source: doc/papers/concurrency/Paper.tex @ bd47f35

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum
Last change on this file since bd47f35 was bd47f35, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Ran longer java experiences for concurrency paper

  • Property mode set to 100644
File size: 180.4 KB
Line 
1\documentclass[AMA,STIX1COL]{WileyNJD-v2}
2
3\articletype{RESEARCH ARTICLE}%
4
5% Referees
6% Doug Lea, dl@cs.oswego.edu, SUNY Oswego
7% Herb Sutter, hsutter@microsoft.com, Microsoft Corp
8% Gor Nishanov, gorn@microsoft.com, Microsoft Corp
9% James Noble, kjx@ecs.vuw.ac.nz, Victoria University of Wellington, School of Engineering and Computer Science
10
11\received{XXXXX}
12\revised{XXXXX}
13\accepted{XXXXX}
14
15\raggedbottom
16
17%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
18
19% Latex packages used in the document.
20
21\usepackage{epic,eepic}
22\usepackage{xspace}
23\usepackage{enumitem}
24\usepackage{comment}
25\usepackage{upquote}                                            % switch curled `'" to straight
26\usepackage{listings}                                           % format program code
27\usepackage[labelformat=simple,aboveskip=0pt,farskip=0pt]{subfig}
28\renewcommand{\thesubfigure}{(\Alph{subfigure})}
29\captionsetup{justification=raggedright,singlelinecheck=false}
30\usepackage{dcolumn}                                            % align decimal points in tables
31\usepackage{capt-of}
32\setlength{\multicolsep}{6.0pt plus 2.0pt minus 1.5pt}
33
34\hypersetup{breaklinks=true}
35\definecolor{OliveGreen}{cmyk}{0.64 0 0.95 0.40}
36\definecolor{Mahogany}{cmyk}{0 0.85 0.87 0.35}
37\definecolor{Plum}{cmyk}{0.50 1 0 0}
38
39\usepackage[pagewise]{lineno}
40\renewcommand{\linenumberfont}{\scriptsize\sffamily}
41
42\renewcommand{\topfraction}{0.8}                        % float must be greater than X of the page before it is forced onto its own page
43\renewcommand{\bottomfraction}{0.8}                     % float must be greater than X of the page before it is forced onto its own page
44\renewcommand{\floatpagefraction}{0.8}          % float must be greater than X of the page before it is forced onto its own page
45\renewcommand{\textfraction}{0.0}                       % the entire page maybe devoted to floats with no text on the page at all
46
47\lefthyphenmin=3                                                        % hyphen only after 4 characters
48\righthyphenmin=3
49
50%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
51
52% Names used in the document.
53
54\newcommand{\CFAIcon}{\textsf{C}\raisebox{\depth}{\rotatebox{180}{\textsf{A}}}\xspace} % Cforall symbolic name
55\newcommand{\CFA}{\protect\CFAIcon}             % safe for section/caption
56\newcommand{\CFL}{\textrm{Cforall}\xspace}      % Cforall symbolic name
57\newcommand{\Celeven}{\textrm{C11}\xspace}      % C11 symbolic name
58\newcommand{\CC}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}\xspace} % C++ symbolic name
59\newcommand{\CCeleven}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}11\xspace} % C++11 symbolic name
60\newcommand{\CCfourteen}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}14\xspace} % C++14 symbolic name
61\newcommand{\CCseventeen}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}17\xspace} % C++17 symbolic name
62\newcommand{\CCtwenty}{\textrm{C}\kern-.1em\hbox{+\kern-.25em+}20\xspace} % C++20 symbolic name
63\newcommand{\Csharp}{C\raisebox{-0.7ex}{\large$^\sharp$}\xspace} % C# symbolic name
64
65%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
66
67\newcommand{\Textbf}[2][red]{{\color{#1}{\textbf{#2}}}}
68\newcommand{\Emph}[2][red]{{\color{#1}\textbf{\emph{#2}}}}
69\newcommand{\uC}{$\mu$\CC}
70\newcommand{\TODO}[1]{{\Textbf{#1}}}
71
72%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
73
74% Default underscore is too low and wide. Cannot use lstlisting "literate" as replacing underscore
75% removes it as a variable-name character so keywords in variables are highlighted. MUST APPEAR
76% AFTER HYPERREF.
77%\DeclareTextCommandDefault{\textunderscore}{\leavevmode\makebox[1.2ex][c]{\rule{1ex}{0.1ex}}}
78\renewcommand{\textunderscore}{\leavevmode\makebox[1.2ex][c]{\rule{1ex}{0.075ex}}}
79
80\renewcommand*{\thefootnote}{\Alph{footnote}} % hack because fnsymbol does not work
81%\renewcommand*{\thefootnote}{\fnsymbol{footnote}}
82
83\makeatletter
84% parindent is relative, i.e., toggled on/off in environments like itemize, so store the value for
85% use rather than use \parident directly.
86\newlength{\parindentlnth}
87\setlength{\parindentlnth}{\parindent}
88
89\newcommand{\LstBasicStyle}[1]{{\lst@basicstyle{\lst@basicstyle{#1}}}}
90\newcommand{\LstKeywordStyle}[1]{{\lst@basicstyle{\lst@keywordstyle{#1}}}}
91\newcommand{\LstCommentStyle}[1]{{\lst@basicstyle{\lst@commentstyle{#1}}}}
92
93\newlength{\gcolumnposn}                                        % temporary hack because lstlisting does not handle tabs correctly
94\newlength{\columnposn}
95\setlength{\gcolumnposn}{3.5in}
96\setlength{\columnposn}{\gcolumnposn}
97
98\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}}}}
99\newcommand{\CRT}{\global\columnposn=\gcolumnposn}
100
101% Denote newterms in particular font and index them without particular font and in lowercase, \eg \newterm{abc}.
102% The option parameter provides an index term different from the new term, \eg \newterm[\texttt{abc}]{abc}
103% The star version does not lowercase the index information, e.g., \newterm*{IBM}.
104\newcommand{\newtermFontInline}{\emph}
105\newcommand{\newterm}{\@ifstar\@snewterm\@newterm}
106\newcommand{\@newterm}[2][\@empty]{\lowercase{\def\temp{#2}}{\newtermFontInline{#2}}\ifx#1\@empty\index{\temp}\else\index{#1@{\protect#2}}\fi}
107\newcommand{\@snewterm}[2][\@empty]{{\newtermFontInline{#2}}\ifx#1\@empty\index{#2}\else\index{#1@{\protect#2}}\fi}
108
109% Latin abbreviation
110\newcommand{\abbrevFont}{\textit}                       % set empty for no italics
111\@ifundefined{eg}{
112\newcommand{\EG}{\abbrevFont{e}\abbrevFont{g}}
113\newcommand*{\eg}{%
114        \@ifnextchar{,}{\EG}%
115                {\@ifnextchar{:}{\EG}%
116                        {\EG,\xspace}}%
117}}{}%
118\@ifundefined{ie}{
119\newcommand{\IE}{\abbrevFont{i}\abbrevFont{e}}
120\newcommand*{\ie}{%
121        \@ifnextchar{,}{\IE}%
122                {\@ifnextchar{:}{\IE}%
123                        {\IE,\xspace}}%
124}}{}%
125\@ifundefined{etc}{
126\newcommand{\ETC}{\abbrevFont{etc}}
127\newcommand*{\etc}{%
128        \@ifnextchar{.}{\ETC}%
129                {\ETC.\xspace}%
130}}{}%
131\@ifundefined{etal}{
132\newcommand{\ETAL}{\abbrevFont{et}~\abbrevFont{al}}
133\newcommand*{\etal}{%
134        \@ifnextchar{.}{\ETAL}%
135                {\ETAL.\xspace}%
136}}{}%
137\@ifundefined{viz}{
138\newcommand{\VIZ}{\abbrevFont{viz}}
139\newcommand*{\viz}{%
140        \@ifnextchar{.}{\VIZ}%
141                {\VIZ.\xspace}%
142}}{}%
143\makeatother
144
145\newenvironment{cquote}
146               {\list{}{\lstset{resetmargins=true,aboveskip=0pt,belowskip=0pt}\topsep=3pt\parsep=0pt\leftmargin=\parindentlnth\rightmargin\leftmargin}%
147                \item\relax}
148               {\endlist}
149
150%\newenvironment{cquote}{%
151%\list{}{\lstset{resetmargins=true,aboveskip=0pt,belowskip=0pt}\topsep=3pt\parsep=0pt\leftmargin=\parindentlnth\rightmargin\leftmargin}%
152%\item\relax%
153%}{%
154%\endlist%
155%}% cquote
156
157% CFA programming language, based on ANSI C (with some gcc additions)
158\lstdefinelanguage{CFA}[ANSI]{C}{
159        morekeywords={
160                _Alignas, _Alignof, __alignof, __alignof__, asm, __asm, __asm__, __attribute, __attribute__,
161                auto, _Bool, catch, catchResume, choose, _Complex, __complex, __complex__, __const, __const__,
162                coroutine, disable, dtype, enable, exception, __extension__, fallthrough, fallthru, finally,
163                __float80, float80, __float128, float128, forall, ftype, generator, _Generic, _Imaginary, __imag, __imag__,
164                inline, __inline, __inline__, __int128, int128, __label__, monitor, mutex, _Noreturn, one_t, or,
165                otype, restrict, resume, __restrict, __restrict__, __signed, __signed__, _Static_assert, suspend, thread,
166                _Thread_local, throw, throwResume, timeout, trait, try, ttype, typeof, __typeof, __typeof__,
167                virtual, __volatile, __volatile__, waitfor, when, with, zero_t},
168        moredirectives={defined,include_next},
169        % replace/adjust listing characters that look bad in sanserif
170        literate={-}{\makebox[1ex][c]{\raisebox{0.5ex}{\rule{0.8ex}{0.1ex}}}}1 {^}{\raisebox{0.6ex}{$\scriptstyle\land\,$}}1
171                {~}{\raisebox{0.3ex}{$\scriptstyle\sim\,$}}1 % {`}{\ttfamily\upshape\hspace*{-0.1ex}`}1
172                {<}{\textrm{\textless}}1 {>}{\textrm{\textgreater}}1
173                {<-}{$\leftarrow$}2 {=>}{$\Rightarrow$}2 {->}{\makebox[1ex][c]{\raisebox{0.5ex}{\rule{0.8ex}{0.075ex}}}\kern-0.2ex{\textrm{\textgreater}}}2,
174}
175
176\lstset{
177language=CFA,
178columns=fullflexible,
179basicstyle=\linespread{0.9}\sf,                                                 % reduce line spacing and use sanserif font
180stringstyle=\tt,                                                                                % use typewriter font
181tabsize=5,                                                                                              % N space tabbing
182xleftmargin=\parindentlnth,                                                             % indent code to paragraph indentation
183%mathescape=true,                                                                               % LaTeX math escape in CFA code $...$
184escapechar=\$,                                                                                  % LaTeX escape in CFA code
185keepspaces=true,                                                                                %
186showstringspaces=false,                                                                 % do not show spaces with cup
187showlines=true,                                                                                 % show blank lines at end of code
188aboveskip=4pt,                                                                                  % spacing above/below code block
189belowskip=3pt,
190moredelim=**[is][\color{red}]{`}{`},
191}% lstset
192
193% uC++ programming language, based on ANSI C++
194\lstdefinelanguage{uC++}[ANSI]{C++}{
195        morekeywords={
196                _Accept, _AcceptReturn, _AcceptWait, _Actor, _At, _CatchResume, _Cormonitor, _Coroutine, _Disable,
197                _Else, _Enable, _Event, _Finally, _Monitor, _Mutex, _Nomutex, _PeriodicTask, _RealTimeTask,
198                _Resume, _Select, _SporadicTask, _Task, _Timeout, _When, _With, _Throw},
199}
200
201% Go programming language: https://github.com/julienc91/listings-golang/blob/master/listings-golang.sty
202\lstdefinelanguage{Golang}{
203        morekeywords=[1]{package,import,func,type,struct,return,defer,panic,recover,select,var,const,iota,},
204        morekeywords=[2]{string,uint,uint8,uint16,uint32,uint64,int,int8,int16,int32,int64,
205                bool,float32,float64,complex64,complex128,byte,rune,uintptr, error,interface},
206        morekeywords=[3]{map,slice,make,new,nil,len,cap,copy,close,true,false,delete,append,real,imag,complex,chan,},
207        morekeywords=[4]{for,break,continue,range,goto,switch,case,fallthrough,if,else,default,},
208        morekeywords=[5]{Println,Printf,Error,},
209        sensitive=true,
210        morecomment=[l]{//},
211        morecomment=[s]{/*}{*/},
212        morestring=[b]',
213        morestring=[b]",
214        morestring=[s]{`}{`},
215        % replace/adjust listing characters that look bad in sanserif
216        literate={-}{\makebox[1ex][c]{\raisebox{0.4ex}{\rule{0.8ex}{0.1ex}}}}1 {^}{\raisebox{0.6ex}{$\scriptstyle\land\,$}}1
217                {~}{\raisebox{0.3ex}{$\scriptstyle\sim\,$}}1 % {`}{\ttfamily\upshape\hspace*{-0.1ex}`}1
218                {<}{\textrm{\textless}}1 {>}{\textrm{\textgreater}}1
219                {<-}{\makebox[2ex][c]{\textrm{\textless}\raisebox{0.5ex}{\rule{0.8ex}{0.075ex}}}}2,
220}
221
222\lstnewenvironment{cfa}[1][]
223{\lstset{#1}}
224{}
225\lstnewenvironment{C++}[1][]                            % use C++ style
226{\lstset{language=C++,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
227{}
228\lstnewenvironment{uC++}[1][]
229{\lstset{language=uC++,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
230{}
231\lstnewenvironment{Go}[1][]
232{\lstset{language=Golang,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
233{}
234\lstnewenvironment{python}[1][]
235{\lstset{language=python,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
236{}
237\lstnewenvironment{java}[1][]
238{\lstset{language=java,moredelim=**[is][\protect\color{red}]{`}{`},#1}\lstset{#1}}
239{}
240
241% inline code @...@
242\lstMakeShortInline@%
243
244\newcommand{\commenttd}[1]{{\color{red}{Thierry : #1}}}
245
246\let\OLDthebibliography\thebibliography
247\renewcommand\thebibliography[1]{
248  \OLDthebibliography{#1}
249  \setlength{\parskip}{0pt}
250  \setlength{\itemsep}{4pt plus 0.3ex}
251}
252
253\newsavebox{\myboxA}
254\newsavebox{\myboxB}
255\newsavebox{\myboxC}
256\newsavebox{\myboxD}
257
258\title{\texorpdfstring{Advanced Control-flow and Concurrency in \protect\CFA}{Advanced Control-flow in Cforall}}
259
260\author[1]{Thierry Delisle}
261\author[1]{Peter A. Buhr*}
262\authormark{DELISLE \textsc{et al.}}
263
264\address[1]{\orgdiv{Cheriton School of Computer Science}, \orgname{University of Waterloo}, \orgaddress{\state{Waterloo, ON}, \country{Canada}}}
265
266\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}}
267
268% \fundingInfo{Natural Sciences and Engineering Research Council of Canada}
269
270\abstract[Summary]{
271\CFA is a polymorphic, non-object-oriented, concurrent, backwards compatible extension of the C programming language.
272This paper discusses the design philosophy and implementation of its advanced control-flow and concurrent/parallel features, along with the supporting runtime written in \CFA.
273These features are created from scratch as ISO C has only low-level and/or unimplemented concurrency, so C programmers continue to rely on library approaches like pthreads.
274\CFA introduces modern language-level control-flow mechanisms, like generators, coroutines, user-level threading, and monitors for mutual exclusion and synchronization.
275% Library extension for executors, futures, and actors are built on these basic mechanisms.
276The runtime provides significant programmer simplification and safety by eliminating spurious wakeup and monitor barging.
277The runtime also ensures multiple monitors can be safely acquired in a deadlock-free way, and this feature is fully integrated with all monitor synchronization mechanisms.
278All control-flow features integrate with the \CFA polymorphic type-system and exception handling, while respecting the expectations and style of C programmers.
279Experimental results show comparable performance of the new features with similar mechanisms in other concurrent programming languages.
280}%
281
282\keywords{generator, coroutine, concurrency, parallelism, thread, monitor, runtime, C, \CFA (Cforall)}
283
284
285\begin{document}
286\linenumbers                            % comment out to turn off line numbering
287
288\maketitle
289
290
291\section{Introduction}
292
293\CFA~\cite{Moss18,Cforall} is a modern, polymorphic, non-object-oriented\footnote{
294\CFA has object-oriented features, such as constructors, destructors, and simple trait/interface inheritance.
295% Go interfaces, Rust traits, Swift Protocols, Haskell Type Classes and Java Interfaces.
296% "Trait inheritance" works for me. "Interface inheritance" might also be a good choice, and distinguish clearly from implementation inheritance.
297% You'll want to be a little bit careful with terms like "structural" and "nominal" inheritance as well. CFA has structural inheritance (I think Go as well) -- it's inferred based on the structure of the code.
298% Java, Rust, and Haskell (not sure about Swift) have nominal inheritance, where there needs to be a specific statement that "this type inherits from this type".
299However, functions \emph{cannot} be nested in structures and there is no mechanism to designate a function parameter as a receiver, \lstinline@this@, parameter.},
300backwards-compatible extension of the C programming language.
301In many ways, \CFA is to C as Scala~\cite{Scala} is to Java, providing a vehicle for new typing and control-flow capabilities on top of a highly popular programming language\footnote{
302The TIOBE index~\cite{TIOBE} for May 2020 ranks the top five \emph{popular} programming languages as C 17\%, Java 16\%, Python 9\%, \CC 6\%, and \Csharp 4\% = 52\%, and over the past 30 years, C has always ranked either first or second in popularity.}
303allowing immediate dissemination.
304This paper discusses the design philosophy and implementation of \CFA's advanced control-flow and concurrent/parallel features, along with the supporting runtime written in \CFA.
305
306% The call/return extensions retain state between callee and caller versus losing the callee's state on return;
307% the concurrency extensions allow high-level management of threads.
308
309The \CFA control-flow framework extends ISO \Celeven~\cite{C11} with new call/return and concurrent/parallel control-flow.
310Call/return control-flow with argument and parameter passing appeared in the first programming languages.
311Over the past 50 years, call/return has been augmented with features like static and dynamic call, exceptions (multi-level return) and generators/coroutines (see Section~\ref{s:StatefulFunction}).
312While \CFA has mechanisms for dynamic call (algebraic effects~\cite{Zhang19}) and exceptions\footnote{
313\CFA exception handling will be presented in a separate paper.
314The key feature that dovetails with this paper is nonlocal 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++}}, this work only discusses retaining state between calls via generators and coroutines.
315\newterm{Coroutining} was introduced by Conway~\cite{Conway63} (1963), discussed by Knuth~\cite[\S~1.4.2]{Knuth73V1}, implemented in Simula67~\cite{Simula67}, formalized by Marlin~\cite{Marlin80}, and is now popular and appears in old and new programming languages: CLU~\cite{CLU}, \Csharp~\cite{Csharp}, Ruby~\cite{Ruby}, Python~\cite{Python}, JavaScript~\cite{JavaScript}, Lua~\cite{Lua}, \CCtwenty~\cite{C++20Coroutine19}.
316Coroutining is sequential execution requiring direct handoff among coroutines, \ie only the programmer is controlling execution order.
317If coroutines transfer to an internal event-engine for scheduling the next coroutines (as in async-await), the program transitions into the realm of concurrency~\cite[\S~3]{Buhr05a}.
318Coroutines are only a stepping stone towards concurrency where the commonality is that coroutines and threads retain state between calls.
319
320\Celeven and \CCeleven define concurrency~\cite[\S~7.26]{C11}, but it is largely wrappers for a subset of the pthreads library~\cite{Pthreads}.\footnote{Pthreads concurrency is based on simple thread fork and join in a function and mutex or condition locks, which is low-level and error-prone}
321Interestingly, almost a decade after the \Celeven standard, the most recent versions of gcc, clang, and msvc do not support the \Celeven include @threads.h@, indicating no interest in the C11 concurrency approach (possibly because of the recent effort to add concurrency to \CC).
322While the \Celeven standard does not state a threading model, the historical association with pthreads suggests implementations would adopt kernel-level threading (1:1)~\cite{ThreadModel}, as for \CC.
323In contrast, there has been a renewed interest during the past decade in user-level (M:N, green) threading in old and new programming languages.
324As multi-core hardware became available in the 1980/90s, both user and kernel threading were examined.
325Kernel 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}.
326Libraries like pthreads were developed for C, and the Solaris operating-system switched from user (JDK 1.1~\cite{JDK1.1}) to kernel threads.
327As a result, many languages adopt the 1:1 kernel-threading model, like Java (Scala), Objective-C~\cite{obj-c-book}, \CCeleven~\cite{C11}, C\#~\cite{Csharp} and Rust~\cite{Rust}, with a variety of presentation mechanisms.
328From 2000 onwards, several language implementations have championed the M:N user-threading model, like Go~\cite{Go}, Erlang~\cite{Erlang}, Haskell~\cite{Haskell}, D~\cite{D}, and \uC~\cite{uC++,uC++book}, including putting green threads back into Java~\cite{Quasar}, and many user-threading libraries have appeared~\cite{Qthreads,MPC,Marcel}.
329The main argument for user-level threading is that it is lighter weight than kernel threading because locking and context switching do not cross the kernel boundary, so there is less restriction on programming styles that encourages large numbers of threads performing medium-sized work to facilitate load balancing by the runtime~\cite{Verch12}.
330As well, user-threading facilitates a simpler concurrency approach using thread objects that leverage sequential patterns versus events with call-backs~\cite{Adya02,vonBehren03}.
331Finally, performant user-threading implementations, both in time and space, meet or exceed direct kernel-threading implementations, while achieving the programming advantages of high concurrency levels and safety.
332
333A further effort over the past two decades is the development of language memory models to deal with the conflict between language features and compiler/hardware optimizations, \eg some language features are unsafe in the presence of aggressive sequential optimizations~\cite{Buhr95a,Boehm05}.
334The consequence is that a language must provide sufficient tools to program around safety issues, as inline and library code is compiled as sequential without any explicit concurrent directive.
335One solution is low-level qualifiers and functions, \eg @volatile@ and atomics, allowing \emph{programmers} to explicitly write safe, race-free~\cite{Boehm12} programs.
336A safer solution is high-level language constructs so the \emph{compiler} knows the concurrency boundaries, \ie where mutual exclusion and synchronization are acquired and released, and provide implicit safety at and across these boundaries.
337While the optimization problem is best known with respect to concurrency, it applies to other complex control-flow, like exceptions and coroutines.
338As well, language solutions allow matching the language paradigm with the approach, \eg matching the functional paradigm with data-flow programming or the imperative paradigm with thread programming.
339
340Finally, it is important for a language to provide safety over performance \emph{as the default}, allowing careful reduction of safety for performance when necessary.
341Two concurrency violations of this philosophy are \emph{spurious} or \emph{random wakeup}~\cite[\S~9]{Buhr05a}, and \emph{barging}\footnote{
342Barging is competitive succession instead of direct handoff, \ie after a lock is released both arriving and preexisting waiter threads compete to acquire the lock.
343Hence, an arriving thread can temporally \emph{barge} ahead of threads already waiting for an event, which can repeat indefinitely leading to starvation of waiter threads.
344} or signals-as-hints~\cite[\S~8]{Buhr05a}, where one is a consequence of the other, \ie once there is spurious wakeup, barging follows.
345(Author experience teaching concurrency is that students are confused by these semantics.)
346However, spurious wakeup is \emph{not} a foundational concurrency property~\cite[\S~9]{Buhr05a};
347it is a performance design choice.
348We argue removing spurious wakeup and signals-as-hints make concurrent programming simpler and safer as there is less local non-determinism to manage.
