; ) {
909                sout | p1 | " " | p2 | nl | " $" | money;
910                status += 1;
911                receipt = payment( p, money );
912                sout | " #" | receipt;
913                money += 1;
914        }
915        sout | "cons stops";
916}
917int delivery( Cons & cons, int p1, int p2 ) {
918        cons.[p1, p2] = [p1, p2];
919        `resume( cons );`
920        return cons.status;
921}
922void stop( Cons & cons ) {
923        cons.done = true;
924        `resume( cons );`
925}
926\end{cfa}
927\end{tabular}
928\caption{Producer / consumer: resume-resume cycle, bi-directional communication}
929\label{f:ProdCons}
930\end{figure}
931
932Figure~\ref{f:ProdCons} shows a producer/consumer symmetric-coroutine performing bi-directional communication.
933Since the solution involves a full-coroutining cycle, the program main creates one coroutine in isolation, passes this coroutine to its partner, and closes the cycle at the call to @start@.
934The @start@ routine communicates both the number of elements to be produced and the consumer into the producer's coroutine-structure.
935Then the @resume@ to @prod@ creates @prod@'s stack with a frame for @prod@'s coroutine main at the top, and context switches to it.
936@prod@'s coroutine main starts, creates local variables that are retained between coroutine activations, and executes $N$ iterations, each generating two random values, calling the consumer to deliver the values, and printing the status returned from the consumer.
937
938The producer call to @delivery@ transfers values into the consumer's communication variables, resumes the consumer, and returns the consumer status.
939For the first resume, @cons@'s stack is initialized, creating local variables retained between subsequent activations of the coroutine.
940The consumer iterates until the @done@ flag is set, prints the values delivered by the producer, increments status, and calls back to the producer via @payment@, and on return from @payment@, prints the receipt from the producer and increments @money@ (inflation).
941The call from the consumer to @payment@ introduces the cycle between producer and consumer.
942When @payment@ is called, the consumer copies values into the producer's communication variable and a resume is executed.
943The context switch restarts the producer at the point where it last context switched, so it continues in @delivery@ after the resume.
944
945@delivery@ returns the status value in @prod@'s coroutine main, where the status is printed.
946The loop then repeats calling @delivery@, where each call resumes the consumer coroutine.
947The context switch to the consumer continues in @payment@.
948The consumer increments and returns the receipt to the call in @cons@'s coroutine main.
949The loop then repeats calling @payment@, where each call resumes the producer coroutine.
950
951After iterating $N$ times, the producer calls @stop@.
952The @done@ flag is set to stop the consumer's execution and a resume is executed.
953The context switch restarts @cons@ in @payment@ and it returns with the last receipt.
954The consumer terminates its loops because @done@ is true, its @main@ terminates, so @cons@ transitions from a coroutine back to an object, and @prod@ reactivates after the resume in @stop@.
955@stop@ returns and @prod@'s coroutine main terminates.
956The program main restarts after the resume in @start@.
957@start@ returns and the program main terminates.
958
959One \emph{killer} application for a coroutine is device drivers, which at one time caused 70\%-85\% of failures in Windows/Linux~\cite{Swift05}.
960Many device drivers are a finite state-machine parsing a protocol, e.g.:
961\begin{tabbing}
962\ldots STX \= \ldots message \ldots \= ESC \= ETX \= \ldots message \ldots  \= ETX \= 2-byte crc \= \ldots      \kill
963\ldots STX \> \ldots message \ldots \> ESC \> ETX \> \ldots message \ldots  \> ETX \> 2-byte crc \> \ldots
964\end{tabbing}
965where a network message begins with the control character STX and ends with an ETX, followed by a 2-byte cyclic-redundancy check.
966Control characters may appear in a message if preceded by an ESC.
967Because FSMs can be complex and occur frequently in important domains, direct support of the coroutine is crucial in a systems programminglanguage.
968
969\begin{figure}
970\begin{cfa}
971enum Status { CONT, MSG, ESTX, ELNTH, ECRC };
972`coroutine` Driver {
973        Status status;
974        char * msg, byte;
975};
976void ?{}( Driver & d, char * m ) { d.msg = m; }         $\C[3.0in]{// constructor}$
977Status next( Driver & d, char b ) with( d ) {           $\C{// 'with' opens scope}$
978        byte = b; `resume( d );` return status;
979}
980void main( Driver & d ) with( d ) {
981        enum { STX = '\002', ESC = '\033', ETX = '\003', MaxMsg = 64 };
982        unsigned short int crc;                                                 $\C{// error checking}$
983  msg: for () {                                                                         $\C{// parse message}$
984                status = CONT;
985                unsigned int lnth = 0, sum = 0;
986                while ( byte != STX ) `suspend();`
987          emsg: for () {
988                        `suspend();`                                                    $\C{// process byte}$
989                        choose ( byte ) {                                               $\C{// switch with default break}$
990                          case STX:
991                                status = ESTX; `suspend();` continue msg;
992                          case ETX:
993                                break emsg;
994                          case ESC:
995                                suspend();
996                        } // choose
997                        if ( lnth >= MaxMsg ) {                                 $\C{// buffer full ?}$
998                                status = ELNTH; `suspend();` continue msg; }
999                        msg[lnth++] = byte;
1000                        sum += byte;
1001                } // for
1002                msg[lnth] = '\0';                                                       $\C{// terminate string}\CRT$
1003                `suspend();`
1004                crc = (unsigned char)byte << 8; // prevent sign extension for signed char
1005                `suspend();`
1006                status = (crc | (unsigned char)byte) == sum ? MSG : ECRC;
1007                `suspend();`
1008        } // for
1009}
1010\end{cfa}
1011\caption{Device driver for simple communication protocol}
1012\end{figure}
1013
1014
1015\subsection{Coroutine Implementation}
1016
1017A significant implementation challenge for coroutines (and threads, see Section~\ref{threads}) is adding extra fields and executing code after/before the coroutine constructor/destructor and coroutine main to create/initialize/de-initialize/destroy extra fields and the stack.
1018There are several solutions to this problem and the chosen option forced the \CFA coroutine design.
1019
1020Object-oriented inheritance provides extra fields and code in a restricted context, but it requires programmers to explicitly perform the inheritance:
1021\begin{cfa}[morekeywords={class,inherits}]
1022class mycoroutine inherits baseCoroutine { ... }
1023\end{cfa}
1024and the programming language (and possibly its tool set, \eg debugger) may need to understand @baseCoroutine@ because of the stack.
1025Furthermore, the execution of constructors/destructors is in the wrong order for certain operations.
1026For example, for threads if the thread is implicitly started, it must start \emph{after} all constructors, because the thread relies on a completely initialized object, but the inherited constructor runs \emph{before} the derived.
1027
1028An alternative is composition:
1029\begin{cfa}
1030struct mycoroutine {
1031        ... // declarations
1032        baseCoroutine dummy; // composition, last declaration
1033}
1034\end{cfa}
1035which also requires an explicit declaration that must be the last one to ensure correct initialization order.
1036However, there is nothing preventing wrong placement or multiple declarations.
1037
1038For coroutines as for threads, many implementations are based on routine pointers or routine objects~\cite{Butenhof97, C++14, MS:VisualC++, BoostCoroutines15}.
1039For example, Boost implements coroutines in terms of four functor object-types:
1040\begin{cfa}
1041asymmetric_coroutine<>::pull_type
1042asymmetric_coroutine<>::push_type
1043symmetric_coroutine<>::call_type
1044symmetric_coroutine<>::yield_type
1045\end{cfa}
1046Similarly, the canonical threading paradigm is often based on routine pointers, \eg @pthreads@~\cite{Butenhof97}, \Csharp~\cite{Csharp}, Go~\cite{Go}, and Scala~\cite{Scala}.
1047However, the generic thread-handle (identifier) is limited (few operations), unless it is wrapped in a custom type.
1048\begin{cfa}
1049void mycor( coroutine_t cid, void * arg ) {
1050        int * value = (int *)arg;                               $\C{// type unsafe, pointer-size only}$
1051        // Coroutine body
1052}
1053int main() {
1054        int input = 0, output;
1055        coroutine_t cid = coroutine_create( &mycor, (void *)&input ); $\C{// type unsafe, pointer-size only}$
1056        coroutine_resume( cid, (void *)input, (void **)&output ); $\C{// type unsafe, pointer-size only}$
1057}
1058\end{cfa}
1059Since the custom type is simple to write in \CFA and solves several issues, added support for routine/lambda-based coroutines adds very little.
1060
1061Note, the type @coroutine_t@ must be an abstract handle to the coroutine, because the coroutine descriptor and its stack are non-copyable.
1062Copying the coroutine descriptor results in copies being out of date with the current state of the stack.
1063Correspondingly, copying the stack results is copies being out of date with the coroutine descriptor, and pointers in the stack being out of date to data on the stack.
1064(There is no mechanism in C to find all stack-specific pointers and update them as part of a copy.)
1065
1066The selected approach is to use language support by introducing a new kind of aggregate (structure):
1067\begin{cfa}
1068coroutine Fibonacci {
1069        int fn; // communication variables
1070};
1071\end{cfa}
1072The @coroutine@ keyword means the compiler (and tool set) can find and inject code where needed.
1073The downside of this approach is that it makes coroutine a special case in the language.
1074Users wanting to extend coroutines or build their own for various reasons can only do so in ways offered by the language.
1075Furthermore, implementing coroutines without language supports also displays the power of a programming language.
1076While this is ultimately the option used for idiomatic \CFA code, coroutines and threads can still be constructed without language support.
1077The reserved keyword simply eases use for the common case.
1078
1079Part of the mechanism to generalize coroutines is using a \CFA trait, which defines a coroutine as anything satisfying the trait @is_coroutine@, and this trait restricts the available set of coroutine-manipulation routines:
1080\begin{cfa}
1081trait is_coroutine( `dtype` T ) {
1082        void main( T & );
1083        coroutine_desc * get_coroutine( T & );
1084};
1085forall( `dtype` T | is_coroutine(T) ) void suspend( T & );
1086forall( `dtype` T | is_coroutine(T) ) void resume( T & );
1087\end{cfa}
1088The @dtype@ property provides no implicit copying operations and the @is_coroutine@ trait provides no explicit copying operations, so all coroutines must be passed by reference (pointer).
1089The routine definitions ensures there is a statically-typed @main@ routine that is the starting point (first stack frame) of a coroutine, and a mechanism to get (read) the currently executing coroutine handle.
1090The @main@ routine has no return value or additional parameters because the coroutine type allows an arbitrary number of interface routines with corresponding arbitrary typed input/output values versus fixed ones.
1091The advantage of this approach is that users can easily create different types of coroutines, \eg changing the memory layout of a coroutine is trivial when implementing the @get_coroutine@ routine, and possibly redefining @suspend@ and @resume@.
1092The \CFA keyword @coroutine@ implicitly implements the getter and forward declarations required for implementing the coroutine main:
1093\begin{cquote}
1094\begin{tabular}{@{}ccc@{}}
1095\begin{cfa}
1096coroutine MyCor {
1097        int value;
1098
1099};
1100\end{cfa}
1101&
1102{\Large $\Rightarrow$}
1103&
1104\begin{tabular}{@{}ccc@{}}
1105\begin{cfa}
1106struct MyCor {
1107        int value;
1108        coroutine_desc cor;
1109};
1110\end{cfa}
1111&
1112\begin{cfa}
1113static inline coroutine_desc *
1114get_coroutine( MyCor & this ) {
1115        return &this.cor;
1116}
1117\end{cfa}
1118&
1119\begin{cfa}
1120void main( MyCor * this );
1121
1122
1123
1124\end{cfa}
1125\end{tabular}
1126\end{tabular}
1127\end{cquote}
1128The combination of these two approaches allows an easy and concise specification to coroutining (and concurrency) for normal users, while more advanced users have tighter control on memory layout and initialization.
1129
1130
1131\section{Concurrency}
1132\label{s:Concurrency}
1133
1134At its core, concurrency is based on multiple call-stacks and scheduling threads executing on these stacks.
1135Multiple call stacks (or contexts) and a single thread of execution, called \newterm{coroutining}~\cite{Conway63,Marlin80}, does \emph{not} imply concurrency~\cite[\S~2]{Buhr05a}.
