% Why direct style instead of continuation passing style? %% Student project ideas: %% * high-level optimizations like procedure inlining, etc. %% * closure optimization %% * adding letrec to the language %% (Thought: in the book and regular course, replace top-level defines %% with letrec.) %% * alternative back ends (ARM, LLVM) %% * alternative calling convention (a la Dybvig) %% * lazy evaluation %% * gradual typing %% * continuations (frames in heap a la SML or segmented stack a la Dybvig) %% * exceptions %% * self hosting %% * I/O %% * foreign function interface %% * quasi-quote and unquote %% * macros (too difficult?) %% * alternative garbage collector %% * alternative register allocator %% * parametric polymorphism %% * type classes (too difficulty?) %% * loops (too easy? combine with something else?) %% * loop optimization (fusion, etc.) %% * deforestation %% * records and subtyping %% * object-oriented features %% - objects, object types, and structural subtyping (e.g. Abadi & Cardelli) %% - class-based objects and nominal subtyping (e.g. Featherweight Java) %% * multi-threading, fork join, futures, implicit parallelism %% * dataflow analysis, type analysis and specialization \documentclass[11pt]{book} \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage{lmodern} \usepackage{hyperref} \usepackage{graphicx} \usepackage[english]{babel} \usepackage{listings} \usepackage{amsmath} \usepackage{amsthm} \usepackage{amssymb} \usepackage{natbib} \usepackage{stmaryrd} \usepackage{xypic} \usepackage{semantic} \usepackage{wrapfig} \usepackage{tcolorbox} \usepackage{multirow} \usepackage{color} \usepackage{upquote} \usepackage{makeidx} \makeindex \definecolor{lightgray}{gray}{1} \newcommand{\black}[1]{{\color{black} #1}} %\newcommand{\gray}[1]{{\color{lightgray} #1}} \newcommand{\gray}[1]{{\color{gray} #1}} %% For pictures \usepackage{tikz} \usetikzlibrary{arrows.meta} \tikzset{baseline=(current bounding box.center), >/.tip={Triangle[scale=1.4]}} % Computer Modern is already the default. -Jeremy %\renewcommand{\ttdefault}{cmtt} \definecolor{comment-red}{rgb}{0.8,0,0} \if01 \newcommand{\rn}[1]{{\color{comment-red}{(RRN: #1)}}} \newcommand{\margincomment}[1]{\marginpar{\color{comment-red}\tiny #1}} \else \newcommand{\rn}[1]{} \newcommand{\margincomment}[1]{} \fi \lstset{% language=Lisp, basicstyle=\ttfamily\small, morekeywords={seq,assign,program,block,define,lambda,match,goto,if,else,then,struct,Integer,Boolean,Vector,Void,Any,while,begin,define,public,override,class}, deletekeywords={read,mapping,vector}, escapechar=|, columns=flexible, moredelim=[is][\color{red}]{~}{~}, showstringspaces=false } \newtheorem{theorem}{Theorem} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{constraint}[theorem]{Constraint} \newtheorem{definition}[theorem]{Definition} \newtheorem{exercise}[theorem]{Exercise} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % 'dedication' environment: To add a dedication paragraph at the start of book % % Source: http://www.tug.org/pipermail/texhax/2010-June/015184.html % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \newenvironment{dedication} { \cleardoublepage \thispagestyle{empty} \vspace*{\stretch{1}} \hfill\begin{minipage}[t]{0.66\textwidth} \raggedright } { \end{minipage} \vspace*{\stretch{3}} \clearpage } %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Chapter quote at the start of chapter % % Source: http://tex.stackexchange.com/a/53380 % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \makeatletter \renewcommand{\@chapapp}{}% Not necessary... \newenvironment{chapquote}[2][2em] {\setlength{\@tempdima}{#1}% \def\chapquote@author{#2}% \parshape 1 \@tempdima \dimexpr\textwidth-2\@tempdima\relax% \itshape} {\par\normalfont\hfill--\ \chapquote@author\hspace*{\@tempdima}\par\bigskip} \makeatother \input{defs} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \title{\Huge \textbf{Essentials of Compilation} \\ \huge The Incremental, Nano-Pass Approach} \author{\textsc{Jeremy G. Siek} \\ %\thanks{\url{http://homes.soic.indiana.edu/jsiek/}} \\ Indiana University \\ \\ with contributions from: \\ Carl Factora \\ Andre Kuhlenschmidt \\ Ryan R. Newton \\ Ryan Scott \\ Cameron Swords \\ Michael M. Vitousek \\ Michael Vollmer } \begin{document} \frontmatter \maketitle \begin{dedication} This book is dedicated to the programming language wonks at Indiana University. \end{dedication} \tableofcontents \listoffigures %\listoftables \mainmatter %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter*{Preface} The tradition of compiler writing at Indiana University goes back to research and courses on programming languages by Professor Daniel Friedman in the 1970's and 1980's. Friedman conducted research on lazy evaluation~\citep{Friedman:1976aa} in the context of Lisp~\citep{McCarthy:1960dz} and then studied continuations~\citep{Felleisen:kx} and macros~\citep{Kohlbecker:1986dk} in the context of the Scheme~\citep{Sussman:1975ab}, a dialect of Lisp. One of the students of those courses, Kent Dybvig, went on to build Chez Scheme~\citep{Dybvig:2006aa}, a production-quality and efficient compiler for Scheme. After completing his Ph.D. at the University of North Carolina, he returned to teach at Indiana University. Throughout the 1990's and 2000's, Professor Dybvig continued development of Chez Scheme and taught the compiler course. The compiler course evolved to incorporate novel pedagogical ideas while also including elements of effective real-world compilers. One of Friedman's ideas was to split the compiler into many small ``passes'' so that the code for each pass would be easy to understood in isolation. In contrast, most compilers of the time were organized into only a few monolithic passes for reasons of compile-time efficiency. Another idea, called ``the game'', was to test the code generated by each pass on interpreters for each intermediate language, thereby helping to pinpoint errors in individual passes. % Dybvig, with later help from his students Dipanwita Sarkar and Andrew Keep, developed infrastructure to support this approach and evolved the course, first to use smaller micro-passes and then into even smaller nano-passes~\citep{Sarkar:2004fk,Keep:2012aa}. I was a student in this compiler course in the early 2000's as part of my Ph.D. studies at Indiana University. Needless to say, I enjoyed the course immensely! During that time, another graduate student named Abdulaziz Ghuloum observed that the front-to-back organization of the course made it difficult for students to understand the rationale for the compiler design. Ghuloum proposed an incremental approach in which the students build the compiler in stages; they start by implementing a complete compiler for a very small subset of the input language and in each subsequent stage they add a language feature and add or modify passes to handle the new feature~\citep{Ghuloum:2006bh}. In this way, the students see how the language features motivate aspects of the compiler design. After graduating from Indiana University in 2005, I went on to teach at the University of Colorado. I adapted the nano-pass and incremental approaches to compiling a subset of the Python language~\citep{Siek:2012ab}. Python and Scheme are quite different on the surface but there is a large overlap in the compiler techniques required for the two languages. Thus, I was able to teach much of the same content from the Indiana compiler course. I very much enjoyed teaching the course organized in this way, and even better, many of the students learned a lot and got excited about compilers. I returned to Indiana University in 2013. In my absence the compiler course had switched from the front-to-back organization to a back-to-front organization~\cite{Dybvig:2010aa}. Seeing how well the incremental approach worked at Colorado, I started porting and adapting the structure of the Colorado course back into the land of Scheme. In the meantime Indiana University had moved on from Scheme to Racket~\citep{plt-tr}, so the course is now about compiling a subset of Racket (and Typed Racket) to the x86 assembly language. This is the textbook for the incremental version of the compiler course at Indiana University (Spring 2016 - present). With this book I hope to make the Indiana compiler course available to people that have not had the chance to study compilers at Indiana University. %% I have captured what %% I think are the most important topics from \cite{Dybvig:2010aa} but %% have omitted topics that are less interesting conceptually. I have %% also made simplifications to reduce complexity. In this way, this %% book leans more towards pedagogy than towards the efficiency of the %% generated code. Also, the book differs in places where we I the %% opportunity to make the topics more fun, such as in relating register %% allocation to Sudoku (Chapter~\ref{ch:register-allocation-r1}). \section*{Prerequisites} The material in this book is challenging but rewarding. It is meant to prepare students for a lifelong career in programming languages. The book uses the Racket language both for the implementation of the compiler and for the language that is compiled, so a student should be proficient with Racket (or Scheme) prior to reading this book. There are many excellent resources for learning Scheme and Racket~\citep{Dybvig:1987aa,Abelson:1996uq,Friedman:1996aa,Felleisen:2001aa,Felleisen:2013aa,Flatt:2014aa}. It is helpful but not necessary for the student to have prior exposure to the x86 assembly language~\citep{Intel:2015aa}, as one might obtain from a computer systems course~\citep{Bryant:2010aa}. This book introduces the parts of x86-64 assembly language that are needed. % We follow the System V calling conventions~\citep{Bryant:2005aa,Matz:2013aa}, which means that the assembly code that we generate will work properly with our runtime system (written in C) when it is compiled using the GNU C compiler (\code{gcc}) on the Linux and MacOS operating systems. (Minor adjustments are needed for MacOS which we note as they arise.) % When running on the Microsoft Windows operating system, the GNU C compiler follows the Microsoft x64 calling convention~\citep{Microsoft:2018aa,Microsoft:2020aa}. So the assembly code that we generate will \emph{not} work properly with our runtime system on Windows. One option to consider for using a Windows computer is to run a virtual machine with Linux as the guest operating system. %\section*{Structure of book} % You might want to add short description about each chapter in this book. %\section*{About the companion website} %The website\footnote{\url{https://github.com/amberj/latex-book-template}} for %this file contains: %\begin{itemize} % \item A link to (freely downlodable) latest version of this document. % \item Link to download LaTeX source for this document. % \item Miscellaneous material (e.g. suggested readings etc). %\end{itemize} \section*{Acknowledgments} Many people have contributed to the ideas, techniques, and organization of this book and have taught courses based on it. Many of the compiler design decisions in this book are drawn from the assignment descriptions of \cite{Dybvig:2010aa}. We also would like to thank John Clements, Bor-Yuh Evan Chang, Daniel P. Friedman, Ronald Garcia, Abdulaziz Ghuloum, Jay McCarthy, Nate Nystrom, Dipanwita Sarkar, Oscar Waddell, and Michael Wollowski. \mbox{}\\ \noindent Jeremy G. Siek \\ \noindent \url{http://homes.soic.indiana.edu/jsiek} \\ %\noindent Spring 2016 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Preliminaries} \label{ch:trees-recur} In this chapter we review the basic tools that are needed to implement a compiler. Programs are typically input by a programmer as text, i.e., a sequence of characters. The program-as-text representation is called \emph{concrete syntax}. We use concrete syntax to concisely write down and talk about programs. Inside the compiler, we use \emph{abstract syntax trees} (ASTs) to represent programs in a way that efficiently supports the operations that the compiler needs to perform.