349If barging acquisition is allowed, its specialized performance advantage should be available as an option not the default.
350
351\CFA embraces language extensions for advanced control-flow, user-level threading, and safety as the default.
352We present comparative examples to support our argument that the \CFA control-flow extensions are as expressive and safe as those in other concurrent imperative programming languages, and perform experiments to show the \CFA runtime is competitive with other similar mechanisms.
353The main contributions of this work are:
354\begin{itemize}[topsep=3pt,itemsep=0pt]
355\item
356a set of fundamental execution properties that dictate which language-level control-flow features need to be supported,
357
358\item
359integration of these language-level control-flow features, while respecting the style and expectations of C programmers,
360
361\item
362monitor synchronization without barging, and the ability to safely acquiring multiple monitors in a deadlock-free way, while seamlessly integrating these capabilities with all monitor synchronization mechanisms,
363
364\item
365providing statically type-safe interfaces that integrate with the \CFA polymorphic type-system and other language features,
366
367% \item
368% library extensions for executors, futures, and actors built on the basic mechanisms.
369
370\item
371a runtime system without spurious wake-up and no performance loss,
372
373\item
374a dynamic partitioning mechanism to segregate groups of executing user and kernel threads performing specialized work, \eg web-server or compute engine, or requiring different scheduling, \eg NUMA or real-time.
375
376% \item
377% a non-blocking I/O library
378
379\item
380experimental results showing comparable performance of the \CFA features with similar mechanisms in other languages.
381\end{itemize}
382
383Section~\ref{s:FundamentalExecutionProperties} presents the compositional hierarchy of execution properties directing the design of control-flow features in \CFA.
384Section~\ref{s:StatefulFunction} begins advanced control by introducing sequential functions that retain data and execution state between calls producing constructs @generator@ and @coroutine@.
385Section~\ref{s:Concurrency} begins concurrency, or how to create (fork) and destroy (join) a thread producing the @thread@ construct.
386Section~\ref{s:MutualExclusionSynchronization} discusses the two mechanisms to restricted nondeterminism when controlling shared access to resources, called mutual exclusion, and timing relationships among threads, called synchronization.
387Section~\ref{s:Monitor} shows how both mutual exclusion and synchronization are safely embedded in the @monitor@ and @thread@ constructs.
388Section~\ref{s:CFARuntimeStructure} describes the large-scale mechanism to structure threads and virtual processors (kernel threads).
389Section~\ref{s:Performance} uses microbenchmarks to compare \CFA threading with pthreads, Java 11.0.6, Go 1.12.6, Rust 1.37.0, Python 3.7.6, Node.js v12.18.0, and \uC 7.0.0.
390
391
392\section{Fundamental Execution Properties}
393\label{s:FundamentalExecutionProperties}
394
395The features in a programming language should be composed of a set of fundamental properties rather than an ad hoc collection chosen by the designers.
396To this end, the control-flow features created for \CFA are based on the fundamental properties of any language with function-stack control-flow (see also \uC~\cite[pp.~140-142]{uC++}).
397The fundamental properties are execution state, thread, and mutual-exclusion/synchronization.
398These independent properties can be used to compose different language features, forming a compositional hierarchy, where the combination of all three is the most advanced feature, called a thread.
399While it is possible for a language to only provide threads for composing programs~\cite{Hermes90}, this unnecessarily complicates and makes inefficient solutions to certain classes of problems.
400As is shown, each of the non-rejected composed language features solves a particular set of problems, and hence, has a defensible position in a programming language.
401If a compositional feature is missing, a programmer has too few fundamental properties resulting in a complex and/or inefficient solution.
402
403In detail, the fundamental properties are:
404\begin{description}[leftmargin=\parindent,topsep=3pt,parsep=0pt]
405\item[\newterm{execution state}:]
406is the state information needed by a control-flow feature to initialize and manage both compute data and execution location(s), and de-initialize.
407For example, calling a function initializes a stack frame including contained objects with constructors, manages local data in blocks and return locations during calls, and de-initializes the frame by running any object destructors and management operations.
408State is retained in fixed-sized aggregate structures (objects) and dynamic-sized stack(s), often allocated in the heap(s) managed by the runtime system.
409The lifetime of state varies with the control-flow feature, where longer life-time and dynamic size provide greater power but also increase usage complexity and cost.
410Control-flow transfers among execution states in multiple ways, such as function call, context switch, asynchronous await, etc.
411Because the programming language determines what constitutes an execution state, implicitly manages this state, and defines movement mechanisms among states, execution state is an elementary property of the semantics of a programming language.
412% An execution-state is related to the notion of a process continuation \cite{Hieb90}.
413
414\item[\newterm{threading}:]
415is execution of code that occurs independently of other execution, where an individual thread's execution is sequential.
416Multiple threads provide \emph{concurrent execution};
417concurrent execution becomes parallel when run on multiple processing units, \eg hyper-threading, cores, or sockets.
418A programmer needs mechanisms to create, block and unblock, and join with a thread, even if these basic mechanisms are supplied indirectly through high-level features.
419
420\item[\newterm{mutual-exclusion / synchronization (MES)}:]
421is the concurrency mechanism to perform an action without interruption and establish timing relationships among multiple threads.
422We contented these two properties are independent, \ie mutual exclusion cannot provide synchronization and vice versa without introducing additional threads~\cite[\S~4]{Buhr05a}.
423Limiting MES functionality results in contrived solutions and inefficiency on multi-core von Neumann computers where shared memory is a foundational aspect of its design.
424\end{description}
425These properties are fundamental as they cannot be built from existing language features, \eg a basic programming language like C99~\cite{C99} cannot create new control-flow features, concurrency, or provide MES without (atomic) hardware mechanisms.
426
427
428\subsection{Structuring Execution Properties}
429
430Programming languages seldom present the fundamental execution properties directly to programmers.
431Instead, the properties are packaged into higher-level constructs that encapsulate details and provide safety to these low-level mechanisms.
432Interestingly, language designers often pick and choose among these execution properties proving a varying subset of constructs.
433
434Table~\ref{t:ExecutionPropertyComposition} shows all combinations of the three fundamental execution properties available to language designers.
435(When doing combination case-analysis, not all combinations are meaningful.)
436The combinations of state, thread, and MES compose a hierarchy of control-flow features all of which have appeared in prior programming languages, where each of these languages have found the feature useful.
437To understand the table, it is important to review the basic von Neumann execution requirement of at least one thread and execution state providing some form of call stack.
438For table entries missing these minimal components, the property is borrowed from the invoker (caller).
439Each entry in the table, numbered \textbf{1}--\textbf{12}, is discussed with respect to how the execution properties combine to generate a high-level language construct.
440
441\begin{table}
442\caption{Execution property composition}
443\centering
444\label{t:ExecutionPropertyComposition}
445\renewcommand{\arraystretch}{1.25}
446%\setlength{\tabcolsep}{5pt}
447\vspace*{-5pt}
448\begin{tabular}{c|c||l|l}
449\multicolumn{2}{c||}{execution properties} & \multicolumn{2}{c}{mutual exclusion / synchronization} \\
450\hline
451stateful                        & thread        & \multicolumn{1}{c|}{No} & \multicolumn{1}{c}{Yes} \\
452\hline
453\hline
454No                                      & No            & \textbf{1}\ \ \ @struct@                              & \textbf{2}\ \ \ @mutex@ @struct@              \\
455\hline
456Yes (stackless)         & No            & \textbf{3}\ \ \ @generator@                   & \textbf{4}\ \ \ @mutex@ @generator@   \\
457\hline
458Yes (stackful)          & No            & \textbf{5}\ \ \ @coroutine@                   & \textbf{6}\ \ \ @mutex@ @coroutine@   \\
459\hline
460No                                      & Yes           & \textbf{7}\ \ \ {\color{red}rejected} & \textbf{8}\ \ \ {\color{red}rejected} \\
461\hline
462Yes (stackless)         & Yes           & \textbf{9}\ \ \ {\color{red}rejected} & \textbf{10}\ \ \ {\color{red}rejected} \\
463\hline
464Yes (stackful)          & Yes           & \textbf{11}\ \ \ @thread@                             & \textbf{12}\ \ @mutex@ @thread@               \\
465\end{tabular}
466\vspace*{-8pt}
467\end{table}
468
469Case 1 is a structure where access functions borrow local state (stack frame/activation) and thread from the invoker and retain this state across \emph{callees}, \ie function local-variables are retained on the borrowed stack during calls.
470Structures are a foundational mechanism for data organization, and access functions provide interface abstraction and code sharing in all programming languages.
471Case 2 is case 1 with thread safety to a structure's state where access functions provide serialization (mutual exclusion) and scheduling among calling threads (synchronization).
472A @mutex@ structure, often called a \newterm{monitor}, provides a high-level interface for race-free access of shared data in concurrent programming-languages.
473Case 3 is case 1 where the structure can implicitly retain execution state and access functions use this execution state to resume/suspend across \emph{callers}, but resume/suspend does not retain a function's local state.
474A stackless structure, often called a \newterm{generator} or \emph{iterator}, is \newterm{stackless} because it still borrows the caller's stack and thread, but the stack is used only to preserve state across its callees not callers.
475Generators provide the first step toward directly solving problems like finite-state machines that retain data and execution state between calls, whereas normal functions restart on each call.
476Case 4 is cases 2 and 3 with thread safety during execution of the generator's access functions.
477A @mutex@ generator extends generators into the concurrent domain.
478Cases 5 and 6 are like cases 3 and 4 where the structure is extended with an implicit separate stack, so only the thread is borrowed by access functions.
479A stackful generator, often called a \newterm{coroutine}, is \newterm{stackful} because resume/suspend now context switch to/from the caller's and coroutine's stack.
480A coroutine extends the state retained between calls beyond the generator's structure to arbitrary call depth in the access functions.
481Cases 7, 8, 9 and 10 are rejected because a new thread must have its own stack, where the thread begins and stack frames are stored for calls, \ie it is unrealistic for a thread to borrow a stack.
482For cases 9 and 10, the stackless frame is not growable, precluding accepting nested calls, making calls, blocking as it requires calls, or preemption as it requires pushing an interrupt frame, all of which compound to require an unknown amount of execution state.
483Hence, if this kind of uninterruptable thread exists, it must execute to completion, \ie computation only, which severely restricts runtime management.
484Cases 11 and 12 are a stackful thread with and without safe access to shared state.
485A thread is the language mechanism to start another thread of control in a program with growable execution state for call/return execution.
486In general, language constructs with more execution properties increase the cost of creation and execution along with complexity of usage.
487
488Given the execution-properties taxonomy, programmers now ask three basic questions: is state necessary across callers and how much, is a separate thread necessary, is thread safety necessary.
489Table~\ref{t:ExecutionPropertyComposition} then suggests the optimal language feature needed for implementing a programming problem.
490The following sections describe how \CFA fills in \emph{all} the non-rejected table entries with language features, while other programming languages may only provide a subset of the table.
491
492
493\subsection{Design Requirements}
494
495The following design requirements largely stem from building \CFA on top of C.
496\begin{itemize}[topsep=3pt,parsep=0pt]
497\item
498All communication must be statically type checkable for early detection of errors and efficient code generation.
499This requirement is consistent with the fact that C is a statically-typed programming-language.
500
501\item
502Direct interaction among language features must be possible allowing any feature to be selected without restricting comm\-unication.
503For example, many concurrent languages do not provide direct communication calls among threads, \ie threads only communicate indirectly through monitors, channels, messages, and/or futures.
504Indirect communication increases the number of objects, consuming more resources, and requires additional synchronization and possibly data transfer.
505
506\item
507All communication is performed using function calls, \ie data is transmitted from argument to parameter and results are returned from function calls.
508Alternative forms of communication, such as call-backs, message passing, channels, or communication ports, step outside of C's normal form of communication.
509
510\item
511All stateful features must follow the same declaration scopes and lifetimes as other language data.
512For C that means at program startup, during block and function activation, and on demand using dynamic allocation.
513
514\item
515MES must be available implicitly in language constructs, \eg Java built-in monitors, as well as explicitly for specialized requirements, \eg @java.util.concurrent@, because requiring programmers to build MES using low-level locks often leads to incorrect programs.
516Furthermore, reducing synchronization scope by encapsulating it within language constructs further reduces errors in concurrent programs.
517
518\item
519Both synchronous and asynchronous communication are needed.
520However, we believe the best way to provide asynchrony, such as call-buffering/chaining and/or returning futures~\cite{multilisp}, is building it from expressive synchronous features.
521
522\item
523Synchronization must be able to control the service order of requests including prioritizing selection from different kinds of outstanding requests, and postponing a request for an unspecified time while continuing to accept new requests.
524Otherwise, certain concurrency problems are difficult, \eg web server, disk scheduling, and the amount of concurrency is inhibited~\cite{Gentleman81}.
525\end{itemize}
526We have satisfied these requirements in \CFA while maintaining backwards compatibility with the huge body of legacy C programs.
527% In contrast, other new programming languages must still access C programs (\eg operating-system service routines), but do so through fragile C interfaces.
528
529
530\subsection{Asynchronous Await / Call}
531
532Asynchronous await/call is a caller mechanism for structuring programs and/or increasing concurrency, where the caller (client) postpones an action into the future, which is subsequently executed by a callee (server).
533The caller detects the action's completion through a \newterm{future} or \newterm{promise}.
534The benefit is asynchronous caller execution with respect to the callee until future resolution.
535For single-threaded languages like JavaScript, an asynchronous call passes a callee action, which is queued in the event-engine, and continues execution with a promise.
536When the caller needs the promise to be fulfilled, it executes @await@.
537A promise-completion call-back can be part of the callee action or the caller is rescheduled;
538in either case, the call back is executed after the promise is fulfilled.
539While asynchronous calls generate new callee (server) events, we contend this mechanism is insufficient for advanced control-flow mechanisms like generators or coroutines, which are discussed next.
540Specifically, control between caller and callee occurs indirectly through the event-engine precluding direct handoff and cycling among events, and requires complex resolution of a control promise and data.
541Note, @async-await@ is just syntactic-sugar over the event engine so it does not solve these deficiencies.
542For multi-threaded languages like Java, the asynchronous call queues a callee action with an executor (server), which subsequently executes the work by a thread in the executor thread-pool.
543The problem is when concurrent work-units need to interact and/or block as this effects the executor by stopping threads.
544While it is possible to extend this approach to support the necessary mechanisms, \eg message passing in Actors, we show monitors and threads provide an equally competitive approach that does not deviate from normal call communication and can be used to build asynchronous call, as is done in Java.
545
546
547\section{Stateful Function}
548\label{s:StatefulFunction}
549
550A \emph{stateful function} has the ability to remember state between calls, where state can be either data or execution, \eg plugin, device driver, finite-state machine (FSM).
551A simple technique to retain data state between calls is @static@ declarations within a function, which is often implemented by hoisting the declarations to the global scope but hiding the names within the function using name mangling.
552However, each call starts the function at the top making it difficult to determine the last point of execution in an algorithm, and requiring multiple flag variables and testing to reestablish the continuation point.
553Hence, the next step of generalizing function state is implicitly remembering the return point between calls and reentering the function at this point rather than the top, called \emph{generators}\,/\,\emph{iterators} or \emph{stackless coroutines}.
554For example, a Fibonacci generator retains data and execution state allowing it to remember prior values needed to generate the next value and the location in the algorithm to compute that value.
555The next step of generalization is instantiating the function to allow multiple named instances, \eg multiple Fibonacci generators, where each instance has its own state, and hence, can generate an independent sequence of values.
556Note, a subset of generator state is a function \emph{closure}, \ie the technique of capturing lexical references when returning a nested function.
557A further generalization is adding a stack to a generator's state, called a \emph{coroutine}, so it can suspend outside of itself, \eg call helper functions to arbitrary depth before suspending back to its resumer without unwinding these calls.
558For example, a coroutine iterator for a binary tree can stop the traversal at the visit point (pre, infix, post traversal), return the node value to the caller, and then continue the recursive traversal from the current node on the next call.
559
560There are two styles of activating a stateful function, \emph{asymmetric} or \emph{symmetric}, identified by resume/suspend (no cycles) and resume/resume (cycles).
561These styles \emph{do not} cause incremental stack growth, \eg a million resume/suspend or resume/resume cycles do not remember each cycle just the last resumer for each cycle.
562Selecting between stackless/stackful semantics and asymmetric/symmetric style is a tradeoff between programming requirements, performance, and design, where stackless is faster and smaller using modified call/return between closures, stackful is more general but slower and larger using context switching between distinct stacks, and asymmetric is simpler control-flow than symmetric.
563Additionally, storage management for the closure/stack must be factored into design and performance, especially in unmanaged languages without garbage collection.
564Note, creation cost (closure/stack) is amortized across usage, so activation cost (resume/suspend) is usually the dominant factor.
565
566% The stateful function is an old idea~\cite{Conway63,Marlin80} that is new again~\cite{C++20Coroutine19}, where execution is temporarily suspended and later resumed, \eg plugin, device driver, finite-state machine.
567% Hence, a stateful function may not end when it returns to its caller, allowing it to be restarted with the data and execution location present at the point of suspension.
568% If the closure is fixed size, we call it a \emph{generator} (or \emph{stackless}), and its control flow is restricted, \eg suspending outside the generator is prohibited.
569% If the closure is variable size, we call it a \emph{coroutine} (or \emph{stackful}), and as the names implies, often implemented with a separate stack with no programming restrictions.
570% Hence, refactoring a stackless coroutine may require changing it to stackful.
571% A foundational property of all \emph{stateful functions} is that resume/suspend \emph{do not} cause incremental stack growth, \ie resume/suspend operations are remembered through the closure not the stack.
572% As well, activating a stateful function is \emph{asymmetric} or \emph{symmetric}, identified by resume/suspend (no cycles) and resume/resume (cycles).
573% A fixed closure activated by modified call/return is faster than a variable closure activated by context switching.
574% Additionally, any storage management for the closure (especially in unmanaged languages, \ie no garbage collection) must also be factored into design and performance.
575% Therefore, selecting between stackless and stackful semantics is a tradeoff between programming requirements and performance, where stackless is faster and stackful is more general.
576% nppNote, creation cost is amortized across usage, so activation cost is usually the dominant factor.
577
578For example, Python presents asymmetric generators as a function object, \uC presents symmetric coroutines as a \lstinline[language=C++]|class|-like object, and many languages present threading using function pointers, @pthreads@~\cite{Butenhof97}, \Csharp~\cite{Csharp}, Go~\cite{Go}, and Scala~\cite{Scala}.
579\begin{center}
580\begin{tabular}{@{}l|l|l@{}}
581\multicolumn{1}{@{}c|}{Python asymmetric generator} & \multicolumn{1}{c|}{\uC symmetric coroutine} & \multicolumn{1}{c@{}}{Pthreads thread} \\
582\hline
583\begin{python}
584`def Gen():` $\LstCommentStyle{\color{red}// function}$
585        ... yield val ...
586gen = Gen()
587for i in range( 10 ):
588        print( next( gen ) )
589\end{python}
590&
591\begin{uC++}
592`_Coroutine Cycle {` $\LstCommentStyle{\color{red}// class}$
593        Cycle * p;
594        void main() { p->cycle(); }
595        void cycle() { resume(); }  `};`
596Cycle c1, c2; c1.p=&c2; c2.p=&c1; c1.cycle();
597\end{uC++}
598&
599\begin{cfa}
600void * `rtn`( void * arg ) { ... }
601int i = 3, rc;
602pthread_t t; $\C{// thread id}$
603$\LstCommentStyle{\color{red}// function pointer}$
604rc=pthread_create(&t, `rtn`, (void *)i);
605\end{cfa}
606\end{tabular}
607\end{center}
608\CFA's preferred presentation model for generators/coroutines/threads is a hybrid of functions and classes, giving an object-oriented flavour.
609Essentially, the generator/coroutine/thread function is semantically coupled with a generator/coroutine/thread custom type via the type's name.
610The custom type solves several issues, while accessing the underlying mechanisms used by the custom types is still allowed for flexibility reasons.
611Each custom type is discussed in detail in the following sections.
612
613
614\subsection{Generator}
615
616Stackless generators (Table~\ref{t:ExecutionPropertyComposition} case 3) have the potential to be very small and fast, \ie as small and fast as function call/return for both creation and execution.
617The \CFA goal is to achieve this performance target, possibly at the cost of some semantic complexity.
618A series of different kinds of generators and their implementation demonstrate how this goal is accomplished.\footnote{
619The \CFA operator syntax uses \lstinline|?| to denote operands, which allows precise definitions for pre, post, and infix operators, \eg \lstinline|?++|, \lstinline|++?|, and \lstinline|?+?|, in addition \lstinline|?\{\}| denotes a constructor, as in \lstinline|foo `f` = `\{`...`\}`|, \lstinline|^?\{\}| denotes a destructor, and \lstinline|?()| is \CC function call \lstinline|operator()|.
620Operator \lstinline+|+ is overloaded for printing, like bit-shift \lstinline|<<| in \CC.
621The \CFA \lstinline|with| clause opens an aggregate scope making its fields directly accessible, like Pascal \lstinline|with|, but using parallel semantics;
622multiple aggregates may be opened.
623\CFA has rebindable references \lstinline|int i, & ip = i, j; `&ip = &j;`| and non-rebindable references \lstinline|int i, & `const` ip = i, j; `&ip = &j;` // disallowed|.