1136In coroutining, the single thread is self-scheduling across the stacks, so execution is deterministic, \ie the execution path from input to output is fixed and predictable.
1137A \newterm{stackless} coroutine executes on the caller's stack~\cite{Python} but this approach is restrictive, \eg preventing modularization and supporting only iterator/generator-style programming;
1138a \newterm{stackful} coroutine executes on its own stack, allowing full generality.
1139Only stackful coroutines are a stepping stone to concurrency.
1140
1141The transition to concurrency, even for execution with a single thread and multiple stacks, occurs when coroutines also context switch to a \newterm{scheduling oracle}, introducing non-determinism from the coroutine perspective~\cite[\S~3]{Buhr05a}.
1142Therefore, a minimal concurrency system is possible using coroutines (see Section \ref{coroutine}) in conjunction with a scheduler to decide where to context switch next.
1143The resulting execution system now follows a cooperative threading-model, called \newterm{non-preemptive scheduling}.
1144
1145Because the scheduler is special, it can either be a stackless or stackful coroutine.
1146For stackless, the scheduler performs scheduling on the stack of the current coroutine and switches directly to the next coroutine, so there is one context switch.
1147For stackful, the current coroutine switches to the scheduler, which performs scheduling, and it then switches to the next coroutine, so there are two context switches.
1148A stackful scheduler is often used for simplicity and security.
1149
1150Regardless of the approach used, a subset of concurrency related challenges start to appear.
1151For the complete set of concurrency challenges to occur, the missing feature is \newterm{preemption}, where context switching occurs randomly between any two instructions, often based on a timer interrupt, called \newterm{preemptive scheduling}.
1152While a scheduler introduces uncertainty in the order of execution, preemption introduces uncertainty about where context switches occur.
1153Interestingly, uncertainty is necessary for the runtime (operating) system to give the illusion of parallelism on a single processor and increase performance on multiple processors.
1154The reason is that only the runtime has complete knowledge about resources and how to best utilized them.
1155However, the introduction of unrestricted non-determinism results in the need for \newterm{mutual exclusion} and \newterm{synchronization} to restrict non-determinism for correctness;
1156otherwise, it is impossible to write meaningful programs.
1157Optimal performance in concurrent applications is often obtained by having as much non-determinism as correctness allows.
1158
1159An important missing feature in C is threading\footnote{While the C11 standard defines a \protect\lstinline@threads.h@ header, it is minimal and defined as optional.
1160As such, library support for threading is far from widespread.
1161At the time of writing the paper, neither \protect\lstinline@gcc@ nor \protect\lstinline@clang@ support \protect\lstinline@threads.h@ in their standard libraries.}.
1162In modern programming languages, a lack of threading is unacceptable~\cite{Sutter05, Sutter05b}, and therefore existing and new programming languages must have tools for writing efficient concurrent programs to take advantage of parallelism.
1163As an extension of C, \CFA needs to express these concepts in a way that is as natural as possible to programmers familiar with imperative languages.
1164Furthermore, because C is a system-level language, programmers expect to choose precisely which features they need and which cost they are willing to pay.
1165Hence, concurrent programs should be written using high-level mechanisms, and only step down to lower-level mechanisms when performance bottlenecks are encountered.
1166
1167
1168\subsection{Thread Interface}
1169\label{threads}
1170
1171Both user and kernel threads are supported, where user threads provide concurrency and kernel threads provide parallelism.
1172Like coroutines and for the same design reasons, the selected approach for user threads is to use language support by introducing a new kind of aggregate (structure) and a \CFA trait:
1173\begin{cquote}
1174\begin{tabular}{@{}c@{\hspace{3\parindentlnth}}c@{}}
1175\begin{cfa}
1176thread myThread {
1177        // communication variables
1178};
1179
1180
1181\end{cfa}
1182&
1183\begin{cfa}
1184trait is_thread( `dtype` T ) {
1185      void main( T & );
1186      thread_desc * get_thread( T & );
1187      void ^?{}( T & `mutex` );
1188};
1189\end{cfa}
1190\end{tabular}
1191\end{cquote}
1192(The qualifier @mutex@ for the destructor parameter is discussed in Section~\ref{s:Monitor}.)
1193Like a coroutine, the statically-typed @main@ routine is the starting point (first stack frame) of a user thread.
1194The difference is that a coroutine borrows a thread from its caller, so the first thread resuming a coroutine creates an instance of @main@;
1195whereas, a user thread receives its own thread from the runtime system, which starts in @main@ as some point after the thread constructor is run.\footnote{
1196The \lstinline@main@ routine is already a special routine in C, \ie where the program's initial thread begins, so it is a natural extension of this semantics to use overloading to declare \lstinline@main@s for user coroutines and threads.}
1197No return value or additional parameters are necessary for this routine because the task type allows an arbitrary number of interface routines with corresponding arbitrary typed input/output values.
1198
1199\begin{comment} % put in appendix with coroutine version ???
1200As such the @main@ routine of a thread can be defined as
1201\begin{cfa}
1202thread foo {};
1203
1204void main(foo & this) {
1205        sout | "Hello World!";
1206}
1207\end{cfa}
1208
1209In this example, threads of type @foo@ start execution in the @void main(foo &)@ routine, which prints @"Hello World!".@ While this paper encourages this approach to enforce strongly typed programming, users may prefer to use the routine-based thread semantics for the sake of simplicity.
1210With the static semantics it is trivial to write a thread type that takes a routine pointer as a parameter and executes it on its stack asynchronously.
1211\begin{cfa}
1212typedef void (*voidRtn)(int);
1213
1214thread RtnRunner {
1215        voidRtn func;
1216        int arg;
1217};
1218
1219void ?{}(RtnRunner & this, voidRtn inRtn, int arg) {
1220        this.func = inRtn;
1221        this.arg  = arg;
1222}
1223
1224void main(RtnRunner & this) {
1225        // thread starts here and runs the routine
1226        this.func( this.arg );
1227}
1228
1229void hello(/*unused*/ int) {
1230        sout | "Hello World!";
1231}
1232
1233int main() {
1234        RtnRunner f = {hello, 42};
1235        return 0?
1236}
1237\end{cfa}
1238A consequence of the strongly typed approach to main is that memory layout of parameters and return values to/from a thread are now explicitly specified in the \textbf{API}.
1239\end{comment}
1240
1241For user threads to be useful, it must be possible to start and stop the underlying thread, and wait for it to complete execution.
1242While using an API such as @fork@ and @join@ is relatively common, such an interface is awkward and unnecessary.
1243A simple approach is to use allocation/deallocation principles, and have threads implicitly @fork@ after construction and @join@ before destruction.
1244\begin{cfa}
1245thread World {};
1246void main( World & this ) {
1247        sout | "World!";
1248}
1249int main() {
1250        World w`[10]`;                                                  $\C{// implicit forks after creation}$
1251        sout | "Hello ";                                        $\C{// "Hello " and 10 "World!" printed concurrently}$
1252}                                                                                       $\C{// implicit joins before destruction}$
1253\end{cfa}
1254This semantics ensures a thread is started and stopped exactly once, eliminating some programming error, and scales to multiple threads for basic (termination) synchronization.
1255This tree-structure (lattice) create/delete from C block-structure is generalized by using dynamic allocation, so threads can outlive the scope in which they are created, much like dynamically allocating memory lets objects outlive the scope in which they are created.
1256\begin{cfa}
1257int main() {
1258        MyThread * heapLive;
1259        {
1260                MyThread blockLive;                                     $\C{// fork block-based thread}$
1261                heapLive = `new`( MyThread );           $\C{// fork heap-based thread}$
1262                ...
1263        }                                                                               $\C{// join block-based thread}$
1264        ...
1265        `delete`( heapLive );                                   $\C{// join heap-based thread}$
1266}
1267\end{cfa}
1268The heap-based approach allows arbitrary thread-creation topologies, with respect to fork/join-style concurrency.
1269
1270Figure~\ref{s:ConcurrentMatrixSummation} shows concurrently adding the rows of a matrix and then totalling the subtotals sequentially, after all the row threads have terminated.
1271The program uses heap-based threads because each thread needs different constructor values.
1272(Python provides a simple iteration mechanism to initialize array elements to different values allowing stack allocation.)
1273The allocation/deallocation pattern appears unusual because allocated objects are immediately deallocated without any intervening code.
1274However, for threads, the deletion provides implicit synchronization, which is the intervening code.
1275While the subtotals are added in linear order rather than completion order, which slightly inhibits concurrency, the computation is restricted by the critical-path thread (\ie the thread that takes the longest), and so any inhibited concurrency is very small as totalling the subtotals is trivial.
1276
1277\begin{figure}
1278\begin{cfa}
1279`thread` Adder { int * row, cols, & subtotal; } $\C{// communication variables}$
1280void ?{}( Adder & adder, int row[], int cols, int & subtotal ) {
1281    adder.[ row, cols, &subtotal ] = [ row, cols, &subtotal ];
1282}
1283void main( Adder & adder ) with( adder ) {
1284    subtotal = 0;
1285    for ( int c = 0; c < cols; c += 1 ) { subtotal += row[c]; }
1286}
1287int main() {
1288    const int rows = 10, cols = 1000;
1289    int matrix[rows][cols], subtotals[rows], total = 0;
1290    // read matrix
1291    Adder * adders[rows];
1292    for ( int r = 0; r < rows; r += 1 ) {       $\C{// start threads to sum rows}$
1293                adders[r] = `new( matrix[r], cols, &subtotals[r] );`
1294    }
1295    for ( int r = 0; r < rows; r += 1 ) {       $\C{// wait for threads to finish}$
1296                `delete( adders[r] );`                          $\C{// termination join}$
1297                total += subtotals[r];                          $\C{// total subtotal}$
1298    }
1299    sout | total;
1300}
1301\end{cfa}
1302\caption{Concurrent Matrix Summation}
1303\label{s:ConcurrentMatrixSummation}
1304\end{figure}
1305
1306
1307\section{Mutual Exclusion / Synchronization}
1308
1309Uncontrolled non-deterministic execution is meaningless.
1310To reestablish meaningful execution requires mechanisms to reintroduce determinism, \ie restrict non-determinism, called mutual exclusion and synchronization, where mutual exclusion is an access-control mechanism on data shared by threads, and synchronization is a timing relationship among threads~\cite[\S~4]{Buhr05a}.
1311Since many deterministic challenges appear with the use of mutable shared state, some languages/libraries disallow it, \eg Erlang~\cite{Erlang}, Haskell~\cite{Haskell}, Akka~\cite{Akka} (Scala).
1312In these paradigms, interaction among concurrent objects is performed by stateless message-passing~\cite{Thoth,Harmony,V-Kernel} or other paradigms closely related to networking concepts, \eg channels~\cite{CSP,Go}.
1313However, in call/return-based languages, these approaches force a clear distinction, \ie introduce a new programming paradigm between regular and concurrent computation, \eg routine call versus message passing.
1314Hence, a programmer must learn and manipulate two sets of design patterns.
1315While this distinction can be hidden away in library code, effective use of the library still has to take both paradigms into account.
1316In contrast, approaches based on stateful models more closely resemble the standard call/return programming-model, resulting in a single programming paradigm.
1317
1318At the lowest level, concurrent control is implemented by atomic operations, upon which different kinds of locking mechanisms are constructed, \eg semaphores~\cite{Dijkstra68b}, barriers, and path expressions~\cite{Campbell74}.
1319However, for productivity it is always desirable to use the highest-level construct that provides the necessary efficiency~\cite{Hochstein05}.
1320A newer approach for restricting non-determinism is transactional memory~\cite{Herlihy93}.
1321While this approach is pursued in hardware~\cite{Nakaike15} and system languages, like \CC~\cite{Cpp-Transactions}, the performance and feature set is still too restrictive to be the main concurrency paradigm for system languages, which is why it is rejected as the core paradigm for concurrency in \CFA.
1322
1323One of the most natural, elegant, and efficient mechanisms for mutual exclusion and synchronization for shared-memory systems is the \emph{monitor}.
1324First proposed by Brinch Hansen~\cite{Hansen73} and later described and extended by C.A.R.~Hoare~\cite{Hoare74}, many concurrent programming-languages provide monitors as an explicit language construct: \eg Concurrent Pascal~\cite{ConcurrentPascal}, Mesa~\cite{Mesa}, Modula~\cite{Modula-2}, Turing~\cite{Turing:old}, Modula-3~\cite{Modula-3}, NeWS~\cite{NeWS}, Emerald~\cite{Emerald}, \uC~\cite{Buhr92a} and Java~\cite{Java}.