\index{concrete syntax}\index{abstract syntax}\index{abstract syntax tree}\index{AST}\index{program}\index{parse} The translation from concrete syntax to abstract syntax is a process called \emph{parsing}~\citep{Aho:1986qf}. We do not cover the theory and implementation of parsing in this book. A parser is provided in the supporting materials for translating from concrete to abstract syntax. ASTs can be represented in many different ways inside the compiler, depending on the programming language used to write the compiler. % We use Racket's \href{https://docs.racket-lang.org/guide/define-struct.html}{\code{struct}} feature to represent ASTs (Section~\ref{sec:ast}). We use grammars to define the abstract syntax of programming languages (Section~\ref{sec:grammar}) and pattern matching to inspect individual nodes in an AST (Section~\ref{sec:pattern-matching}). We use recursive functions to construct and deconstruct entire ASTs (Section~\ref{sec:recursion}). This chapter provides an brief introduction to these ideas. \index{struct} \section{Abstract Syntax Trees and Racket Structures} \label{sec:ast} Compilers use abstract syntax trees to represent programs because they often need to ask questions like: for a given part of a program, what kind of language feature is it? What are its sub-parts? Consider the program on the left and its AST on the right. This program is an addition and it has two sub-parts, a read operation and a negation. The negation has another sub-part, the integer constant \code{8}. By using a tree to represent the program, we can easily follow the links to go from one part of a program to its sub-parts. \begin{center} \begin{minipage}{0.4\textwidth} \begin{lstlisting} (+ (read) (- 8)) \end{lstlisting} \end{minipage} \begin{minipage}{0.4\textwidth} \begin{equation} \begin{tikzpicture} \node[draw, circle] (plus) at (0 , 0) {\key{+}}; \node[draw, circle] (read) at (-1, -1.5) {{\footnotesize\key{read}}}; \node[draw, circle] (minus) at (1 , -1.5) {$\key{-}$}; \node[draw, circle] (8) at (1 , -3) {\key{8}}; \draw[->] (plus) to (read); \draw[->] (plus) to (minus); \draw[->] (minus) to (8); \end{tikzpicture} \label{eq:arith-prog} \end{equation} \end{minipage} \end{center} We use the standard terminology for trees to describe ASTs: each circle above is called a \emph{node}. The arrows connect a node to its \emph{children} (which are also nodes). The top-most node is the \emph{root}. Every node except for the root has a \emph{parent} (the node it is the child of). If a node has no children, it is a \emph{leaf} node. Otherwise it is an \emph{internal} node. \index{node} \index{children} \index{root} \index{parent} \index{leaf} \index{internal node} %% Recall that an \emph{symbolic expression} (S-expression) is either %% \begin{enumerate} %% \item an atom, or %% \item a pair of two S-expressions, written $(e_1 \key{.} e_2)$, %% where $e_1$ and $e_2$ are each an S-expression. %% \end{enumerate} %% An \emph{atom} can be a symbol, such as \code{`hello}, a number, the %% null value \code{'()}, etc. We can create an S-expression in Racket %% simply by writing a backquote (called a quasi-quote in Racket) %% followed by the textual representation of the S-expression. It is %% quite common to use S-expressions to represent a list, such as $a, b %% ,c$ in the following way: %% \begin{lstlisting} %% `(a . (b . (c . ()))) %% \end{lstlisting} %% Each element of the list is in the first slot of a pair, and the %% second slot is either the rest of the list or the null value, to mark %% the end of the list. Such lists are so common that Racket provides %% special notation for them that removes the need for the periods %% and so many parenthesis: %% \begin{lstlisting} %% `(a b c) %% \end{lstlisting} %% The following expression creates an S-expression that represents AST %% \eqref{eq:arith-prog}. %% \begin{lstlisting} %% `(+ (read) (- 8)) %% \end{lstlisting} %% When using S-expressions to represent ASTs, the convention is to %% represent each AST node as a list and to put the operation symbol at %% the front of the list. The rest of the list contains the children. So %% in the above case, the root AST node has operation \code{`+} and its %% two children are \code{`(read)} and \code{`(- 8)}, just as in the %% diagram \eqref{eq:arith-prog}. %% To build larger S-expressions one often needs to splice together %% several smaller S-expressions. Racket provides the comma operator to %% splice an S-expression into a larger one. For example, instead of %% creating the S-expression for AST \eqref{eq:arith-prog} all at once, %% we could have first created an S-expression for AST %% \eqref{eq:arith-neg8} and then spliced that into the addition %% S-expression. %% \begin{lstlisting} %% (define ast1.4 `(- 8)) %% (define ast1.1 `(+ (read) ,ast1.4)) %% \end{lstlisting} %% In general, the Racket expression that follows the comma (splice) %% can be any expression that produces an S-expression. We define a Racket \code{struct} for each kind of node. For this chapter we require just two kinds of nodes: one for integer constants and one for primitive operations. The following is the \code{struct} definition for integer constants. \begin{lstlisting} (struct Int (value)) \end{lstlisting} An integer node includes just one thing: the integer value. To create a AST node for the integer $8$, we write \code{(Int 8)}. \begin{lstlisting} (define eight (Int 8)) \end{lstlisting} We say that the value created by \code{(Int 8)} is an \emph{instance} of the \code{Int} structure. The following is the \code{struct} definition for primitives operations. \begin{lstlisting} (struct Prim (op args)) \end{lstlisting} A primitive operation node includes an operator symbol \code{op} and a list of children \code{args}. For example, to create an AST that negates the number $8$, we write \code{(Prim '- (list eight))}. \begin{lstlisting} (define neg-eight (Prim '- (list eight))) \end{lstlisting} Primitive operations may have zero or more children. The \code{read} operator has zero children: \begin{lstlisting} (define rd (Prim 'read '())) \end{lstlisting} whereas the addition operator has two children: \begin{lstlisting} (define ast1.1 (Prim '+ (list rd neg-eight))) \end{lstlisting} We have made a design choice regarding the \code{Prim} structure. Instead of using one structure for many different operations (\code{read}, \code{+}, and \code{-}), we could have instead defined a structure for each operation, as follows. \begin{lstlisting} (struct Read ()) (struct Add (left right)) (struct Neg (value)) \end{lstlisting} The reason we choose to use just one structure is that in many parts of the compiler the code for the different primitive operators is the same, so we might as well just write that code once, which is enabled by using a single structure. When compiling a program such as \eqref{eq:arith-prog}, we need to know that the operation associated with the root node is addition and we need to be able to access its two children. Racket provides pattern matching over structures to support these kinds of queries, as we see in Section~\ref{sec:pattern-matching}. In this book, we often write down the concrete syntax of a program even when we really have in mind the AST because the concrete syntax is more concise. We recommend that, in your mind, you always think of programs as abstract syntax trees. \section{Grammars} \label{sec:grammar} \index{integer} \index{literal} \index{constant} A programming language can be thought of as a \emph{set} of programs. The set is typically infinite (one can always create larger and larger programs), so one cannot simply describe a language by listing all of the programs in the language. Instead we write down a set of rules, a \emph{grammar}, for building programs. Grammars are often used to define the concrete syntax of a language, but they can also be used to describe the abstract syntax. We write our rules in a variant of Backus-Naur Form (BNF)~\citep{Backus:1960aa,Knuth:1964aa}. \index{Backus-Naur Form}\index{BNF} As an example, we describe a small language, named $R_0$, that consists of integers and arithmetic operations. \index{grammar} The first grammar rule for the abstract syntax of $R_0$ says that an instance of the \code{Int} structure is an expression: \begin{equation} \Exp ::= \INT{\Int} \label{eq:arith-int} \end{equation} % Each rule has a left-hand-side and a right-hand-side. The way to read a rule is that if you have an AST node that matches the right-hand-side, then you can categorize it according to the left-hand-side. % A name such as $\Exp$ that is defined by the grammar rules is a \emph{non-terminal}. \index{non-terminal} % The name $\Int$ is a also a non-terminal, but instead of defining it with a grammar rule, we define it with the following explanation. We make the simplifying design decision that all of the languages in this book only handle machine-representable integers. On most modern machines this corresponds to integers represented with 64-bits, i.e., the in range $-2^{63}$ to $2^{63}-1$. We restrict this range further to match the Racket \texttt{fixnum} datatype, which allows 63-bit integers on a 64-bit machine. So an $\Int$ is a sequence of decimals ($0$ to $9$), possibly starting with $-$ (for negative integers), such that the sequence of decimals represent an integer in range $-2^{62}$ to $2^{62}-1$. The second grammar rule is the \texttt{read} operation that receives an input integer from the user of the program. \begin{equation} \Exp ::= \READ{} \label{eq:arith-read} \end{equation} The third rule says that, given an $\Exp$ node, you can build another $\Exp$ node by negating it. \begin{equation} \Exp ::= \NEG{\Exp} \label{eq:arith-neg} \end{equation} Symbols in typewriter font such as \key{-} and \key{read} are \emph{terminal} symbols and must literally appear in the program for the rule to be applicable. \index{terminal} We can apply these rules to build ASTs in the $R_0$ language. By rule \eqref{eq:arith-int}, \texttt{(Int 8)} is an $\Exp$, then by rule \eqref{eq:arith-neg}, the following AST is an $\Exp$. \begin{center} \begin{minipage}{0.4\textwidth} \begin{lstlisting} (Prim '- (list (Int 8))) \end{lstlisting} \end{minipage} \begin{minipage}{0.25\textwidth} \begin{equation} \begin{tikzpicture} \node[draw, circle] (minus) at (0, 0) {$\text{--}$}; \node[draw, circle] (8) at (0, -1.2) {$8$}; \draw[->] (minus) to (8); \end{tikzpicture} \label{eq:arith-neg8} \end{equation} \end{minipage} \end{center} The next grammar rule defines addition expressions: \begin{equation} \Exp ::= \ADD{\Exp}{\Exp} \label{eq:arith-add} \end{equation} We can now justify that the AST \eqref{eq:arith-prog} is an $\Exp$ in $R_0$. We know that \lstinline{(Prim 'read '())} is an $\Exp$ by rule \eqref{eq:arith-read} and we have already categorized \code{(Prim '- (list (Int 8)))} as an $\Exp$, so we apply rule \eqref{eq:arith-add} to show that \begin{lstlisting} (Prim '+ (list (Prim 'read '()) (Prim '- (list (Int 8))))) \end{lstlisting} is an $\Exp$ in the $R_0$ language. If you have an AST for which the above rules do not apply, then the AST is not in $R_0$. For example, the program \code{(- (read) (+ 8))} is not in $R_0$ because there are no rules for \code{+} with only one argument, nor for \key{-} with two arguments. Whenever we define a language with a grammar, the language only includes those programs that are justified by the rules. The last grammar rule for $R_0$ states that there is a \code{Program} node to mark the top of the whole program: \[ R_0 ::= \PROGRAM{\code{'()}}{\Exp} \] The \code{Program} structure is defined as follows \begin{lstlisting} (struct Program (info body)) \end{lstlisting} where \code{body} is an expression. In later chapters, the \code{info} part will be used to store auxiliary information but for now it is just the empty list. It is common to have many grammar rules with the same left-hand side but different right-hand sides, such as the rules for $\Exp$ in the grammar of $R_0$. As a short-hand, a vertical bar can be used to combine several right-hand-sides into a single rule. We collect all of the grammar rules for the abstract syntax of $R_0$ in Figure~\ref{fig:r0-syntax}. The concrete syntax for $R_0$ is defined in Figure~\ref{fig:r0-concrete-syntax}. The \code{read-program} function provided in \code{utilities.rkt} of the support materials reads a program in from a file (the sequence of characters in the concrete syntax of Racket) and parses it into an abstract syntax tree. See the description of \code{read-program} in Appendix~\ref{appendix:utilities} for more details. \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{rcl} \begin{array}{rcl} \Exp &::=& \Int \mid \LP\key{read}\RP \mid \LP\key{-}\;\Exp\RP \mid \LP\key{+} \; \Exp\;\Exp\RP\\ R_0 &::=& \Exp \end{array} \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_0$.} \label{fig:r0-concrete-syntax} \end{figure} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{rcl} \Exp &::=& \INT{\Int} \mid \READ{} \mid \NEG{\Exp} \\ &\mid& \ADD{\Exp}{\Exp} \\ R_0 &::=& \PROGRAM{\code{'()}}{\Exp} \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_0$.} \label{fig:r0-syntax} \end{figure} \section{Pattern Matching} \label{sec:pattern-matching} As mentioned in Section~\ref{sec:ast}, compilers often need to access the parts of an AST node. Racket provides the \texttt{match} form to access the parts of a structure. Consider the following example and the output on the right. \index{match} \index{pattern matching} \begin{center} \begin{minipage}{0.5\textwidth} \begin{lstlisting} (match ast1.1 [(Prim op (list child1 child2)) (print op)]) \end{lstlisting} \end{minipage} \vrule \begin{minipage}{0.25\textwidth} \begin{lstlisting} '+ \end{lstlisting} \end{minipage} \end{center} In the above example, the \texttt{match} form takes the AST \eqref{eq:arith-prog} and binds its parts to the three pattern variables \texttt{op}, \texttt{child1}, and \texttt{child2}. In general, a match clause consists of a \emph{pattern} and a \emph{body}. \index{pattern} Patterns are recursively defined to be either a pattern variable, a structure name followed by a pattern for each of the structure's arguments, or an S-expression (symbols, lists, etc.). (See Chapter 12 of The Racket Guide\footnote{\url{https://docs.racket-lang.org/guide/match.html}} and Chapter 9 of The Racket Reference\footnote{\url{https://docs.racket-lang.org/reference/match.html}} for a complete description of \code{match}.) % The body of a match clause may contain arbitrary Racket code. The pattern variables can be used in the scope of the body, such as \code{op} in \code{(print op)}. A \code{match} form may contain several clauses, as in the following function \code{leaf?} that recognizes when an $R_0$ node is a leaf in the AST. The \code{match} proceeds through the clauses in order, checking whether the pattern can match the input AST. The body of the first clause that matches is executed. The output of \code{leaf?} for several ASTs is shown on the right. \begin{center} \begin{minipage}{0.6\textwidth} \begin{lstlisting} (define (leaf? arith) (match arith [(Int n) #t] [(Prim 'read '()) #t] [(Prim '- (list e1)) #f] [(Prim '+ (list e1 e2)) #f])) (leaf? (Prim 'read '())) (leaf? (Prim '- (list (Int 8)))) (leaf? (Int 8)) \end{lstlisting} \end{minipage} \vrule \begin{minipage}{0.25\textwidth} \begin{lstlisting} #t #f #t \end{lstlisting} \end{minipage} \end{center} When writing a \code{match}, we refer to the grammar definition to identify which non-terminal we are expecting to match against, then we make sure that 1) we have one clause for each alternative of that non-terminal and 2) that the pattern in each clause corresponds to the corresponding right-hand side of a grammar rule. For the \code{match} in the \code{leaf?} function, we refer to the grammar for $R_0$ in Figure~\ref{fig:r0-syntax}. The $\Exp$ non-terminal has 4 alternatives, so the \code{match} has 4 clauses. The pattern in each clause corresponds to the right-hand side of a grammar rule. For example, the pattern \code{(Prim '+ (list e1 e2))} corresponds to the right-hand side $\ADD{\Exp}{\Exp}$. When translating from grammars to patterns, replace non-terminals such as $\Exp$ with pattern variables of your choice (e.g. \code{e1} and \code{e2}). \section{Recursive Functions} \label{sec:recursion} \index{recursive function} Programs are inherently recursive. For example, an $R_0$ expression is often made of smaller expressions. Thus, the natural way to process an entire program is with a recursive function. As a first example of such a recursive function, we define \texttt{exp?} below, which takes an arbitrary value and determines whether or not it is an $R_0$ expression. % We say that a function is defined by \emph{structural recursion} when it is defined using a sequence of match clauses that correspond to a grammar, and the body of each clause makes a recursive call on each child node.\footnote{This principle of structuring code according to the data definition is advocated in the book \emph{How to Design Programs}\url{http://www.ccs.neu.edu/home/matthias/HtDP2e/}.}. Below we also define a second function, named \code{R0?}, that determines whether an AST is an $R_0$ program. In general we can expect to write one recursive function to handle each non-terminal in a grammar.\index{structural recursion} % \begin{center} \begin{minipage}{0.7\textwidth} \begin{lstlisting} (define (exp? ast) (match ast [(Int n) #t] [(Prim 'read '()) #t] [(Prim '- (list e)) (exp? e)] [(Prim '+ (list e1 e2)) (and (exp? e1) (exp? e2))] [else #f])) (define (R0? ast) (match ast [(Program '() e) (exp? e)] [else #f])) (R0? (Program '() ast1.1) (R0? (Program '() (Prim '- (list (Prim 'read '()) (Prim '+ (list (Num 8))))))) \end{lstlisting} \end{minipage} \vrule \begin{minipage}{0.25\textwidth} \begin{lstlisting} #t #f \end{lstlisting} \end{minipage} \end{center} You may be tempted to merge the two functions into one, like this: \begin{center} \begin{minipage}{0.5\textwidth} \begin{lstlisting} (define (R0? ast) (match ast [(Int n) #t] [(Prim 'read '()) #t] [(Prim '- (list e)) (R0? e)] [(Prim '+ (list e1 e2)) (and (R0? e1) (R0? e2))] [(Program '() e) (R0? e)] [else #f])) \end{lstlisting} \end{minipage} \end{center} % Sometimes such a trick will save a few lines of code, especially when it comes to the \code{Program} wrapper. Yet this style is generally \emph{not} recommended because it can get you into trouble. % For example, the above function is subtly wrong: \lstinline{(R0? (Program '() (Program '() (Int 3))))} would return true, when it should return false. \section{Interpreters} \label{sec:interp-R0} \index{interpreter} In general, the intended behavior of a program is defined by the specification of the language. For example, the Scheme language is defined in the report by \cite{SPERBER:2009aa}. The Racket language is defined in its reference manual~\citep{plt-tr}. In this book we use interpreters to specify each language that we consider. An interpreter that is designated as the definition of a language is called a \emph{definitional interpreter}~\citep{reynolds72:_def_interp}. \index{definitional interpreter} We warm up by creating a definitional interpreter for the $R_0$ language, which serves as a second example of structural recursion. The \texttt{interp-R0} function is defined in Figure~\ref{fig:interp-R0}. The body of the function is a match on the input program followed by a call to the \lstinline{interp-exp} helper function, which in turn has one match clause per grammar rule for $R_0$ expressions. \begin{figure}[tp] \begin{lstlisting} (define (interp-exp e) (match e [(Int n) n] [(Prim 'read '()) (define r (read)) (cond [(fixnum? r) r] [else (error 'interp-R0 "expected an integer" r)])] [(Prim '- (list e)) (define v (interp-exp e)) (fx- 0 v)] [(Prim '+ (list e1 e2)) (define v1 (interp-exp e1)) (define v2 (interp-exp e2)) (fx+ v1 v2)] )) (define (interp-R0 p) (match p [(Program '() e) (interp-exp e)] )) \end{lstlisting} \caption{Interpreter for the $R_0$ language.} \label{fig:interp-R0} \end{figure} Let us consider the result of interpreting a few $R_0$ programs. The following program adds two integers. \begin{lstlisting} (+ 10 32) \end{lstlisting} The result is \key{42}. We wrote the above program in concrete syntax, whereas the parsed abstract syntax is: \begin{lstlisting} (Program '() (Prim '+ (list (Int 10) (Int 32)))) \end{lstlisting} The next example demonstrates that expressions may be nested within each other, in this case nesting several additions and negations. \begin{lstlisting} (+ 10 (- (+ 12 20))) \end{lstlisting} What is the result of the above program? As mentioned previously, the $R_0$ language does not support arbitrarily-large integers, but only $63$-bit integers, so we interpret the arithmetic operations of $R_0$ using fixnum arithmetic in Racket. Suppose \[ n = 999999999999999999 \] which indeed fits in $63$-bits. What happens when we run the following program in our interpreter? \begin{lstlisting} (+ (+ (+ |$n$| |$n$|) (+ |$n$| |$n$|)) (+ (+ |$n$| |$n$|) (+ |$n$| |$n$|))))) \end{lstlisting} It produces an error: \begin{lstlisting} fx+: result is not a fixnum \end{lstlisting} We establish the convention that if running the definitional interpreter on a program produces an error other than \code{trapped-error}, then the meaning of that program is \emph{unspecified}\index{unspecified behavior}. That means a compiler for the language is under no obligations regarding that program; it may or may not produce an executable, and if it does, that executable can do anything. On the other hand, if the error is a \code{trapped-error}, then the compiled program is also required to report that an error occurred. To signal an error, exit with a return code of \code{255}. The interpreters in chapters \ref{ch:type-dynamic} and \ref{ch:gradual-typing} use \code{trapped-error}. %% This convention applies to the languages defined in this %% book, as a way to simplify the student's task of implementing them, %% but this convention is not applicable to all programming languages. %% Moving on to the last feature of the $R_0$ language, the \key{read} operation prompts the user of the program for an integer. Recall that program \eqref{eq:arith-prog} performs a \key{read} and then subtracts \code{8}. So if we run \begin{lstlisting} (interp-R0 (Program '() ast1.1)) \end{lstlisting} and if the input is \code{50}, then we get the answer to life, the universe, and everything: \code{42}!