624}%
625
626\begin{figure}
627\centering
628\begin{lrbox}{\myboxA}
629\begin{cfa}[aboveskip=0pt,belowskip=0pt]
630typedef struct {
631        int fn1, fn;
632} Fib;
633#define FibCtor { 1, 0 }
634int fib( Fib * f ) {
635
636
637
638
639
640        int fn = f->fn; f->fn = f->fn1;
641                f->fn1 = f->fn + fn;
642        return fn;
643}
644int main() {
645        Fib f1 = FibCtor, f2 = FibCtor;
646        for ( int i = 0; i < 10; i += 1 )
647                printf( "%d %d\n",
648                           fib( &f1 ), fib( &f2 ) );
649}
650\end{cfa}
651\end{lrbox}
652
653\begin{lrbox}{\myboxB}
654\begin{cfa}[aboveskip=0pt,belowskip=0pt]
655`generator` Fib {
656        int fn1, fn;
657};
658
659void `main(Fib & fib)` with(fib) {
660
661
662        [fn1, fn] = [1, 0];
663        for () {
664                `suspend;`
665                [fn1, fn] = [fn, fn + fn1];
666
667        }
668}
669int main() {
670        Fib f1, f2;
671        for ( 10 )
672                sout | `resume( f1 )`.fn
673                         | `resume( f2 )`.fn;
674}
675\end{cfa}
676\end{lrbox}
677
678\begin{lrbox}{\myboxC}
679\begin{cfa}[aboveskip=0pt,belowskip=0pt]
680typedef struct {
681        int `restart`, fn1, fn;
682} Fib;
683#define FibCtor { `0`, 1, 0 }
684Fib * comain( Fib * f ) {
685        `static void * states[] = {&&s0, &&s1};`
686        `goto *states[f->restart];`
687  s0: f->`restart` = 1;
688        for ( ;; ) {
689                return f;
690          s1:; int fn = f->fn + f->fn1;
691                f->fn1 = f->fn; f->fn = fn;
692        }
693}
694int main() {
695        Fib f1 = FibCtor, f2 = FibCtor;
696        for ( int i = 0; i < 10; i += 1 )
697                printf("%d %d\n",comain(&f1)->fn,
698                                 comain(&f2)->fn);
699}
700\end{cfa}
701\end{lrbox}
702
703\subfloat[C]{\label{f:CFibonacci}\usebox\myboxA}
704\hspace{3pt}
705\vrule
706\hspace{3pt}
707\subfloat[\CFA]{\label{f:CFAFibonacciGen}\usebox\myboxB}
708\hspace{3pt}
709\vrule
710\hspace{3pt}
711\subfloat[C generated code for \CFA version]{\label{f:CFibonacciSim}\usebox\myboxC}
712\caption{Fibonacci output asymmetric generator}
713\label{f:FibonacciAsymmetricGenerator}
714
715\bigskip
716
717\begin{lrbox}{\myboxA}
718\begin{cfa}[aboveskip=0pt,belowskip=0pt]
719`generator Fmt` {
720        char ch;
721        int g, b;
722};
723void ?{}( Fmt & fmt ) { `resume(fmt);` } // constructor
724void ^?{}( Fmt & f ) with(f) { $\C[2.25in]{// destructor}$
725        if ( g != 0 || b != 0 ) sout | nl; }
726void `main( Fmt & f )` with(f) {
727        for () { $\C{// until destructor call}$
728                for ( ; g < 5; g += 1 ) { $\C{// groups}$
729                        for ( ; b < 4; b += 1 ) { $\C{// blocks}$
730                                do { `suspend;` $\C{// wait for character}$
731                                while ( ch == '\n' ); // ignore newline
732                                sout | ch;                      $\C{// print character}$
733                        } sout | " ";  $\C{// block separator}$
734                } sout | nl; $\C{// group separator}$
735        }
736}
737int main() {
738        Fmt fmt; $\C{// fmt constructor called}$
739        for () {
740                sin | fmt.ch; $\C{// read into generator}$
741          if ( eof( sin ) ) break;
742                `resume( fmt );`
743        }
744
745} $\C{// fmt destructor called}\CRT$
746\end{cfa}
747\end{lrbox}
748
749\begin{lrbox}{\myboxB}
750\begin{cfa}[aboveskip=0pt,belowskip=0pt]
751typedef struct {
752        int `restart`, g, b;
753        char ch;
754} Fmt;
755void comain( Fmt * f ) {
756        `static void * states[] = {&&s0, &&s1};`
757        `goto *states[f->restart];`
758  s0: f->`restart` = 1;
759        for ( ;; ) {
760                for ( f->g = 0; f->g < 5; f->g += 1 ) {
761                        for ( f->b = 0; f->b < 4; f->b += 1 ) {
762                                do { return;  s1: ;
763                                } while ( f->ch == '\n' );
764                                printf( "%c", f->ch );
765                        } printf( " " );
766                } printf( "\n" );
767        }
768}
769int main() {
770        Fmt fmt = { `0` };  comain( &fmt ); // prime
771        for ( ;; ) {
772                scanf( "%c", &fmt.ch );
773          if ( feof( stdin ) ) break;
774                comain( &fmt );
775        }
776        if ( fmt.g != 0 || fmt.b != 0 ) printf( "\n" );
777}
778\end{cfa}
779\end{lrbox}
780
781\subfloat[\CFA]{\label{f:CFAFormatGen}\usebox\myboxA}
782\hspace{35pt}
783\vrule
784\hspace{3pt}
785\subfloat[C generated code for \CFA version]{\label{f:CFormatGenImpl}\usebox\myboxB}
786\hspace{3pt}
787\caption{Formatter input asymmetric generator}
788\label{f:FormatterAsymmetricGenerator}
789\end{figure}
790
791Figure~\ref{f:FibonacciAsymmetricGenerator} shows an unbounded asymmetric generator for an infinite sequence of Fibonacci numbers written left to right in C, \CFA, and showing the underlying C implementation for the \CFA version.
792This generator is an \emph{output generator}, producing a new result on each resumption.
793To compute Fibonacci, the previous two values in the sequence are retained to generate the next value, \ie @fn1@ and @fn@, plus the execution location where control restarts when the generator is resumed, \ie top or middle.
794An additional requirement is the ability to create an arbitrary number of generators of any kind, \ie retaining one state in global variables is insufficient;
795hence, state is retained in a closure between calls.
796Figure~\ref{f:CFibonacci} shows the C approach of manually creating the closure in structure @Fib@, and multiple instances of this closure provide multiple Fibonacci generators.
797The C version only has the middle execution state because the top execution state is declaration initialization.
798Figure~\ref{f:CFAFibonacciGen} shows the \CFA approach, which also has a manual closure, but replaces the structure with a custom \CFA @generator@ type.
799Each generator type must have a function named \lstinline|main|,
800% \footnote{
801% The name \lstinline|main| has special meaning in C, specifically the function where a program starts execution.
802% Leveraging starting semantics to this name for generator/coroutine/thread is a logical extension.}
803called a \emph{generator main} (leveraging the starting semantics for program @main@ in C), which is connected to the generator type via its single reference parameter.
804The generator main contains @suspend@ statements that suspend execution without ending the generator versus @return@.
805For the Fibonacci generator-main,
806the top initialization state appears at the start and the middle execution state is denoted by statement @suspend@.
807Any local variables in @main@ \emph{are not retained} between calls;
808hence local variables are only for temporary computations \emph{between} suspends.
809All retained state \emph{must} appear in the generator's type.
810As well, generator code containing a @suspend@ cannot be refactored into a helper function called by the generator, because @suspend@ is implemented via @return@, so a return from the helper function goes back to the current generator not the resumer.
811The generator is started by calling function @resume@ with a generator instance, which begins execution at the top of the generator main, and subsequent @resume@ calls restart the generator at its point of last suspension.
812Resuming an ended (returned) generator is undefined.
813Function @resume@ returns its argument generator so it can be cascaded in an expression, in this case to print the next Fibonacci value @fn@ computed in the generator instance.
814Figure~\ref{f:CFibonacciSim} shows the C implementation of the \CFA asymmetric generator.
815Only one execution-state field, @restart@, is needed to subscript the suspension points in the generator.
816At the start of the generator main, the @static@ declaration, @states@, is initialized to the N suspend points in the generator, where operator @&&@ dereferences or references a label~\cite{gccValueLabels}.
817Next, the computed @goto@ selects the last suspend point and branches to it.
818The  cost of setting @restart@ and branching via the computed @goto@ adds very little cost to the suspend and resume calls.
819
820An advantage of the \CFA explicit generator type is the ability to allow multiple type-safe interface functions taking and returning arbitrary types.
821\begin{cfa}
822int ?()( Fib & fib ) { return `resume( fib )`.fn; } $\C[3.9in]{// function-call interface}$
823int ?()( Fib & fib, int N ) { for ( N - 1 ) `fib()`; return `fib()`; } $\C{// add parameter to skip N values}$
824double ?()( Fib & fib ) { return (int)`fib()` / 3.14159; } $\C{// different return type, cast prevents recursive call}$
825Fib f;  int i;  double d;
826i = f();  i = f( 2 );  d = f();                                         $\C{// alternative interfaces}\CRT$
827\end{cfa}
828Now, the generator can be a separately compiled opaque-type only accessed through its interface functions.
829For contrast, Figure~\ref{f:PythonFibonacci} shows the equivalent Python Fibonacci generator, which does not use a generator type, and hence only has a single interface, but an implicit closure.
830
831\begin{figure}
832%\centering
833\newbox\myboxA
834\begin{lrbox}{\myboxA}
835\begin{python}[aboveskip=0pt,belowskip=0pt]
836def Fib():
837        fn1, fn = 0, 1
838        while True:
839                `yield fn1`
840                fn1, fn = fn, fn1 + fn
841f1 = Fib()
842f2 = Fib()
843for i in range( 10 ):
844        print( next( f1 ), next( f2 ) )
845
846
847
848
849
850
851
852
853
854
855\end{python}
856\end{lrbox}
857
858\newbox\myboxB
859\begin{lrbox}{\myboxB}
860\begin{python}[aboveskip=0pt,belowskip=0pt]
861def Fmt():
862        try:
863                while True:                                             $\C[2.5in]{\# until destructor call}$
864                        for g in range( 5 ):            $\C{\# groups}$
865                                for b in range( 4 ):    $\C{\# blocks}$
866                                        while True:
867                                                ch = (yield)    $\C{\# receive from send}$
868                                                if '\n' not in ch: $\C{\# ignore newline}$
869                                                        break
870                                        print( ch, end='' )     $\C{\# print character}$
871                                print( '  ', end='' )   $\C{\# block separator}$
872                        print()                                         $\C{\# group separator}$
873        except GeneratorExit:                           $\C{\# destructor}$
874                if g != 0 | b != 0:                             $\C{\# special case}$
875                        print()
876fmt = Fmt()
877`next( fmt )`                                                   $\C{\# prime, next prewritten}$
878for i in range( 41 ):
879        `fmt.send( 'a' );`                                      $\C{\# send to yield}$
880\end{python}
881\end{lrbox}
882
883\hspace{30pt}
884\subfloat[Fibonacci]{\label{f:PythonFibonacci}\usebox\myboxA}
885\hspace{3pt}
886\vrule
887\hspace{3pt}
888\subfloat[Formatter]{\label{f:PythonFormatter}\usebox\myboxB}
889\caption{Python generator}
890\label{f:PythonGenerator}
891\end{figure}
892
893Having to manually create the generator closure by moving local-state variables into the generator type is an additional programmer burden (removed by the coroutine in Section~\ref{s:Coroutine}).
894This manual requirement follows from the generality of allowing variable-size local-state, \eg local state with a variable-length array requires dynamic allocation as the array size is unknown at compile time.
895However, dynamic allocation significantly increases the cost of generator creation/destruction and is a showstopper for embedded real-time programming.
896But more importantly, the size of the generator type is tied to the local state in the generator main, which precludes separate compilation of the generator main, \ie a generator must be inlined or local state must be dynamically allocated.
897With respect to safety, we believe static analysis can discriminate persistent generator state from temporary generator-main state and raise a compile-time error for temporary usage spanning suspend points.
898Our experience using generators is that the problems have simple data state, including local state, but complex execution state, so the burden of creating the generator type is small.
899As well, C programmers are not afraid of this kind of semantic programming requirement, if it results in very small and fast generators.
900
901Figure~\ref{f:CFAFormatGen} shows an asymmetric \newterm{input generator}, @Fmt@, for restructuring text into groups of characters of fixed-size blocks, \ie the input on the left is reformatted into the output on the right, where newlines are ignored.
902\begin{center}
903\tt
904\begin{tabular}{@{}l|l@{}}
905\multicolumn{1}{c|}{\textbf{\textrm{input}}} & \multicolumn{1}{c}{\textbf{\textrm{output}}} \\
906\begin{tabular}[t]{@{}ll@{}}
907abcdefghijklmnopqrstuvwxyz \\
908abcdefghijklmnopqrstuvwxyz
909\end{tabular}
910&
911\begin{tabular}[t]{@{}lllll@{}}
912abcd    & efgh  & ijkl  & mnop  & qrst  \\
913uvwx    & yzab  & cdef  & ghij  & klmn  \\
914opqr    & stuv  & wxyz  &               &
915\end{tabular}
916\end{tabular}
917\end{center}
918The example takes advantage of resuming a generator in the constructor to prime the loops so the first character sent for formatting appears inside the nested loops.
919The destructor provides a newline, if formatted text ends with a full line.
920Figure~\ref{f:CFormatGenImpl} shows the C implementation of the \CFA input generator with one additional field and the computed @goto@.
921For contrast, Figure~\ref{f:PythonFormatter} shows the equivalent Python format generator with the same properties as the \CFA format generator.
922
923% https://dl-acm-org.proxy.lib.uwaterloo.ca/
924
925An important application for the asymmetric generator is a device-driver, because device drivers are a significant source of operating-system errors: 85\% in Windows XP~\cite[p.~78]{Swift05} and 51.6\% in Linux~\cite[p.~1358,]{Xiao19}. %\cite{Palix11}
926Swift \etal~\cite[p.~86]{Swift05} restructure device drivers using the Extension Procedure Call (XPC) within the kernel via functions @nooks_driver_call@ and @nooks_kernel_call@, which have coroutine properties context switching to separate stacks with explicit hand-off calls;
927however, the calls do not retain execution state, and hence always start from the top.
928The alternative approach for implementing device drivers is using stack-ripping.
929However, Adya \etal~\cite{Adya02} argue against stack ripping in Section 3.2 and suggest a hybrid approach in Section 4 using cooperatively scheduled \emph{fibers}, which is coroutining.
930
931Figure~\ref{f:DeviceDriverGen} shows the generator advantages in implementing a simple network device-driver with the following protocol:
932\begin{center}
933\ldots\, STX \ldots\, message \ldots\, ESC ETX \ldots\, message \ldots\, ETX 2-byte crc \ldots
934\end{center}
935where the network message begins with the control character STX, ends with an ETX, and is followed by a 2-byte cyclic-redundancy check.
936Control characters may appear in a message if preceded by an ESC.
937When a message byte arrives, it triggers an interrupt, and the operating system services the interrupt by calling the device driver with the byte read from a hardware register.
938The device driver returns a status code of its current state, and when a complete message is obtained, the operating system reads the message accumulated in the supplied buffer.
939Hence, the device driver is an input/output generator, where the cost of resuming the device-driver generator is the same as call and return, so performance in an operating-system kernel is excellent.
940The key benefits of using a generator are correctness, safety, and maintenance because the execution states are transcribed directly into the programming language rather than table lookup or stack ripping.
941% The conclusion is that FSMs are complex and occur in important domains, so direct generator support is important in a system programming language.
942
943\begin{figure}
944\centering
945\begin{tabular}{@{}l|l@{}}
946\begin{cfa}[aboveskip=0pt,belowskip=0pt]
947enum Status { CONT, MSG, ESTX,
948                                ELNTH, ECRC };
949`generator` Driver {
950        Status status;
951        char byte, * msg; // communication
952        int lnth, sum;      // local state
953        short int crc;
954};
955void ?{}( Driver & d, char * m ) { d.msg = m; }
956Status next( Driver & d, char b ) with( d ) {
957        byte = b; `resume( d );` return status;
958}
959void main( Driver & d ) with( d ) {
960        enum { STX = '\002', ESC = '\033',
961                        ETX = '\003', MaxMsg = 64 };
962  msg: for () { // parse message
963                status = CONT;
964                lnth = 0; sum = 0;
965                while ( byte != STX ) `suspend;`
966          emsg: for () {
967                        `suspend;` // process byte
968\end{cfa}
969&
970\begin{cfa}[aboveskip=0pt,belowskip=0pt]
971                        choose ( byte ) { // switch with implicit break
972                          case STX:
973                                status = ESTX; `suspend;` continue msg;
974                          case ETX:
975                                break emsg;
976                          case ESC:
977                                `suspend;`
978                        }
979                        if ( lnth >= MaxMsg ) { // buffer full ?
980                                status = ELNTH; `suspend;` continue msg; }
981                        msg[lnth++] = byte;
982                        sum += byte;
983                }
984                msg[lnth] = '\0'; // terminate string
985                `suspend;`
986                crc = byte << 8;
987                `suspend;`
988                status = (crc | byte) == sum ? MSG : ECRC;
989                `suspend;`
990        }
991}
992\end{cfa}
993\end{tabular}
994\caption{Device-driver generator for communication protocol}
995\label{f:DeviceDriverGen}
996\end{figure}
997
998Generators can also have symmetric activation using resume/resume to create control-flow cycles among generators.
999(The trivial cycle is a generator resuming itself.)
1000This control flow is similar to recursion for functions but without stack growth.
1001Figure~\ref{f:PingPongFullCoroutineSteps} shows the steps for symmetric control-flow using for the ping/pong program in Figure~\ref{f:CFAPingPongGen}.
1002The program starts by creating the generators, @ping@ and @pong@, and then assigns the partners that form the cycle.
1003Constructing the cycle must deal with definition-before-use to close the cycle, \ie, the first generator must know about the last generator, which is not within scope.
1004(This issue occurs for any cyclic data structure.)
1005% (Alternatively, the constructor can assign the partners as they are declared, except the first, and the first-generator partner is set after the last generator declaration to close the cycle.)
1006Once the cycle is formed, the program main resumes one of the generators, @ping@, and the generators can then traverse an arbitrary number of cycles using @resume@ to activate partner generator(s).
1007Terminating the cycle is accomplished by @suspend@ or @return@, both of which go back to the stack frame that started the cycle (program main in the example).
1008Note, the creator and starter may be different, \eg if the creator calls another function that starts the cycle.
1009The starting stack-frame is below the last active generator because the resume/resume cycle does not grow the stack.
1010Also, since local variables are not retained in the generator function, there are no objects with destructors to be called, so the cost is the same as a function return.
1011Destructor cost occurs when the generator instance is deallocated by the creator.
1012
1013\begin{figure}
1014\centering
1015\input{FullCoroutinePhases.pstex_t}
1016\vspace*{-10pt}
1017\caption{Symmetric coroutine steps: Ping / Pong}
1018\label{f:PingPongFullCoroutineSteps}
1019\end{figure}
1020
1021\begin{figure}
1022\centering
1023\begin{lrbox}{\myboxA}
1024\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1025`generator PingPong` {
1026        int N, i;                               // local state
1027        const char * name;
1028        PingPong & partner; // rebindable reference
1029};
1030
1031void `main( PingPong & pp )` with(pp) {
1032
1033
1034        for ( ; i < N; i += 1 ) {
1035                sout | name | i;
1036                `resume( partner );`
1037        }
1038}
1039int main() {
1040        enum { N = 5 };
1041        PingPong ping = {"ping",N,0}, pong = {"pong",N,0};
1042        &ping.partner = &pong;  &pong.partner = &ping;
1043        `resume( ping );`
1044}
1045\end{cfa}
1046\end{lrbox}
1047
1048\begin{lrbox}{\myboxB}
1049\begin{cfa}[escapechar={},aboveskip=0pt,belowskip=0pt]
1050typedef struct PingPong {
1051        int restart, N, i;
1052        const char * name;
1053        struct PingPong * partner;
1054} PingPong;
1055#define PPCtor(name, N) {0, N, 0, name, NULL}
1056void comain( PingPong * pp ) {
1057        static void * states[] = {&&s0, &&s1};
1058        goto *states[pp->restart];
1059  s0: pp->restart = 1;
1060        for ( ; pp->i < pp->N; pp->i += 1 ) {
1061                printf( "%s %d\n", pp->name, pp->i );
1062                asm( "mov  %0,%%rdi" : "=m" (pp->partner) );
1063                asm( "mov  %rdi,%rax" );
1064                asm( "add  $16, %rsp" );
1065                asm( "popq %rbp" );
1066                asm( "jmp  comain" );
1067          s1: ;
1068        }
1069}
1070\end{cfa}
1071\end{lrbox}
1072
1073\subfloat[\CFA symmetric generator]{\label{f:CFAPingPongGen}\usebox\myboxA}
1074\hspace{3pt}
1075\vrule
1076\hspace{3pt}
1077\subfloat[C generator simulation]{\label{f:CPingPongSim}\usebox\myboxB}
1078\hspace{3pt}
1079\caption{Ping-Pong symmetric generator}
1080\label{f:PingPongSymmetricGenerator}
1081\end{figure}
1082
1083Figure~\ref{f:CPingPongSim} shows the C implementation of the \CFA symmetric generator, where there is still only one additional field, @restart@, but @resume@ is more complex because it does a forward rather than backward jump.
1084Before the jump, the parameter for the next call @partner@ is placed into the register used for the first parameter, @rdi@, and the remaining registers are reset for a return.
1085The @jmp comain@ restarts the function but with a different parameter, so the new call's behaviour depends on the state of the coroutine type, i.e., branch to restart location with different data state.
1086While the semantics of call forward is a tail-call optimization, which compilers perform, the generator state is different on each call rather a common state for a tail-recursive function (i.e., the parameter to the function never changes during the forward calls.
1087However, this assembler code depends on what entry code is generated, specifically if there are local variables and the level of optimization.
1088Hence, internal compiler support is necessary for any forward call or backwards return, \eg LLVM has various coroutine support~\cite{CoroutineTS}, and \CFA can leverage this support should it eventually fork @clang@.
1089For this reason, \CFA does not support general symmetric generators at this time, but, it is possible to hand generate any symmetric generators, as in Figure~\ref{f:CPingPongSim}, for proof of concept and performance testing.
1090
1091Finally, part of this generator work was inspired by the recent \CCtwenty coroutine proposal~\cite{C++20Coroutine19}, which uses the general term coroutine to mean generator.
1092Our work provides the same high-performance asymmetric generators as \CCtwenty, and extends their work with symmetric generators.
1093An additional \CCtwenty generator feature allows @suspend@ and @resume@ to be followed by a restricted compound statement that is executed after the current generator has reset its stack but before calling the next generator, specified with \CFA syntax:
1094\begin{cfa}
1095... suspend`{ ... }`;
1096... resume( C )`{ ... }` ...
1097\end{cfa}
1098Since the current generator's stack is released before calling the compound statement, the compound statement can only reference variables in the generator's type.
1099This feature is useful when a generator is used in a concurrent context to ensure it is stopped before releasing a lock in the compound statement, which might immediately allow another thread to resume the generator.
1100Hence, this mechanism provides a general and safe handoff of the generator among competing threads.
1101
1102
1103\subsection{Coroutine}
1104\label{s:Coroutine}
1105
1106Stackful coroutines (Table~\ref{t:ExecutionPropertyComposition} case 5) extend generator semantics with an implicit closure and @suspend@ may appear in a helper function called from the coroutine main because of the separate stack.
1107Note, simulating coroutines with stacks of generators, \eg Python with @yield from@ cannot handle symmetric control-flow.
1108Furthermore, all stack components must be of generators, so it is impossible to call a library function passing a generator that yields.
1109Creating a generator copy of the library function maybe impossible because the library function is opaque.
1110
1111A \CFA coroutine is specified by replacing @generator@ with @coroutine@ for the type.
1112Coroutine generality results in higher cost for creation, due to dynamic stack allocation, for execution, due to context switching among stacks, and for terminating, due to possible stack unwinding and dynamic stack deallocation.
1113A series of different kinds of coroutines and their implementations demonstrate how coroutines extend generators.
1114
1115First, the previous generator examples are converted to their coroutine counterparts, allowing local-state variables to be moved from the generator type into the coroutine main.
1116Now the coroutine type only contains communication variables between interface functions and the coroutine main.
1117\begin{center}
1118\begin{tabular}{@{}l|l|l|l@{}}
1119\multicolumn{1}{c|}{Fibonacci} & \multicolumn{1}{c|}{Formatter} & \multicolumn{1}{c|}{Device Driver} & \multicolumn{1}{c}{PingPong} \\
1120\hline
1121\begin{cfa}[xleftmargin=0pt]
1122void main( Fib & fib ) ...
1123        `int fn1;`
1124
1125
1126\end{cfa}
1127&
1128\begin{cfa}[xleftmargin=0pt]
1129for ( `g`; 5 ) {
1130        for ( `b`; 4 ) {
1131
1132
1133\end{cfa}
1134&
1135\begin{cfa}[xleftmargin=0pt]
1136status = CONT;
1137`int lnth = 0, sum = 0;`
1138...
1139`short int crc = byte << 8;`
1140\end{cfa}
1141&
1142\begin{cfa}[xleftmargin=0pt]
1143void main( PingPong & pp ) ...
1144        for ( `i`; N ) {
1145
1146
1147\end{cfa}
1148\end{tabular}
1149\end{center}
1150It is also possible to refactor code containing local-state and @suspend@ statements into a helper function, like the computation of the CRC for the device driver.
1151\begin{cfa}
1152int Crc() {
1153        `suspend;`  short int crc = byte << 8;
1154        `suspend;`  status = (crc | byte) == sum ? MSG : ECRC;
1155        return crc;
1156}
1157\end{cfa}
1158A call to this function is placed at the end of the device driver's coroutine-main.
1159For complex finite-state machines, refactoring is part of normal program abstraction, especially when code is used in multiple places.
1160Again, this complexity is usually associated with execution state rather than data state.
1161
1162\begin{comment}
1163Figure~\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 functions, \eg @restart@.