1325In addition, operating-system kernels and device drivers have a monitor-like structure, although they often use lower-level primitives such as mutex locks or semaphores to simulate monitors.
1326For these reasons, \CFA selected monitors as the core high-level concurrency-construct, upon which higher-level approaches can be easily constructed.
1327
1328
1329\subsection{Mutual Exclusion}
1330
1331A group of instructions manipulating a specific instance of shared data that must be performed atomically is called an (individual) \newterm{critical-section}~\cite{Dijkstra65}.
1332The generalization is called a \newterm{group critical-section}~\cite{Joung00}, where multiple tasks with the same session may use the resource simultaneously, but different sessions may not use the resource simultaneously.
1333The readers/writer problem~\cite{Courtois71} is an instance of a group critical-section, where readers have the same session and all writers have a unique session.
1334\newterm{Mutual exclusion} enforces that the correct kind and number of threads are using a critical section.
1335
1336However, many solutions exist for mutual exclusion, which vary in terms of performance, flexibility and ease of use.
1337Methods range from low-level locks, which are fast and flexible but require significant attention for correctness, to higher-level concurrency techniques, which sacrifice some performance to improve ease of use.
1338Ease of use comes by either guaranteeing some problems cannot occur, \eg deadlock free, or by offering a more explicit coupling between shared data and critical section.
1339For example, the \CC @std::atomic<T>@ offers an easy way to express mutual-exclusion on a restricted set of operations, \eg reading/writing, for numerical types.
1340However, a significant challenge with locks is composability because it takes careful organization for multiple locks to be used while preventing deadlock.
1341Easing composability is another feature higher-level mutual-exclusion mechanisms can offer.
1342
1343
1344\subsection{Synchronization}
1345
1346Synchronization enforces relative ordering of execution, and synchronization tools provide numerous mechanisms to establish these timing relationships.
1347Low-level synchronization primitives offer good performance and flexibility at the cost of ease of use;
1348higher-level mechanisms often simplify usage by adding better coupling between synchronization and data, \eg message passing, or offering a simpler solution to otherwise involved challenges, \eg barrier lock.
1349Often synchronization is used to order access to a critical section, \eg ensuring a reader thread is the next kind of thread to enter a critical section.
1350If a writer thread is scheduled for next access, but another reader thread acquires the critical section first, that reader \newterm{barged}.
1351Barging can result in staleness/freshness problems, where a reader barges ahead of a writer and reads temporally stale data, or a writer barges ahead of another writer overwriting data with a fresh value preventing the previous value from ever being read (lost computation).
1352Preventing or detecting barging is an involved challenge with low-level locks, which can be made much easier by higher-level constructs.
1353This challenge is often split into two different approaches: barging avoidance and barging prevention.
1354Algorithms that allow a barger, but divert it until later using current synchronization state (flags), are avoiding the barger;
1355algorithms that preclude a barger from entering during synchronization in the critical section prevent barging completely.
1356Techniques like baton-passing locks~\cite{Andrews89} between threads instead of unconditionally releasing locks is an example of barging prevention.
1357
1358
1359\section{Monitor}
1360\label{s:Monitor}
1361
1362A \textbf{monitor} is a set of routines that ensure mutual exclusion when accessing shared state.
1363More precisely, a monitor is a programming technique that binds mutual exclusion to routine scope, as opposed to locks, where mutual-exclusion is defined by acquire/release calls, independent of lexical context (analogous to block and heap storage allocation).
1364The strong association with the call/return paradigm eases programmability, readability and maintainability, at a slight cost in flexibility and efficiency.
1365
1366Note, like coroutines/threads, both locks and monitors require an abstract handle to reference them, because at their core, both mechanisms are manipulating non-copyable shared-state.
1367Copying a lock is insecure because it is possible to copy an open lock and then use the open copy when the original lock is closed to simultaneously access the shared data.
1368Copying a monitor is secure because both the lock and shared data are copies, but copying the shared data is meaningless because it no longer represents a unique entity.
1369As for coroutines/tasks, the @dtype@ property provides no implicit copying operations and the @is_monitor@ trait provides no explicit copying operations, so all locks/monitors must be passed by reference (pointer).
1370\begin{cfa}
1371trait is_monitor( `dtype` T ) {
1372        monitor_desc * get_monitor( T & );
1373        void ^?{}( T & mutex );
1374};
1375\end{cfa}
1376
1377
1378\subsection{Mutex Acquisition}
1379\label{s:MutexAcquisition}
1380
1381While correctness implies a monitor's mutual exclusion is acquired and released, there are implementation options about when and where the locking/unlocking occurs.
1382(Much of this discussion also applies to basic locks.)
1383For example, a monitor may need to be passed through multiple helper routines before it becomes necessary to acquire the monitor mutual-exclusion.
1384\begin{cfa}[morekeywords=nomutex]
1385monitor Aint { int cnt; };                                      $\C{// atomic integer counter}$
1386void ?{}( Aint & `nomutex` this ) with( this ) { cnt = 0; } $\C{// constructor}$
1387int ?=?( Aint & `mutex`$\(_{opt}\)$ lhs, int rhs ) with( lhs ) { cnt = rhs; } $\C{// conversions}$
1388void ?{}( int & this, Aint & `mutex`$\(_{opt}\)$ v ) { this = v.cnt; }
1389int ?=?( int & lhs, Aint & `mutex`$\(_{opt}\)$ rhs ) with( rhs ) { lhs = cnt; }
1390int ++?( Aint & `mutex`$\(_{opt}\)$ this ) with( this ) { return ++cnt; } $\C{// increment}$
1391\end{cfa}
1392The @Aint@ constructor, @?{}@, uses the \lstinline[morekeywords=nomutex]@nomutex@ qualifier indicating mutual exclusion is unnecessary during construction because an object is inaccessible (private) until after it is initialized.
1393(While a constructor may publish its address into a global variable, doing so generates a race-condition.)
1394The conversion operators for initializing and assigning with a normal integer only need @mutex@, if reading/writing the implementation type is not atomic.
1395Finally, the prefix increment operato, @++?@, is normally @mutex@ to protect the incrementing from race conditions, unless there is an atomic increment instruction for the implementation type.
1396
1397The atomic counter is used without any explicit mutual-exclusion and provides thread-safe semantics, which is similar to the \CC template @std::atomic@.
1398\begin{cfa}
1399Aint x, y, z;
1400++x; ++y; ++z;                                                          $\C{// safe increment by multiple threads}$
1401x = 2; y = 2; z = 2;                                            $\C{// conversions}$
1402int i = x, j = y, k = z;
1403i = x; j = y; k = z;
1404\end{cfa}
1405
1406For maximum usability, monitors have \newterm{multi-acquire} semantics allowing a thread to acquire it multiple times without deadlock.
1407\begin{cfa}
1408monitor M { ... } m;
1409void foo( M & mutex m ) { ... }                         $\C{// acquire mutual exclusion}$
1410void bar( M & mutex m ) {                                       $\C{// acquire mutual exclusion}$
1411        ... `foo( m );` ...                                             $\C{// reacquire mutual exclusion}$
1412}
1413`bar( m );`                                                                     $\C{// nested monitor call}$
1414\end{cfa}
1415
1416The benefit of mandatory monitor qualifiers is self-documentation, but requiring both @mutex@ and \lstinline[morekeywords=nomutex]@nomutex@ for all monitor parameters is redundant.
1417Instead, the semantics have one qualifier as the default, and the other required.
1418For example, make the safe @mutex@ qualifier the default because assuming \lstinline[morekeywords=nomutex]@nomutex@ may cause subtle errors.
1419Alternatively, make the unsafe \lstinline[morekeywords=nomutex]@nomutex@ qualifier the default because it is the \emph{normal} parameter semantics while @mutex@ parameters are rare.
1420Providing a default qualifier implies knowing whether a parameter is a monitor.
1421Since \CFA relies heavily on traits as an abstraction mechanism, the distinction between a type that is a monitor and a type that looks like a monitor can become blurred.
1422For this reason, \CFA requires programmers to identify the kind of parameter with the @mutex@ keyword and uses no keyword to mean \lstinline[morekeywords=nomutex]@nomutex@.
1423
1424The next semantic decision is establishing which parameter \emph{types} may be qualified with @mutex@.
1425Given:
1426\begin{cfa}
1427monitor M { ... }
1428int f1( M & mutex m );
1429int f2( M * mutex m );
1430int f3( M * mutex m[] );
1431int f4( stack( M * ) & mutex m );
1432\end{cfa}
1433the issue is that some of these parameter types are composed of multiple objects.
1434For @f1@, there is only a single parameter object.
1435Adding indirection in @f2@ still identifies a single object.
1436However, the matrix in @f3@ introduces multiple objects.
1437While shown shortly, multiple acquisition is possible;
1438however array lengths are often unknown in C.
1439This issue is exacerbated in @f4@, where the data structure must be safely traversed to acquire all of its elements.
1440
1441To make the issue tractable, \CFA only acquires one monitor per parameter with at most one level of indirection.
1442However, there is an ambiguity in the C type-system with respects to arrays.
1443Is the argument for @f2@ a single object or an array of objects?
1444If it is an array, only the first element of the array is acquired, which seems unsafe;
1445hence, @mutex@ is disallowed for array parameters.
1446\begin{cfa}
1447int f1( M & mutex m );                                          $\C{// allowed: recommended case}$
1448int f2( M * mutex m );                                          $\C{// disallowed: could be an array}$
1449int f3( M mutex m[$\,$] );                                      $\C{// disallowed: array length unknown}$
1450int f4( M ** mutex m );                                         $\C{// disallowed: could be an array}$
1451int f5( M * mutex m[$\,$] );                            $\C{// disallowed: array length unknown}$
1452\end{cfa}
1453% Note, not all array routines have distinct types: @f2@ and @f3@ have the same type, as do @f4@ and @f5@.
1454% However, even if the code generation could tell the difference, the extra information is still not sufficient to extend meaningfully the monitor call semantic.
1455
1456For object-oriented monitors, calling a mutex member \emph{implicitly} acquires mutual exclusion of the receiver object, @`rec`.foo(...)@.
1457\CFA has no receiver, and hence, must use an explicit mechanism to specify which object acquires mutual exclusion.
1458A positive consequence of this design decision is the ability to support multi-monitor routines.
1459\begin{cfa}
1460int f( M & mutex x, M & mutex y );              $\C{// multiple monitor parameter of any type}$
1461M m1, m2;
1462f( m1, m2 );
1463\end{cfa}
1464(While object-oriented monitors can be extended with a mutex qualifier for multiple-monitor members, no prior example of this feature could be found.)
1465In practice, writing multi-locking routines that do not deadlock is tricky.
1466Having language support for such a feature is therefore a significant asset for \CFA.
1467
1468The capability to acquire multiple locks before entering a critical section is called \newterm{bulk acquire} (see Section~\ref{s:Implementation} for implementation details).
1469In the previous example, \CFA guarantees the order of acquisition is consistent across calls to different routines using the same monitors as arguments.
1470This consistent ordering means acquiring multiple monitors is safe from deadlock.
1471However, users can force the acquiring order.
1472For example, notice the use of @mutex@/\lstinline[morekeywords=nomutex]@nomutex@ and how this affects the acquiring order:
1473\begin{cfa}
1474void foo( M & mutex m1, M & mutex m2 );         $\C{// acquire m1 and m2}$
1475void bar( M & mutex m1, M & /* nomutex */ m2 ) { $\C{// acquire m1}$
1476        ... foo( m1, m2 ); ...                                  $\C{// acquire m2}$
1477}
1478void baz( M & /* nomutex */ m1, M & mutex m2 ) { $\C{// acquire m2}$
1479        ... foo( m1, m2 ); ...                                  $\C{// acquire m1}$
1480}
1481\end{cfa}
1482The multi-acquire semantics allows @bar@ or @baz@ to acquire a monitor lock and reacquire it in @foo@.
1483In the calls to @bar@ and @baz@, the monitors are acquired in opposite order.