\footnote{\emph{The Hitchhiker's Guide to the Galaxy} by Douglas Adams.} We include the \key{read} operation in $R_0$ so a clever student cannot implement a compiler for $R_0$ that simply runs the interpreter during compilation to obtain the output and then generates the trivial code to produce the output. (Yes, a clever student did this in the first instance of this course.) The job of a compiler is to translate a program in one language into a program in another language so that the output program behaves the same way as the input program does according to its definitional interpreter. This idea is depicted in the following diagram. Suppose we have two languages, $\mathcal{L}_1$ and $\mathcal{L}_2$, and an interpreter for each language. Suppose that the compiler translates program $P_1$ in language $\mathcal{L}_1$ into program $P_2$ in language $\mathcal{L}_2$. Then interpreting $P_1$ and $P_2$ on their respective interpreters with input $i$ should yield the same output $o$. \begin{equation} \label{eq:compile-correct} \begin{tikzpicture}[baseline=(current bounding box.center)] \node (p1) at (0, 0) {$P_1$}; \node (p2) at (3, 0) {$P_2$}; \node (o) at (3, -2.5) {$o$}; \path[->] (p1) edge [above] node {compile} (p2); \path[->] (p2) edge [right] node {interp-$\mathcal{L}_2$($i$)} (o); \path[->] (p1) edge [left] node {interp-$\mathcal{L}_1$($i$)} (o); \end{tikzpicture} \end{equation} In the next section we see our first example of a compiler. \section{Example Compiler: a Partial Evaluator} \label{sec:partial-evaluation} In this section we consider a compiler that translates $R_0$ programs into $R_0$ programs that may be more efficient, that is, this compiler is an optimizer. This optimizer eagerly computes the parts of the program that do not depend on any inputs, a process known as \emph{partial evaluation}~\cite{Jones:1993uq}. \index{partial evaluation} For example, given the following program \begin{lstlisting} (+ (read) (- (+ 5 3))) \end{lstlisting} our compiler will translate it into the program \begin{lstlisting} (+ (read) -8) \end{lstlisting} Figure~\ref{fig:pe-arith} gives the code for a simple partial evaluator for the $R_0$ language. The output of the partial evaluator is an $R_0$ program. In Figure~\ref{fig:pe-arith}, the structural recursion over $\Exp$ is captured in the \code{pe-exp} function whereas the code for partially evaluating the negation and addition operations is factored into two separate helper functions: \code{pe-neg} and \code{pe-add}. The input to these helper functions is the output of partially evaluating the children. \begin{figure}[tp] \begin{lstlisting} (define (pe-neg r) (match r [(Int n) (Int (fx- 0 n))] [else (Prim '- (list r))])) (define (pe-add r1 r2) (match* (r1 r2) [((Int n1) (Int n2)) (Int (fx+ n1 n2))] [(_ _) (Prim '+ (list r1 r2))])) (define (pe-exp e) (match e [(Int n) (Int n)] [(Prim 'read '()) (Prim 'read '())] [(Prim '- (list e1)) (pe-neg (pe-exp e1))] [(Prim '+ (list e1 e2)) (pe-add (pe-exp e1) (pe-exp e2))] )) (define (pe-R0 p) (match p [(Program '() e) (Program '() (pe-exp e))] )) \end{lstlisting} \caption{A partial evaluator for $R_0$ expressions.} \label{fig:pe-arith} \end{figure} The \texttt{pe-neg} and \texttt{pe-add} functions check whether their arguments are integers and if they are, perform the appropriate arithmetic. Otherwise, they create an AST node for the operation (either negation or addition). To gain some confidence that the partial evaluator is correct, we can test whether it produces programs that get the same result as the input programs. That is, we can test whether it satisfies Diagram \eqref{eq:compile-correct}. The following code runs the partial evaluator on several examples and tests the output program. The \texttt{parse-program} and \texttt{assert} functions are defined in Appendix~\ref{appendix:utilities}.\\ \begin{minipage}{1.0\textwidth} \begin{lstlisting} (define (test-pe p) (assert "testing pe-R0" (equal? (interp-R0 p) (interp-R0 (pe-R0 p))))) (test-pe (parse-program `(program () (+ 10 (- (+ 5 3)))))) (test-pe (parse-program `(program () (+ 1 (+ 3 1))))) (test-pe (parse-program `(program () (- (+ 3 (- 5)))))) \end{lstlisting} \end{minipage} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Integers and Variables} \label{ch:int-exp} This chapter is about compiling the subset of Racket that includes integer arithmetic and local variable binding, which we name $R_1$, to x86-64 assembly code~\citep{Intel:2015aa}. Henceforth we refer to x86-64 simply as x86. The chapter begins with a description of the $R_1$ language (Section~\ref{sec:s0}) followed by a description of x86 (Section~\ref{sec:x86}). The x86 assembly language is large, so we discuss only what is needed for compiling $R_1$. We introduce more of x86 in later chapters. Once we have introduced $R_1$ and x86, we reflect on their differences and come up with a plan to break down the translation from $R_1$ to x86 into a handful of steps (Section~\ref{sec:plan-s0-x86}). The rest of the sections in this chapter give detailed hints regarding each step (Sections~\ref{sec:uniquify-s0} through \ref{sec:patch-s0}). We hope to give enough hints that the well-prepared reader, together with a few friends, can implement a compiler from $R_1$ to x86 in a couple weeks while at the same time leaving room for some fun and creativity. To give the reader a feeling for the scale of this first compiler, the instructor solution for the $R_1$ compiler is less than 500 lines of code. \section{The $R_1$ Language} \label{sec:s0} \index{variable} The $R_1$ language extends the $R_0$ language with variable definitions. The concrete syntax of the $R_1$ language is defined by the grammar in Figure~\ref{fig:r1-concrete-syntax} and the abstract syntax is defined in Figure~\ref{fig:r1-syntax}. The non-terminal \Var{} may be any Racket identifier. As in $R_0$, \key{read} is a nullary operator, \key{-} is a unary operator, and \key{+} is a binary operator. Similar to $R_0$, the abstract syntax of $R_1$ includes the \key{Program} struct to mark the top of the program. %% The $\itm{info}$ %% field of the \key{Program} structure contains an \emph{association %% list} (a list of key-value pairs) that is used to communicate %% auxiliary data from one compiler pass the next. Despite the simplicity of the $R_1$ language, it is rich enough to exhibit several compilation techniques. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{rcl} \Exp &::=& \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp}\\ &\mid& \Var \mid \CLET{\Var}{\Exp}{\Exp} \\ R_1 &::=& \Exp \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_1$.} \label{fig:r1-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{rcl} \Exp &::=& \INT{\Int} \mid \READ{} \\ &\mid& \NEG{\Exp} \mid \ADD{\Exp}{\Exp} \\ &\mid& \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} \\ R_1 &::=& \PROGRAM{\code{'()}}{\Exp} \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_1$.} \label{fig:r1-syntax} \end{figure} Let us dive further into the syntax and semantics of the $R_1$ language. The \key{Let} feature defines a variable for use within its body and initializes the variable with the value of an expression. The abstract syntax for \key{Let} is defined in Figure~\ref{fig:r1-syntax}. The concrete syntax for \key{Let} is \begin{lstlisting} (let ([|$\itm{var}$| |$\itm{exp}$|]) |$\itm{exp}$|) \end{lstlisting} For example, the following program initializes \code{x} to $32$ and then evaluates the body \code{(+ 10 x)}, producing $42$. \begin{lstlisting} (let ([x (+ 12 20)]) (+ 10 x)) \end{lstlisting} When there are multiple \key{let}'s for the same variable, the closest enclosing \key{let} is used. That is, variable definitions overshadow prior definitions. Consider the following program with two \key{let}'s that define variables named \code{x}. Can you figure out the result? \begin{lstlisting} (let ([x 32]) (+ (let ([x 10]) x) x)) \end{lstlisting} For the purposes of depicting which variable uses correspond to which definitions, the following shows the \code{x}'s annotated with subscripts to distinguish them. Double check that your answer for the above is the same as your answer for this annotated version of the program. \begin{lstlisting} (let ([x|$_1$| 32]) (+ (let ([x|$_2$| 10]) x|$_2$|) x|$_1$|)) \end{lstlisting} The initializing expression is always evaluated before the body of the \key{let}, so in the following, the \key{read} for \code{x} is performed before the \key{read} for \code{y}. Given the input $52$ then $10$, the following produces $42$ (not $-42$). \begin{lstlisting} (let ([x (read)]) (let ([y (read)]) (+ x (- y)))) \end{lstlisting} \subsection{Extensible Interpreters via Method Overriding} To prepare for discussing the interpreter for $R_1$, we need to explain why we choose to implement the interpreter using object-oriented programming, that is, as a collection of methods inside of a class. Throughout this book we define many interpreters, one for each of the languages that we study. Because each language builds on the prior one, there is a lot of commonality between their interpreters. We want to write down those common parts just once instead of many times. A naive approach would be to have, for example, the interpreter for $R_2$ handle all of the new features in that language and then have a default case that dispatches to the interpreter for $R_1$. The following code sketches this idea. \begin{center} \begin{minipage}{0.45\textwidth} \begin{lstlisting} (define (interp-R1 e) (match e [(Prim '- (list e)) (define v (interp-R1 e)) (fx- 0 v)] ... )) \end{lstlisting} \end{minipage} \begin{minipage}{0.45\textwidth} \begin{lstlisting} (define (interp-R2 e) (match e [(If cnd thn els) (define b (interp-R2 cnd)) (match b [#t (interp-R2 thn)] [#f (interp-R2 els)])] ... [else (interp-R1 e)] )) \end{lstlisting} \end{minipage} \end{center} The problem with this approach is that it does not handle situations in which an $R_2$ feature, like \code{If}, is nested inside an $R_1$ feature, like the \code{-} operator, as in the following program. \begin{lstlisting} (Prim '- (list (If (Bool #t) (Int 42) (Int 0)))) \end{lstlisting} If we invoke \code{interp-R2} on this program, it dispatches to \code{interp-R1} to handle the \code{-} operator, but then it recurisvely calls \code{interp-R1} again on the argument of \code{-}, which is an \code{If}. But there is no case for \code{If} in \code{interp-R1}, so we get an error! To make our intepreters extensible we need something called \emph{open recursion}\index{open recursion}. That is, a recursive call should always invoke the ``top'' interpreter, even if the recursive call is made from interpreters that are lower down. Object-oriented languages provide open recursion in the form of method overriding\index{method overriding}. The following code sketches this idea for interpreting $R_1$ and $R_2$ using the \href{https://docs.racket-lang.org/guide/classes.html}{\code{class}} \index{class} feature of Racket. We define one class for each language and place a method for interpreting expressions inside each class. The class for $R_2$ inherits from the class for $R_1$ and the method \code{interp-exp} for $R_2$ overrides the \code{interp-exp} for $R_1$. Note that the default case in \code{interp-exp} for $R_2$ uses \code{super} to invoke \code{interp-exp}, and because $R_2$ inherits from $R_1$, that dispatches to the \code{interp-exp} for $R_1$. \begin{center} \begin{minipage}{0.45\textwidth} \begin{lstlisting} (define interp-R1-class (class object% (define/public (interp-exp e) (match e [(Prim '- (list e)) (define v (interp-exp e)) (fx- 0 v)] ... )) ... )) \end{lstlisting} \end{minipage} \begin{minipage}{0.45\textwidth} \begin{lstlisting} (define interp-R2-class (class interp-R1-class (define/override (interp-exp e) (match e [(If cnd thn els) (define b (interp-exp cnd)) (match b [#t (interp-exp thn)] [#f (interp-exp els)])] ... [else (super interp-exp e)] )) ... )) \end{lstlisting} \end{minipage} \end{center} Getting back to the troublesome example, repeated here: \begin{lstlisting} (define e0 (Prim '- (list (If (Bool #t) (Int 42) (Int 0))))) \end{lstlisting} We can invoke the \code{interp-exp} method for $R_2$ on this expression by creating an object of the $R_2$ class and sending it the \code{interp-exp} method with the argument \code{e0}. \begin{lstlisting} (send (new interp-R2-class) interp-exp e0) \end{lstlisting} This will again hit the default case of \code{interp-exp} in $R_2$ and dispatch to the \code{interp-exp} method for $R_1$, which will handle the \code{-} operator. But then for the recursive method call, it will dispatch back to \code{interp-exp} for $R_2$, where the \code{If} will be correctly handled. Thus, method overriding gives us the open recursion that we need to implement our interpreters in an extensible way. \newpage \subsection{Definitional Interpreter for $R_1$} \begin{wrapfigure}[25]{r}[1.0in]{0.6\textwidth} \small \begin{tcolorbox}[title=Association Lists as Dictionaries] An \emph{association list} (alist) is a list of key-value pairs. For example, we can map people to their ages with an alist. \index{alist}\index{association list} \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define ages '((jane . 25) (sam . 24) (kate . 45))) \end{lstlisting} The \emph{dictionary} interface is for mapping keys to values. Every alist implements this interface. \index{dictionary} The package \href{https://docs.racket-lang.org/reference/dicts.html}{\code{racket/dict}} provides many functions for working with dictionaries. Here are a few of them: \begin{description} \item[$\LP\key{dict-ref}\,\itm{dict}\,\itm{key}\RP$] returns the value associated with the given $\itm{key}$. \item[$\LP\key{dict-set}\,\itm{dict}\,\itm{key}\,\itm{val}\RP$] returns a new dictionary that maps $\itm{key}$ to $\itm{val}$ but otherwise is the same as $\itm{dict}$. \item[$\LP\code{in-dict}\,\itm{dict}\RP$] returns the \href{https://docs.racket-lang.org/reference/sequences.html}{sequence} of keys and values in $\itm{dict}$. For example, the following creates a new alist in which the ages are incremented. \end{description} \vspace{-10pt} \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (for/list ([(k v) (in-dict ages)]) (cons k (add1 v))) \end{lstlisting} \end{tcolorbox} \end{wrapfigure} Now that we have explained why we use classes and methods to implement interpreters, we turn to the discussion of the actual interpreter for $R_1$. Figure~\ref{fig:interp-R1} shows the definitional interpreter for the $R_1$ language. It is similar to the interpreter for $R_0$ but it adds two new \key{match} clauses for variables and for \key{let}. For \key{let}, we need a way to communicate the value of a variable to all the uses of a variable. To accomplish this, we maintain a mapping from variables to values. Throughout the compiler we often need to map variables to information about them. We refer to these mappings as \emph{environments}\index{environment} \footnote{Another common term for environment in the compiler literature is \emph{symbol table}\index{symbol table}.}. For simplicity, we use an association list (alist) to represent the environment. The sidebar to the right gives a brief introduction to alists and the \code{racket/dict} package. The \code{interp-R1} function takes the current environment, \code{env}, as an extra parameter. When the interpreter encounters a variable, it finds the corresponding value using the \code{dict-ref} function. When the interpreter encounters a \key{Let}, it evaluates the initializing expression, extends the environment with the result value bound to the variable, using \code{dict-set}, then evaluates the body of the \key{Let}. \begin{figure}[tp] \begin{lstlisting} (define interp-R1-class (class object% (super-new) (define/public ((interp-exp env) e) (match e [(Int n) n] [(Prim 'read '()) (define r (read)) (cond [(fixnum? r) r] [else (error 'interp-exp "expected an integer" r)])] [(Prim '- (list e)) (define v ((interp-exp env) e)) (fx- 0 v)] [(Prim '+ (list e1 e2)) (define v1 ((interp-exp env) e1)) (define v2 ((interp-exp env) e2)) (fx+ v1 v2)] [(Var x) (dict-ref env x)] [(Let x e body) (define new-env (dict-set env x ((interp-exp env) e))) ((interp-exp new-env) body)] )) (define/public (interp-program p) (match p [(Program '() e) ((interp-exp '()) e)] )) )) (define (interp-R1 p) (send (new interp-R1-class) interp-program p)) \end{lstlisting} \caption{Interpreter for the $R_1$ language.} \label{fig:interp-R1} \end{figure} The goal for this chapter is to implement a compiler that translates any program $P_1$ written in the $R_1$ language into an x86 assembly program $P_2$ such that $P_2$ exhibits the same behavior when run on a computer as the $P_1$ program interpreted by \code{interp-R1}. That is, they both output the same integer $n$. We depict this correctness criteria in the following diagram. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (p1) at (0, 0) {$P_1$}; \node (p2) at (4, 0) {$P_2$}; \node (o) at (4, -2) {$n$}; \path[->] (p1) edge [above] node {\footnotesize compile} (p2); \path[->] (p1) edge [left] node {\footnotesize interp-$R_1$} (o); \path[->] (p2) edge [right] node {\footnotesize interp-x86} (o); \end{tikzpicture} \] In the next section we introduce enough of the x86 assembly language to compile $R_1$. \section{The x86$_0$ Assembly Language} \label{sec:x86} \index{x86} Figure~\ref{fig:x86-0-concrete} defines the concrete syntax for the subset of the x86 assembly language needed for this chapter, which we call x86$_0$. % An x86 program begins with a \code{main} label followed by a sequence of instructions. In the grammar, ellipses such as $\ldots$ are used to indicate a sequence of items, e.g., $\Instr \ldots$ is a sequence of instructions.\index{instruction} % An x86 program is stored in the computer's memory and the computer has a \emph{program counter} (PC)\index{program counter}\index{PC} that points to the address of the next instruction to be executed. For most instructions, once the instruction is executed, the program counter is incremented to point to the immediately following instruction in memory. Most x86 instructions take two operands, where each operand is either an integer constant (called \emph{immediate value}\index{immediate value}), a \emph{register}\index{register}, or a memory location. A register is a special kind of variable. Each one holds a 64-bit value; there are 16 registers in the computer and their names are given in Figure~\ref{fig:x86-0-concrete}. The computer's memory as a mapping of 64-bit addresses to 64-bit values% \footnote{This simple story suffices for describing how sequential programs access memory but is not sufficient for multi-threaded programs. However, multi-threaded execution is beyond the scope of this book.}. % We use the AT\&T syntax expected by the GNU assembler, which comes with the \key{gcc} compiler that we use for compiling assembly code to machine code. % Appendix~\ref{sec:x86-quick-reference} is a quick-reference for all of the x86 instructions used in this book. % to do: finish treatment of imulq % it's needed for vector's in R6/R7 \newcommand{\allregisters}{\key{rsp} \mid \key{rbp} \mid \key{rax} \mid \key{rbx} \mid \key{rcx} \mid \key{rdx} \mid \key{rsi} \mid \key{rdi} \mid \\ && \key{r8} \mid \key{r9} \mid \key{r10} \mid \key{r11} \mid \key{r12} \mid \key{r13} \mid \key{r14} \mid \key{r15}} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Reg &::=& \allregisters{} \\ \Arg &::=& \key{\$}\Int \mid \key{\%}\Reg \mid \Int\key{(}\key{\%}\Reg\key{)}\\ \Instr &::=& \key{addq} \; \Arg\key{,} \Arg \mid \key{subq} \; \Arg\key{,} \Arg \mid \key{negq} \; \Arg \mid \key{movq} \; \Arg\key{,} \Arg \mid \\ && \key{callq} \; \mathit{label} \mid \key{pushq}\;\Arg \mid \key{popq}\;\Arg \mid \key{retq} \mid \key{jmp}\,\itm{label} \\ && \itm{label}\key{:}\; \Instr \\ x86_0 &::= & \key{.globl main}\\ & & \key{main:} \; \Instr\ldots \end{array} \] \end{minipage} } \caption{The syntax of the x86$_0$ assembly language (AT\&T syntax).} \label{fig:x86-0-concrete} \end{figure} An immediate value is written using the notation \key{\$}$n$ where $n$ is an integer. % A register is written with a \key{\%} followed by the register name, such as \key{\%rax}. % An access to memory is specified using the syntax $n(\key{\%}r)$, which obtains the address stored in register $r$ and then adds $n$ bytes to the address. The resulting address is used to either load or store to memory depending on whether it occurs as a source or destination argument of an instruction. An arithmetic instruction such as $\key{addq}\,s\key{,}\,d$ reads from the source $s$ and destination $d$, applies the arithmetic operation, then writes the result back to the destination $d$. % The move instruction $\key{movq}\,s\key{,}\,d$ reads from $s$ and stores the result in $d$. % The $\key{callq}\,\itm{label}$ instruction executes the procedure specified by the label and $\key{retq}$ returns from a procedure to its caller. The abstract syntax for \code{callq} includes an extra integer field that represents the arity (number of parameters) of the function being called. % We discuss procedure calls in more detail later in this chapter and in Chapter~\ref{ch:functions}. The $\key{jmp}\,\itm{label}$ instruction updates the program counter to the address of the instruction after the specified label. Figure~\ref{fig:p0-x86} depicts an x86 program that is equivalent to \code{(+ 10 32)}. The \key{globl} directive says that the \key{main} procedure is externally visible, which is necessary so that the operating system can call it. The label \key{main:} indicates the beginning of the \key{main} procedure which is where the operating system starts executing this program. The instruction \lstinline{movq $10, %rax} puts $10$ into register \key{rax}. The following instruction \lstinline{addq $32, %rax} adds $32$ to the $10$ in \key{rax} and puts the result, $42$, back into \key{rax}. % The last instruction, \key{retq}, finishes the \key{main} function by returning the integer in \key{rax} to the operating system. The operating system interprets this integer as the program's exit code. By convention, an exit code of 0 indicates that a program completed successfully, and all other exit codes indicate various errors. Nevertheless, we return the result of the program as the exit code. %\begin{wrapfigure}{r}{2.25in} \begin{figure}[tbp] \begin{lstlisting} .globl main main: movq $10, %rax addq $32, %rax retq \end{lstlisting} \caption{An x86 program equivalent to \code{(+ 10 32)}.