1164Like 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.
1165The 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@.
1166The interface function @restart@, takes a Fibonacci instance and context switches to it using @resume@;
1167on restart, the Fibonacci field, @fn@, contains the next value in the sequence, which is returned.
1168The first @resume@ is special because it allocates the coroutine stack and cocalls its coroutine main on that stack;
1169when the coroutine main returns, its stack is deallocated.
1170Hence, @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.
1171Figure~\ref{f:Coroutine1State} shows the coroutine version of the C version in Figure~\ref{f:ExternalState}.
1172Coroutine generators are called \newterm{output coroutines} because values are only returned.
1173
1174\begin{figure}
1175\centering
1176\newbox\myboxA
1177% \begin{lrbox}{\myboxA}
1178% \begin{cfa}[aboveskip=0pt,belowskip=0pt]
1179% `int fn1, fn2, state = 1;`   // single global variables
1180% int fib() {
1181%       int fn;
1182%       `switch ( state )` {  // explicit execution state
1183%         case 1: fn = 0;  fn1 = fn;  state = 2;  break;
1184%         case 2: fn = 1;  fn2 = fn1;  fn1 = fn;  state = 3;  break;
1185%         case 3: fn = fn1 + fn2;  fn2 = fn1;  fn1 = fn;  break;
1186%       }
1187%       return fn;
1188% }
1189% int main() {
1190%
1191%       for ( int i = 0; i < 10; i += 1 ) {
1192%               printf( "%d\n", fib() );
1193%       }
1194% }
1195% \end{cfa}
1196% \end{lrbox}
1197\begin{lrbox}{\myboxA}
1198\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1199#define FibCtor { 0, 1 }
1200typedef struct { int fn1, fn; } Fib;
1201int fib( Fib * f ) {
1202
1203        int ret = f->fn1;
1204        f->fn1 = f->fn;
1205        f->fn = ret + f->fn;
1206        return ret;
1207}
1208
1209
1210
1211int main() {
1212        Fib f1 = FibCtor, f2 = FibCtor;
1213        for ( int i = 0; i < 10; i += 1 ) {
1214                printf( "%d %d\n",
1215                                fib( &f1 ), fib( &f2 ) );
1216        }
1217}
1218\end{cfa}
1219\end{lrbox}
1220
1221\newbox\myboxB
1222\begin{lrbox}{\myboxB}
1223\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1224`coroutine` Fib { int fn1; };
1225void main( Fib & fib ) with( fib ) {
1226        int fn;
1227        [fn1, fn] = [0, 1];
1228        for () {
1229                `suspend;`
1230                [fn1, fn] = [fn, fn1 + fn];
1231        }
1232}
1233int ?()( Fib & fib ) with( fib ) {
1234        return `resume( fib )`.fn1;
1235}
1236int main() {
1237        Fib f1, f2;
1238        for ( 10 ) {
1239                sout | f1() | f2();
1240}
1241
1242
1243\end{cfa}
1244\end{lrbox}
1245
1246\newbox\myboxC
1247\begin{lrbox}{\myboxC}
1248\begin{python}[aboveskip=0pt,belowskip=0pt]
1249
1250def Fib():
1251
1252        fn1, fn = 0, 1
1253        while True:
1254                `yield fn1`
1255                fn1, fn = fn, fn1 + fn
1256
1257
1258// next prewritten
1259
1260
1261f1 = Fib()
1262f2 = Fib()
1263for i in range( 10 ):
1264        print( next( f1 ), next( f2 ) )
1265
1266
1267
1268\end{python}
1269\end{lrbox}
1270
1271\subfloat[C]{\label{f:GlobalVariables}\usebox\myboxA}
1272\hspace{3pt}
1273\vrule
1274\hspace{3pt}
1275\subfloat[\CFA]{\label{f:ExternalState}\usebox\myboxB}
1276\hspace{3pt}
1277\vrule
1278\hspace{3pt}
1279\subfloat[Python]{\label{f:ExternalState}\usebox\myboxC}
1280\caption{Fibonacci generator}
1281\label{f:C-fibonacci}
1282\end{figure}
1283
1284\bigskip
1285
1286\newbox\myboxA
1287\begin{lrbox}{\myboxA}
1288\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1289`coroutine` Fib { int fn; };
1290void main( Fib & fib ) with( fib ) {
1291        fn = 0;  int fn1 = fn; `suspend`;
1292        fn = 1;  int fn2 = fn1;  fn1 = fn; `suspend`;
1293        for () {
1294                fn = fn1 + fn2; fn2 = fn1; fn1 = fn; `suspend`; }
1295}
1296int next( Fib & fib ) with( fib ) { `resume( fib );` return fn; }
1297int main() {
1298        Fib f1, f2;
1299        for ( 10 )
1300                sout | next( f1 ) | next( f2 );
1301}
1302\end{cfa}
1303\end{lrbox}
1304\newbox\myboxB
1305\begin{lrbox}{\myboxB}
1306\begin{python}[aboveskip=0pt,belowskip=0pt]
1307
1308def Fibonacci():
1309        fn = 0; fn1 = fn; `yield fn`  # suspend
1310        fn = 1; fn2 = fn1; fn1 = fn; `yield fn`
1311        while True:
1312                fn = fn1 + fn2; fn2 = fn1; fn1 = fn; `yield fn`
1313
1314
1315f1 = Fibonacci()
1316f2 = Fibonacci()
1317for i in range( 10 ):
1318        print( `next( f1 )`, `next( f2 )` ) # resume
1319
1320\end{python}
1321\end{lrbox}
1322\subfloat[\CFA]{\label{f:Coroutine3States}\usebox\myboxA}
1323\qquad
1324\subfloat[Python]{\label{f:Coroutine1State}\usebox\myboxB}
1325\caption{Fibonacci input coroutine, 3 states, internal variables}
1326\label{f:cfa-fibonacci}
1327\end{figure}
1328\end{comment}
1329
1330\begin{figure}
1331\centering
1332\begin{tabular}{@{}l@{\hspace{2\parindentlnth}}l@{}}
1333\begin{cfa}
1334`coroutine` Prod {
1335        Cons & c;                       $\C[1.5in]{// communication}$
1336        int N, money, receipt;
1337};
1338void main( Prod & prod ) with( prod ) {
1339        for ( i; N ) {          $\C{// 1st resume}\CRT$
1340                int p1 = random( 100 ), p2 = random( 100 );
1341                int status = delivery( c, p1, p2 );
1342                receipt += 1;
1343        }
1344        stop( c );
1345}
1346int payment( Prod & prod, int money ) {
1347        prod.money = money;
1348        `resume( prod );`
1349        return prod.receipt;
1350}
1351void start( Prod & prod, int N, Cons &c ) {
1352        &prod.c = &c;
1353        prod.[N, receipt] = [N, 0];
1354        `resume( prod );`
1355}
1356int main() {
1357        Prod prod;
1358        Cons cons = { prod };
1359        start( prod, 5, cons );
1360}
1361\end{cfa}
1362&
1363\begin{cfa}
1364`coroutine` Cons {
1365        Prod & p;                       $\C[1.5in]{// communication}$
1366        int p1, p2, status;
1367        bool done;
1368};
1369void ?{}( Cons & cons, Prod & p ) {
1370        &cons.p = &p;           $\C{// reassignable reference}$
1371        cons.[status, done ] = [0, false];
1372}
1373void main( Cons & cons ) with( cons ) {
1374        int money = 1, receipt; $\C{// 1st resume}\CRT$
1375        for ( ; ! done; ) {
1376                status += 1;
1377                receipt = payment( p, money );
1378                money += 1;
1379        }
1380}
1381int delivery( Cons & cons, int p1, int p2 ) {
1382        cons.[p1, p2] = [p1, p2];
1383        `resume( cons );`
1384        return cons.status;
1385}
1386void stop( Cons & cons ) {
1387        cons.done = true;
1388        `resume( cons );`
1389}
1390
1391\end{cfa}
1392\end{tabular}
1393\caption{Producer / consumer: resume-resume cycle, bidirectional communication}
1394\label{f:ProdCons}
1395\end{figure}
1396
1397Figure~\ref{f:ProdCons} shows the ping-pong example in Figure~\ref{f:CFAPingPongGen} extended into a producer/consumer symmetric-coroutine performing bidirectional communication.
1398This example is illustrative because both producer and consumer have two interface functions with @resume@s that suspend execution in these interface functions.
1399The program main creates the producer coroutine, passes it to the consumer coroutine in its initialization, and closes the cycle at the call to @start@ along with the number of items to be produced.
1400The call to @start@ is the first @resume@ of @prod@, which remembers the program main as the starter and creates @prod@'s stack with a frame for @prod@'s coroutine main at the top, and context switches to it.
1401@prod@'s coroutine main starts, creates local-state variables that are retained between coroutine activations, and executes $N$ iterations, each generating two random values, calling the consumer's @deliver@ function to transfer the values, and printing the status returned from the consumer.
1402The producer's call to @delivery@ transfers values into the consumer's communication variables, resumes the consumer, and returns the consumer status.
1403Similarly on the first resume, @cons@'s stack is created and initialized, holding local-state variables retained between subsequent activations of the coroutine.
1404The symmetric coroutine cycle forms when the consumer calls the producer's @payment@ function, which resumes the producer in the consumer's delivery function.
1405When the producer calls @delivery@ again, it resumes the consumer in the @payment@ function.
1406Both interface functions then return to their corresponding coroutine-main functions for the next cycle.
1407Figure~\ref{f:ProdConsRuntimeStacks} shows the runtime stacks of the program main, and the coroutine mains for @prod@ and @cons@ during the cycling.
1408As a consequence of a coroutine retaining its last resumer for suspending back, these reverse pointers allow @suspend@ to cycle \emph{backwards} around a symmetric coroutine cycle.
1409
1410\begin{figure}
1411\begin{center}
1412\input{FullProdConsStack.pstex_t}
1413\end{center}
1414\vspace*{-10pt}
1415\caption{Producer / consumer runtime stacks}
1416\label{f:ProdConsRuntimeStacks}
1417\end{figure}
1418
1419Terminating a coroutine cycle is more complex than a generator cycle, because it requires context switching to the program main's \emph{stack} to shutdown the program, whereas generators started by the program main run on its stack.
1420Furthermore, each deallocated coroutine must execute all destructors for objects allocated in the coroutine type \emph{and} allocated on the coroutine's stack at the point of suspension, which can be arbitrarily deep.
1421In the example, termination begins with the producer's loop stopping after N iterations and calling the consumer's @stop@ function, which sets the @done@ flag, resumes the consumer in function @payment@, terminating the call, and the consumer's loop in its coroutine main.
1422% (Not shown is having @prod@ raise a nonlocal @stop@ exception at @cons@ after it finishes generating values and suspend back to @cons@, which catches the @stop@ exception to terminate its loop.)
1423When the consumer's main ends, its stack is already unwound so any stack allocated objects with destructors are finalized.
1424The question now is where does control continue?
1425
1426The na\"{i}ve semantics for coroutine-cycle termination is to context switch to the last resumer, like executing a @suspend@ or @return@ in a generator.
1427However, for coroutines, the last resumer is \emph{not} implicitly below the current stack frame, as for generators, because each coroutine's stack is independent.
1428Unfortunately, it is impossible to determine statically if a coroutine is in a cycle and unrealistic to check dynamically (graph-cycle problem).
1429Hence, a compromise solution is necessary that works for asymmetric (acyclic) and symmetric (cyclic) coroutines.
1430Our solution is to retain a coroutine's starter (first resumer), and context switch back to the starter when the coroutine ends.
1431Hence, the consumer restarts its first resumer, @prod@, in @stop@, and when the producer ends, it restarts its first resumer, program main, in @start@ (see dashed lines from the end of the coroutine mains in Figure~\ref{f:ProdConsRuntimeStacks}).
1432This semantics works well for the most common asymmetric and symmetric coroutine usage patterns.
1433For asymmetric coroutines, it is common for the first resumer (starter) coroutine to be the only resumer;
1434for symmetric coroutines, it is common for the cycle creator to persist for the lifetime of the cycle.
1435For other scenarios, it is always possible to devise a solution with additional programming effort, such as forcing the cycle forward or backward to a safe point before starting termination.
1436
1437Note, the producer/consumer example does not illustrate the full power of the starter semantics because @cons@ always ends first.
1438Assume generator @PingPong@ in Figure~\ref{f:PingPongSymmetricGenerator} is converted to a coroutine.
1439Unlike generators, coroutines have a starter structure with multiple levels, where the program main starts @ping@ and @ping@ starts @pong@.
1440By adjusting $N$ for either @ping@ or @pong@, it is possible to have either finish first.
1441If @pong@ ends first, it resumes its starter @ping@ in its coroutine main, then @ping@ ends and resumes its starter the program main on return;
1442if @ping@ ends first, it resumes its starter the program main on return.
1443Regardless of the cycle complexity, the starter structure always leads back to the program main, but the path can be entered at an arbitrary point.
1444Once back at the program main (creator), coroutines @ping@ and @pong@ are deallocated, running any destructors for objects within the coroutine and possibly deallocating any coroutine stacks for non-terminated coroutines, where stack deallocation implies stack unwinding to find destructors for allocated objects on the stack.
1445Hence, the \CFA termination semantics for the generator and coroutine ensure correct deallocation semantics, regardless of the coroutine's state (terminated or active), like any other aggregate object.
1446
1447
1448\subsection{Generator / Coroutine Implementation}
1449
1450A significant implementation challenge for generators and coroutines (and threads in Section~\ref{s:threads}) is adding extra fields to the custom types and related functions, \eg inserting code after/before the coroutine constructor/destructor and @main@ to create/initialize/de-initialize/destroy any extra fields, \eg the coroutine stack.
1451There are several solutions to this problem, which follow from the object-oriented flavour of adopting custom types.
1452
1453For object-oriented languages, inheritance is used to provide extra fields and code via explicit inheritance:
1454\begin{cfa}[morekeywords={class,inherits}]
1455class myCoroutine inherits baseCoroutine { ... }
1456\end{cfa}
1457% The problem is that the programming language and its tool chain, \eg debugger, @valgrind@, need to understand @baseCoroutine@ because it infers special property, so type @baseCoroutine@ becomes a de facto keyword and all types inheriting from it are implicitly custom types.
1458The problem is that some special properties are not handled by existing language semantics, \eg the execution of constructors and destructors is in the wrong order to implicitly start threads because the thread must start \emph{after} all constructors as it relies on a completely initialized object, but the inherited constructor runs \emph{before} the derived.
1459Alternatives, such as explicitly starting threads as in Java, are repetitive and forgetting to call start is a common source of errors.
1460An alternative is composition:
1461\begin{cfa}
1462struct myCoroutine {
1463        ... // declaration/communication variables
1464        baseCoroutine dummy; // composition, last declaration
1465}
1466\end{cfa}
1467which also requires an explicit declaration that must be last to ensure correct initialization order.
1468However, there is nothing preventing wrong placement or multiple declarations.
1469
1470\CFA custom types make any special properties explicit to the language and its tool chain, \eg the language code-generator knows where to inject code
1471% and when it is unsafe to perform certain optimizations,
1472and IDEs using simple parsing can find and manipulate types with special properties.
1473The downside of this approach is that it makes custom types a special case in the language.
1474Users wanting to extend custom types or build their own can only do so in ways offered by the language.
1475Furthermore, implementing custom types without language support may display the power of a programming language.
1476\CFA blends the two approaches, providing custom type for idiomatic \CFA code, while extending and building new custom types is still possible, similar to Java concurrency with builtin and library (@java.util.concurrent@) monitors.
1477
1478Part of the mechanism to generalize custom types is the \CFA trait~\cite[\S~2.3]{Moss18}, \eg the definition for custom-type @coroutine@ is anything satisfying the trait @is_coroutine@, and this trait both enforces and restricts the coroutine-interface functions.
1479\begin{cfa}
1480trait is_coroutine( `dtype` T ) {
1481        void main( T & );
1482        coroutine_desc * get_coroutine( T & );
1483};
1484forall( `dtype` T | is_coroutine(T) ) void $suspend$( T & ), resume( T & );
1485\end{cfa}
1486Note, copying generators, coroutines, and threads is undefined because multiple objects cannot execute on a shared stack and stack copying does not work in unmanaged languages (no garbage collection), like C, because the stack may contain pointers to objects within it that require updating for the copy.
1487The \CFA @dtype@ property provides no \emph{implicit} copying operations and the @is_coroutine@ trait provides no \emph{explicit} copying operations, so all coroutines must be passed by reference or pointer.
1488The function definitions ensure there is a statically typed @main@ function that is the starting point (first stack frame) of a coroutine, and a mechanism to read the coroutine descriptor from its handle.
1489The @main@ function has no return value or additional parameters because the coroutine type allows an arbitrary number of interface functions with arbitrary typed input and output values versus fixed ones.
1490The 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@ function, and possibly redefining \textsf{suspend} and @resume@.
1491
1492The \CFA custom-type @coroutine@ implicitly implements the getter and forward declarations for the coroutine main.
1493\begin{cquote}
1494\begin{tabular}{@{}ccc@{}}
1495\begin{cfa}
1496coroutine MyCor {
1497        int value;
1498
1499};
1500\end{cfa}
1501&
1502{\Large $\Rightarrow$}
1503&
1504\begin{tabular}{@{}ccc@{}}
1505\begin{cfa}
1506struct MyCor {
1507        int value;
1508        coroutine_desc cor;
1509};
1510\end{cfa}
1511&
1512\begin{cfa}
1513static inline coroutine_desc *
1514get_coroutine( MyCor & this ) {
1515        return &this.cor;
1516}
1517\end{cfa}
1518&
1519\begin{cfa}
1520void main( MyCor * this );
1521
1522
1523
1524\end{cfa}
1525\end{tabular}
1526\end{tabular}
1527\end{cquote}
1528The combination of custom types and fundamental @trait@ description of these types allows a concise specification for programmers and tools, while more advanced programmers can have tighter control over memory layout and initialization.
1529
1530Figure~\ref{f:CoroutineMemoryLayout} shows different memory-layout options for a coroutine (where a thread is similar).
1531The coroutine handle is the @coroutine@ instance containing programmer specified type global and communication variables across interface functions.
1532The coroutine descriptor contains all implicit declarations needed by the runtime, \eg @suspend@/@resume@, and can be part of the coroutine handle or separate.
1533The coroutine stack can appear in a number of locations and be fixed or variable sized.
1534Hence, the coroutine's stack could be a variable-length structure (VLS)
1535% \footnote{
1536% We are examining VLSs, where fields can be variable-sized structures or arrays.
1537% Once allocated, a VLS is fixed sized.}
1538on the allocating stack, provided the allocating stack is large enough.
1539For a VLS stack allocation and deallocation is an inexpensive adjustment of the stack pointer, modulo any stack constructor costs to initial frame setup.
1540For stack allocation in the heap, allocation and deallocation is an expensive allocation, where the heap can be a shared resource, modulo any stack constructor costs.
1541It is also possible to use a split or segmented stack calling convention, available with gcc and clang, allowing a variable-sized stack via a set of connected blocks in the heap.
1542Currently, \CFA supports stack and heap allocated descriptors but only fixed-sized heap allocated stacks.
1543In \CFA debug-mode, the fixed-sized stack is terminated with a write-only page, which catches most stack overflows.
1544Experience teaching concurrency with \uC~\cite{CS343} shows fixed-sized stacks are rarely an issue for students.
1545Split-stack allocation is under development but requires recompilation of legacy code, which is not always possible.
1546
1547\begin{figure}
1548\centering
1549\input{corlayout.pstex_t}
1550\caption{Coroutine memory layout}
1551\label{f:CoroutineMemoryLayout}
1552\end{figure}
1553
1554
1555\section{Concurrency}
1556\label{s:Concurrency}
1557
1558Concurrency is nondeterministic scheduling of independent sequential execution paths (threads), where each thread has its own stack.
1559A single thread with multiple stacks, \ie coroutining, does \emph{not} imply concurrency~\cite[\S~3]{Buhr05a}.
1560Coroutining self-schedule the thread across stacks so execution is deterministic.
1561(It is \emph{impossible} to generate a concurrency error when coroutining.)
1562
1563The transition to concurrency, even for a single thread with multiple stacks, occurs when coroutines context switch to a \newterm{scheduling coroutine}, introducing non-determinism from the coroutine perspective~\cite[\S~3]{Buhr05a}.
1564Therefore, a minimal concurrency system requires coroutines \emph{in conjunction with a nondeterministic scheduler}.
1565The resulting execution system now follows a cooperative threading-model~\cite{Adya02,libdill} because context-switching points to the scheduler are known, but the next unblocking point is unknown due to the scheduler.
1566Adding \newterm{preemption} introduces \newterm{non-cooperative} or \newterm{preemptive} scheduling, where context switching points to the scheduler are unknown as they can occur randomly between any two instructions often based on a timer interrupt.
1567Uncertainty gives the illusion of parallelism on a single processor and provides a mechanism to access and increase performance on multiple processors.
1568The reason is that the scheduler and runtime have complete knowledge about resources and how to best utilized them.
1569However, the introduction of unrestricted nondeterminism results in the need for \newterm{mutual exclusion} and \newterm{synchronization}~\cite[\S~4]{Buhr05a}, which restrict nondeterminism for correctness;
1570otherwise, it is impossible to write meaningful concurrent programs.
1571Optimal concurrent performance is often obtained by having as much nondeterminism as mutual exclusion and synchronization correctness allow.
1572
1573A scheduler can also be stackless or stackful.
1574For 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.
1575For 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.
1576The \CFA runtime uses a stackful scheduler for uniformity and security.
1577
1578
1579\subsection{Thread}
1580\label{s:threads}
1581
1582Threading (Table~\ref{t:ExecutionPropertyComposition} case 11) needs the ability to start a thread and wait for its completion, where a common API is @fork@ and @join@.
1583\vspace{4pt}
1584\par\noindent
1585\begin{tabular}{@{}l|l|l@{}}
1586\multicolumn{1}{c|}{\textbf{Java}} & \multicolumn{1}{c|}{\textbf{\Celeven}} & \multicolumn{1}{c}{\textbf{pthreads}} \\
1587\hline
1588\begin{cfa}
1589class MyThread extends Thread {...}
1590mythread t = new MyThread(...);
1591`t.start();` // start
1592// concurrency
1593`t.join();` // wait
1594\end{cfa}
1595&
1596\begin{cfa}
1597class MyThread { ... } // functor
1598MyThread mythread;
1599`thread t( mythread, ... );` // start
1600// concurrency
1601`t.join();` // wait
1602\end{cfa}
1603&
1604\begin{cfa}
1605void * rtn( void * arg ) {...}
1606pthread_t t;  int i = 3;
1607`pthread_create( &t, rtn, (void *)i );` // start
1608// concurrency
1609`pthread_join( t, NULL );` // wait
1610\end{cfa}
1611\end{tabular}
1612\vspace{1pt}
1613\par\noindent
1614\CFA has a simpler approach using a custom @thread@ type and leveraging declaration semantics, allocation and deallocation, where threads implicitly @fork@ after construction and @join@ before destruction.
1615\begin{cfa}
1616thread MyThread {};
1617void main( MyThread & this ) { ... }
1618int main() {
1619        MyThread team`[10]`; $\C[2.5in]{// allocate stack-based threads, implicit start after construction}$
1620        // concurrency
1621} $\C{// deallocate stack-based threads, implicit joins before destruction}$
1622\end{cfa}
1623This semantic ensures a thread is started and stopped exactly once, eliminating some programming error, and scales to multiple threads for basic termination synchronization.
1624For block allocation to arbitrary depth, including recursion, threads are created and destroyed in a lattice structure (tree with top and bottom).
1625Arbitrary topologies are possible using dynamic allocation, allowing threads to outlive their declaration scope, identical to normal dynamic allocation.