1484
1485However, such use leads to lock acquiring order problems resulting in deadlock~\cite{Lister77}, where detecting it requires dynamic tracking of monitor calls, and dealing with it requires rollback semantics~\cite{Dice10}.
1486In \CFA, a safety aid is provided by using bulk acquire of all monitors to shared objects, whereas other monitor systems provide no aid.
1487While \CFA provides only a partial solution, it handles many useful cases, \eg:
1488\begin{cfa}
1489monitor BankAccount { ... };
1490void deposit( BankAccount & `mutex` b, int deposit );
1491void transfer( BankAccount & `mutex` my, BankAccount & `mutex` your, int me2you ) {
1492        deposit( my, `-`me2you );                               $\C{// debit}$
1493        deposit( your, me2you );                                $\C{// credit}$
1494}
1495\end{cfa}
1496This example shows a trivial solution to the bank-account transfer problem.
1497Without multi- and bulk acquire, the solution to this problem requires careful engineering.
1498
1499
1500\subsection{\protect\lstinline@mutex@ statement}
1501\label{mutex-stmt}
1502
1503The monitor call-semantics associate all locking semantics to routines.
1504Like Java, \CFA offers an alternative @mutex@ statement to reduce refactoring and naming.
1505\begin{cquote}
1506\begin{tabular}{@{}l@{\hspace{3\parindentlnth}}l@{}}
1507\begin{cfa}
1508monitor M { ... };
1509void foo( M & mutex m1, M & mutex m2 ) {
1510        // critical section
1511}
1512void bar( M & m1, M & m2 ) {
1513        foo( m1, m2 );
1514}
1515\end{cfa}
1516&
1517\begin{cfa}
1518
1519void bar( M & m1, M & m2 ) {
1520        mutex( m1, m2 ) {       // remove refactoring and naming
1521                // critical section
1522        }
1523}
1524
1525\end{cfa}
1526\\
1527\multicolumn{1}{c}{\textbf{routine call}} & \multicolumn{1}{c}{\lstinline@mutex@ \textbf{statement}}
1528\end{tabular}
1529\end{cquote}
1530
1531
1532\section{Scheduling}
1533\label{s:Scheduling}
1534
1535While monitor mutual-exclusion provides safe access to shared data, the monitor data may indicate that a thread accessing it cannot proceed.
1536For example, Figure~\ref{f:GenericBoundedBuffer} shows a bounded buffer that may be full/empty so produce/consumer threads must block.
1537Leaving the monitor and trying again (busy waiting) is impractical for high-level programming.
1538Monitors eliminate busy waiting by providing synchronization to schedule threads needing access to the shared data, where threads block versus spinning.
1539Synchronization is generally achieved with internal~\cite{Hoare74} or external~\cite[\S~2.9.2]{uC++} scheduling, where \newterm{scheduling} defines which thread acquires the critical section next.
1540\newterm{Internal scheduling} is characterized by each thread entering the monitor and making an individual decision about proceeding or blocking, while \newterm{external scheduling} is characterized by an entering thread making a decision about proceeding for itself and on behalf of other threads attempting entry.
1541
1542Figure~\ref{f:BBInt} shows a \CFA generic bounded-buffer with internal scheduling, where producers/consumers enter the monitor, see the buffer is full/empty, and block on an appropriate condition lock, @full@/@empty@.
1543The @wait@ routine atomically blocks the calling thread and implicitly releases the monitor lock(s) for all monitors in the routine's parameter list.
1544The appropriate condition lock is signalled to unblock an opposite kind of thread after an element is inserted/removed from the buffer.
1545Signalling is unconditional, because signalling an empty condition lock does nothing.
1546
1547Signalling semantics cannot have the signaller and signalled thread in the monitor simultaneously, which means:
1548\begin{enumerate}
1549\item
1550The signalling thread returns immediately, and the signalled thread continues.
1551\item
1552The signalling thread continues and the signalled thread is marked for urgent unblocking at the next scheduling point (exit/wait).
1553\item
1554The signalling thread blocks but is marked for urgrent unblocking at the next scheduling point and the signalled thread continues.
1555\end{enumerate}
1556The first approach is too restrictive, as it precludes solving a reasonable class of problems, \eg dating service (see Figure~\ref{f:DatingService}).
1557\CFA supports the next two semantics as both are useful.
1558Finally, while it is common to store a @condition@ as a field of the monitor, in \CFA, a @condition@ variable can be created/stored independently.
1559Furthermore, a condition variable is tied to a \emph{group} of monitors on first use, called \newterm{branding}, which means that using internal scheduling with distinct sets of monitors requires one condition variable per set of monitors.
1560
1561\begin{figure}
1562\centering
1563\newbox\myboxA
1564\begin{lrbox}{\myboxA}
1565\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1566forall( otype T ) { // distribute forall
1567        monitor Buffer {
1568                `condition` full, empty;
1569                int front, back, count;
1570                T elements[10];
1571        };
1572        void ?{}( Buffer(T) & buffer ) with(buffer) {
1573                [front, back, count] = 0;
1574        }
1575
1576        void insert( Buffer(T) & mutex buffer, T elem )
1577                                with(buffer) {
1578                if ( count == 10 ) `wait( empty )`;
1579                // insert elem into buffer
1580                `signal( full )`;
1581        }
1582        T remove( Buffer(T) & mutex buffer ) with(buffer) {
1583                if ( count == 0 ) `wait( full )`;
1584                // remove elem from buffer
1585                `signal( empty )`;
1586                return elem;
1587        }
1588}
1589\end{cfa}
1590\end{lrbox}
1591
1592\newbox\myboxB
1593\begin{lrbox}{\myboxB}
1594\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1595forall( otype T ) { // distribute forall
1596        monitor Buffer {
1597
1598                int front, back, count;
1599                T elements[10];
1600        };
1601        void ?{}( Buffer(T) & buffer ) with(buffer) {
1602                [front, back, count] = 0;
1603        }
1604        T remove( Buffer(T) & mutex buffer ); // forward
1605        void insert( Buffer(T) & mutex buffer, T elem )
1606                                with(buffer) {
1607                if ( count == 10 ) `waitfor( remove, buffer )`;
1608                // insert elem into buffer
1609
1610        }
1611        T remove( Buffer(T) & mutex buffer ) with(buffer) {
1612                if ( count == 0 ) `waitfor( insert, buffer )`;
1613                // remove elem from buffer
1614
1615                return elem;
1616        }
1617}
1618\end{cfa}
1619\end{lrbox}
1620
1621\subfloat[Internal Scheduling]{\label{f:BBInt}\usebox\myboxA}
1622%\qquad
1623\subfloat[External Scheduling]{\label{f:BBExt}\usebox\myboxB}
1624\caption{Generic Bounded-Buffer}
1625\label{f:GenericBoundedBuffer}
1626\end{figure}
1627
1628Figure~\ref{f:BBExt} shows a \CFA generic bounded-buffer with external scheduling, where producers/consumers detecting a full/empty buffer block and prevent more producers/consumers from entering the monitor until there is a free/empty slot in the buffer.
1629External scheduling is controlled by the @waitfor@ statement, which atomically blocks the calling thread, releases the monitor lock, and restricts the routine calls that can next acquire mutual exclusion.
1630If the buffer is full, only calls to @remove@ can acquire the buffer, and if the buffer is empty, only calls to @insert@ can acquire the buffer.
1631Threads making calls to routines that are currently excluded, block outside of (external to) the monitor on a calling queue, versus blocking on condition queues inside of (internal to) the monitor.
1632External scheduling allows users to wait for events from other threads without concern of unrelated events occurring.
1633The mechnaism can be done in terms of control flow, \eg Ada @accept@ or \uC @_Accept@, or in terms of data, \eg Go channels.
1634While both mechanisms have strengths and weaknesses, this project uses a control-flow mechanism to stay consistent with other language semantics.
1635Two challenges specific to \CFA for external scheduling are loose object-definitions (see Section~\ref{s:LooseObjectDefinitions}) and multiple-monitor routines (see Section~\ref{s:Multi-MonitorScheduling}).
1636
1637For internal scheduling, non-blocking signalling (as in the producer/consumer example) is used when the signaller is providing the cooperation for a waiting thread;
1638the signaller enters the monitor and changes state, detects a waiting threads that can use the state, performs a non-blocking signal on the condition queue for the waiting thread, and exits the monitor to run concurrently.
1639The waiter unblocks next from the urgent queue, uses/takes the state, and exits the monitor.
1640Blocking signalling is the reverse, where the waiter is providing the cooperation for the signalling thread;
1641the signaller enters the monitor, detects a waiting thread providing the necessary state, performs a blocking signal to place it on the urgent queue and unblock the waiter.
1642The waiter changes state and exits the monitor, and the signaller unblocks next from the urgent queue to use/take the state.
1643
1644Figure~\ref{f:DatingService} shows a dating service demonstrating non-blocking and blocking signalling.
1645The dating service matches girl and boy threads with matching compatibility codes so they can exchange phone numbers.
1646A thread blocks until an appropriate partner arrives.
1647The complexity is exchanging phone numbers in the monitor because of the mutual-exclusion property.
1648For signal scheduling, the @exchange@ condition is necessary to block the thread finding the match, while the matcher unblocks to take the opposite number, post its phone number, and unblock the partner.
1649For signal-block scheduling, the implicit urgent-queue replaces the explict @exchange@-condition and @signal_block@ puts the finding thread on the urgent condition and unblocks the matcher.
1650
1651The dating service is an example of a monitor that cannot be written using external scheduling because it requires knowledge of calling parameters to make scheduling decisions, and parameters of waiting threads are unavailable;
1652as well, an arriving thread may not find a partner and must wait, which requires a condition variable, and condition variables imply internal scheduling.
1653
1654\begin{figure}
1655\centering
1656\newbox\myboxA
1657\begin{lrbox}{\myboxA}
1658\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1659enum { CCodes = 20 };
1660monitor DS {
1661        int GirlPhNo, BoyPhNo;
1662        condition Girls[CCodes], Boys[CCodes];
1663        condition exchange;
1664};
1665int girl( DS & mutex ds, int phNo, int ccode ) {
1666        if ( is_empty( Boys[ccode] ) ) {
1667                wait( Girls[ccode] );
1668                GirlPhNo = phNo;
1669                `signal( exchange );`
1670        } else {
1671                GirlPhNo = phNo;
1672                `signal( Boys[ccode] );`
1673                `wait( exchange );`
1674        } // if
1675        return BoyPhNo;
1676}
1677int boy( DS & mutex ds, int phNo, int ccode ) {
1678        // as above with boy/girl interchanged
1679}
1680\end{cfa}
1681\end{lrbox}
1682
1683\newbox\myboxB
1684\begin{lrbox}{\myboxB}
1685\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1686
1687monitor DS {
1688        int GirlPhNo, BoyPhNo;
1689        condition Girls[CCodes], Boys[CCodes];
1690
1691};
1692int girl( DS & mutex ds, int phNo, int ccode ) {
1693        if ( is_empty( Boys[ccode] ) ) { // no compatible
1694                wait( Girls[ccode] ); // wait for boy
1695                GirlPhNo = phNo; // make phone number available
1696
1697        } else {
1698                GirlPhNo = phNo; // make phone number available
1699                `signal_block( Boys[ccode] );` // restart boy
1700
1701        } // if
1702        return BoyPhNo;
1703}
1704int boy( DS & mutex ds, int phNo, int ccode ) {
1705        // as above with boy/girl interchanged
1706}
1707\end{cfa}
1708\end{lrbox}
1709
1710\subfloat[\lstinline@signal@]{\label{f:DatingSignal}\usebox\myboxA}
1711\qquad
1712\subfloat[\lstinline@signal_block@]{\label{f:DatingSignalBlock}\usebox\myboxB}
1713\caption{Dating service. }
1714\label{f:DatingService}
1715\end{figure}
1716
1717Both internal and external scheduling extend to multiple monitors in a natural way.
1718\begin{cquote}
1719\begin{tabular}{@{}l@{\hspace{3\parindentlnth}}l@{}}
1720\begin{cfa}
1721monitor M { `condition e`; ... };
1722void foo( M & mutex m1, M & mutex m2 ) {
1723        ... wait( `e` ); ...   // wait( e, m1, m2 )
1724        ... wait( `e, m1` ); ...
1725        ... wait( `e, m2` ); ...