} \label{fig:p0-x86} %\end{wrapfigure} \end{figure} Unfortunately, x86 varies in a couple ways depending on what operating system it is assembled in. The code examples shown here are correct on Linux and most Unix-like platforms, but when assembled on Mac OS X, labels like \key{main} must be prefixed with an underscore, as in \key{\_main}. We exhibit the use of memory for storing intermediate results in the next example. Figure~\ref{fig:p1-x86} lists an x86 program that is equivalent to \code{(+ 52 (- 10))}. This program uses a region of memory called the \emph{procedure call stack} (or \emph{stack} for short). \index{stack}\index{procedure call stack} The stack consists of a separate \emph{frame}\index{frame} for each procedure call. The memory layout for an individual frame is shown in Figure~\ref{fig:frame}. The register \key{rsp} is called the \emph{stack pointer}\index{stack pointer} and points to the item at the top of the stack. The stack grows downward in memory, so we increase the size of the stack by subtracting from the stack pointer. In the context of a procedure call, the \emph{return address}\index{return address} is the instruction after the call instruction on the caller side. The function call instruction, \code{callq}, pushes the return address onto the stack. The register \key{rbp} is the \emph{base pointer}\index{base pointer} and is used to access variables associated with the current procedure call. The base pointer of the caller is pushed onto the stack after the return address. We number the variables from $1$ to $n$. Variable $1$ is stored at address $-8\key{(\%rbp)}$, variable $2$ at $-16\key{(\%rbp)}$, etc. \begin{figure}[tbp] \begin{lstlisting} start: movq $10, -8(%rbp) negq -8(%rbp) movq -8(%rbp), %rax addq $52, %rax jmp conclusion .globl main main: pushq %rbp movq %rsp, %rbp subq $16, %rsp jmp start conclusion: addq $16, %rsp popq %rbp retq \end{lstlisting} \caption{An x86 program equivalent to \code{(+ 10 32)}.} \label{fig:p1-x86} \end{figure} \begin{figure}[tbp] \centering \begin{tabular}{|r|l|} \hline Position & Contents \\ \hline 8(\key{\%rbp}) & return address \\ 0(\key{\%rbp}) & old \key{rbp} \\ -8(\key{\%rbp}) & variable $1$ \\ -16(\key{\%rbp}) & variable $2$ \\ \ldots & \ldots \\ 0(\key{\%rsp}) & variable $n$\\ \hline \end{tabular} \caption{Memory layout of a frame.} \label{fig:frame} \end{figure} Getting back to the program in Figure~\ref{fig:p1-x86}, consider how control is transferred from the operating system to the \code{main} function. The operating system issues a \code{callq main} instruction which pushes its return address on the stack and then jumps to \code{main}. In x86-64, the stack pointer \code{rsp} must be divisible by 16 bytes prior to the execution of any \code{callq} instruction, so when control arrives at \code{main}, the \code{rsp} is 8 bytes out of alignment (because the \code{callq} pushed the return address). The first three instructions are the typical \emph{prelude}\index{prelude} for a procedure. The instruction \code{pushq \%rbp} saves the base pointer for the caller onto the stack and subtracts $8$ from the stack pointer. At this point the stack pointer is back to being 16-byte aligned. The second instruction \code{movq \%rsp, \%rbp} changes the base pointer so that it points the location of the old base pointer. The instruction \code{subq \$16, \%rsp} moves the stack pointer down to make enough room for storing variables. This program needs one variable ($8$ bytes) but we round up to 16 bytes to maintain the 16-byte alignment of the \code{rsp}. With the \code{rsp} aligned, we are ready to make calls to other functions. The last instruction of the prelude is \code{jmp start}, which transfers control to the instructions that were generated from the Racket expression \code{(+ 10 32)}. The four instructions under the label \code{start} carry out the work of computing \code{(+ 52 (- 10)))}. % The first instruction \code{movq \$10, -8(\%rbp)} stores $10$ in variable $1$. % The instruction \code{negq -8(\%rbp)} changes variable $1$ to $-10$. % The following instruction moves the $-10$ from variable $1$ into the \code{rax} register. Finally, \code{addq \$52, \%rax} adds $52$ to the value in \code{rax}, updating its contents to $42$. The three instructions under the label \code{conclusion} are the typical \emph{conclusion}\index{conclusion} of a procedure. The first two instructions are necessary to get the state of the machine back to where it was at the beginning of the procedure. The instruction \key{addq \$16, \%rsp} moves the stack pointer back to point at the old base pointer. The amount added here needs to match the amount that was subtracted in the prelude of the procedure. Then \key{popq \%rbp} returns the old base pointer to \key{rbp} and adds $8$ to the stack pointer. The last instruction, \key{retq}, jumps back to the procedure that called this one and adds 8 to the stack pointer, which returns the stack pointer to where it was prior to the procedure call. The compiler needs a convenient representation for manipulating x86 programs, so we define an abstract syntax for x86 in Figure~\ref{fig:x86-0-ast}. We refer to this language as x86$_0$ with a subscript $0$ because later we introduce extended versions of this assembly language. The main difference compared to the concrete syntax of x86 (Figure~\ref{fig:x86-0-concrete}) is that it does not allow labeled instructions to appear anywhere, but instead organizes instructions into a group called a \emph{block}\index{block}\index{basic block} and associates a label with every block, which is why the \key{CFG} struct (for control-flow graph) includes an alist mapping labels to blocks. The reason for this organization becomes apparent in Chapter~\ref{ch:bool-types} when we introduce conditional branching. The \code{Block} structure includes an $\itm{info}$ field that is not needed for this chapter, but will become useful in Chapter~\ref{ch:register-allocation-r1}. For now, the $\itm{info}$ field should just contain an empty list. Also, regarding the abstract syntax for \code{callq}, the \code{Callq} struct includes an integer for representing the arity of the function, i.e., the number of arguments, which is helpful to know during register allocation (Chapter~\ref{ch:register-allocation-r1}). \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Reg &::=& \allregisters{} \\ \Arg &::=& \IMM{\Int} \mid \REG{\Reg} \mid \DEREF{\Reg}{\Int} \\ \Instr &::=& \BININSTR{\code{'addq}}{\Arg}{\Arg} \mid \BININSTR{\code{'subq}}{\Arg}{\Arg} \\ &\mid& \BININSTR{\code{'movq}}{\Arg}{\Arg} \mid \UNIINSTR{\code{'negq}}{\Arg}\\ &\mid& \CALLQ{\itm{label}}{\itm{int}} \mid \RETQ{} \mid \PUSHQ{\Arg} \mid \POPQ{\Arg} \mid \JMP{\itm{label}} \\ \Block &::= & \BLOCK{\itm{info}}{\Instr\ldots} \\ x86_0 &::= & \PROGRAM{\itm{info}}{\CFG{\key{(}\itm{label} \,\key{.}\, \Block \key{)}\ldots}} \end{array} \] \end{minipage} } \caption{The abstract syntax of x86$_0$ assembly.} \label{fig:x86-0-ast} \end{figure} \section{Planning the trip to x86 via the $C_0$ language} \label{sec:plan-s0-x86} To compile one language to another it helps to focus on the differences between the two languages because the compiler will need to bridge those differences. What are the differences between $R_1$ and x86 assembly? Here are some of the most important ones: \begin{enumerate} \item[(a)] x86 arithmetic instructions typically have two arguments and update the second argument in place. In contrast, $R_1$ arithmetic operations take two arguments and produce a new value. An x86 instruction may have at most one memory-accessing argument. Furthermore, some instructions place special restrictions on their arguments. \item[(b)] An argument of an $R_1$ operator can be any expression, whereas x86 instructions restrict their arguments to be integers constants, registers, and memory locations. \item[(c)] The order of execution in x86 is explicit in the syntax: a sequence of instructions and jumps to labeled positions, whereas in $R_1$ the order of evaluation is a left-to-right depth-first traversal of the abstract syntax tree. \item[(d)] An $R_1$ program can have any number of variables whereas x86 has 16 registers and the procedure calls stack. \item[(e)] Variables in $R_1$ can overshadow other variables with the same name. The registers and memory locations of x86 all have unique names or addresses. \end{enumerate} We ease the challenge of compiling from $R_1$ to x86 by breaking down the problem into several steps, dealing with the above differences one at a time. Each of these steps is called a \emph{pass} of the compiler.\index{pass}\index{compiler pass} % This terminology comes from each step traverses (i.e. passes over) the AST of the program. % We begin by sketching how we might implement each pass, and give them names. We then figure out an ordering of the passes and the input/output language for each pass. The very first pass has $R_1$ as its input language and the last pass has x86 as its output language. In between we can choose whichever language is most convenient for expressing the output of each pass, whether that be $R_1$, x86, or new \emph{intermediate languages} of our own design. Finally, to implement each pass we write one recursive function per non-terminal in the grammar of the input language of the pass. \index{intermediate language} \begin{description} \item[Pass \key{select-instructions}] To handle the difference between $R_1$ operations and x86 instructions we convert each $R_1$ operation to a short sequence of instructions that accomplishes the same task. \item[Pass \key{remove-complex-opera*}] To ensure that each subexpression (i.e. operator and operand, and hence the name \key{opera*}) is an \emph{atomic} expression (a variable or integer), we introduce temporary variables to hold the results of subexpressions.\index{atomic expression} \item[Pass \key{explicate-control}] To make the execution order of the program explicit, we convert from the abstract syntax tree representation into a control-flow graph in which each node contains a sequence of statements and the edges between nodes say where to go at the end of the sequence. \item[Pass \key{assign-homes}] To handle the difference between the variables in $R_1$ versus the registers and stack locations in x86, we map each variable to a register or stack location. \item[Pass \key{uniquify}] This pass deals with the shadowing of variables by renaming every variable to a unique name, so that shadowing no longer occurs. \end{description} The next question is: in what order should we apply these passes? This question can be challenging because it is difficult to know ahead of time which orders will be better (easier to implement, produce more efficient code, etc.) so oftentimes trial-and-error is involved. Nevertheless, we can try to plan ahead and make educated choices regarding the ordering. Let us consider the ordering of \key{uniquify} and \key{remove-complex-opera*}. The assignment of subexpressions to temporary variables involves introducing new variables and moving subexpressions, which might change the shadowing of variables and inadvertently change the behavior of the program. But if we apply \key{uniquify} first, this will not be an issue. Of course, this means that in \key{remove-complex-opera*}, we need to ensure that the temporary variables that it creates are unique. What should be the ordering of \key{explicate-control} with respect to \key{uniquify}? The \key{uniquify} pass should come first because \key{explicate-control} changes all the \key{let}-bound variables to become local variables whose scope is the entire program, which would confuse variables with the same name. % Likewise, we place \key{explicate-control} after \key{remove-complex-opera*} because \key{explicate-control} removes the \key{let} form, but it is