1626\begin{cfa}
1627MyThread * factory( int N ) { ... return `anew( N )`; } $\C{// allocate heap-based threads, implicit start after construction}$
1628int main() {
1629        MyThread * team = factory( 10 );
1630        // concurrency
1631        `adelete( team );` $\C{// deallocate heap-based threads, implicit joins before destruction}\CRT$
1632}
1633\end{cfa}
1634
1635Figure~\ref{s:ConcurrentMatrixSummation} shows concurrently adding the rows of a matrix and then totalling the subtotals sequentially, after all the row threads have terminated.
1636The program uses heap-based threads because each thread needs different constructor values.
1637(Python provides a simple iteration mechanism to initialize array elements to different values allowing stack allocation.)
1638The allocation/deallocation pattern appears unusual because allocated objects are immediately deallocated without any intervening code.
1639However, for threads, the deletion provides implicit synchronization, which is the intervening code.
1640% While 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.
1641
1642\begin{figure}
1643\begin{cfa}
1644`thread` Adder { int * row, cols, & subtotal; } $\C{// communication variables}$
1645void ?{}( Adder & adder, int row[], int cols, int & subtotal ) {
1646        adder.[ row, cols, &subtotal ] = [ row, cols, &subtotal ];
1647}
1648void main( Adder & adder ) with( adder ) {
1649        subtotal = 0;
1650        for ( c; cols ) { subtotal += row[c]; }
1651}
1652int main() {
1653        const int rows = 10, cols = 1000;
1654        int matrix[rows][cols], subtotals[rows], total = 0;
1655        // read matrix
1656        Adder * adders[rows];
1657        for ( r; rows; ) { $\C{// start threads to sum rows}$
1658                adders[r] = `new( matrix[r], cols, &subtotals[r] );`
1659        }
1660        for ( r; rows ) { $\C{// wait for threads to finish}$
1661                `delete( adders[r] );` $\C{// termination join}$
1662                total += subtotals[r]; $\C{// total subtotal}$
1663        }
1664        sout | total;
1665}
1666\end{cfa}
1667\caption{Concurrent matrix summation}
1668\label{s:ConcurrentMatrixSummation}
1669\end{figure}
1670
1671
1672\subsection{Thread Implementation}
1673
1674Threads in \CFA are user level run by runtime kernel threads (see Section~\ref{s:CFARuntimeStructure}), where user threads provide concurrency and kernel threads provide parallelism.
1675Like coroutines, and for the same design reasons, \CFA provides a custom @thread@ type and a @trait@ to enforce and restrict the thread-interface functions.
1676\begin{cquote}
1677\begin{tabular}{@{}c@{\hspace{3\parindentlnth}}c@{}}
1678\begin{cfa}
1679thread myThread {
1680        ... // declaration/communication variables
1681};
1682
1683
1684\end{cfa}
1685&
1686\begin{cfa}
1687trait is_thread( `dtype` T ) {
1688        void main( T & );
1689        thread_desc * get_thread( T & );
1690        void ^?{}( T & `mutex` );
1691};
1692\end{cfa}
1693\end{tabular}
1694\end{cquote}
1695Like coroutines, the @dtype@ property prevents \emph{implicit} copy operations and the @is_thread@ trait provides no \emph{explicit} copy operations, so threads must be passed by reference or pointer.
1696Similarly, the function definitions ensure there is a statically typed @main@ function that is the thread starting point (first stack frame), a mechanism to read the thread descriptor from its handle, and a special destructor to prevent deallocation while the thread is executing.
1697(The qualifier @mutex@ for the destructor parameter is discussed in Section~\ref{s:Monitor}.)
1698The difference between the coroutine and thread is that a coroutine borrows a thread from its caller, so the first thread resuming a coroutine creates the coroutine's stack and starts running the coroutine main on the stack;
1699whereas, a thread is scheduling for execution in @main@ immediately after its constructor is run.
1700No return value or additional parameters are necessary for this function because the @thread@ type allows an arbitrary number of interface functions with corresponding arbitrary typed input and output values.
1701
1702
1703\section{Mutual Exclusion / Synchronization}
1704\label{s:MutualExclusionSynchronization}
1705
1706Unrestricted nondeterminism is meaningless as there is no way to know when a result is completed and safe to access.
1707To produce meaningful execution requires clawing back some determinism using mutual exclusion and synchronization, where mutual exclusion provides access control for threads using shared data, and synchronization is a timing relationship among threads~\cite[\S~4]{Buhr05a}.
1708The shared data protected by mutual exclusion is called a \newterm{critical section}~\cite{Dijkstra65}, and the protection can be simple, only 1 thread, or complex, only N kinds of threads, \eg group~\cite{Joung00} or readers/writer~\cite{Courtois71} problems.
1709Without synchronization control in a critical section, an arriving thread can barge ahead of preexisting waiter threads resulting in short/long-term starvation, staleness and freshness problems, and incorrect transfer of data.
1710Preventing or detecting barging is a challenge with low-level locks, but made easier through higher-level constructs.
1711This challenge is often split into two different approaches: barging \emph{avoidance} and \emph{prevention}.
1712Approaches that unconditionally releasing a lock for competing threads to acquire must use barging avoidance with flag/counter variable(s) to force barging threads to wait;
1713approaches that conditionally hold locks during synchronization, \eg baton-passing~\cite{Andrews89}, prevent barging completely.
1714
1715At the lowest level, concurrent control is provided by atomic operations, upon which different kinds of locking mechanisms are constructed, \eg spin locks, semaphores~\cite{Dijkstra68b}, barriers, and path expressions~\cite{Campbell74}.
1716However, for productivity it is always desirable to use the highest-level construct that provides the necessary efficiency~\cite{Hochstein05}.
1717A significant challenge with locks is composability because it takes careful organization for multiple locks to be used while preventing deadlock.
1718Easing composability is another feature higher-level mutual-exclusion mechanisms can offer.
1719Some concurrent systems eliminate mutable shared-state by switching to non-shared communication like message passing~\cite{Thoth,Harmony,V-Kernel,MPI} (Erlang, MPI), channels~\cite{CSP} (CSP,Go), actors~\cite{Akka} (Akka, Scala), or functional techniques (Haskell).
1720However, these approaches introduce a new communication mechanism for concurrency different from the standard communication using function call/return.
1721Hence, a programmer must learn and manipulate two sets of design and programming patterns.
1722While this distinction can be hidden away in library code, effective use of the library still has to take both paradigms into account.
1723In contrast, approaches based on shared-state models more closely resemble the standard call and return programming model, resulting in a single programming paradigm.
1724Finally, a newer approach for restricting non-determinism is transactional memory~\cite{Herlihy93}.
1725While 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~\cite{Cascaval08,Boehm09} to be the main concurrency paradigm for system languages.
1726
1727
1728\section{Monitor}
1729\label{s:Monitor}
1730
1731One of the most natural, elegant, efficient, high-level mechanisms for mutual exclusion and synchronization for shared-memory systems is the \emph{monitor} (Table~\ref{t:ExecutionPropertyComposition} case 2).
1732First 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}.
1733In 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 manually implement a monitor.
1734For these reasons, \CFA selected monitors as the core high-level concurrency construct, upon which higher-level approaches can be easily constructed.
1735
1736Figure~\ref{f:AtomicCounter} compares a \CFA and Java monitor implementing an atomic counter.
1737(Like other concurrent programming languages, \CFA and Java have performant specializations for the basic types using atomic instructions.)
1738A \newterm{monitor} is a set of functions that ensure mutual exclusion when accessing shared state.
1739(Note, in \CFA, @monitor@ is short-hand for @mutex struct@.)
1740More precisely, a monitor is a programming technique that implicitly binds mutual exclusion to static function scope by call and return, as opposed to locks, where mutual exclusion is defined by acquire/release calls, independent of lexical context (analogous to block and heap storage allocation).
1741Restricting acquire and release points eases programming, comprehension, and maintenance, at a slight cost in flexibility and efficiency.
1742As for other special types, \CFA has a custom @monitor@ type.
1743
1744\begin{figure}
1745\centering
1746
1747\begin{lrbox}{\myboxA}
1748\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1749`monitor` Aint { // atomic integer counter
1750        int cnt;
1751};
1752int ++?( Aint & `mutex` this ) with(this) { return ++cnt; }
1753int ?=?( Aint & `mutex` lhs, int rhs ) with(lhs) { cnt = rhs; }
1754int ?=?(int & lhs, Aint & rhs) with(rhs) { lhs = cnt; }
1755
1756int i = 0, j = 0, k = 5;
1757Aint x = { 0 }, y = { 0 }, z = { 5 }; // no mutex
1758++x; ++y; ++z;     // mutex
1759x = 2; y = i; z = k;  // mutex
1760i = x; j = y; k = z;  // no mutex
1761\end{cfa}
1762\end{lrbox}
1763
1764\begin{lrbox}{\myboxB}
1765\begin{java}[aboveskip=0pt,belowskip=0pt]
1766class Aint {
1767    private int cnt;
1768    public Aint( int init ) { cnt = init; }
1769    `synchronized` public int inc() { return ++cnt; }
1770    `synchronized` public void set( int rhs ) {cnt=rhs;}
1771    public int get() { return cnt; }
1772}
1773int i = 0, j = 0, k = 5;
1774Aint x=new Aint(0), y=new Aint(0), z=new Aint(5);
1775x.inc(); y.inc(); z.inc();
1776x.set( 2 ); y.set( i ); z.set( k );
1777i = x.get(); j = y.get(); k = z.get();
1778\end{java}
1779\end{lrbox}
1780
1781\subfloat[\CFA]{\label{f:AtomicCounterCFA}\usebox\myboxA}
1782\hspace{3pt}
1783\vrule
1784\hspace{3pt}
1785\subfloat[Java]{\label{f:AtomicCounterJava}\usebox\myboxB}
1786\caption{Atomic counter}
1787\label{f:AtomicCounter}
1788\end{figure}
1789
1790Like Java, \CFA monitors have \newterm{multi-acquire} semantics so the thread in the monitor may acquire it multiple times without deadlock, allowing recursion and calling other interface functions.
1791% \begin{cfa}
1792% monitor M { ... } m;
1793% void foo( M & mutex m ) { ... } $\C{// acquire mutual exclusion}$
1794% void bar( M & mutex m ) { $\C{// acquire mutual exclusion}$
1795%       ... `bar( m );` ... `foo( m );` ... $\C{// reacquire mutual exclusion}$
1796% }
1797% \end{cfa}
1798\CFA monitors also ensure the monitor lock is released regardless of how an acquiring function ends, normal or exceptional, and returning a shared variable is safe via copying before the lock is released.
1799Similar safety is offered by \emph{explicit} opt-in disciplines like \CC RAII versus the monitor \emph{implicit} language-enforced safety guarantee ensuring no programmer usage errors.
1800However, RAII mechanisms cannot handle complex synchronization within a monitor, where the monitor lock may not be released on function exit because it is passed to an unblocking thread;
1801RAII is purely a mutual-exclusion mechanism (see Section~\ref{s:Scheduling}).
1802
1803Both Java and \CFA use a keyword @mutex@/\lstinline[language=java]|synchronized| to designate functions that implicitly acquire/release the monitor lock on call/return providing mutual exclusion to the stared data.
1804Non-designated functions provide no mutual exclusion for read-only access or as an interface to a multi-step protocol requiring several steps of acquiring and releasing the monitor.
1805Monitor objects can be passed through multiple helper functions without acquiring mutual exclusion, until a designated function associated with the object is called.
1806\CFA designated functions are marked by an explicitly parameter-only pointer/reference qualifier @mutex@ (discussed further in Section\ref{s:MutexAcquisition}).
1807Whereas, Java designated members are marked with \lstinline[language=java]|synchronized| that applies to the implicit reference parameter @this@.
1808In the example, the increment and setter operations need mutual exclusion while the read-only getter operation can be non-mutex if reading the implementation is atomic.
1809
1810
1811\subsection{Monitor Implementation}
1812
1813For the same design reasons, \CFA provides a custom @monitor@ type and a @trait@ to enforce and restrict the monitor-interface functions.
1814\begin{cquote}
1815\begin{tabular}{@{}c@{\hspace{3\parindentlnth}}c@{}}
1816\begin{cfa}
1817monitor M {
1818        ... // shared data
1819};
1820
1821\end{cfa}
1822&
1823\begin{cfa}
1824trait is_monitor( `dtype` T ) {
1825        monitor_desc * get_monitor( T & );
1826        void ^?{}( T & mutex );
1827};
1828\end{cfa}
1829\end{tabular}
1830\end{cquote}
1831The @dtype@ property prevents \emph{implicit} copy operations and the @is_monitor@ trait provides no \emph{explicit} copy operations, so monitors must be passed by reference or pointer.
1832Similarly, the function definitions ensure there is a mechanism to read the monitor descriptor from its handle, and a special destructor to prevent deallocation if a thread is using the shared data.
1833The custom monitor type also inserts any locks needed to implement the mutual exclusion semantics.
1834\CFA relies heavily on traits as an abstraction mechanism, so the @mutex@ qualifier prevents coincidentally matching of a monitor trait with a type that is not a monitor, similar to coincidental inheritance where a shape and playing card can both be drawable.
1835
1836
1837\subsection{Mutex Acquisition}
1838\label{s:MutexAcquisition}
1839
1840For object-oriented programming languages, the mutex property applies to one object, the implicit pointer/reference to the monitor type.
1841Because \CFA uses a pointer qualifier, other possibilities exist, \eg:
1842\begin{cfa}
1843monitor M { ... };
1844int f1( M & mutex m ); $\C{// single parameter object}$
1845int f2( M * mutex m ); $\C{// single or multiple parameter object}$
1846int f3( M * mutex m[$\,$] ); $\C{// multiple parameter object}$
1847int f4( stack( M * ) & mutex m ); $\C{// multiple parameters object}$
1848\end{cfa}
1849Function @f1@ has a single object parameter, while functions @f2@ to @f4@ can be a single or multi-element parameter with statically unknown size.
1850Because of the statically unknown size, \CFA only supports a single reference @mutex@ parameter, @f1@.
1851
1852The \CFA @mutex@ qualifier does allow the ability to support multi-monitor functions,\footnote{
1853While object-oriented monitors can be extended with a mutex qualifier for multiple-monitor members, no prior example of this feature could be found.}
1854where the number of acquisitions is statically known, called \newterm{bulk acquire}.
1855\CFA guarantees bulk acquisition order is consistent across calls to @mutex@ functions using the same monitors as arguments, so acquiring multiple monitors in a bulk acquire is safe from deadlock.
1856Figure~\ref{f:BankTransfer} shows a trivial solution to the bank transfer problem~\cite{BankTransfer}, where two resources must be locked simultaneously, using \CFA monitors with implicit locking and \CC with explicit locking.
1857A \CFA programmer only has to manage when to acquire mutual exclusion;
1858a \CC programmer must select the correct lock and acquisition mechanism from a panoply of locking options.
1859Making good choices for common cases in \CFA simplifies the programming experience and enhances safety.
1860
1861\begin{figure}
1862\centering
1863\begin{lrbox}{\myboxA}
1864\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1865monitor BankAccount {
1866
1867        int balance;
1868} b1 = { 0 }, b2 = { 0 };
1869void deposit( BankAccount & `mutex` b,
1870                        int deposit ) with(b) {
1871        balance += deposit;
1872}
1873void transfer( BankAccount & `mutex` my,
1874        BankAccount & `mutex` your, int me2you ) {
1875        // bulk acquire
1876        deposit( my, -me2you ); // debit
1877        deposit( your, me2you ); // credit
1878}
1879`thread` Person { BankAccount & b1, & b2; };
1880void main( Person & person ) with(person) {
1881        for ( 10_000_000 ) {
1882                if ( random() % 3 ) deposit( b1, 3 );
1883                if ( random() % 3 ) transfer( b1, b2, 7 );
1884        }
1885}
1886int main() {
1887        `Person p1 = { b1, b2 }, p2 = { b2, b1 };`
1888
1889} // wait for threads to complete
1890\end{cfa}
1891\end{lrbox}
1892
1893\begin{lrbox}{\myboxB}
1894\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1895struct BankAccount {
1896        `recursive_mutex m;`
1897        int balance = 0;
1898} b1, b2;
1899void deposit( BankAccount & b, int deposit ) {
1900        `scoped_lock lock( b.m );`
1901        b.balance += deposit;
1902}
1903void transfer( BankAccount & my,
1904                        BankAccount & your, int me2you ) {
1905        `scoped_lock lock( my.m, your.m );` // bulk acquire
1906        deposit( my, -me2you ); // debit
1907        deposit( your, me2you ); // credit
1908}
1909
1910void person( BankAccount & b1, BankAccount & b2 ) {
1911        for ( int i = 0; i < 10$'$000$'$000; i += 1 ) {
1912                if ( random() % 3 ) deposit( b1, 3 );
1913                if ( random() % 3 ) transfer( b1, b2, 7 );
1914        }
1915}
1916int main() {
1917        `thread p1(person, ref(b1), ref(b2)), p2(person, ref(b2), ref(b1));`
1918        `p1.join(); p2.join();`
1919}
1920\end{cfa}
1921\end{lrbox}
1922
1923\subfloat[\CFA]{\label{f:CFABank}\usebox\myboxA}
1924\hspace{3pt}
1925\vrule
1926\hspace{3pt}
1927\subfloat[\CC]{\label{f:C++Bank}\usebox\myboxB}
1928\hspace{3pt}
1929\caption{Bank transfer problem}
1930\label{f:BankTransfer}
1931\end{figure}
1932
1933Users can still force the acquiring order by using or not using @mutex@.
1934\begin{cfa}
1935void foo( M & mutex m1, M & mutex m2 ); $\C{// acquire m1 and m2}$
1936void bar( M & mutex m1, M & m2 ) { $\C{// only acquire m1}$
1937        ... foo( m1, m2 ); ... $\C{// acquire m2}$
1938}
1939void baz( M & m1, M & mutex m2 ) { $\C{// only acquire m2}$
1940        ... foo( m1, m2 ); ... $\C{// acquire m1}$
1941}
1942\end{cfa}
1943The bulk-acquire semantics allow @bar@ or @baz@ to acquire a monitor lock and reacquire it in @foo@.
1944The calls to @bar@ and @baz@ acquired the monitors in opposite order, possibly resulting in deadlock.
1945However, this case is the simplest instance of the \emph{nested-monitor problem}~\cite{Lister77}, where monitors are acquired in sequence versus bulk.
1946Detecting the nested-monitor problem requires dynamic tracking of monitor calls, and dealing with it requires rollback semantics~\cite{Dice10}.
1947\CFA does not deal with this fundamental problem.
1948
1949Finally, like Java, \CFA offers an alternative @mutex@ statement to reduce refactoring and naming.
1950\begin{cquote}
1951\renewcommand{\arraystretch}{0.0}
1952\begin{tabular}{@{}l@{\hspace{3\parindentlnth}}l@{}}
1953\multicolumn{1}{c}{\textbf{\lstinline@mutex@ call}} & \multicolumn{1}{c}{\lstinline@mutex@ \textbf{statement}} \\
1954\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1955monitor M { ... };
1956void foo( M & mutex m1, M & mutex m2 ) {
1957        // critical section
1958}
1959void bar( M & m1, M & m2 ) {
1960        foo( m1, m2 );
1961}
1962\end{cfa}
1963&
1964\begin{cfa}[aboveskip=0pt,belowskip=0pt]
1965
1966void bar( M & m1, M & m2 ) {
1967        mutex( m1, m2 ) {       // remove refactoring and naming
1968                // critical section
1969        }
1970}
1971
1972\end{cfa}
1973\end{tabular}
1974\end{cquote}
1975
1976
1977\subsection{Scheduling}
1978\label{s:Scheduling}
1979
1980% There are many aspects of scheduling in a concurrency system, all related to resource utilization by waiting threads, \ie which thread gets the resource next.
1981% Different forms of scheduling include access to processors by threads (see Section~\ref{s:RuntimeStructureCluster}), another is access to a shared resource by a lock or monitor.
1982This section discusses scheduling for waiting threads eligible for monitor entry~\cite{Buhr95b}, \ie which user thread gets the shared resource next. (See Section~\ref{s:RuntimeStructureCluster} for scheduling kernel threads on virtual processors.)
1983While monitor mutual-exclusion provides safe access to its shared data, the data may indicate a thread cannot proceed, \eg a bounded buffer may be full/\-empty so produce/consumer threads must block.
1984Leaving the monitor and retrying (busy waiting) is impractical for high-level programming.
1985
1986Monitors eliminate busy waiting by providing synchronization within the monitor critical-section to schedule threads needing access to the shared data, where threads block versus spin.
1987Synchronization is generally achieved with internal~\cite{Hoare74} or external~\cite[\S~2.9.2]{uC++} scheduling.
1988\newterm{Internal} largely schedules threads located \emph{inside} the monitor and is accomplished using condition variables with signal and wait.
1989\newterm{External} largely schedules threads located \emph{outside} the monitor and is accomplished with the @waitfor@ statement.
1990Note, internal scheduling has a small amount of external scheduling and vice versa, so the naming denotes where the majority of the block threads reside (inside or outside) for scheduling.
1991For complex scheduling, the approaches can be combined, so there are threads waiting inside and outside.
1992
1993\CFA monitors do not allow calling threads to barge ahead of signalled threads via barging prevention, which simplifies synchronization among threads in the monitor and increases correctness.
1994A direct consequence of this semantics is that unblocked waiting threads are not required to recheck the waiting condition, \ie waits are not in a starvation-prone busy-loop as required by the signals-as-hints style with barging.
1995Preventing barging comes directly from Hoare's semantics in the seminal paper on monitors~\cite[p.~550]{Hoare74}.
1996% \begin{cquote}
1997% However, 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.
1998% It 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}
1999% \end{cquote}
2000Furthermore, \CFA concurrency has no spurious wakeup~\cite[\S~9]{Buhr05a}, which eliminates an implicit self barging.
2001
2002Monitor mutual-exclusion means signalling cannot have the signaller and signalled thread in the monitor simultaneously, so only the signaller or signallee can proceed and the other waits on an implicit urgent list~\cite[p.~551]{Hoare74}.
2003Figure~\ref{f:MonitorScheduling} shows internal and external scheduling for the bounded-buffer examples in Figure~\ref{f:GenericBoundedBuffer}.
2004For internal scheduling in Figure~\ref{f:BBInt}, the @signal@ moves the signallee, front thread of the specified condition queue, to the urgent list (see Figure~\ref{f:MonitorScheduling}) and the signaller continues (solid line).
2005Multiple signals move multiple signallees to urgent until the condition queue is empty.
2006When the signaller exits or waits, a thread is implicitly unblocked from urgent, if available, before unblocking a calling thread to prevent barging.
2007(Java conceptually moves the signalled thread to the calling queue, and hence, allows barging.)
2008Signal is used when the signaller is providing the cooperation needed by the signallee, \eg creating an empty slot in a buffer for a producer, and the signaller immediately exits the monitor to run concurrently consuming the buffer element, and passes control of the monitor to the signalled thread, which can immediately take advantage of the state change.
2009Specifically, the @wait@ function atomically blocks the calling thread and implicitly releases the monitor lock(s) for all monitors in the function's parameter list.
2010Signalling is unconditional because signalling an empty condition queue does nothing.
2011It is common to declare condition queues as monitor fields to prevent shared access, hence no locking is required for access as the queues are protected by the monitor lock.
2012In \CFA, a condition queue can be created and stored independently.