1726}
1727\end{cfa}
1728&
1729\begin{cfa}
1730void rtn$\(_1\)$( M & mutex m1, M & mutex m2 );
1731void rtn$\(_2\)$( M & mutex m1 );
1732void bar( M & mutex m1, M & mutex m2 ) {
1733        ... waitfor( `rtn` ); ...       // $\LstCommentStyle{waitfor( rtn\(_1\), m1, m2 )}$
1734        ... waitfor( `rtn, m1` ); ... // $\LstCommentStyle{waitfor( rtn\(_2\), m1 )}$
1735}
1736\end{cfa}
1737\end{tabular}
1738\end{cquote}
1739For @wait( e )@, the default semantics is to atomically block the signaller and release all acquired mutex types in the parameter list, \ie @wait( e, m1, m2 )@.
1740To override the implicit multi-monitor wait, specific mutex parameter(s) can be specified, \eg @wait( e, m1 )@.
1741Wait statically verifies the released monitors are the acquired mutex-parameters so unconditional release is safe.
1742Finally, a signaller,
1743\begin{cfa}
1744void baz( M & mutex m1, M & mutex m2 ) {
1745        ... signal( e ); ...
1746}
1747\end{cfa}
1748must have acquired at least the same locks as the waiting thread signalled from the condition queue.
1749
1750Similarly, for @waitfor( rtn )@, the default semantics is to atomically block the acceptor and release all acquired mutex types in the parameter list, \ie @waitfor( rtn, m1, m2 )@.
1751To override the implicit multi-monitor wait, specific mutex parameter(s) can be specified, \eg @waitfor( rtn, m1 )@.
1752@waitfor@ statically verifies the released monitors are the same as the acquired mutex-parameters of the given routine or routine pointer.
1753To statically verify the released monitors match with the accepted routine's mutex parameters, the routine (pointer) prototype must be accessible.
1754% When an overloaded routine appears in an @waitfor@ statement, calls to any routine with that name are accepted.
1755% The rationale is that members with the same name should perform a similar function, and therefore, all should be eligible to accept a call.
1756Overloaded routines can be disambiguated using a cast:
1757\begin{cfa}
1758void rtn( M & mutex m );
1759`int` rtn( M & mutex m );
1760waitfor( (`int` (*)( M & mutex ))rtn, m );
1761\end{cfa}
1762
1763The ability to release a subset of acquired monitors can result in a \newterm{nested monitor}~\cite{Lister77} deadlock.
1764\begin{cfa}
1765void foo( M & mutex m1, M & mutex m2 ) {
1766        ... wait( `e, m1` ); ...                                $\C{// release m1, keeping m2 acquired )}$
1767void bar( M & mutex m1, M & mutex m2 ) {        $\C{// must acquire m1 and m2 )}$
1768        ... signal( `e` ); ...
1769\end{cfa}
1770The @wait@ only releases @m1@ so the signalling thread cannot acquire both @m1@ and @m2@ to  enter @bar@ to get to the @signal@.
1771While deadlock issues can occur with multiple/nesting acquisition, this issue results from the fact that locks, and by extension monitors, are not perfectly composable.
1772
1773Finally, an important aspect of monitor implementation is barging, \ie can calling threads barge ahead of signalled threads?
1774If barging is allowed, synchronization between a signaller and signallee is difficult, often requiring multiple unblock/block cycles (looping around a wait rechecking if a condition is met).
1775In fact, signals-as-hints is completely opposite from that proposed by Hoare in the seminal paper on monitors:
1776\begin{quote}
1777However, we decree that a signal operation be followed immediately by resumption of a waiting program, without possibility of an intervening procedure call from yet a third program.
1778It is only in this way that a waiting program has an absolute guarantee that it can acquire the resource just released by the signalling program without any danger that a third program will interpose a monitor entry and seize the resource instead.~\cite[p.~550]{Hoare74}
1779\end{quote}
1780\CFA scheduling \emph{precludes} barging, which simplifies synchronization among threads in the monitor and increases correctness.
1781Furthermore, \CFA concurrency has no spurious wakeup~\cite[\S~9]{Buhr05a}, which eliminates an implict form of barging.
1782For example, there are no loops in either bounded buffer solution in Figure~\ref{f:GenericBoundedBuffer}.
1783Supporting barging prevention as well as extending internal scheduling to multiple monitors is the main source of complexity in the design and implementation of \CFA concurrency.
1784
1785
1786\subsection{Barging Prevention}
1787
1788Figure~\ref{f:BargingPrevention} shows \CFA code where bulk acquire adds complexity to the internal-signalling semantics.
1789The complexity begins at the end of the inner @mutex@ statement, where the semantics of internal scheduling need to be extended for multiple monitors.
1790The problem is that bulk acquire is used in the inner @mutex@ statement where one of the monitors is already acquired.
1791When the signalling thread reaches the end of the inner @mutex@ statement, it should transfer ownership of @m1@ and @m2@ to the waiting threads to prevent barging into the outer @mutex@ statement by another thread.
1792However, both the signalling and waiting thread W1 still need monitor @m1@.
1793
1794\begin{figure}
1795\newbox\myboxA
1796\begin{lrbox}{\myboxA}
1797\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1798monitor M m1, m2;
1799condition c;
1800mutex( m1 ) { // $\LstCommentStyle{\color{red}outer}$
1801        ...
1802        mutex( m1, m2 ) { // $\LstCommentStyle{\color{red}inner}$
1803                ... `signal( c )`; ...
1804                // m1, m2 acquired
1805        } // $\LstCommentStyle{\color{red}release m2}$
1806        // m1 acquired
1807} // release m1
1808\end{cfa}
1809\end{lrbox}
1810
1811\newbox\myboxB
1812\begin{lrbox}{\myboxB}
1813\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1814
1815
1816mutex( m1 ) {
1817        ...
1818        mutex( m1, m2 ) {
1819                ... `wait( c )`; // block and release m1, m2
1820                // m1, m2 acquired
1821        } // $\LstCommentStyle{\color{red}release m2}$
1822        // m1 acquired
1823} // release m1
1824\end{cfa}
1825\end{lrbox}
1826
1827\newbox\myboxC
1828\begin{lrbox}{\myboxC}
1829\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1830
1831
1832mutex( m2 ) {
1833        ... `wait( c )`; ...
1834        // m2 acquired
1835} // $\LstCommentStyle{\color{red}release m2}$
1836
1837
1838
1839
1840\end{cfa}
1841\end{lrbox}
1842
1843\begin{cquote}
1844\subfloat[Signalling Thread]{\label{f:SignallingThread}\usebox\myboxA}
1845\hspace{2\parindentlnth}
1846\subfloat[Waiting Thread (W1)]{\label{f:WaitingThread}\usebox\myboxB}
1847\hspace{2\parindentlnth}
1848\subfloat[Waiting Thread (W2)]{\label{f:OtherWaitingThread}\usebox\myboxC}
1849\end{cquote}
1850\caption{Barging Prevention}
1851\label{f:BargingPrevention}
1852\end{figure}
1853
1854One scheduling solution is for the signaller to keep ownership of all locks until the last lock is ready to be transferred, because this semantics fits most closely to the behaviour of single-monitor scheduling.
1855However, Figure~\ref{f:OtherWaitingThread} shows this solution is complex depending on other waiters, resulting in options when the signaller finishes the inner mutex-statement.
1856The signaller can retain @m2@ until completion of the outer mutex statement and pass the locks to waiter W1, or it can pass @m2@ to waiter W2 after completing the inner mutex-statement, while continuing to hold @m1@.
1857In the latter case, waiter W2 must eventually pass @m2@ to waiter W1, which is complex because W1 may have waited before W2, so W2 is unaware of it.
1858Furthermore, there is an execution sequence where the signaller always finds waiter W2, and hence, waiter W1 starves.
1859
1860While a number of approaches were examined~\cite[\S~4.3]{Delisle18}, the solution chosen for \CFA is a novel techique called \newterm{partial signalling}.
1861Signalled threads are moved to the urgent queue and the waiter at the front defines the set of monitors necessary for it to unblock.
1862Partial signalling transfers ownership of monitors to the front waiter.
1863When the signaller thread exits or waits in the monitor, the front waiter is unblocked if all its monitors are released.
1864The benefit is encapsulating complexity into only two actions: passing monitors to the next owner when they should be released and conditionally waking threads if all conditions are met.
1865
1866
1867\subsection{Loose Object Definitions}
1868\label{s:LooseObjectDefinitions}
1869
1870In an object-oriented programming-language, a class includes an exhaustive list of operations.
1871However, new members can be added via static inheritance or dynamic members, \eg JavaScript~\cite{JavaScript}.
1872Similarly, monitor routines can be added at any time in \CFA, making it less clear for programmers and more difficult to implement.
1873\begin{cfa}
1874monitor M { ... };
1875void `f`( M & mutex m );
1876void g( M & mutex m ) { waitfor( `f` ); }       $\C{// clear which f}$
1877void `f`( M & mutex m, int );                           $\C{// different f}$
1878void h( M & mutex m ) { waitfor( `f` ); }       $\C{// unclear which f}$
1879\end{cfa}
1880Hence, the cfa-code for entering a monitor looks like:
1881\begin{cfa}
1882if ( $\textrm{\textit{monitor is free}}$ ) $\LstCommentStyle{// \color{red}enter}$
1883else if ( $\textrm{\textit{already own monitor}}$ ) $\LstCommentStyle{// \color{red}continue}$
1884else if ( $\textrm{\textit{monitor accepts me}}$ ) $\LstCommentStyle{// \color{red}enter}$
1885else $\LstCommentStyle{// \color{red}block}$
1886\end{cfa}
1887For the first two conditions, it is easy to implement a check that can evaluate the condition in a few instructions.
1888However, a fast check for \emph{monitor accepts me} is much harder to implement depending on the constraints put on the monitors.
1889Figure~\ref{fig:ClassicalMonitor} shows monitors are often expressed as an entry (calling) queue, some acceptor queues, and an urgent stack/queue.
1890
1891\begin{figure}
1892\centering
1893\subfloat[Classical monitor] {
1894\label{fig:ClassicalMonitor}
1895{\resizebox{0.45\textwidth}{!}{\input{monitor.pstex_t}}}
1896}% subfloat
1897\quad
1898\subfloat[Bulk acquire monitor] {
1899\label{fig:BulkMonitor}
1900{\resizebox{0.45\textwidth}{!}{\input{ext_monitor.pstex_t}}}
1901}% subfloat
1902\caption{Monitor Implementation}
1903\label{f:MonitorImplementation}
1904\end{figure}
1905
1906For a fixed (small) number of mutex routines (\eg 128), the accept check reduces to a bitmask of allowed callers, which can be checked with a single instruction.
1907This approach requires a unique dense ordering of routines with a small upper-bound and the ordering must be consistent across translation units.
1908For object-oriented languages these constraints are common, but \CFA mutex routines can be added in any scope and are only visible in certain translation unit, precluding program-wide dense-ordering among mutex routines.
1909
1910Figure~\ref{fig:BulkMonitor} shows the \CFA monitor implementation.
1911The mutex routine called is associated with each thread on the entry queue, while a list of acceptable routines is kept separately.
1912The accepted list is a variable-sized array of accepted routine pointers, so the single instruction bitmask comparison is replaced by dereferencing a pointer followed by a (usually short) linear search.
1913
1914
1915\subsection{Multi-Monitor Scheduling}
1916\label{s:Multi-MonitorScheduling}
1917
1918External scheduling, like internal scheduling, becomes significantly more complex for multi-monitor semantics.
1919Even in the simplest case, new semantics needs to be established.
1920\newpage
1921\begin{cfa}
1922monitor M { ... };
1923void f( M & mutex m1 );
1924void g( M & mutex m1, M & mutex m2 ) {
1925        waitfor( f );                                                   $\C{\color{red}// pass m1 or m2 to f?}$
1926}
1927\end{cfa}
1928The solution is for the programmer to disambiguate:
1929\begin{cfa}
1930        waitfor( f, m2 );                                               $\C{\color{red}// wait for call to f with argument m2}$
1931\end{cfa}
1932Both locks are acquired by routine @g@, so when routine @f@ is called, the lock for monitor @m2@ is passed from @g@ to @f@, while @g@ still holds lock @m1@.