convenient to use \key{let} in the output of \key{remove-complex-opera*}. % Regarding \key{assign-homes}, it is helpful to place \key{explicate-control} first because \key{explicate-control} changes \key{let}-bound variables into program-scope variables. This means that the \key{assign-homes} pass can read off the variables from the $\itm{info}$ of the \key{Program} AST node instead of traversing the entire program in search of \key{let}-bound variables. Last, we need to decide on the ordering of \key{select-instructions} and \key{assign-homes}. These two passes are intertwined, creating a Gordian Knot. To do a good job of assigning homes, it is helpful to have already determined which instructions will be used, because x86 instructions have restrictions about which of their arguments can be registers versus stack locations. One might want to give preferential treatment to variables that occur in register-argument positions. On the other hand, it may turn out to be impossible to make sure that all such variables are assigned to registers, and then one must redo the selection of instructions. Some compilers handle this problem by iteratively repeating these two passes until a good solution is found. We use a simpler approach in which \key{select-instructions} comes first, followed by the \key{assign-homes}, then a third pass named \key{patch-instructions} that uses a reserved register to patch-up outstanding problems regarding instructions with too many memory accesses. The disadvantage of this approach is some programs may not execute as efficiently as they would if we used the iterative approach and used all of the registers for variables. \begin{figure}[tbp] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R1) at (0,2) {\large $R_1$}; \node (R1-2) at (3,2) {\large $R_1$}; \node (R1-3) at (6,2) {\large $R_1^{\dagger}$}; %\node (C0-1) at (6,0) {\large $C_0$}; \node (C0-2) at (3,0) {\large $C_0$}; \node (x86-2) at (3,-2) {\large $\text{x86}^{*}_0$}; \node (x86-3) at (6,-2) {\large $\text{x86}^{*}_0$}; \node (x86-4) at (9,-2) {\large $\text{x86}_0$}; \node (x86-5) at (12,-2) {\large $\text{x86}^{\dagger}_0$}; \path[->,bend left=15] (R1) edge [above] node {\ttfamily\footnotesize uniquify} (R1-2); \path[->,bend left=15] (R1-2) edge [above] node {\ttfamily\footnotesize remove-complex.} (R1-3); \path[->,bend left=15] (R1-3) edge [right] node {\ttfamily\footnotesize explicate-control} (C0-2); \path[->,bend right=15] (C0-2) edge [left] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend left=15] (x86-2) edge [above] node {\ttfamily\footnotesize assign-homes} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [above] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Overview of the passes for compiling $R_1$. } \label{fig:R1-passes} \end{figure} Figure~\ref{fig:R1-passes} presents the ordering of the compiler passes in the form of a graph. Each pass is an edge and the input/output language of each pass is a node in the graph. The output of \key{uniquify} and \key{remove-complex-opera*} are programs that are still in the $R_1$ language, but the output of the pass \key{explicate-control} is in a different language $C_0$ that is designed to make the order of evaluation explicit in its syntax, which we introduce in the next section. The \key{select-instruction} pass translates from $C_0$ to a variant of x86. The \key{assign-homes} and \key{patch-instructions} passes input and output variants of x86 assembly. The last pass in Figure~\ref{fig:R1-passes} is \key{print-x86}, which converts from the abstract syntax of $\text{x86}_0$ to the concrete syntax of x86. In the next sections we discuss the $C_0$ language and the $\text{x86}^{*}_0$ and $\text{x86}^{\dagger}_0$ dialects of x86. The remainder of this chapter gives hints regarding the implementation of each of the compiler passes in Figure~\ref{fig:R1-passes}. \subsection{The $C_0$ Intermediate Language} The output of \key{explicate-control} is similar to the $C$ language~\citep{Kernighan:1988nx} in that it has separate syntactic categories for expressions and statements, so we name it $C_0$. The abstract syntax for $C_0$ is defined in Figure~\ref{fig:c0-syntax}. (The concrete syntax for $C_0$ is in the Appendix, Figure~\ref{fig:c0-concrete-syntax}.) % The $C_0$ language supports the same operators as $R_1$ but the arguments of operators are restricted to atomic expressions (variables and integers), thanks to the \key{remove-complex-opera*} pass. Instead of \key{Let} expressions, $C_0$ has assignment statements which can be executed in sequence using the \key{Seq} form. A sequence of statements always ends with \key{Return}, a guarantee that is baked into the grammar rules for the \itm{tail} non-terminal. The naming of this non-terminal comes from the term \emph{tail position}\index{tail position}, which refers to an expression that is the last one to execute within a function. (An expression in tail position may contain subexpressions, and those may or may not be in tail position depending on the kind of expression.) A $C_0$ program consists of a control-flow graph (represented as an alist mapping labels to tails). This is more general than necessary for the present chapter, as we do not yet need to introduce \key{goto} for jumping to labels, but it saves us from having to change the syntax of the program construct in Chapter~\ref{ch:bool-types}. For now there will be just one label, \key{start}, and the whole program is its tail. % The $\itm{info}$ field of the \key{Program} form, after the \key{explicate-control} pass, contains a mapping from the symbol \key{locals} to a list of variables, that is, a list of all the variables used in the program. At the start of the program, these variables are uninitialized; they become initialized on their first assignment. \begin{figure}[tbp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Atm &::=& \INT{\Int} \mid \VAR{\Var} \\ \Exp &::=& \Atm \mid \READ{} \mid \NEG{\Atm} \\ &\mid& \ADD{\Atm}{\Atm}\\ \Stmt &::=& \ASSIGN{\VAR{\Var}}{\Exp} \\ \Tail &::= & \RETURN{\Exp} \mid \SEQ{\Stmt}{\Tail} \\ C_0 & ::= & \PROGRAM{\itm{info}}{\CFG{\key{(}\itm{label}\,\key{.}\,\Tail\key{)}\ldots}} \end{array} \] \end{minipage} } \caption{The abstract syntax of the $C_0$ intermediate language.} \label{fig:c0-syntax} \end{figure} \subsection{The dialects of x86} The x86$^{*}_0$ language, pronounced ``pseudo x86'', is the output of the pass \key{select-instructions}. It extends x86$_0$ with an unbounded number of program-scope variables and has looser rules regarding instruction arguments. The x86$^{\dagger}$ language, the output of \key{print-x86}, is the concrete syntax for x86. \section{Uniquify Variables} \label{sec:uniquify-s0} The \code{uniquify} pass compiles arbitrary $R_1$ programs into $R_1$ programs in which every \key{let} uses a unique variable name. For example, the \code{uniquify} pass should translate the program on the left into the program on the right. \\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([x 32]) (+ (let ([x 10]) x) x)) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([x.1 32]) (+ (let ([x.2 10]) x.2) x.1)) \end{lstlisting} \end{minipage} \end{tabular} \\ % The following is another example translation, this time of a program with a \key{let} nested inside the initializing expression of another \key{let}.\\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([x (let ([x 4]) (+ x 1))]) (+ x 2)) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([x.2 (let ([x.1 4]) (+ x.1 1))]) (+ x.2 2)) \end{lstlisting} \end{minipage} \end{tabular} We recommend implementing \code{uniquify} by creating a function named \code{uniquify-exp} that is structurally recursive function and mostly just copies the input program. However, when encountering a \key{let}, it should generate a unique name for the variable (the Racket function \code{gensym} is handy for this) and associate the old name with the new unique name in an alist. The \code{uniquify-exp} function will need to access this alist when it gets to a variable reference, so we add another parameter to \code{uniquify-exp} for the alist. The skeleton of the \code{uniquify-exp} function is shown in Figure~\ref{fig:uniquify-s0}. The function is curried so that it is convenient to partially apply it to a symbol table and then apply it to different expressions, as in the last clause for primitive operations in Figure~\ref{fig:uniquify-s0}. The \href{https://docs.racket-lang.org/reference/for.html#%28form._%28%28lib._racket%2Fprivate%2Fbase..rkt%29._for%2Flist%29%29}{\key{for/list}} form is useful for applying a function to each element of a list to produce a new list. \index{for/list} \begin{exercise} \normalfont % I don't like the italics for exercises. -Jeremy Complete the \code{uniquify} pass by filling in the blanks, that is, implement the clauses for variables and for the \key{let} form. \end{exercise} \begin{figure}[tbp] \begin{lstlisting} (define (uniquify-exp symtab) (lambda (e) (match e [(Var x) ___] [(Int n) (Int n)] [(Let x e body) ___] [(Prim op es) (Prim op (for/list ([e es]) ((uniquify-exp symtab) e)))] ))) (define (uniquify p) (match p [(Program '() e) (Program '() ((uniquify-exp '()) e))] ))) \end{lstlisting} \caption{Skeleton for the \key{uniquify} pass.} \label{fig:uniquify-s0} \end{figure} \begin{exercise} \normalfont % I don't like the italics for exercises. -Jeremy Test your \key{uniquify} pass by creating five example $R_1$ programs and checking whether the output programs produce the same result as the input programs. The $R_1$ programs should be designed to test the most interesting parts of the \key{uniquify} pass, that is, the programs should include \key{let} forms, variables, and variables that overshadow each other. The five programs should be in a subdirectory named \key{tests} and they should have the same file name except for a different integer at the end of the name, followed by the ending \key{.rkt}. Use the \key{interp-tests} function (Appendix~\ref{appendix:utilities}) from \key{utilities.rkt} to test your \key{uniquify} pass on the example programs. See the \key{run-tests.rkt} script in the support code for an example of how to use \key{interp-tests}. The support code is in a \code{github} repository at the following URL: \begin{center}\footnotesize \url{https://github.com/IUCompilerCourse/public-student-support-code} \end{center} \end{exercise} \section{Remove Complex Operands} \label{sec:remove-complex-opera-R1} The \code{remove-complex-opera*} pass compiles $R_1$ programs into $R_1$ programs in which the arguments of operations are atomic expressions. Put another way, this pass removes complex operands\index{complex operand}, such as the expression \code{(- 10)} in the program below. This is accomplished by introducing a new \key{let}-bound variable, binding the complex operand to the new variable, and then using the new variable in place of the complex operand, as shown in the output of \code{remove-complex-opera*} on the right.\\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} % s0_19.rkt \begin{lstlisting} (+ 52 (- 10)) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([tmp.1 (- 10)]) (+ 52 tmp.1)) \end{lstlisting} \end{minipage} \end{tabular} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{rcl} \Atm &::=& \INT{\Int} \mid \VAR{\Var} \\ \Exp &::=& \Atm \mid \READ{} \\ &\mid& \NEG{\Atm} \mid \ADD{\Atm}{\Atm} \\ &\mid& \LET{\Var}{\Exp}{\Exp} \\ R^{\dagger}_1 &::=& \PROGRAM{\code{'()}}{\Exp} \end{array} \] \end{minipage} } \caption{$R_1^{\dagger}$ is $R_1$ in administrative normal form (ANF).