2013
2014\begin{figure}
2015\centering
2016% \subfloat[Scheduling Statements] {
2017% \label{fig:SchedulingStatements}
2018% {\resizebox{0.45\textwidth}{!}{\input{CondSigWait.pstex_t}}}
2019\input{CondSigWait.pstex_t}
2020% }% subfloat
2021% \quad
2022% \subfloat[Bulk acquire monitor] {
2023% \label{fig:BulkMonitor}
2024% {\resizebox{0.45\textwidth}{!}{\input{ext_monitor.pstex_t}}}
2025% }% subfloat
2026\caption{Monitor Scheduling}
2027\label{f:MonitorScheduling}
2028\end{figure}
2029
2030\begin{figure}
2031\centering
2032\newbox\myboxA
2033\begin{lrbox}{\myboxA}
2034\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2035forall( otype T ) { // distribute forall
2036        monitor Buffer {
2037                `condition` full, empty;
2038                int front, back, count;
2039                T elements[10];
2040        };
2041        void ?{}( Buffer(T) & buf ) with(buf) {
2042                front = back = count = 0;
2043        }
2044
2045        void insert(Buffer(T) & mutex buf, T elm) with(buf){
2046                if ( count == 10 ) `wait( empty )`; // full ?
2047                // insert elm into buf
2048                `signal( full )`;
2049        }
2050        T remove( Buffer(T) & mutex buf ) with(buf) {
2051                if ( count == 0 ) `wait( full )`; // empty ?
2052                // remove elm from buf
2053                `signal( empty )`;
2054                return elm;
2055        }
2056}
2057\end{cfa}
2058\end{lrbox}
2059
2060\newbox\myboxB
2061\begin{lrbox}{\myboxB}
2062\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2063forall( otype T ) { // distribute forall
2064        monitor Buffer {
2065
2066                int front, back, count;
2067                T elements[10];
2068        };
2069        void ?{}( Buffer(T) & buf ) with(buf) {
2070                front = back = count = 0;
2071        }
2072        T remove( Buffer(T) & mutex buf ); // forward
2073        void insert(Buffer(T) & mutex buf, T elm) with(buf){
2074                if ( count == 10 ) `waitfor( remove : buf )`;
2075                // insert elm into buf
2076
2077        }
2078        T remove( Buffer(T) & mutex buf ) with(buf) {
2079                if ( count == 0 ) `waitfor( insert : buf )`;
2080                // remove elm from buf
2081
2082                return elm;
2083        }
2084}
2085\end{cfa}
2086\end{lrbox}
2087
2088\subfloat[Internal scheduling]{\label{f:BBInt}\usebox\myboxA}
2089\hspace{1pt}
2090\vrule
2091\hspace{3pt}
2092\subfloat[External scheduling]{\label{f:BBExt}\usebox\myboxB}
2093
2094\caption{Generic bounded buffer}
2095\label{f:GenericBoundedBuffer}
2096\end{figure}
2097
2098The @signal_block@ provides the opposite unblocking order, where the signaller is moved to urgent and the signallee continues and a thread is implicitly unblocked from urgent when the signallee exits or waits (dashed line)~\cite[p.~551]{Hoare74}.
2099Signal block is used when the signallee is providing the cooperation needed by the signaller, \eg if the buffer is removed and a producer hands off an item to a consumer as in Figure~\ref{f:DatingSignalBlock}, so the signaller must wait until the signallee unblocks, provides the cooperation, exits the monitor to run concurrently, and passes control of the monitor to the signaller, which can immediately take advantage of the state change.
2100Using @signal@ or @signal_block@ can be a dynamic decision based on whether the thread providing the cooperation arrives before or after the thread needing the cooperation.
2101
2102For external scheduling in Figure~\ref{f:BBExt}, the internal scheduling is replaced, eliminating condition queues and @signal@/@wait@ (cases where it cannot are discussed shortly), and has existed in the programming language Ada for almost 40 years with variants in other languages~\cite{SR,ConcurrentC++,uC++}.
2103While prior languages use external scheduling solely for thread interaction, \CFA generalizes it to both monitors and threads.
2104External scheduling allows waiting for events from other threads while restricting unrelated events, that would otherwise have to wait on condition queues in the monitor.
2105Scheduling is controlled by the @waitfor@ statement, which atomically blocks the calling thread, releases the monitor lock, and restricts the function calls that can next acquire mutual exclusion.
2106Specifically, a thread calling the monitor is unblocked directly from the calling queue based on function names that can fulfill the cooperation required by the signaller.
2107(The linear search through the calling queue to locate a particular call can be reduced to $O(1)$.)
2108Hence, the @waitfor@ has the same semantics as @signal_block@, where the signallee thread from the calling queue executes before the signaller, which waits on urgent.
2109Now when a producer/consumer detects a full/empty buffer, the necessary cooperation for continuation is specified by indicating the next function call that can occur.
2110For example, a producer detecting a full buffer must have cooperation from a consumer to remove an item so function @remove@ is accepted, which prevents producers from entering the monitor, and after a consumer calls @remove@, the producer waiting on urgent is \emph{implicitly} unblocked because it can now continue its insert operation.
2111Hence, this mechanism is done in terms of control flow, next call, versus in terms of data, channels, as in Go and Rust @select@.
2112While both mechanisms have strengths and weaknesses, \CFA uses the control-flow mechanism to be consistent with other language features.
2113
2114Figure~\ref{f:ReadersWriterLock} shows internal and external scheduling for a readers/writer lock with no barging and threads are serviced in FIFO order to eliminate staleness and freshness among the reader/writer threads.
2115For internal scheduling in Figure~\ref{f:RWInt}, the readers and writers wait on the same condition queue in FIFO order, making it impossible to tell if a waiting thread is a reader or writer.
2116To clawback the kind of thread, a \CFA condition can store user data in the node for a blocking thread at the @wait@, \ie whether the thread is a @READER@ or @WRITER@.
2117An unblocked reader thread checks if the thread at the front of the queue is a reader and unblock it, \ie the readers daisy-chain signal the next group of readers demarcated by the next writer or end of the queue.
2118For external scheduling in Figure~\ref{f:RWExt}, a waiting reader checks if a writer is using the resource, and if so, restricts further calls until the writer exits by calling @EndWrite@.
2119The writer does a similar action for each reader or writer using the resource.
2120Note, no new calls to @StartRead@/@StartWrite@ may occur when waiting for the call to @EndRead@/@EndWrite@.
2121
2122\begin{figure}
2123\centering
2124\newbox\myboxA
2125\begin{lrbox}{\myboxA}
2126\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2127enum RW { READER, WRITER };
2128monitor ReadersWriter {
2129        int rcnt, wcnt; // readers/writer using resource
2130        `condition RWers;`
2131};
2132void ?{}( ReadersWriter & rw ) with(rw) {
2133        rcnt = wcnt = 0;
2134}
2135void EndRead( ReadersWriter & mutex rw ) with(rw) {
2136        rcnt -= 1;
2137        if ( rcnt == 0 ) `signal( RWers )`;
2138}
2139void EndWrite( ReadersWriter & mutex rw ) with(rw) {
2140        wcnt = 0;
2141        `signal( RWers );`
2142}
2143void StartRead( ReadersWriter & mutex rw ) with(rw) {
2144        if ( wcnt !=0 || ! empty( RWers ) )
2145                `wait( RWers, READER )`;
2146        rcnt += 1;
2147        if ( ! empty(RWers) && `front(RWers) == READER` )
2148                `signal( RWers )`;  // daisy-chain signalling
2149}
2150void StartWrite( ReadersWriter & mutex rw ) with(rw) {
2151        if ( wcnt != 0 || rcnt != 0 ) `wait( RWers, WRITER )`;
2152
2153        wcnt = 1;
2154}
2155\end{cfa}
2156\end{lrbox}
2157
2158\newbox\myboxB
2159\begin{lrbox}{\myboxB}
2160\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2161
2162monitor ReadersWriter {
2163        int rcnt, wcnt; // readers/writer using resource
2164
2165};
2166void ?{}( ReadersWriter & rw ) with(rw) {
2167        rcnt = wcnt = 0;
2168}
2169void EndRead( ReadersWriter & mutex rw ) with(rw) {
2170        rcnt -= 1;
2171
2172}
2173void EndWrite( ReadersWriter & mutex rw ) with(rw) {
2174        wcnt = 0;
2175
2176}
2177void StartRead( ReadersWriter & mutex rw ) with(rw) {
2178        if ( wcnt > 0 ) `waitfor( EndWrite : rw );`
2179
2180        rcnt += 1;
2181
2182
2183}
2184void StartWrite( ReadersWriter & mutex rw ) with(rw) {
2185        if ( wcnt > 0 ) `waitfor( EndWrite : rw );`
2186        else while ( rcnt > 0 ) `waitfor( EndRead : rw );`
2187        wcnt = 1;
2188}
2189\end{cfa}
2190\end{lrbox}
2191
2192\subfloat[Internal scheduling]{\label{f:RWInt}\usebox\myboxA}
2193\hspace{1pt}
2194\vrule
2195\hspace{3pt}
2196\subfloat[External scheduling]{\label{f:RWExt}\usebox\myboxB}
2197
2198\caption{Readers / writer lock}
2199\label{f:ReadersWriterLock}
2200\end{figure}
2201
2202Finally, external scheduling requires urgent to be a stack, because the signaller expects to execute immediately after the specified monitor call has exited or waited.
2203Internal schedulling performing multiple signalling results in unblocking from urgent in the reverse order from signalling.
2204It is rare for the unblocking order to be important as an unblocked thread can be time-sliced immediately after leaving the monitor.
2205If the unblocking order is important, multiple signalling can be restructured into daisy-chain signalling, where each thread signals the next thread.
2206Hence, \CFA uses a single urgent stack to correctly handle @waitfor@ and adequately support both forms of signalling.
2207(Advanced @waitfor@ features are discussed in Section~\ref{s:ExtendedWaitfor}.)
2208
2209\begin{figure}
2210\centering
2211\newbox\myboxA
2212\begin{lrbox}{\myboxA}
2213\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2214enum { CCodes = 20 };
2215monitor DS {
2216        int GirlPhNo, BoyPhNo;
2217        condition Girls[CCodes], Boys[CCodes];
2218        `condition exchange;`
2219};
2220int girl( DS & mutex ds, int phNo, int ccode ) {
2221        if ( empty( Boys[ccode] ) ) {
2222                wait( Girls[ccode] );
2223                GirlPhNo = phNo;
2224                `signal( exchange );`
2225        } else {
2226                GirlPhNo = phNo;
2227                `signal( Boys[ccode] );`
2228                `wait( exchange );`
2229        }
2230        return BoyPhNo;
2231}
2232int boy( DS & mutex ds, int phNo, int ccode ) {
2233        // as above with boy/girl interchanged
2234}
2235\end{cfa}
2236\end{lrbox}
2237
2238\newbox\myboxB
2239\begin{lrbox}{\myboxB}
2240\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2241
2242monitor DS {
2243        int GirlPhNo, BoyPhNo;
2244        condition Girls[CCodes], Boys[CCodes];
2245
2246};
2247int girl( DS & mutex ds, int phNo, int ccode ) {
2248        if ( empty( Boys[ccode] ) ) { // no compatible
2249                wait( Girls[ccode] ); // wait for boy
2250                GirlPhNo = phNo; // make phone number available
2251
2252        } else {
2253                GirlPhNo = phNo; // make phone number available
2254                `signal_block( Boys[ccode] );` // restart boy
2255
2256        } // if
2257        return BoyPhNo;
2258}
2259int boy( DS & mutex ds, int phNo, int ccode ) {
2260        // as above with boy/girl interchanged
2261}
2262\end{cfa}
2263\end{lrbox}
2264
2265\subfloat[\lstinline@signal@]{\label{f:DatingSignal}\usebox\myboxA}
2266\qquad
2267\subfloat[\lstinline@signal_block@]{\label{f:DatingSignalBlock}\usebox\myboxB}
2268\caption{Dating service Monitor}
2269\label{f:DatingServiceMonitor}
2270\end{figure}
2271
2272Figure~\ref{f:DatingServiceMonitor} shows a dating service demonstrating non-blocking and blocking signalling.
2273The dating service matches girl and boy threads with matching compatibility codes so they can exchange phone numbers.
2274A thread blocks until an appropriate partner arrives.
2275The complexity is exchanging phone numbers in the monitor because of the mutual-exclusion property.
2276For 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.
2277For signal-block scheduling, the implicit urgent-queue replaces the explicit @exchange@-condition and @signal_block@ puts the finding thread on the urgent stack and unblocks the matcher.
2278Note, barging corrupts the dating service during an exchange because a barger may also match and change the phone numbers, invalidating the previous exchange phone number.
2279This situation shows rechecking the waiting condition and waiting again (signals-as-hints) fails, requiring significant restructured to account for barging.
2280
2281Given external and internal scheduling, what guidelines can a programmer use to select between them?
2282In general, external scheduling is easier to understand and code because only the next logical action (mutex function(s)) is stated, and the monitor implicitly handles all the details.
2283Therefore, there are no condition variables, and hence, no wait and signal, which reduces coding complexity and synchronization errors.
2284If external scheduling is simpler than internal, why not use it all the time?
2285Unfortunately, external scheduling cannot be used if: scheduling depends on parameter value(s) or scheduling must block across an unknown series of calls on a condition variable, \ie internal scheduling.
2286For example, the dating service cannot be written using external scheduling.
2287First, scheduling requires knowledge of calling parameters to make matching decisions and parameters of calling threads are unavailable within the monitor.
2288Specifically, a thread within the monitor cannot examine the @ccode@ of threads waiting on the calling queue to determine if there is a matching partner.
2289(Similarly, if the bounded buffer or readers/writer are restructured with a single interface function with a parameter denoting producer/consumer or reader/write, they cannot be solved with external scheduling.)
2290Second, a scheduling decision may be delayed across an unknown number of calls when there is no immediate match so the thread in the monitor must block on a condition.
2291Specifically, if a thread determines there is no opposite calling thread with the same @ccode@, it must wait an unknown period until a matching thread arrives.
2292For complex synchronization, both external and internal scheduling can be used to take advantage of best of properties of each.
2293
2294Finally, both internal and external scheduling extend to multiple monitors in a natural way.
2295\begin{cquote}
2296\begin{tabular}{@{}l@{\hspace{2\parindentlnth}}l@{}}
2297\begin{cfa}
2298monitor M { `condition e`; ... };
2299void foo( M & mutex m1, M & mutex m2 ) {
2300        ... wait( `e` ); ...   // wait( e, m1, m2 )
2301        ... wait( `e, m1` ); ...
2302        ... wait( `e, m2` ); ...
2303}
2304\end{cfa}
2305&
2306\begin{cfa}
2307void rtn$\(_1\)$( M & mutex m1, M & mutex m2 ); // overload rtn
2308void rtn$\(_2\)$( M & mutex m1 );
2309void bar( M & mutex m1, M & mutex m2 ) {
2310        ... waitfor( `rtn`${\color{red}\(_1\)}$ ); ...       // $\LstCommentStyle{waitfor( rtn\(_1\) : m1, m2 )}$
2311        ... waitfor( `rtn${\color{red}\(_2\)}$ : m1` ); ...
2312}
2313\end{cfa}
2314\end{tabular}
2315\end{cquote}
2316For @wait( e )@, the default semantics is to atomically block the signaller and release all acquired mutex parameters, \ie @wait( e, m1, m2 )@.
2317To override the implicit multi-monitor wait, specific mutex parameter(s) can be specified, \eg @wait( e, m1 )@.
2318Wait cannot statically verify the released monitors are the acquired mutex-parameters without disallowing separately compiled helper functions calling @wait@.
2319While \CC supports bulk locking, @wait@ only accepts a single lock for a condition queue, so bulk locking with condition queues is asymmetric.
2320Finally, a signaller,
2321\begin{cfa}
2322void baz( M & mutex m1, M & mutex m2 ) {
2323        ... signal( e ); ...
2324}
2325\end{cfa}
2326must have acquired at least the same locks as the waiting thread signalled from a condition queue to allow the locks to be passed, and hence, prevent barging.
2327
2328Similarly, for @waitfor( rtn )@, the default semantics is to atomically block the acceptor and release all acquired mutex parameters, \ie @waitfor( rtn : m1, m2 )@.
2329To override the implicit multi-monitor wait, specific mutex parameter(s) can be specified, \eg @waitfor( rtn : m1 )@.
2330@waitfor@ does statically verify the monitor types passed are the same as the acquired mutex-parameters of the given function or function pointer, hence the prototype must be accessible.
2331% When an overloaded function appears in an @waitfor@ statement, calls to any function with that name are accepted.
2332% The rationale is that functions with the same name should perform a similar actions, and therefore, all should be eligible to accept a call.
2333Overloaded functions can be disambiguated using a cast
2334\begin{cfa}
2335void rtn( M & mutex m );
2336`int` rtn( M & mutex m );
2337waitfor( (`int` (*)( M & mutex ))rtn : m );
2338\end{cfa}
2339
2340The ability to release a subset of acquired monitors can result in a \newterm{nested monitor}~\cite{Lister77} deadlock (see Section~\ref{s:MutexAcquisition}).
2341\begin{cfa}
2342void foo( M & mutex m1, M & mutex m2 ) {
2343        ... wait( `e, m1` ); ...                                $\C{// release m1, keeping m2 acquired}$
2344void bar( M & mutex m1, M & mutex m2 ) {        $\C{// must acquire m1 and m2}$
2345        ... signal( `e` ); ...
2346\end{cfa}
2347The @wait@ only releases @m1@ so the signalling thread cannot acquire @m1@ and @m2@ to enter @bar@ and @signal@ the condition.
2348While deadlock can occur with multiple/nesting acquisition, this is a consequence of locks, and by extension monitor locking is not perfectly composable.
2349
2350
2351\subsection{\texorpdfstring{Extended \protect\lstinline@waitfor@}{Extended waitfor}}
2352\label{s:ExtendedWaitfor}
2353
2354Figure~\ref{f:ExtendedWaitfor} shows the extended form of the @waitfor@ statement to conditionally accept one of a group of mutex functions, with an optional statement to be performed \emph{after} the mutex function finishes.
2355For a @waitfor@ clause to be executed, its @when@ must be true and an outstanding call to its corresponding function(s) must exist.
2356The \emph{conditional-expression} of a @when@ may call a function, but the function must not block or context switch.
2357If there are multiple acceptable mutex calls, selection is prioritized top-to-bottom among the @waitfor@ clauses, whereas some programming languages with similar mechanisms accept nondeterministically for this case, \eg Go \lstinline[morekeywords=select]@select@.
2358If some accept guards are true and there are no outstanding calls to these functions, the acceptor is blocked until a call to one of these functions is made.
2359If there is a @timeout@ clause, it provides an upper bound on waiting.
2360If 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.
2361Hence, the terminating @else@ clause allows a conditional attempt to accept a call without blocking.
2362If both @timeout@ and @else@ clause are present, the @else@ must be conditional, or the @timeout@ is never triggered.
2363% There is also a traditional future wait queue (not shown) (\eg Microsoft @WaitForMultipleObjects@), to wait for a specified number of future elements in the queue.
2364Finally, there is a shorthand for specifying multiple functions using the same set of monitors: @waitfor( f, g, h : m1, m2, m3 )@.
2365
2366\begin{figure}
2367\centering
2368\begin{cfa}
2369`when` ( $\emph{conditional-expression}$ )      $\C{// optional guard}$
2370        waitfor( $\emph{mutex-function-name}$ ) $\emph{statement}$ $\C{// action after call}$
2371`or` `when` ( $\emph{conditional-expression}$ ) $\C{// any number of functions}$
2372        waitfor( $\emph{mutex-function-name}$ ) $\emph{statement}$
2373`or`    ...
2374`when` ( $\emph{conditional-expression}$ ) $\C{// optional guard}$
2375        `timeout` $\emph{statement}$ $\C{// optional terminating timeout clause}$
2376`when` ( $\emph{conditional-expression}$ ) $\C{// optional guard}$
2377        `else`  $\emph{statement}$ $\C{// optional terminating clause}$
2378\end{cfa}
2379\caption{Extended \protect\lstinline@waitfor@}
2380\label{f:ExtendedWaitfor}
2381\end{figure}
2382
2383Note, a group of conditional @waitfor@ clauses is \emph{not} the same as a group of @if@ statements, \eg:
2384\begin{cfa}
2385if ( C1 ) waitfor( mem1 );                       when ( C1 ) waitfor( mem1 );
2386else if ( C2 ) waitfor( mem2 );         or when ( C2 ) waitfor( mem2 );
2387\end{cfa}
2388The left example only accepts @mem1@ if @C1@ is true or only @mem2@ if @C2@ is true.
2389The right example accepts either @mem1@ or @mem2@ if @C1@ and @C2@ are true.
2390Hence, the @waitfor@ has parallel semantics, accepting any true @when@ clause.
2391
2392An interesting use of @waitfor@ is accepting the @mutex@ destructor to know when an object is deallocated, \eg assume the bounded buffer is restructured from a monitor to a thread with the following @main@.
2393\begin{cfa}
2394void main( Buffer(T) & buffer ) with(buffer) {
2395        for () {
2396                `waitfor( ^?{} : buffer )` break;
2397                or when ( count != 20 ) waitfor( insert : buffer ) { ... }
2398                or when ( count != 0 ) waitfor( remove : buffer ) { ... }
2399        }
2400        // clean up
2401}
2402\end{cfa}
2403When the program main deallocates the buffer, it first calls the buffer's destructor, which is accepted, the destructor runs, and the buffer is deallocated.
2404However, the buffer thread cannot continue after the destructor call because the object is gone;
2405hence, clean up in @main@ cannot occur, which means destructors for local objects are not run.
2406To make this useful capability work, the semantics for accepting the destructor is the same as @signal@, \ie the destructor call is placed on urgent and the acceptor continues execution, which ends the loop, cleans up, and the thread terminates.
2407Then, the destructor caller unblocks from urgent to deallocate the object.
2408Accepting the destructor is the idiomatic way in \CFA to terminate a thread performing direct communication.
2409
2410
2411\subsection{Bulk Barging Prevention}
2412
2413Figure~\ref{f:BulkBargingPrevention} shows \CFA code where bulk acquire adds complexity to the internal-signalling semantics.
2414The complexity begins at the end of the inner @mutex@ statement, where the semantics of internal scheduling need to be extended for multiple monitors.
2415The problem is that bulk acquire is used in the inner @mutex@ statement where one of the monitors is already acquired.
2416When 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.
2417However, both the signalling and waiting threads W1 and W2 need some subset of monitors @m1@ and @m2@.
2418\begin{cquote}
2419condition c: (order 1) W2(@m2@), W1(@m1@,@m2@)\ \ \ or\ \ \ (order 2) W1(@m1@,@m2@), W2(@m2@) \\
2420S: acq. @m1@ $\rightarrow$ acq. @m1,m2@ $\rightarrow$ @signal(c)@ $\rightarrow$ rel. @m2@ $\rightarrow$ pass @m2@ unblock W2 (order 2) $\rightarrow$ rel. @m1@ $\rightarrow$ pass @m1,m2@ unblock W1 \\
2421\hspace*{2.75in}$\rightarrow$ rel. @m1@ $\rightarrow$ pass @m1,m2@ unblock W1 (order 1)
2422\end{cquote}
2423
2424\begin{figure}
2425\newbox\myboxA
2426\begin{lrbox}{\myboxA}
2427\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2428monitor M m1, m2;
2429condition c;
2430mutex( m1 ) { // $\LstCommentStyle{\color{red}outer}$
2431        ...
2432        mutex( m1, m2 ) { // $\LstCommentStyle{\color{red}inner}$
2433                ... `signal( c )`; ...
2434                // m1, m2 still acquired
2435        } // $\LstCommentStyle{\color{red}release m2}$
2436        // m1 acquired
2437} // release m1
2438\end{cfa}
2439\end{lrbox}
2440
2441\newbox\myboxB
2442\begin{lrbox}{\myboxB}
2443\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2444
2445
2446mutex( m1 ) {
2447        ...