1933This behaviour can be extended to the multi-monitor @waitfor@ statement.
1934\begin{cfa}
1935monitor M { ... };
1936void f( M & mutex m1, M & mutex m2 );
1937void g( M & mutex m1, M & mutex m2 ) {
1938        waitfor( f, m1, m2 );                                   $\C{\color{red}// wait for call to f with arguments m1 and m2}$
1939}
1940\end{cfa}
1941Again, the set of monitors passed to the @waitfor@ statement must be entirely contained in the set of monitors already acquired by the accepting routine.
1942Also, the order of the monitors in a @waitfor@ statement is unimportant.
1943
1944Figure~\ref{f:UnmatchedMutexSets} shows an example where, for internal and external scheduling with multiple monitors, a signalling or accepting thread must match exactly, \ie partial matching results in waiting.
1945For both examples, the set of monitors is disjoint so unblocking is impossible.
1946
1947\begin{figure}
1948\lstDeleteShortInline@%
1949\begin{tabular}{@{}l@{\hspace{\parindentlnth}}|@{\hspace{\parindentlnth}}l@{}}
1950\begin{cfa}
1951monitor M1 {} m11, m12;
1952monitor M2 {} m2;
1953condition c;
1954void f( M1 & mutex m1, M2 & mutex m2 ) {
1955        signal( c );
1956}
1957void g( M1 & mutex m1, M2 & mutex m2 ) {
1958        wait( c );
1959}
1960g( `m11`, m2 ); // block on wait
1961f( `m12`, m2 ); // cannot fulfil
1962\end{cfa}
1963&
1964\begin{cfa}
1965monitor M1 {} m11, m12;
1966monitor M2 {} m2;
1967
1968void f( M1 & mutex m1, M2 & mutex m2 ) {
1969
1970}
1971void g( M1 & mutex m1, M2 & mutex m2 ) {
1972        waitfor( f, m1, m2 );
1973}
1974g( `m11`, m2 ); // block on accept
1975f( `m12`, m2 ); // cannot fulfil
1976\end{cfa}
1977\end{tabular}
1978\lstMakeShortInline@%
1979\caption{Unmatched \protect\lstinline@mutex@ sets}
1980\label{f:UnmatchedMutexSets}
1981\end{figure}
1982
1983
1984\subsection{Extended \protect\lstinline@waitfor@}
1985
1986Figure~\ref{f:ExtendedWaitfor} show the extended form of the @waitfor@ statement to conditionally accept one of a group of mutex routines, with a specific action to be performed \emph{after} the mutex routine finishes.
1987For a @waitfor@ clause to be executed, its @when@ must be true and an outstanding call to its corresponding member(s) must exist.
1988The \emph{conditional-expression} of a @when@ may call a routine, but the routine must not block or context switch.
1989If there are multiple acceptable mutex calls, selection occurs top-to-bottom (prioritized) in the @waitfor@ clauses, whereas some programming languages with similar mechanisms accept non-deterministically for this case, \eg Go \lstinline[morekeywords=select]@select@.
1990If some accept guards are true and there are no outstanding calls to these members, the acceptor is accept-blocked until a call to one of these members is made.
1991If all the accept guards are false, the statement does nothing, unless there is a terminating @else@ clause with a true guard, which is executed instead.
1992Hence, the terminating @else@ clause allows a conditional attempt to accept a call without blocking.
1993If there is a @timeout@ clause, it provides an upper bound on waiting.
1994If both a @timeout@ clause and an @else@ clause are present, the @else@ must be conditional, or the @timeout@ is never triggered.
1995In all cases, the statement following is executed \emph{after} a clause is executed to know which of the clauses executed.
1996
1997\begin{figure}
1998\begin{cfa}
1999`when` ( $\emph{conditional-expression}$ )      $\C{// optional guard}$
2000        waitfor( $\emph{mutex-member-name}$ )
2001                $\emph{statement}$                                      $\C{// action after call}$
2002`or` `when` ( $\emph{conditional-expression}$ ) $\C{// optional guard}$
2003        waitfor( $\emph{mutex-member-name}$ )
2004                $\emph{statement}$                                      $\C{// action after call}$
2005`or`    ...                                                                     $\C{// list of waitfor clauses}$
2006`when` ( $\emph{conditional-expression}$ )      $\C{// optional guard}$
2007        `timeout`                                                               $\C{// optional terminating timeout clause}$
2008                $\emph{statement}$                                      $\C{// action after timeout}$
2009`when` ( $\emph{conditional-expression}$ )      $\C{// optional guard}$
2010        `else`                                                                  $\C{// optional terminating clause}$
2011                $\emph{statement}$                                      $\C{// action when no immediate calls}$
2012\end{cfa}
2013\caption{Extended \protect\lstinline@waitfor@}
2014\label{f:ExtendedWaitfor}
2015\end{figure}
2016
2017Note, a group of conditional @waitfor@ clauses is \emph{not} the same as a group of @if@ statements, e.g.:
2018\begin{cfa}
2019if ( C1 ) waitfor( mem1 );                       when ( C1 ) waitfor( mem1 );
2020else if ( C2 ) waitfor( mem2 );         or when ( C2 ) waitfor( mem2 );
2021\end{cfa}
2022The left example accepts only @mem1@ if @C1@ is true or only @mem2@ if @C2@ is true.
2023The right example accepts either @mem1@ or @mem2@ if @C1@ and @C2@ are true.
2024
2025An interesting use of @waitfor@ is accepting the @mutex@ destructor to know when an object is deallocated.
2026\begin{cfa}
2027void insert( Buffer(T) & mutex buffer, T elem ) with( buffer ) {
2028        if ( count == 10 )
2029                waitfor( remove, buffer ) {
2030                        // insert elem into buffer
2031                } or `waitfor( ^?{}, buffer )` throw insertFail;
2032}
2033\end{cfa}
2034When the buffer is deallocated, the current waiter is unblocked and informed, so it can perform an appropriate action.
2035However, the basic @waitfor@ semantics do not support this functionality, since using an object after its destructor is called is undefined.
2036Therefore, to make this useful capability work, the semantics for accepting the destructor is the same as @signal@, \ie the call to the destructor is placed on the urgent queue and the acceptor continues execution, which throws an exception to the acceptor and then the caller is unblocked from the urgent queue to deallocate the object.
2037Accepting the destructor is an idiomatic way to terminate a thread in \CFA.
2038
2039
2040\subsection{\protect\lstinline@mutex@ Threads}
2041
2042Threads in \CFA are monitors to allow direct communication among threads, \ie threads can have mutex routines that are called by other threads.
2043Hence, all monitor features are available when using threads.
2044The following shows an example of two threads directly calling each other and accepting calls from each other in a cycle.
2045\begin{cfa}
2046thread Ping {} pi;
2047thread Pong {} po;
2048void ping( Ping & mutex ) {}
2049void pong( Pong & mutex ) {}
2050int main() {}
2051\end{cfa}
2052\vspace{-0.8\baselineskip}
2053\begin{cquote}
2054\begin{tabular}{@{}l@{\hspace{3\parindentlnth}}l@{}}
2055\begin{cfa}
2056void main( Ping & pi ) {
2057        for ( int i = 0; i < 10; i += 1 ) {
2058                `waitfor( ping, pi );`
2059                `pong( po );`
2060        }
2061}
2062\end{cfa}
2063&
2064\begin{cfa}
2065void main( Pong & po ) {
2066        for ( int i = 0; i < 10; i += 1 ) {
2067                `ping( pi );`
2068                `waitfor( pong, po );`
2069        }
2070}
2071\end{cfa}
2072\end{tabular}
2073\end{cquote}
2074% \lstMakeShortInline@%
2075% \caption{Threads ping/pong using external scheduling}
2076% \label{f:pingpong}
2077% \end{figure}
2078Note, the ping/pong threads are globally declared, @pi@/@po@, and hence, start (and possibly complete) before the program main starts.
2079
2080
2081\subsection{Low-level Locks}
2082
2083For completeness and efficiency, \CFA provides a standard set of low-level locks: recursive mutex, condition, semaphore, barrier, \etc, and atomic instructions: @fetchAssign@, @fetchAdd@, @testSet@, @compareSet@, \etc.
2084
2085
2086\section{Parallelism}
2087
2088Historically, computer performance was about processor speeds.
2089However, with heat dissipation being a direct consequence of speed increase, parallelism is the new source for increased performance~\cite{Sutter05, Sutter05b}.
2090Now, high-performance applications must care about parallelism, which requires concurrency.
2091The lowest-level approach of parallelism is to use \newterm{kernel threads} in combination with semantics like @fork@, @join@, \etc.
2092However, kernel threads are better as an implementation tool because of complexity and higher cost.
2093Therefore, different abstractions are often layered onto kernel threads to simplify them, \eg pthreads.
2094
2095
2096\subsection{User Threads with Preemption}
2097
2098A direct improvement on kernel threads is user threads, \eg Erlang~\cite{Erlang} and \uC~\cite{uC++book}.
2099This approach provides an interface that matches the language paradigms, more control over concurrency by the language runtime, and an abstract (and portable) interface to the underlying kernel threads across operating systems.
2100In many cases, user threads can be used on a much larger scale (100,000 threads).
2101Like kernel threads, user threads support preemption, which maximizes nondeterminism, but introduces the same concurrency errors: race, livelock, starvation, and deadlock.
2102\CFA adopts user-threads as they represent the truest realization of concurrency and can build any other concurrency approach, \eg thread pools and actors~\cite{Actors}.
2103
2104
2105\subsection{User Threads without Preemption (Fiber)}
2106\label{s:fibers}
2107
2108A variant of user thread is \newterm{fibers}, which removes preemption, \eg Go~\cite{Go} @goroutine@s.
2109Like functional programming, which removes mutation and its associated problems, removing preemption from concurrency reduces nondeterminism, making race and deadlock errors more difficult to generate.
2110However, preemption is necessary for concurrency that relies on spinning, so there are a class of problems that cannot be programmed without preemption.
2111
2112
2113\subsection{Thread Pools}
2114
2115In contrast to direct threading is indirect \newterm{thread pools}, where small jobs (work units) are inserted into a work pool for execution.
2116If the jobs are dependent, \ie interact, there is an implicit/explicit dependency graph that ties them together.
2117While removing direct concurrency, and hence the amount of context switching, thread pools significantly limit the interaction that can occur among jobs.
2118Indeed, jobs should not block because that also blocks the underlying thread, which effectively means the CPU utilization, and therefore throughput, suffers.
2119While it is possible to tune the thread pool with sufficient threads, it becomes difficult to obtain high throughput and good core utilization as job interaction increases.
2120As well, concurrency errors return, which threads pools are suppose to mitigate.
2121
2122
2123\section{\protect\CFA Runtime Structure}
2124
2125Figure~\ref{f:RunTimeStructure} illustrates the runtime structure of a \CFA program.
2126In addition to the new kinds of objects introduced by \CFA, there are two more runtime entities used to control parallel execution: cluster and (virtual) processor.
2127An executing thread is illustrated by its containment in a processor.
2128
2129\begin{figure}
2130\centering
2131\input{RunTimeStructure}
2132\caption{\CFA Runtime Structure}
2133\label{f:RunTimeStructure}
2134\end{figure}
2135
2136
2137\subsection{Cluster}
2138\label{s:RuntimeStructureCluster}
2139
2140A \newterm{cluster} is a collection of threads and virtual processors (abstract kernel-thread) that execute the threads (like a virtual machine).
2141The purpose of a cluster is to control the amount of parallelism that is possible among threads, plus scheduling and other execution defaults.
2142The default cluster-scheduler is single-queue multi-server, which provides automatic load-balancing of threads on processors.
2143However, the scheduler is pluggable, supporting alternative schedulers.
2144If several clusters exist, both threads and virtual processors, can be explicitly migrated from one cluster to another.
2145No automatic load balancing among clusters is performed by \CFA.
2146
2147When a \CFA program begins execution, it creates a user cluster with a single processor and a special processor to handle preemption that does not execute user threads.
2148The user cluster is created to contain the application user-threads.
2149Having all threads execute on the one cluster often maximizes utilization of processors, which minimizes runtime.
2150However, because of limitations of the underlying operating system, heterogeneous hardware, or scheduling requirements (real-time), multiple clusters are sometimes necessary.