2448        mutex( m1, m2 ) {
2449                ... `wait( c )`; // release m1, m2
2450                // m1, m2 reacquired
2451        } // $\LstCommentStyle{\color{red}release m2}$
2452        // m1 acquired
2453} // release m1
2454\end{cfa}
2455\end{lrbox}
2456
2457\newbox\myboxC
2458\begin{lrbox}{\myboxC}
2459\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2460
2461
2462mutex( m2 ) {
2463        ... `wait( c )`; // release m2
2464        // m2 reacquired
2465} // $\LstCommentStyle{\color{red}release m2}$
2466
2467
2468
2469
2470\end{cfa}
2471\end{lrbox}
2472
2473\begin{cquote}
2474\subfloat[Signalling Thread (S)]{\label{f:SignallingThread}\usebox\myboxA}
2475\hspace{3\parindentlnth}
2476\subfloat[Waiting Thread (W1)]{\label{f:WaitingThread}\usebox\myboxB}
2477\hspace{2\parindentlnth}
2478\subfloat[Waiting Thread (W2)]{\label{f:OtherWaitingThread}\usebox\myboxC}
2479\end{cquote}
2480\caption{Bulk Barging Prevention}
2481\label{f:BulkBargingPrevention}
2482\end{figure}
2483
2484One scheduling solution is for the signaller S 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.
2485However, this solution is inefficient if W2 waited first and immediate passed @m2@ when released, while S retains @m1@ until completion of the outer mutex statement.
2486If W1 waited first, the signaller must retain @m1@ amd @m2@ until completion of the outer mutex statement and then pass both to W1.
2487% Furthermore, there is an execution sequence where the signaller always finds waiter W2, and hence, waiter W1 starves.
2488To support these efficient semantics and prevent barging, the implementation maintains a list of monitors acquired for each blocked thread.
2489When a signaller exits or waits in a mutex function or statement, the front waiter on urgent is unblocked if all its monitors are released.
2490Implementing a fast subset check for the necessarily released monitors is important and discussed in the following sections.
2491% The 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.
2492
2493
2494\subsection{\texorpdfstring{\protect\lstinline@waitfor@ Implementation}{waitfor Implementation}}
2495\label{s:waitforImplementation}
2496
2497In a statically-typed object-oriented programming language, a class has an exhaustive list of members, even when members are added via static inheritance (see Figure~\ref{f:uCinheritance}).
2498Knowing all members at compilation, even separate compilation, allows uniquely numbered them so the accept-statement implementation can use a fast and compact bit mask with $O(1)$ compare.
2499
2500\begin{figure}
2501\centering
2502\begin{lrbox}{\myboxA}
2503\begin{uC++}[aboveskip=0pt,belowskip=0pt]
2504$\emph{translation unit 1}$
2505_Monitor B { // common type in .h file
2506        _Mutex virtual void `f`( ... );
2507        _Mutex virtual void `g`( ... );
2508        _Mutex virtual void w1( ... ) { ... _Accept(`f`, `g`); ... }
2509};
2510$\emph{translation unit 2}$
2511// include B
2512_Monitor D : public B { // inherit
2513        _Mutex void `h`( ... ); // add
2514        _Mutex void w2( ... ) { ... _Accept(`f`, `h`); ... }
2515};
2516\end{uC++}
2517\end{lrbox}
2518
2519\begin{lrbox}{\myboxB}
2520\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2521$\emph{translation unit 1}$
2522monitor M { ... }; // common type in .h file
2523void `f`( M & mutex m, ... );
2524void `g`( M & mutex m, ... );
2525void w1( M & mutex m, ... ) { ... waitfor(`f`, `g` : m); ... }
2526
2527$\emph{translation unit 2}$
2528// include M
2529extern void `f`( M & mutex m, ... ); // import f but not g
2530void `h`( M & mutex m ); // add
2531void w2( M & mutex m, ... ) { ... waitfor(`f`, `h` : m); ... }
2532
2533\end{cfa}
2534\end{lrbox}
2535
2536\subfloat[\uC]{\label{f:uCinheritance}\usebox\myboxA}
2537\hspace{3pt}
2538\vrule
2539\hspace{3pt}
2540\subfloat[\CFA]{\label{f:CFinheritance}\usebox\myboxB}
2541\caption{Member / Function visibility}
2542\label{f:MemberFunctionVisibility}
2543\end{figure}
2544
2545However, the @waitfor@ statement in translation unit 2 (see Figure~\ref{f:CFinheritance}) cannot see function @g@ in translation unit 1 precluding a unique numbering for a bit-mask because the monitor type only carries the protected shared-data.
2546(A possible way to construct a dense mapping is at link or load-time.)
2547Hence, function pointers are used to identify the functions listed in the @waitfor@ statement, stored in a variable-sized array.
2548Then, the same implementation approach used for the urgent stack (see Section~\ref{s:Scheduling}) is used for the calling queue.
2549Each caller has a list of monitors acquired, and the @waitfor@ statement performs a short linear search matching functions in the @waitfor@ list with called functions, and then verifying the associated mutex locks can be transferred.
2550
2551
2552\subsection{Multi-Monitor Scheduling}
2553\label{s:Multi-MonitorScheduling}
2554
2555External scheduling, like internal scheduling, becomes significantly more complex for multi-monitor semantics.
2556Even in the simplest case, new semantics need to be established.
2557\begin{cfa}
2558monitor M { ... };
2559void f( M & mutex m1 );
2560void g( M & mutex m1, M & mutex m2 ) { `waitfor( f );` } $\C{// pass m1 or m2 to f?}$
2561\end{cfa}
2562The solution is for the programmer to disambiguate:
2563\begin{cfa}
2564waitfor( f : `m2` ); $\C{// wait for call to f with argument m2}$
2565\end{cfa}
2566Both locks are acquired by function @g@, so when function @f@ is called, the lock for monitor @m2@ is passed from @g@ to @f@, while @g@ still holds lock @m1@.
2567This behaviour can be extended to the multi-monitor @waitfor@ statement.
2568\begin{cfa}
2569monitor M { ... };
2570void f( M & mutex m1, M & mutex m2 );
2571void g( M & mutex m1, M & mutex m2 ) { waitfor( f : `m1, m2` ); $\C{// wait for call to f with arguments m1 and m2}$
2572\end{cfa}
2573Again, the set of monitors passed to the @waitfor@ statement must be entirely contained in the set of monitors already acquired by the accepting function.
2574% Also, the order of the monitors in a @waitfor@ statement must match the order of the mutex parameters.
2575
2576Figure~\ref{f:UnmatchedMutexSets} shows internal and external scheduling with multiple monitors that must match exactly with a signalling or accepting thread, \ie partial matching results in waiting.
2577In both cases, the set of monitors is disjoint so unblocking is impossible.
2578
2579\begin{figure}
2580\centering
2581\begin{lrbox}{\myboxA}
2582\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2583monitor M1 {} m11, m12;
2584monitor M2 {} m2;
2585condition c;
2586void f( M1 & mutex m1, M2 & mutex m2 ) {
2587        signal( c );
2588}
2589void g( M1 & mutex m1, M2 & mutex m2 ) {
2590        wait( c );
2591}
2592g( `m11`, m2 ); // block on wait
2593f( `m12`, m2 ); // cannot fulfil
2594\end{cfa}
2595\end{lrbox}
2596
2597\begin{lrbox}{\myboxB}
2598\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2599monitor M1 {} m11, m12;
2600monitor M2 {} m2;
2601
2602void f( M1 & mutex m1, M2 & mutex m2 ) {
2603
2604}
2605void g( M1 & mutex m1, M2 & mutex m2 ) {
2606        waitfor( f : m1, m2 );
2607}
2608g( `m11`, m2 ); // block on accept
2609f( `m12`, m2 ); // cannot fulfil
2610\end{cfa}
2611\end{lrbox}
2612\subfloat[Internal scheduling]{\label{f:InternalScheduling}\usebox\myboxA}
2613\hspace{3pt}
2614\vrule
2615\hspace{3pt}
2616\subfloat[External scheduling]{\label{f:ExternalScheduling}\usebox\myboxB}
2617\caption{Unmatched \protect\lstinline@mutex@ sets}
2618\label{f:UnmatchedMutexSets}
2619\end{figure}
2620
2621\begin{figure}
2622\centering
2623\begin{lrbox}{\myboxA}
2624\begin{cfa}[aboveskip=0pt,belowskip=0pt]
2625
2626struct Msg { int i, j; };
2627mutex thread GoRtn { int i;  float f;  Msg m; };
2628void mem1( GoRtn & mutex gortn, int i ) { gortn.i = i; }
2629void mem2( GoRtn & mutex gortn, float f ) { gortn.f = f; }
2630void mem3( GoRtn & mutex gortn, Msg m ) { gortn.m = m; }
2631void ^?{}( GoRtn & mutex ) {}
2632
2633void main( GoRtn & mutex gortn ) with(gortn) { // thread starts
2634
2635        for () {
2636
2637                `waitfor( mem1 : gortn )` sout | i;  // wait for calls
2638                or `waitfor( mem2 : gortn )` sout | f;
2639                or `waitfor( mem3 : gortn )` sout | m.i | m.j;
2640                or `waitfor( ^?{} : gortn )` break; // low priority
2641
2642        }
2643
2644}
2645int main() {
2646        GoRtn gortn; $\C[2.0in]{// start thread}$
2647        `mem1( gortn, 0 );` $\C{// different calls}\CRT$
2648        `mem2( gortn, 2.5 );`
2649        `mem3( gortn, (Msg){1, 2} );`
2650
2651
2652} // wait for completion
2653\end{cfa}
2654\end{lrbox}
2655
2656\begin{lrbox}{\myboxB}
2657\begin{Go}[aboveskip=0pt,belowskip=0pt]
2658func main() {
2659        type Msg struct{ i, j int }
2660
2661        ch1 := make( chan int )
2662        ch2 := make( chan float32 )
2663        ch3 := make( chan Msg )
2664        hand := make( chan string )
2665        shake := make( chan string )
2666        gortn := func() { $\C[1.5in]{// thread starts}$
2667                var i int;  var f float32;  var m Msg
2668                L: for {
2669                        select { $\C{// wait for messages}$
2670                          case `i = <- ch1`: fmt.Println( i )
2671                          case `f = <- ch2`: fmt.Println( f )
2672                          case `m = <- ch3`: fmt.Println( m )
2673                          case `<- hand`: break L $\C{// sentinel}$
2674                        }
2675                }
2676                `shake <- "SHAKE"` $\C{// completion}$
2677        }
2678
2679        go gortn() $\C{// start thread}$
2680        `ch1 <- 0` $\C{// different messages}$
2681        `ch2 <- 2.5`
2682        `ch3 <- Msg{1, 2}`
2683        `hand <- "HAND"` $\C{// sentinel value}$
2684        `<- shake` $\C{// wait for completion}\CRT$
2685}
2686\end{Go}
2687\end{lrbox}
2688
2689\subfloat[\CFA]{\label{f:CFAwaitfor}\usebox\myboxA}
2690\hspace{3pt}
2691\vrule
2692\hspace{3pt}
2693\subfloat[Go]{\label{f:Gochannel}\usebox\myboxB}
2694\caption{Direct versus indirect communication}
2695\label{f:DirectCommunicationComparison}
2696
2697\medskip
2698
2699\begin{cfa}
2700mutex thread DatingService {
2701        condition Girls[CompCodes], Boys[CompCodes];
2702        int girlPhoneNo, boyPhoneNo, ccode;
2703};
2704int girl( DatingService & mutex ds, int phoneno, int code ) with( ds ) {
2705        girlPhoneNo = phoneno;  ccode = code;
2706        `wait( Girls[ccode] );`                                                         $\C{// wait for boy}$
2707        girlPhoneNo = phoneno;  return boyPhoneNo;
2708}
2709int boy( DatingService & mutex ds, int phoneno, int code ) with( ds ) {
2710        boyPhoneNo = phoneno;  ccode = code;
2711        `wait( Boys[ccode] );`                                                          $\C{// wait for girl}$
2712        boyPhoneNo = phoneno;  return girlPhoneNo;
2713}
2714void main( DatingService & ds ) with( ds ) {                    $\C{// thread starts, ds defaults to mutex}$
2715        for () {
2716                waitfor( ^?{} ) break;                                                  $\C{// high priority}$
2717                or waitfor( girl )                                                              $\C{// girl called, compatible boy ? restart boy then girl}$
2718                        if ( ! is_empty( Boys[ccode] ) ) { `signal_block( Boys[ccode] );  signal_block( Girls[ccode] );` }
2719                or waitfor( boy ) {                                                             $\C{// boy called, compatible girl ? restart girl then boy}$
2720                        if ( ! is_empty( Girls[ccode] ) ) { `signal_block( Girls[ccode] );  signal_block( Boys[ccode] );` }
2721        }
2722}
2723\end{cfa}
2724\caption{Direct communication dating service}
2725\label{f:DirectCommunicationDatingService}
2726\end{figure}
2727
2728\begin{comment}
2729The following shows an example of two threads directly calling each other and accepting calls from each other in a cycle.
2730\begin{cfa}
2731\end{cfa}
2732\vspace{-0.8\baselineskip}
2733\begin{cquote}
2734\begin{tabular}{@{}l@{\hspace{3\parindentlnth}}l@{}}
2735\begin{cfa}
2736thread Ping {} pi;
2737void ping( Ping & mutex ) {}
2738void main( Ping & pi ) {
2739        for ( 10 ) {
2740                `waitfor( ping : pi );`
2741                `pong( po );`
2742        }
2743}
2744int main() {}
2745\end{cfa}
2746&
2747\begin{cfa}
2748thread Pong {} po;
2749void pong( Pong & mutex ) {}
2750void main( Pong & po ) {
2751        for ( 10 ) {
2752                `ping( pi );`
2753                `waitfor( pong : po );`
2754        }
2755}
2756
2757\end{cfa}
2758\end{tabular}
2759\end{cquote}
2760% \lstMakeShortInline@%
2761% \caption{Threads ping/pong using external scheduling}
2762% \label{f:pingpong}
2763% \end{figure}
2764Note, the ping/pong threads are globally declared, @pi@/@po@, and hence, start and possibly complete before the program main starts.
2765\end{comment}
2766
2767
2768\subsection{\texorpdfstring{\protect\lstinline@mutex@ Generators / Coroutines / Threads}{monitor Generators / Coroutines / Threads}}
2769
2770\CFA generators, coroutines, and threads can also be @mutex@ (Table~\ref{t:ExecutionPropertyComposition} cases 4, 6, 12) allowing safe \emph{direct communication} with threads, \ie the custom types can have mutex functions that are called by other threads.
2771All monitor features are available within these mutex functions.
2772For example, if the formatter generator or coroutine equivalent in Figure~\ref{f:CFAFormatGen} is extended with the monitor property and this interface function is used to communicate with the formatter:
2773\begin{cfa}
2774void fmt( Fmt & mutex fmt, char ch ) { fmt.ch = ch; resume( fmt ) }
2775\end{cfa}
2776multiple threads can safely pass characters for formatting.
2777
2778Figure~\ref{f:DirectCommunicationComparison} shows a comparison of direct call-communication in \CFA versus indirect channel-communication in Go.
2779(Ada has a similar mechanism to \CFA direct communication.)
2780% The thread main function is by default @mutex@, so the @mutex@ qualifier for the thread parameter is optional.
2781% The reason is that the thread logically starts instantaneously in the thread main acquiring its mutual exclusion, so it starts before any calls to prepare for synchronizing these calls.
2782The \CFA program @main@ uses the call/return paradigm to directly communicate with the @GoRtn main@, whereas Go switches to the unbuffered channel paradigm to indirectly communicate with the goroutine.
2783Communication by multiple threads is safe for the @gortn@ thread via mutex calls in \CFA or channel assignment in Go.
2784The difference between call and channel send occurs for buffered channels making the send asynchronous.
2785In \CFA, asynchronous call and multiple buffers are provided using an administrator and worker threads~\cite{Gentleman81} and/or futures (not discussed).
2786
2787Figure~\ref{f:DirectCommunicationDatingService} shows the dating-service problem in Figure~\ref{f:DatingServiceMonitor} extended from indirect monitor communication to direct thread communication.
2788When converting a monitor to a thread (server), the coding pattern is to move as much code as possible from the accepted functions into the thread main so it does as much work as possible.
2789Notice, the dating server is postponing requests for an unspecified time while continuing to accept new requests.
2790For complex servers, \eg web-servers, there can be hundreds of lines of code in the thread main and safe interaction with clients can be complex.
2791
2792
2793\subsection{Low-level Locks}
2794
2795For 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.
2796Some of these low-level mechanisms are used to build the \CFA runtime, but we always advocate using high-level mechanisms whenever possible.
2797
2798
2799% \section{Parallelism}
2800% \label{s:Parallelism}
2801%
2802% Historically, computer performance was about processor speeds.
2803% However, with heat dissipation being a direct consequence of speed increase, parallelism is the new source for increased performance~\cite{Sutter05, Sutter05b}.
2804% Therefore, high-performance applications must care about parallelism, which requires concurrency.
2805% The lowest-level approach of parallelism is to use \newterm{kernel threads} in combination with semantics like @fork@, @join@, \etc.
2806% However, kernel threads are better as an implementation tool because of complexity and higher cost.
2807% Therefore, different abstractions are often layered onto kernel threads to simplify them, \eg pthreads.
2808%
2809%
2810% \subsection{User Threads}
2811%
2812% A direct improvement on kernel threads is user threads, \eg Erlang~\cite{Erlang} and \uC~\cite{uC++book}.
2813% This approach provides an interface that matches the language paradigms, gives more control over concurrency by the language runtime, and an abstract (and portable) interface to the underlying kernel threads across operating systems.
2814% In many cases, user threads can be used on a much larger scale (100,000 threads).
2815% Like kernel threads, user threads support preemption, which maximizes nondeterminism, but increases the potential for concurrency errors: race, livelock, starvation, and deadlock.
2816% \CFA adopts user-threads to provide more flexibility and a low-cost mechanism to build any other concurrency approach, \eg thread pools and actors~\cite{Actors}.
2817%
2818% A variant of user thread is \newterm{fibres}, which removes preemption, \eg Go~\cite{Go} @goroutine@s.
2819% Like functional programming, which removes mutation and its associated problems, removing preemption from concurrency reduces nondeterminism, making race and deadlock errors more difficult to generate.
2820% However, preemption is necessary for fairness and to reduce tail-latency.
2821% For concurrency that relies on spinning, if all cores spin the system is livelocked, whereas preemption breaks the livelock.
2822
2823
2824\begin{comment}
2825\subsection{Thread Pools}
2826
2827In contrast to direct threading is indirect \newterm{thread pools}, \eg Java @executor@, where small jobs (work units) are inserted into a work pool for execution.
2828If the jobs are dependent, \ie interact, there is an implicit dependency graph that ties them together.
2829While removing direct concurrency, and hence the amount of context switching, thread pools significantly limit the interaction that can occur among jobs.
2830Indeed, jobs should not block because that also blocks the underlying thread, which effectively means the CPU utilization, and therefore throughput, suffers.
2831While 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.
2832As well, concurrency errors return, which threads pools are suppose to mitigate.
2833
2834\begin{figure}
2835\centering
2836\begin{tabular}{@{}l|l@{}}
2837\begin{cfa}
2838struct Adder {
2839        int * row, cols;
2840};
2841int operator()() {
2842        subtotal = 0;
2843        for ( int c = 0; c < cols; c += 1 )
2844                subtotal += row[c];
2845        return subtotal;
2846}
2847void ?{}( Adder * adder, int row[$\,$], int cols, int & subtotal ) {
2848        adder.[rows, cols, subtotal] = [rows, cols, subtotal];
2849}
2850
2851
2852
2853
2854\end{cfa}
2855&
2856\begin{cfa}
2857int main() {
2858        const int rows = 10, cols = 10;
2859        int matrix[rows][cols], subtotals[rows], total = 0;
2860        // read matrix
2861        Executor executor( 4 ); // kernel threads
2862        Adder * adders[rows];
2863        for ( r; rows ) { // send off work for executor
2864                adders[r] = new( matrix[r], cols, &subtotal[r] );
2865                executor.send( *adders[r] );
2866        }
2867        for ( r; rows ) {       // wait for results
2868                delete( adders[r] );
2869                total += subtotals[r];
2870        }
2871        sout | total;
2872}
2873\end{cfa}
2874\end{tabular}
2875\caption{Executor}
2876\end{figure}
2877\end{comment}
2878
2879
2880\section{Runtime Structure}
2881\label{s:CFARuntimeStructure}
2882
2883Figure~\ref{f:RunTimeStructure} illustrates the runtime structure of a \CFA program.
2884In 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.
2885An executing thread is illustrated by its containment in a processor.
2886
2887\begin{figure}
2888\centering
2889\input{RunTimeStructure}
2890\caption{\CFA Runtime structure}
2891\label{f:RunTimeStructure}
2892\end{figure}
2893
2894
2895\subsection{Cluster}
2896\label{s:RuntimeStructureCluster}
2897
2898A \newterm{cluster} is a collection of user and kernel threads, where the kernel threads run the user threads from the cluster's ready queue, and the operating system runs the kernel threads on the processors from its ready queue.
2899The term \newterm{virtual processor} is introduced as a synonym for kernel thread to disambiguate between user and kernel thread.
2900From the language perspective, a virtual processor is an actual processor (core).
2901
2902The purpose of a cluster is to control the amount of parallelism that is possible among threads, plus scheduling and other execution defaults.
2903The default cluster-scheduler is single-queue multi-server, which provides automatic load-balancing of threads on processors.
2904However, the design allows changing the scheduler, \eg multi-queue multi-server with work-stealing/sharing across the virtual processors.
2905If several clusters exist, both threads and virtual processors, can be explicitly migrated from one cluster to another.
2906No automatic load balancing among clusters is performed by \CFA.
2907
2908When 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.
2909The user cluster is created to contain the application user-threads.
2910Having all threads execute on the one cluster often maximizes utilization of processors, which minimizes runtime.
2911However, because of limitations of scheduling requirements (real-time), NUMA architecture, heterogeneous hardware, or issues with the underlying operating system, multiple clusters are sometimes necessary.
2912
2913
2914\subsection{Virtual Processor}
2915\label{s:RuntimeStructureProcessor}
2916
2917A virtual processor is implemented by a kernel thread, \eg UNIX process, which are scheduled for execution on a hardware processor by the underlying operating system.
2918Programs may use more virtual processors than hardware processors.
2919On a multiprocessor, kernel threads are distributed across the hardware processors resulting in virtual processors executing in parallel.
2920(It is possible to use affinity to lock a virtual processor onto a particular hardware processor~\cite{affinityLinux,affinityWindows}, which is used when caching issues occur or for heterogeneous hardware processors.) %, affinityFreebsd, affinityNetbsd, affinityMacosx
2921The \CFA runtime attempts to block unused processors and unblock processors as the system load increases;
2922balancing the workload with processors is difficult because it requires future knowledge, \ie what will the application workload do next.
2923Preemption occurs on virtual processors rather than user threads, via operating-system interrupts.
2924Thus virtual processors execute user threads, where preemption frequency applies to a virtual processor, so preemption occurs randomly across the executed user threads.
2925Turning off preemption transforms user threads into fibres.
2926
2927
2928\begin{comment}
2929\section{Implementation}
2930\label{s:Implementation}
2931
2932A primary implementation challenge is avoiding contention from dynamically allocating memory because of bulk acquire, \eg the internal-scheduling design is almost free of allocations.
2933All 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.
2934Furthermore, several bulk-acquire operations need a variable amount of memory.
2935This 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.
2936
2937In \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.
2938When a mutex call is made, pointers to the concerned monitors are aggregated into a variable-length array and sorted.
2939This array persists for the entire duration of the mutual exclusion and is used extensively for synchronization operations.
2940
2941To improve performance and simplicity, context switching occurs inside a function call, so only callee-saved registers are copied onto the stack and then the stack register is switched;
2942the corresponding registers are then restored for the other context.