2151
2152
2153\subsection{Virtual Processor}
2154\label{s:RuntimeStructureProcessor}
2155
2156A virtual processor is implemented by a kernel thread (\eg UNIX process), which is subsequently scheduled for execution on a hardware processor by the underlying operating system.
2157Programs may use more virtual processors than hardware processors.
2158On a multiprocessor, kernel threads are distributed across the hardware processors resulting in virtual processors executing in parallel.
2159(It is possible to use affinity to lock a virtual processor onto a particular hardware processor~\cite{affinityLinux, affinityWindows, affinityFreebsd, affinityNetbsd, affinityMacosx}, which is used when caching issues occur or for heterogeneous hardware processors.)
2160The \CFA runtime attempts to block unused processors and unblock processors as the system load increases;
2161balancing the workload with processors is difficult.
2162Preemption occurs on virtual processors rather than user threads, via operating-system interrupts.
2163Thus virtual processors execute user threads, where preemption frequency applies to a virtual processor, so preemption occurs randomly across the executed user threads.
2164Turning off preemption transforms user threads into fibers.
2165
2166
2167\subsection{Debug Kernel}
2168
2169There are two versions of the \CFA runtime kernel: debug and non-debug.
2170The debugging version has many runtime checks and internal assertions, \eg stack (non-writable) guard page, and checks for stack overflow whenever context switches occur among coroutines and threads, which catches most stack overflows.
2171After a program is debugged, the non-debugging version can be used to decrease space and increase performance.
2172
2173
2174\section{Implementation}
2175\label{s:Implementation}
2176
2177Currently, \CFA has fixed-sized stacks, where the stack size can be set at coroutine/thread creation but with no subsequent growth.
2178Schemes exist for dynamic stack-growth, such as stack copying and chained stacks.
2179However, stack copying requires pointer adjustment to items on the stack, which is impossible without some form of garbage collection.
2180As well, chained stacks require all modules be recompiled to use this feature, which breaks backward compatibility with existing C libraries.
2181In the long term, it is likely C libraries will migrate to stack chaining to support concurrency, at only a minimal cost to sequential programs.
2182Nevertheless, experience teaching \uC~\cite{CS343} shows fixed-sized stacks are rarely an issue in most concurrent programs.
2183
2184A primary implementation challenge is avoiding contention from dynamically allocating memory because of bulk acquire, \eg the internal-scheduling design is (almost) free of allocations.
2185All blocking operations are made by parking threads onto queues, therefore all queues are designed with intrusive nodes, where each node has preallocated link fields for chaining.
2186Furthermore, several bulk-acquire operations need a variable amount of memory.
2187This storage is allocated at the base of a thread's stack before blocking, which means programmers must add a small amount of extra space for stacks.
2188
2189In \CFA, ordering of monitor acquisition relies on memory ordering to prevent deadlock~\cite{Havender68}, because all objects have distinct non-overlapping memory layouts, and mutual-exclusion for a monitor is only defined for its lifetime.
2190When a mutex call is made, pointers to the concerned monitors are aggregated into a variable-length array and sorted.
2191This array persists for the entire duration of the mutual exclusion and is used extensively for synchronization operations.
2192
2193To improve performance and simplicity, context switching occurs inside a routine call, so only callee-saved registers are copied onto the stack and then the stack register is switched;
2194the corresponding registers are then restored for the other context.
2195Note, the instruction pointer is untouched since the context switch is always inside the same routine.
2196Unlike coroutines, threads do not context switch among each other;
2197they context switch to the cluster scheduler.
2198This method is a 2-step context-switch and provides a clear distinction between user and kernel code, where scheduling and other system operations happen.
2199The alternative 1-step context-switch uses the \emph{from} thread's stack to schedule and then context-switches directly to the \emph{to} thread's stack.
2200Experimental results (not presented) show the performance of these two approaches is virtually equivalent, because both approaches are dominated by locking to prevent a race condition.
2201
2202All kernel threads (@pthreads@) created a stack.
2203Each \CFA virtual processor is implemented as a coroutine and these coroutines run directly on the kernel-thread stack, effectively stealing this stack.
2204The exception to this rule is the program main, \ie the initial kernel thread that is given to any program.
2205In order to respect C expectations, the stack of the initial kernel thread is used by program main rather than the main processor, allowing it to grow dynamically as in a normal C program.
2206
2207Finally, an important aspect for a complete threading system is preemption, which introduces extra non-determinism via transparent interleaving, rather than cooperation among threads for proper scheduling and processor fairness from long-running threads.
2208Because preemption frequency is usually long (1 millisecond) performance cost is negligible.
2209Preemption is normally handled by setting a count-down timer on each virtual processor.
2210When the timer expires, an interrupt is delivered, and the interrupt handler resets the count-down timer, and if the virtual processor is executing in user code, the signal handler performs a user-level context-switch, or if executing in the language runtime-kernel, the preemption is ignored or rolled forward to the point where the runtime kernel context switches back to user code.
2211Multiple signal handlers may be pending.
2212When control eventually switches back to the signal handler, it returns normally, and execution continues in the interrupted user thread, even though the return from the signal handler may be on a different kernel thread than the one where the signal is delivered.
2213The only issue with this approach is that signal masks from one kernel thread may be restored on another as part of returning from the signal handler;
2214therefore, the same signal mask is required for all virtual processors in a cluster.
2215
2216However, on current UNIX systems:
2217\begin{quote}
2218A process-directed signal may be delivered to any one of the threads that does not currently have the signal blocked.
2219If more than one of the threads has the signal unblocked, then the kernel chooses an arbitrary thread to which to deliver the signal.
2220SIGNAL(7) - Linux Programmer's Manual
2221\end{quote}
2222Hence, the timer-expiry signal, which is generated \emph{externally} by the UNIX kernel to the UNIX process, is delivered to any of its UNIX subprocesses (kernel threads).
2223To ensure each virtual processor receives its own preemption signals, a discrete-event simulation is run on a special virtual processor, and only it sets and receives timer events.
2224Virtual processors register an expiration time with the discrete-event simulator, which is inserted in sorted order.
2225The simulation sets the count-down timer to the value at the head of the event list, and when the timer expires, all events less than or equal to the current time are processed.
2226Processing a preemption event sends an \emph{internal} @SIGUSR1@ signal to the registered virtual processor, which is always delivered to that processor.
2227
2228
2229\section{Performance}
2230\label{results}
2231
2232To verify the implementation of the \CFA runtime, a series of microbenchmarks are performed comparing \CFA with other widely used programming languages with concurrency.
2233Table~\ref{t:machine} shows the specifications of the computer used to run the benchmarks, and the versions of the software used in the comparison.
2234
2235\begin{table}
2236\centering
2237\caption{Experiment environment}
2238\label{t:machine}
2239
2240\begin{tabular}{|l|r||l|r|}
2241\hline
2242Architecture            & x86\_64                               & NUMA node(s)  & 8 \\
2243\hline
2244CPU op-mode(s)          & 32-bit, 64-bit                & Model name    & AMD Opteron\texttrademark\ Processor 6380 \\
2245\hline
2246Byte Order                      & Little Endian                 & CPU Freq              & 2.5 GHz \\
2247\hline
2248CPU(s)                          & 64                                    & L1d cache     & 16 KiB \\
2249\hline
2250Thread(s) per core      & 2                                     & L1i cache     & 64 KiB \\
2251\hline
2252Core(s) per socket      & 8                                     & L2 cache              & 2048 KiB \\
2253\hline
2254Socket(s)                       & 4                                     & L3 cache              & 6144 KiB \\
2255\hline
2256\hline
2257Operating system        & Ubuntu 16.04.3 LTS    & Kernel                & Linux 4.4-97-generic \\
2258\hline
2259gcc                                     & 6.3                                   & \CFA                  & 1.0.0 \\
2260\hline
2261Java                            & OpenJDK-9                     & Go                    & 1.9.2 \\
2262\hline
2263\end{tabular}
2264\end{table}
2265
2266All benchmarks are run using the following harness:
2267\begin{cfa}
2268unsigned int N = 10_000_000;
2269#define BENCH( run ) Time before = getTimeNsec(); run; Duration result = (getTimeNsec() - before) / N;
2270\end{cfa}
2271The method used to get time is @clock_gettime( CLOCK_REALTIME )@.
2272Each benchmark is performed @N@ times, where @N@ varies depending on the benchmark;
2273the total time is divided by @N@ to obtain the average time for a benchmark.
2274All omitted tests for other languages are functionally identical to the shown \CFA test.
2275
2276
2277\paragraph{Context-Switching}
2278
2279In procedural programming, the cost of a routine call is important as modularization (refactoring) increases.
2280(In many cases, a compiler inlines routine calls to eliminate this cost.)
2281Similarly, when modularization extends to coroutines/tasks, the time for a context switch becomes a relevant factor.
2282The coroutine context-switch is 2-step using resume/suspend, \ie from resumer to suspender and from suspender to resumer.
2283The thread context switch is 2-step using yield, \ie enter and return from the runtime kernel.
2284Figure~\ref{f:ctx-switch} shows the code for coroutines/threads with all results in Table~\ref{tab:ctx-switch}.
2285The difference in performance between coroutine and thread context-switch is the cost of scheduling for threads, whereas coroutines are self-scheduling.
2286
2287\begin{figure}
2288\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
2289
2290\newbox\myboxA
2291\begin{lrbox}{\myboxA}
2292\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2293coroutine C {} c;
2294void main( C & ) { for ( ;; ) { @suspend();@ } }
2295int main() {
2296        BENCH(
2297                for ( size_t i = 0; i < N; i += 1 ) { @resume( c );@ } )
2298        sout | result`ns;
2299}
2300\end{cfa}
2301\end{lrbox}
2302
2303\newbox\myboxB
2304\begin{lrbox}{\myboxB}
2305\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2306
2307
2308int main() {
2309        BENCH(
2310                for ( size_t i = 0; i < N; i += 1 ) { @yield();@ } )
2311        sout | result`ns;
2312}
2313\end{cfa}
2314\end{lrbox}
2315
2316\subfloat[Coroutine]{\usebox\myboxA}
2317\quad
2318\subfloat[Thread]{\usebox\myboxB}
2319\captionof{figure}{\CFA context-switch benchmark}
2320\label{f:ctx-switch}
2321
2322\centering
2323
2324\captionof{table}{Context switch comparison (nanoseconds)}
2325\label{tab:ctx-switch}
2326\bigskip
2327\begin{tabular}{|r|*{3}{D{.}{.}{3.2}|}}
2328\cline{2-4}
2329\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{Median} &\multicolumn{1}{c|}{Average} & \multicolumn{1}{c|}{Std Dev} \\
2330\hline
2331Kernel Thread   & 333.5 & 332.96        & 4.1 \\
2332\CFA Coroutine  & 49            & 48.68 & 0.47    \\
2333\CFA Thread             & 105           & 105.57        & 1.37 \\
2334\uC Coroutine   & 44            & 44            & 0 \\
2335\uC Thread              & 100           & 99.29 & 0.96 \\
2336Goroutine               & 145           & 147.25        & 4.15 \\
2337Java Thread             & 373.5 & 375.14        & 8.72 \\
2338\hline
2339\end{tabular}
2340
2341\bigskip
2342\bigskip
2343
2344\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
2345\begin{cfa}
2346monitor M { ... } m1/*, m2, m3, m4*/;
2347void __attribute__((noinline)) do_call( M & mutex m/*, m2, m3, m4*/ ) {}
2348int main() {
2349        BENCH( for( size_t i = 0; i < N; i += 1 ) { @do_call( m1/*, m2, m3, m4*/ );@ } )
2350        sout | result`ns;
2351}
2352\end{cfa}
2353\captionof{figure}{\CFA acquire/release mutex benchmark}
2354\label{f:mutex}
2355
2356\centering
2357
2358\captionof{table}{Mutex comparison (nanoseconds)}
2359\label{tab:mutex}
2360\bigskip
2361
2362\begin{tabular}{|r|*{3}{D{.}{.}{3.2}|}}
2363\cline{2-4}
2364\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{Median} &\multicolumn{1}{c|}{Average} & \multicolumn{1}{c|}{Std Dev} \\
2365\hline
2366C routine                                       & 2             & 2             & 0    \\
2367FetchAdd + FetchSub                     & 26            & 26            & 0    \\
2368Pthreads Mutex Lock                     & 31            & 31.71 & 0.97 \\
2369\uC @monitor@ member routine            & 31            & 31            & 0    \\
2370\CFA @mutex@ routine, 1 argument        & 46            & 46.68 & 0.93  \\
2371\CFA @mutex@ routine, 2 argument        & 84            & 85.36 & 1.99 \\
2372\CFA @mutex@ routine, 4 argument        & 158           & 161           & 4.22 \\
2373Java synchronized routine               & 27.5  & 29.79 & 2.93  \\
2374\hline
2375\end{tabular}
2376\end{figure}
2377
2378
2379\paragraph{Mutual-Exclusion}
2380
2381Mutual exclusion is measured by entering/leaving a critical section.
2382For monitors, entering and leaving a monitor routine is measured.
2383Figure~\ref{f:mutex} shows the code for \CFA with all results in Table~\ref{tab:mutex}.
2384To put the results in context, the cost of entering a non-inline routine and the cost of acquiring and releasing a @pthread_mutex@ lock is also measured.
2385Note, the incremental cost of bulk acquire for \CFA, which is largely a fixed cost for small numbers of mutex objects.
2386
2387
2388\paragraph{Internal Scheduling}
2389
2390Internal scheduling is measured by waiting on and signalling a condition variable.
2391Figure~\ref{f:int-sched} shows the code for \CFA, with results in Table~\ref{tab:int-sched}.
2392Note, the incremental cost of bulk acquire for \CFA, which is largely a fixed cost for small numbers of mutex objects.
2393Java scheduling is significantly greater because the benchmark explicitly creates multiple thread in order to prevent the JIT from making the program sequential, \ie removing all locking.
2394
2395\begin{figure}
2396\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
2397\begin{cfa}
2398volatile int go = 0;
2399condition c;
2400monitor M { ... } m;
2401void __attribute__((noinline)) do_call( M & mutex a1 ) { signal( c ); }
2402thread T {};
2403void main( T & this ) {
2404        while ( go == 0 ) { yield(); }  // wait for other thread to start
2405        while ( go == 1 ) { @do_call( m );@ }
2406}
2407int  __attribute__((noinline)) do_wait( M & mutex m ) {
2408        go = 1; // continue other thread
2409        BENCH( for ( size_t i = 0; i < N; i += 1 ) { @wait( c );@ } );
2410        go = 0; // stop other thread
2411        sout | result`ns;
2412}
2413int main() {
2414        T t;
2415        do_wait( m );
2416}
2417\end{cfa}
2418\captionof{figure}{\CFA Internal-scheduling benchmark}
2419\label{f:int-sched}
2420
2421\centering
2422\captionof{table}{Internal-scheduling comparison (nanoseconds)}
2423\label{tab:int-sched}
2424\bigskip
2425
2426\begin{tabular}{|r|*{3}{D{.}{.}{5.2}|}}
2427\cline{2-4}
2428\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{Median} &\multicolumn{1}{c|}{Average} & \multicolumn{1}{c|}{Std Dev} \\
2429\hline
2430Pthreads Condition Variable             & 6005  & 5681.43       & 835.45 \\
2431\uC @signal@                                    & 324           & 325.54        & 3,02   \\
2432\CFA @signal@, 1 @monitor@              & 368.5         & 370.61        & 4.77   \\
2433\CFA @signal@, 2 @monitor@              & 467           & 470.5 & 6.79   \\
2434\CFA @signal@, 4 @monitor@              & 700.5         & 702.46        & 7.23  \\
2435Java @notify@                                   & 15471 & 172511        & 5689 \\
2436\hline
2437\end{tabular}
2438\end{figure}
2439
2440
2441\paragraph{External Scheduling}
2442
2443External scheduling is measured by accepting a call using the @waitfor@ statement (@_Accept@ in \uC).
2444Figure~\ref{f:ext-sched} shows the code for \CFA, with results in Table~\ref{tab:ext-sched}.
2445Note, the incremental cost of bulk acquire for \CFA, which is largely a fixed cost for small numbers of mutex objects.
2446
2447\begin{figure}
2448\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
2449\begin{cfa}
2450volatile int go = 0;
2451monitor M { ... } m;
2452thread T {};
2453void __attribute__((noinline)) do_call( M & mutex ) {}
2454void main( T & ) {
2455        while ( go == 0 ) { yield(); }  // wait for other thread to start
2456        while ( go == 1 ) { @do_call( m );@ }
2457}
2458int __attribute__((noinline)) do_wait( M & mutex m ) {
2459        go = 1; // continue other thread
2460        BENCH( for ( size_t i = 0; i < N; i += 1 ) { @waitfor( do_call, m );@ } )
2461        go = 0; // stop other thread
2462        sout | result`ns;
2463}
2464int main() {
2465        T t;
2466        do_wait( m );
2467}
2468\end{cfa}
2469\captionof{figure}{\CFA external-scheduling benchmark}
2470\label{f:ext-sched}
2471
2472\centering
2473
2474\captionof{table}{External-scheduling comparison (nanoseconds)}
2475\label{tab:ext-sched}
2476\bigskip
2477\begin{tabular}{|r|*{3}{D{.}{.}{3.2}|}}
2478\cline{2-4}
2479\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{Median} &\multicolumn{1}{c|}{Average} & \multicolumn{1}{c|}{Std Dev} \\
2480\hline
2481\uC @_Accept@                           & 358           & 359.11        & 2.53  \\
2482\CFA @waitfor@, 1 @monitor@     & 359           & 360.93        & 4.07  \\
2483\CFA @waitfor@, 2 @monitor@     & 450           & 449.39        & 6.62  \\
2484\CFA @waitfor@, 4 @monitor@     & 652           & 655.64        & 7.73 \\
2485\hline
2486\end{tabular}
2487
2488\bigskip
2489\medskip
2490
2491\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
2492\begin{cfa}
2493thread MyThread {};
2494void main( MyThread & ) {}
2495int main() {
2496        BENCH( for ( size_t i = 0; i < N; i += 1 ) { @MyThread m;@ } )
2497        sout | result`ns;
2498}
2499\end{cfa}
2500\captionof{figure}{\CFA object-creation benchmark}
2501\label{f:creation}
2502
2503\centering
2504
2505\captionof{table}{Creation comparison (nanoseconds)}
2506\label{tab:creation}
2507\bigskip
2508
2509\begin{tabular}{|r|*{3}{D{.}{.}{5.2}|}}
2510\cline{2-4}
2511\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{Median} & \multicolumn{1}{c|}{Average} & \multicolumn{1}{c|}{Std Dev} \\
2512\hline
2513Pthreads                                & 28091         & 28073.39      & 163.1  \\
2514\CFA Coroutine Lazy             & 6                     & 6.07          & 0.26   \\
2515\CFA Coroutine Eager    & 520           & 520.61        & 2.04   \\
2516\CFA Thread                             & 2032  & 2016.29       & 112.07  \\
2517\uC Coroutine                   & 106           & 107.36        & 1.47   \\
2518\uC Thread                              & 536.5 & 537.07        & 4.64   \\
2519Goroutine                               & 3103  & 3086.29       & 90.25  \\
2520Java Thread                             & 103416.5      & 103732.29     & 1137 \\
2521\hline
2522\end{tabular}
2523\end{figure}
2524
2525
2526\paragraph{Object Creation}
2527
2528Object creation is measured by creating/deleting the specific kind of concurrent object.
2529Figure~\ref{f:creation} shows the code for \CFA, with results in Table~\ref{tab:creation}.
2530The only note here is that the call stacks of \CFA coroutines are lazily created, therefore without priming the coroutine to force stack creation, the creation cost is artificially low.
2531
2532
2533\section{Conclusion}
2534
2535This paper demonstrates a concurrency API that is simple, efficient, and able to build higher-level concurrency features.
2536The approach provides concurrency based on a preemptive M:N user-level threading-system, executing in clusters, which encapsulate scheduling of work on multiple kernel threads providing parallelism.
2537The M:N model is judged to be efficient and provide greater flexibility than a 1:1 threading model.
2538High-level objects (monitor/task) are the core mechanism for mutual exclusion and synchronization.
2539A novel aspect is allowing multiple mutex-objects to be accessed simultaneously reducing the potential for deadlock for this complex scenario.
2540These concepts and the entire \CFA runtime-system are written in the \CFA language, demonstrating the expressiveness of the \CFA language.
2541Performance comparisons with other concurrent systems/languages show the \CFA approach is competitive across all low-level operations, which translates directly into good performance in well-written concurrent applications.
2542C programmers should feel comfortable using these mechanisms for developing concurrent applications, with the ability to obtain maximum available performance by mechanisms at the appropriate level.
2543
2544
2545\section{Future Work}
2546
2547While concurrency in \CFA has a strong start, development is still underway and there are missing features.
2548
2549\paragraph{Flexible Scheduling}
2550\label{futur:sched}
2551
2552An important part of concurrency is scheduling.
2553Different scheduling algorithms can affect performance (both in terms of average and variation).
2554However, no single scheduler is optimal for all workloads and therefore there is value in being able to change the scheduler for given programs.
2555One solution is to offer various tweaking options, allowing the scheduler to be adjusted to the requirements of the workload.
2556However, to be truly flexible, a pluggable scheduler is necessary.
2557Currently, the \CFA pluggable scheduler is too simple to handle complex scheduling, \eg quality of service and real-time, where the scheduler must interact with mutex objects to deal with issues like priority inversion.
2558
2559\paragraph{Non-Blocking I/O}
2560\label{futur:nbio}
2561
2562Many modern workloads are not bound by computation but IO operations, a common case being web servers and XaaS~\cite{XaaS} (anything as a service).
2563These types of workloads require significant engineering to amortizing costs of blocking IO-operations.
2564At its core, non-blocking I/O is an operating-system level feature queuing IO operations, \eg network operations, and registering for notifications instead of waiting for requests to complete.
2565Current trends use asynchronous programming like callbacks, futures, and/or promises, \eg Node.js~\cite{NodeJs} for JavaScript, Spring MVC~\cite{SpringMVC} for Java, and Django~\cite{Django} for Python.
2566However, these solutions lead to code that is hard to create, read and maintain.
2567A better approach is to tie non-blocking I/O into the concurrency system to provide ease of use with low overhead, \eg thread-per-connection web-services.
2568A non-blocking I/O library is currently under development for \CFA.
2569
2570\paragraph{Other Concurrency Tools}
2571\label{futur:tools}
2572
2573While monitors offer flexible and powerful concurrency for \CFA, other concurrency tools are also necessary for a complete multi-paradigm concurrency package.
2574Examples of such tools can include futures and promises~\cite{promises}, executors and actors.
2575These additional features are useful when monitors offer a level of abstraction that is inadequate for certain tasks.
2576As well, new \CFA extensions should make it possible to create a uniform interface for virtually all mutual exclusion, including monitors and low-level locks.
2577
2578\paragraph{Implicit Threading}
2579\label{futur:implcit}
2580
2581Basic concurrent (embarrassingly parallel) applications can benefit greatly from implicit concurrency, where sequential programs are converted to concurrent, possibly with some help from pragmas to guide the conversion.
2582This type of concurrency can be achieved both at the language level and at the library level.
2583The canonical example of implicit concurrency is concurrent nested @for@ loops, which are amenable to divide and conquer algorithms~\cite{uC++book}.
2584The \CFA language features should make it possible to develop a reasonable number of implicit concurrency mechanism to solve basic HPC data-concurrency problems.
2585However, implicit concurrency is a restrictive solution with significant limitations, so it can never replace explicit concurrent programming.
2586
2587
2588\section{Acknowledgements}
2589
2590The authors would like to recognize the design assistance of Aaron Moss, Rob Schluntz and Andrew Beach on the features described in this paper.
2591Funding for this project has been provided by Huawei Ltd.\ (\url{http://www.huawei.com}), and Peter Buhr is partially funded by the Natural Sciences and Engineering Research Council of Canada.
2592
2593{%
2594\fontsize{9bp}{12bp}\selectfont%
2595\bibliography{pl,local}
2596}%
2597
2598\end{document}
2599
2600% Local Variables: %
2601% tab-width: 4 %
2602% fill-column: 120 %
2603% compile-command: "make" %
2604% End: %
Note: See TracBrowser for help on using the repository browser.