2943Note, the instruction pointer is untouched since the context switch is always inside the same function.
2944Experimental results (not presented) for a stackless or stackful scheduler (1 versus 2 context switches) (see Section~\ref{s:Concurrency}) show the performance is virtually equivalent, because both approaches are dominated by locking to prevent a race condition.
2945
2946All kernel threads (@pthreads@) created a stack.
2947Each \CFA virtual processor is implemented as a coroutine and these coroutines run directly on the kernel-thread stack, effectively stealing this stack.
2948The exception to this rule is the program main, \ie the initial kernel thread that is given to any program.
2949In 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.
2950\end{comment}
2951
2952
2953\subsection{Preemption}
2954
2955Nondeterministic preemption provides fairness from long-running threads, and forces concurrent programmers to write more robust programs, rather than relying on code between cooperative scheduling to be atomic.
2956This atomic reliance can fail on multi-core machines, because execution across cores is nondeterministic.
2957A different reason for not supporting preemption is that it significantly complicates the runtime system, \eg Windows runtime does not support interrupts and on Linux systems, interrupts are complex (see below).
2958Preemption is normally handled by setting a countdown timer on each virtual processor.
2959When the timer expires, an interrupt is delivered, and its signal handler resets the countdown 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.
2960Multiple signal handlers may be pending.
2961When 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.
2962The 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;
2963therefore, the same signal mask is required for all virtual processors in a cluster.
2964Because preemption interval is usually long (1 millisecond) performance cost is negligible.
2965
2966Linux switched a decade ago from specific to arbitrary virtual-processor signal-delivery for applications with multiple kernel threads.
2967In the new semantics, a virtual-processor directed signal may be delivered to any virtual processor created by the application that does not have the signal blocked.
2968Hence, the timer-expiry signal, which is generated \emph{externally} by the Linux kernel to an application, is delivered to any of its Linux subprocesses (kernel threads).
2969To ensure each virtual processor receives a preemption signal, a discrete-event simulation is run on a special virtual processor, and only it sets and receives timer events.
2970Virtual processors register an expiration time with the discrete-event simulator, which is inserted in sorted order.
2971The simulation sets the countdown 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.
2972Processing a preemption event sends an \emph{internal} @SIGUSR1@ signal to the registered virtual processor, which is always delivered to that processor.
2973
2974
2975\subsection{Debug Kernel}
2976
2977There are two versions of the \CFA runtime kernel: debug and non-debug.
2978The 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.
2979After a program is debugged, the non-debugging version can be used to significantly decrease space and increase performance.
2980
2981
2982\section{Performance}
2983\label{s:Performance}
2984
2985To test the performance of the \CFA runtime, a series of microbenchmarks are used to compare \CFA with pthreads, Java 11.0.6, Go 1.12.6, Rust 1.37.0, Python 3.7.6, Node.js 12.14.1, and \uC 7.0.0.
2986For comparison, the package must be multi-processor (M:N), which excludes libdil and libmil~\cite{libdill} (M:1)), and use a shared-memory programming model, \eg not message passing.
2987The benchmark computer is an AMD Opteron\texttrademark\ 6380 NUMA 64-core, 8 socket, 2.5 GHz processor, running Ubuntu 16.04.6 LTS, and pthreads/\CFA/\uC are compiled with gcc 9.2.1.
2988
2989All benchmarks are run using the following harness. (The Java harness is augmented to circumvent JIT issues.)
2990\begin{cfa}
2991#define BENCH( `run` ) uint64_t start = cputime_ns();  `run;`  double result = (double)(cputime_ns() - start) / N;
2992\end{cfa}
2993where CPU time in nanoseconds is from the appropriate language clock.
2994Each benchmark is performed @N@ times, where @N@ is selected so the benchmark runs in the range of 2--20 seconds for the specific programming language.
2995The total time is divided by @N@ to obtain the average time for a benchmark.
2996Each benchmark experiment is run 13 times and the average appears in the table.
2997All omitted tests for other languages are functionally identical to the \CFA tests and available online~\cite{CforallConcurrentBenchmarks}.
2998% tar --exclude-ignore=exclude -cvhf benchmark.tar benchmark
2999% cp -p benchmark.tar /u/cforall/public_html/doc/concurrent_benchmark.tar
3000
3001\paragraph{Creation}
3002
3003Creation is measured by creating and deleting a specific kind of control-flow object.
3004Figure~\ref{f:creation} shows the code for \CFA with results in Table~\ref{t:creation}.
3005Note, the call stacks of \CFA coroutines are lazily created on the first resume, therefore the cost of creation with and without a stack are presented.
3006
3007\begin{multicols}{2}
3008\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
3009\begin{cfa}
3010@coroutine@ MyCoroutine {};
3011void ?{}( MyCoroutine & this ) {
3012#ifdef EAGER
3013        resume( this );
3014#endif
3015}
3016void main( MyCoroutine & ) {}
3017int main() {
3018        BENCH( for ( N ) { @MyCoroutine c;@ } )
3019        sout | result;
3020}
3021\end{cfa}
3022\captionof{figure}{\CFA creation benchmark}
3023\label{f:creation}
3024
3025\columnbreak
3026
3027\vspace*{-16pt}
3028\captionof{table}{Creation comparison (nanoseconds)}
3029\label{t:creation}
3030
3031\begin{tabular}[t]{@{}r*{3}{D{.}{.}{5.2}}@{}}
3032\multicolumn{1}{@{}c}{} & \multicolumn{1}{c}{Median} & \multicolumn{1}{c}{Average} & \multicolumn{1}{c@{}}{Std Dev} \\
3033\CFA generator                  & 0.6           & 0.6           & 0.0           \\
3034\CFA coroutine lazy             & 13.4          & 13.1          & 0.5           \\
3035\CFA coroutine eager    & 144.7         & 143.9         & 1.5           \\
3036\CFA thread                             & 466.4         & 468.0         & 11.3          \\
3037\uC coroutine                   & 155.6         & 155.7         & 1.7           \\
3038\uC thread                              & 523.4         & 523.9         & 7.7           \\
3039Python generator                & 123.2         & 124.3         & 4.1           \\
3040Node.js generator               & 33.4          & 33.5          & 0.3           \\
3041Goroutine thread                & 751.0         & 750.5         & 3.1           \\
3042Rust tokio thread               & 1860.0        & 1881.1        & 37.6          \\
3043Rust thread                             & 53801.0       & 53896.8       & 274.9         \\
3044Java thread (   10 000)         & 119256.0      & 119679.2      & 2244.0        \\
3045Java thread (1 000 000)         & 123100.0      & 123052.5      & 751.6         \\
3046Pthreads thread                 & 31465.5       & 31419.5       & 140.4
3047\end{tabular}
3048\end{multicols}
3049
3050\vspace*{-10pt}
3051\paragraph{Internal Scheduling}
3052
3053Internal scheduling is measured using a cycle of two threads signalling and waiting.
3054Figure~\ref{f:schedint} shows the code for \CFA, with results in Table~\ref{t:schedint}.
3055Note, the incremental cost of bulk acquire for \CFA, which is largely a fixed cost for small numbers of mutex objects.
3056Java scheduling is significantly greater because the benchmark explicitly creates multiple threads in order to prevent the JIT from making the program sequential, \ie removing all locking.
3057
3058\begin{multicols}{2}
3059\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
3060\begin{cfa}
3061volatile int go = 0;
3062@condition c;@
3063@monitor@ M {} m1/*, m2, m3, m4*/;
3064void call( M & @mutex p1/*, p2, p3, p4*/@ ) {
3065        @signal( c );@
3066}
3067void wait( M & @mutex p1/*, p2, p3, p4*/@ ) {
3068        go = 1; // continue other thread
3069        for ( N ) { @wait( c );@ } );
3070}
3071thread T {};
3072void main( T & ) {
3073        while ( go == 0 ) { yield(); } // waiter must start first
3074        BENCH( for ( N ) { call( m1/*, m2, m3, m4*/ ); } )
3075        sout | result;
3076}
3077int main() {
3078        T t;
3079        wait( m1/*, m2, m3, m4*/ );
3080}
3081\end{cfa}
3082\vspace*{-8pt}
3083\captionof{figure}{\CFA Internal-scheduling benchmark}
3084\label{f:schedint}
3085
3086\columnbreak
3087
3088\vspace*{-16pt}
3089\captionof{table}{Internal-scheduling comparison (nanoseconds)}
3090\label{t:schedint}
3091\bigskip
3092
3093\begin{tabular}{@{}r*{3}{D{.}{.}{5.2}}@{}}
3094\multicolumn{1}{@{}c}{} & \multicolumn{1}{c}{Median} & \multicolumn{1}{c}{Average} & \multicolumn{1}{c@{}}{Std Dev} \\
3095\CFA @signal@, 1 monitor        & 364.4         & 364.2         & 4.4           \\
3096\CFA @signal@, 2 monitor        & 484.4         & 483.9         & 8.8           \\
3097\CFA @signal@, 4 monitor        & 709.1         & 707.7         & 15.0          \\
3098\uC @signal@ monitor            & 328.3         & 327.4         & 2.4           \\
3099Rust cond. variable                     & 7514.0        & 7437.4        & 397.2         \\
3100Java @notify@ monitor (  1 000 000)             & 8717.0        & 8774.1        & 471.8         \\
3101Java @notify@ monitor (100 000 000)             & 8634.0        & 8683.5        & 330.5         \\
3102Pthreads cond. variable         & 5553.7        & 5576.1        & 345.6
3103\end{tabular}
3104\end{multicols}
3105
3106
3107\paragraph{External Scheduling}
3108
3109External scheduling is measured using a cycle of two threads calling and accepting the call using the @waitfor@ statement.
3110Figure~\ref{f:schedext} shows the code for \CFA with results in Table~\ref{t:schedext}.
3111Note, the incremental cost of bulk acquire for \CFA, which is largely a fixed cost for small numbers of mutex objects.
3112
3113\begin{multicols}{2}
3114\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
3115\vspace*{-16pt}
3116\begin{cfa}
3117@monitor@ M {} m1/*, m2, m3, m4*/;
3118void call( M & @mutex p1/*, p2, p3, p4*/@ ) {}
3119void wait( M & @mutex p1/*, p2, p3, p4*/@ ) {
3120        for ( N ) { @waitfor( call : p1/*, p2, p3, p4*/ );@ }
3121}
3122thread T {};
3123void main( T & ) {
3124        BENCH( for ( N ) { call( m1/*, m2, m3, m4*/ ); } )
3125        sout | result;
3126}
3127int main() {
3128        T t;
3129        wait( m1/*, m2, m3, m4*/ );
3130}
3131\end{cfa}
3132\captionof{figure}{\CFA external-scheduling benchmark}
3133\label{f:schedext}
3134
3135\columnbreak
3136
3137\vspace*{-16pt}
3138\captionof{table}{External-scheduling comparison (nanoseconds)}
3139\label{t:schedext}
3140\begin{tabular}{@{}r*{3}{D{.}{.}{3.2}}@{}}
3141\multicolumn{1}{@{}c}{} & \multicolumn{1}{c}{Median} &\multicolumn{1}{c}{Average} & \multicolumn{1}{c@{}}{Std Dev} \\
3142\CFA @waitfor@, 1 monitor       & 367.1 & 365.3 & 5.0   \\
3143\CFA @waitfor@, 2 monitor       & 463.0 & 464.6 & 7.1   \\
3144\CFA @waitfor@, 4 monitor       & 689.6 & 696.2 & 21.5  \\
3145\uC \lstinline[language=uC++]|_Accept| monitor  & 328.2 & 329.1 & 3.4   \\
3146Go \lstinline[language=Golang]|select| channel  & 365.0 & 365.5 & 1.2
3147\end{tabular}
3148\end{multicols}
3149
3150\paragraph{Mutual-Exclusion}
3151
3152Uncontented mutual exclusion, which frequently occurs, is measured by entering and leaving a critical section.
3153For monitors, entering and leaving a mutex function is measured, otherwise the language-appropriate mutex-lock is measured.
3154For comparison, a spinning (versus blocking) test-and-test-set lock is presented.
3155Figure~\ref{f:mutex} shows the code for \CFA with results in Table~\ref{t:mutex}.
3156Note the incremental cost of bulk acquire for \CFA, which is largely a fixed cost for small numbers of mutex objects.
3157
3158\begin{multicols}{2}
3159\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
3160\begin{cfa}
3161@monitor@ M {} m1/*, m2, m3, m4*/;
3162call( M & @mutex p1/*, p2, p3, p4*/@ ) {}
3163int main() {
3164        BENCH( for( N ) call( m1/*, m2, m3, m4*/ ); )
3165        sout | result;
3166}
3167\end{cfa}
3168\captionof{figure}{\CFA acquire/release mutex benchmark}
3169\label{f:mutex}
3170
3171\columnbreak
3172
3173\vspace*{-16pt}
3174\captionof{table}{Mutex comparison (nanoseconds)}
3175\label{t:mutex}
3176\begin{tabular}{@{}r*{3}{D{.}{.}{3.2}}@{}}
3177\multicolumn{1}{@{}c}{} & \multicolumn{1}{c}{Median} &\multicolumn{1}{c}{Average} & \multicolumn{1}{c@{}}{Std Dev} \\
3178test-and-test-set lock                  & 19.1  & 18.9  & 0.4   \\
3179\CFA @mutex@ function, 1 arg.   & 48.3  & 47.8  & 0.9   \\
3180\CFA @mutex@ function, 2 arg.   & 86.7  & 87.6  & 1.9   \\
3181\CFA @mutex@ function, 4 arg.   & 173.4 & 169.4 & 5.9   \\
3182\uC @monitor@ member rtn.               & 54.8  & 54.8  & 0.1   \\
3183Goroutine mutex lock                    & 34.0  & 34.0  & 0.0   \\
3184Rust mutex lock                                 & 33.0  & 33.2  & 0.8   \\
3185Java synchronized method (   100 000 000)               & 31.0  & 30.9  & 0.5   \\
3186Java synchronized method (10 000 000 000)               & 31.0 & 30.2 & 0.9 \\
3187Pthreads mutex Lock                             & 31.0  & 31.1  & 0.4
3188\end{tabular}
3189\end{multicols}
3190
3191\paragraph{Context Switching}
3192
3193In procedural programming, the cost of a function call is important as modularization (refactoring) increases.
3194(In many cases, a compiler inlines function calls to increase the size and number of basic blocks for optimizing.)
3195Similarly, when modularization extends to coroutines and threads, the time for a context switch becomes a relevant factor.
3196The coroutine test is from resumer to suspender and from suspender to resumer, which is two context switches.
3197%For async-await systems, the test is scheduling and fulfilling @N@ empty promises, where all promises are allocated before versus interleaved with fulfillment to avoid garbage collection.
3198For async-await systems, the test measures the cost of the @await@ expression entering the event engine by awaiting @N@ promises, where each created promise is resolved by an immediate event in the engine (using Node.js @setImmediate@).
3199The thread test is using yield to enter and return from the runtime kernel, which is two context switches.
3200The difference in performance between coroutine and thread context-switch is the cost of scheduling for threads, whereas coroutines are self-scheduling.
3201Figure~\ref{f:ctx-switch} shows the \CFA code for a coroutine and thread with results in Table~\ref{t:ctx-switch}.
3202
3203% From: Gregor Richards <gregor.richards@uwaterloo.ca>
3204% To: "Peter A. Buhr" <pabuhr@plg2.cs.uwaterloo.ca>
3205% Date: Fri, 24 Jan 2020 13:49:18 -0500
3206%
3207% I can also verify that the previous version, which just tied a bunch of promises together, *does not* go back to the
3208% event loop at all in the current version of Node. Presumably they're taking advantage of the fact that the ordering of
3209% events is intentionally undefined to just jump right to the next 'then' in the chain, bypassing event queueing
3210% entirely. That's perfectly correct behavior insofar as its difference from the specified behavior isn't observable, but
3211% it isn't typical or representative of much anything useful, because most programs wouldn't have whole chains of eager
3212% promises. Also, it's not representative of *anything* you can do with async/await, as there's no way to encode such an
3213% eager chain that way.
3214
3215\begin{multicols}{2}
3216\lstset{language=CFA,moredelim=**[is][\color{red}]{@}{@},deletedelim=**[is][]{`}{`}}
3217\begin{cfa}[aboveskip=0pt,belowskip=0pt]
3218@coroutine@ C {};
3219void main( C & ) { for () { @suspend;@ } }
3220int main() { // coroutine test
3221        C c;
3222        BENCH( for ( N ) { @resume( c );@ } )
3223        sout | result;
3224}
3225int main() { // thread test
3226        BENCH( for ( N ) { @yield();@ } )
3227        sout | result;
3228}
3229\end{cfa}
3230\captionof{figure}{\CFA context-switch benchmark}
3231\label{f:ctx-switch}
3232
3233\columnbreak
3234
3235\vspace*{-16pt}
3236\captionof{table}{Context switch comparison (nanoseconds)}
3237\label{t:ctx-switch}
3238\begin{tabular}{@{}r*{3}{D{.}{.}{3.2}}@{}}
3239\multicolumn{1}{@{}c}{} & \multicolumn{1}{c}{Median} &\multicolumn{1}{c}{Average} & \multicolumn{1}{c@{}}{Std Dev} \\
3240C function                      & 1.8           & 1.8           & 0.0   \\
3241\CFA generator          & 1.8           & 2.0           & 0.3   \\
3242\CFA coroutine          & 32.5          & 32.9          & 0.8   \\
3243\CFA thread                     & 93.8          & 93.6          & 2.2   \\
3244\uC coroutine           & 50.3          & 50.3          & 0.2   \\
3245\uC thread                      & 97.3          & 97.4          & 1.0   \\
3246Python generator        & 40.9          & 41.3          & 1.5   \\
3247Node.js await           & 1852.2        & 1854.7        & 16.4  \\
3248Node.js generator       & 33.3          & 33.4          & 0.3   \\
3249Goroutine thread        & 143.0         & 143.3         & 1.1   \\
3250Rust async await        & 32.0          & 32.0          & 0.0   \\
3251Rust tokio thread       & 143.0         & 143.0         & 1.7   \\
3252Rust thread                     & 332.0         & 331.4         & 2.4   \\
3253Java thread     (      100 000)         & 405.0         & 415.0         & 17.6  \\
3254Java thread (  100 000 000)                     & 413.0 & 414.2 & 6.2 \\
3255Java thread (5 000 000 000)                     & 415.0 & 415.2 & 6.1 \\
3256Pthreads thread         & 334.3         & 335.2         & 3.9
3257\end{tabular}
3258\end{multicols}
3259
3260
3261\subsection{Discussion}
3262
3263Languages using 1:1 threading based on pthreads can at best meet or exceed, due to language overhead, the pthread results.
3264Note, pthreads has a fast zero-contention mutex lock checked in user space.
3265Languages with M:N threading have better performance than 1:1 because there is no operating-system interactions.
3266Languages with stackful coroutines have higher cost than stackless coroutines because of stack allocation and context switching;
3267however, stackful \uC and \CFA coroutines have approximately the same performance as stackless Python and Node.js generators.
3268The \CFA stackless generator is approximately 25 times faster for suspend/resume and 200 times faster for creation than stackless Python and Node.js generators.
3269
3270
3271\section{Conclusions and Future Work}
3272
3273Advanced control-flow will always be difficult, especially when there is temporal ordering and nondeterminism.
3274However, many systems exacerbate the difficulty through their presentation mechanisms.
3275This paper shows it is possible to understand high-level control-flow using three properties: statefulness, thread, mutual-exclusion/synchronization.
3276Combining these properties creates a number of high-level, efficient, and maintainable control-flow types: generator, coroutine, thread, each of which can be a monitor.
3277Eliminated from \CFA are barging and spurious wakeup, which are nonintuitive and lead to errors, and having to work with a bewildering set of low-level locks and acquisition techniques.
3278\CFA high-level race-free monitors and threads, when used with mutex access function, provide the core mechanisms for mutual exclusion and synchronization, without having to resort to magic qualifiers like @volatile@ or @atomic@.
3279Extending these mechanisms to handle high-level deadlock-free bulk acquire across both mutual exclusion and synchronization is a unique contribution.
3280The \CFA runtime 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.
3281The M:N model is judged to be efficient and provide greater flexibility than a 1:1 threading model.
3282These concepts and the \CFA runtime-system are written in the \CFA language, extensively leveraging the \CFA type-system, which demonstrates the expressiveness of the \CFA language.
3283Performance comparisons with other concurrent systems and languages show the \CFA approach is competitive across all basic operations, which translates directly into good performance in well-written applications with advanced control-flow.
3284C programmers should feel comfortable using these mechanisms for developing complex control-flow in applications, with the ability to obtain maximum available performance by selecting mechanisms at the appropriate level of need using only calling communication.
3285
3286While control flow in \CFA has a strong start, development is still underway to complete a number of missing features.
3287
3288\medskip
3289\textbf{Flexible Scheduling:}
3290An important part of concurrency is scheduling.
3291Different scheduling algorithms can affect performance, both in terms of average and variation.
3292However, no single scheduler is optimal for all workloads and therefore there is value in being able to change the scheduler for given programs.
3293One solution is to offer various tuning options, allowing the scheduler to be adjusted to the requirements of the workload.
3294However, to be truly flexible, a pluggable scheduler is necessary.
3295Currently, 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~\cite{Buhr00b}.
3296
3297\smallskip
3298\textbf{Non-Blocking I/O:}
3299Many modern workloads are not bound by computation but IO operations, common cases being web servers and XaaS~\cite{XaaS} (anything as a service).
3300These types of workloads require significant engineering to amortizing costs of blocking IO-operations.
3301At 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.
3302Current 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.
3303However, these solutions lead to code that is hard to create, read and maintain.
3304A 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.
3305A non-blocking I/O library is currently under development for \CFA.
3306
3307\smallskip
3308\textbf{Other Concurrency Tools:}
3309While monitors offer flexible and powerful concurrency for \CFA, other concurrency tools are also necessary for a complete multi-paradigm concurrency package.
3310Examples of such tools can include futures and promises~\cite{promises}, executors and actors.
3311These additional features are useful for applications that can be constructed without shared data and direct blocking.
3312As well, new \CFA extensions should make it possible to create a uniform interface for virtually all mutual exclusion, including monitors and low-level locks.
3313
3314\smallskip
3315\textbf{Implicit Threading:}
3316Basic \emph{embarrassingly parallel} applications can benefit greatly from implicit concurrency, where sequential programs are converted to concurrent, with some help from pragmas to guide the conversion.
3317This type of concurrency can be achieved both at the language level and at the library level.
3318The canonical example of implicit concurrency is concurrent nested @for@ loops, which are amenable to divide and conquer algorithms~\cite{uC++book}.
3319The \CFA language features should make it possible to develop a reasonable number of implicit concurrency mechanisms to solve basic HPC data-concurrency problems.
3320However, implicit concurrency is a restrictive solution with significant limitations, so it can never replace explicit concurrent programming.
3321
3322
3323\section{Acknowledgements}
3324
3325The authors recognize the design assistance of Aaron Moss, Rob Schluntz, Andrew Beach, and Michael Brooks; David Dice for commenting and helping with the Java benchmarks; and Gregor Richards for helping with the Node.js benchmarks.
3326This research is funded by a grant from Waterloo-Huawei (\url{http://www.huawei.com}) Joint Innovation Lab. %, and Peter Buhr is partially funded by the Natural Sciences and Engineering Research Council of Canada.
3327
3328{%
3329\fontsize{9bp}{11.5bp}\selectfont%
3330\bibliography{pl,local}
3331}%
3332
3333\end{document}
3334
3335% Local Variables: %
3336% tab-width: 4 %
3337% fill-column: 120 %
3338% compile-command: "make" %
3339% End: %
Note: See TracBrowser for help on using the repository browser.