} \label{fig:r1-anf-syntax} \end{figure} Figure~\ref{fig:r1-anf-syntax} presents the grammar for the output of this pass, language $R_1^{\dagger}$. The main difference is that operator arguments are required to be atomic expressions. In the literature, this is called \emph{administrative normal form}, or ANF for short~\citep{Danvy:1991fk,Flanagan:1993cg}. \index{administrative normal form} \index{ANF} We recommend implementing this pass with two mutually recursive functions, \code{rco-atom} and \code{rco-exp}. The idea is to apply \code{rco-atom} to subexpressions that are required to be atomic and to apply \code{rco-exp} to subexpressions that can be atomic or complex (see Figure~\ref{fig:r1-anf-syntax}). Both functions take an $R_1$ expression as input. The \code{rco-exp} function returns an expression. The \code{rco-atom} function returns two things: an atomic expression and alist mapping temporary variables to complex subexpressions. You can return multiple things from a function using Racket's \key{values} form and you can receive multiple things from a function call using the \key{define-values} form. If you are not familiar with these features, review the Racket documentation. Also, the \href{https://docs.racket-lang.org/reference/for.html#%28form._%28%28lib._racket%2Fprivate%2Fbase..rkt%29._for%2Flists%29%29}{\code{for/lists}} form is useful for applying a function to each element of a list, in the case where the function returns multiple values. \index{for/lists} The following shows the output of \code{rco-atom} on the expression \code{(- 10)} (using concrete syntax to be concise). \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} (- 10) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} tmp.1 ((tmp.1 . (- 10))) \end{lstlisting} \end{minipage} \end{tabular} Take special care of programs such as the next one that \key{let}-bind variables with integers or other variables. You should leave them unchanged, as shown in to the program on the right \\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} % s0_20.rkt \begin{lstlisting} (let ([a 42]) (let ([b a]) b)) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([a 42]) (let ([b a]) b)) \end{lstlisting} \end{minipage} \end{tabular} \\ A careless implementation of \key{rco-exp} and \key{rco-atom} might produce the following output.\\ \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([tmp.1 42]) (let ([a tmp.1]) (let ([tmp.2 a]) (let ([b tmp.2]) b)))) \end{lstlisting} \end{minipage} \begin{exercise} \normalfont Implement the \code{remove-complex-opera*} pass. Test the new pass on all of the example programs that you created to test the \key{uniquify} pass and create three new example programs that are designed to exercise the interesting code in the \code{remove-complex-opera*} pass. Use the \key{interp-tests} function (Appendix~\ref{appendix:utilities}) from \key{utilities.rkt} to test your passes on the example programs. \end{exercise} \section{Explicate Control} \label{sec:explicate-control-r1} The \code{explicate-control} pass compiles $R_1$ programs into $C_0$ programs that make the order of execution explicit in their syntax. For now this amounts to flattening \key{let} constructs into a sequence of assignment statements. For example, consider the following $R_1$ program.\\ % s0_11.rkt \begin{minipage}{0.96\textwidth} \begin{lstlisting} (let ([y (let ([x 20]) (+ x (let ([x 22]) x)))]) y) \end{lstlisting} \end{minipage}\\ % The output of the previous pass and of \code{explicate-control} is shown below. Recall that the right-hand-side of a \key{let} executes before its body, so the order of evaluation for this program is to assign \code{20} to \code{x.1}, assign \code{22} to \code{x.2}, assign \code{(+ x.1 x.2)} to \code{y}, then return \code{y}. Indeed, the output of \code{explicate-control} makes this ordering explicit.\\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} (let ([y (let ([x.1 20]) (let ([x.2 22]) (+ x.1 x.2)))]) y) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} start: x.1 = 20; x.2 = 22; y = (+ x.1 x.2); return y; \end{lstlisting} \end{minipage} \end{tabular} We recommend implementing \code{explicate-control} using two mutually recursive functions: \code{explicate-tail} and \code{explicate-assign}. The first function should be applied to expressions in tail position whereas the second should be applied to expressions that occur on the right-hand-side of a \key{let}. % The \code{explicate-tail} function takes an $R_1$ expression as input and produces a $C_0$ $\Tail$ (see Figure~\ref{fig:c0-syntax}). % The \code{explicate-assign} function takes an $R_1$ expression, the variable that it is to be assigned to, and $C_0$ code (a $\Tail$) that should come after the assignment (e.g., the code generated for the body of the \key{let}) and returns a $\Tail$. The \code{explicate-assign} function is in accumulator-passing style in that its third parameter is some $C_0$ code that it adds to and returns. The reader might be tempted to instead organize \code{explicate-assign} in a more direct fashion, without the third parameter and perhaps using \code{append} to combine statements. We warn against that alternative because the accumulator-passing style is key to how we generate high-quality code for conditional expressions in Chapter~\ref{ch:bool-types}. The top-level \code{explicate-control} function should invoke \code{explicate-tail} on the body of the \key{Program} AST node. \section{Select Instructions} \label{sec:select-r1} \index{instruction selection} In the \code{select-instructions} pass we begin the work of translating from $C_0$ to $\text{x86}^{*}_0$. The target language of this pass is a variant of x86 that still uses variables, so we add an AST node of the form $\VAR{\itm{var}}$ to the $\text{x86}_0$ abstract syntax of Figure~\ref{fig:x86-0-ast}. We recommend implementing the \code{select-instructions} in terms of three auxiliary functions, one for each of the non-terminals of $C_0$: $\Atm$, $\Stmt$, and $\Tail$. The cases for $\Atm$ are straightforward, variables stay the same and integer constants are changed to immediates: $\INT{n}$ changes to $\IMM{n}$. Next we consider the cases for $\Stmt$, starting with arithmetic operations. For example, in $C_0$ an addition operation can take the form below, to the left of the $\Rightarrow$. To translate to x86, we need to use the \key{addq} instruction which does an in-place update. So we must first move \code{10} to \code{x}. \\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} x = (+ 10 32); \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} movq $10, x addq $32, x \end{lstlisting} \end{minipage} \end{tabular} \\ % There are cases that require special care to avoid generating needlessly complicated code. If one of the arguments of the addition is the same as the left-hand side of the assignment, then there is no need for the extra move instruction. For example, the following assignment statement can be translated into a single \key{addq} instruction.\\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} x = (+ 10 x); \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} addq $10, x \end{lstlisting} \end{minipage} \end{tabular} \\ The \key{read} operation does not have a direct counterpart in x86 assembly, so we have instead implemented this functionality in the C language~\citep{Kernighan:1988nx}, with the function \code{read\_int} in the file \code{runtime.c}. In general, we refer to all of the functionality in this file as the \emph{runtime system}\index{runtime system}, or simply the \emph{runtime} for short. When compiling your generated x86 assembly code, you need to compile \code{runtime.c} to \code{runtime.o} (an ``object file'', using \code{gcc} option \code{-c}) and link it into the executable. For our purposes of code generation, all you need to do is translate an assignment of \key{read} into some variable $\itm{lhs}$ (for left-hand side) into a call to the \code{read\_int} function followed by a move from \code{rax} to the left-hand side. The move from \code{rax} is needed because the return value from \code{read\_int} goes into \code{rax}, as is the case in general. \\ \begin{tabular}{lll} \begin{minipage}{0.3\textwidth} \begin{lstlisting} |$\itm{var}$| = (read); \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.3\textwidth} \begin{lstlisting} callq read_int movq %rax, |$\itm{var}$| \end{lstlisting} \end{minipage} \end{tabular} \\ There are two cases for the $\Tail$ non-terminal: \key{Return} and \key{Seq}. Regarding \key{Return}, we recommend treating it as an assignment to the \key{rax} register followed by a jump to the conclusion of the program (so the conclusion needs to be labeled). For $\SEQ{s}{t}$, you can translate the statement $s$ and tail $t$ recursively and append the resulting instructions. \begin{exercise} \normalfont Implement the \key{select-instructions} pass and test it on all of the example programs that you created for the previous passes and create three new example programs that are designed to exercise all of the interesting code in this pass. Use the \key{interp-tests} function (Appendix~\ref{appendix:utilities}) from \key{utilities.rkt} to test your passes on the example programs. \end{exercise} \section{Assign Homes} \label{sec:assign-r1} The \key{assign-homes} pass compiles $\text{x86}^{*}_0$ programs to $\text{x86}^{*}_0$ programs that no longer use program variables. Thus, the \key{assign-homes} pass is responsible for placing all of the program variables in registers or on the stack. For runtime efficiency, it is better to place variables in registers, but as there are only 16 registers, some programs must necessarily resort to placing some variables on the stack. In this chapter we focus on the mechanics of placing variables on the stack. We study an algorithm for placing variables in registers in Chapter~\ref{ch:register-allocation-r1}. Consider again the following $R_1$ program. % s0_20.rkt \begin{lstlisting} (let ([a 42]) (let ([b a]) b)) \end{lstlisting} For reference, we repeat the output of \code{select-instructions} on the left and show the output of \code{assign-homes} on the right. % %% Recall that \key{explicate-control} associated the list of %% variables with the \code{locals} symbol in the program's $\itm{info}$ %% field, so \code{assign-homes} has convenient access to the them. % In this example, we assign variable \code{a} to stack location \code{-8(\%rbp)} and variable \code{b} to location \code{-16(\%rbp)}.\\ \begin{tabular}{l} \begin{minipage}{0.4\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] locals-types: a : 'Integer, b : 'Integer start: movq $42, a movq a, b movq b, %rax jmp conclusion \end{lstlisting} \end{minipage} {$\Rightarrow$} \begin{minipage}{0.4\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] stack-space: 16 start: movq $42, -8(%rbp) movq -8(%rbp), -16(%rbp) movq -16(%rbp), %rax jmp conclusion \end{lstlisting} \end{minipage} \end{tabular} \\ In the output of \code{select-instructions}, there is a entry for \code{locals-types} in the $\itm{info}$ of the \code{Program} node, which is needed here so that we have the list of variables that should be assigned to homes. The support code computes the \code{locals-types} entry. In particular, \code{type-check-C0} installs it in the $\itm{info}$ field of the \code{Program} node. When using \code{interp-tests} or \code{compiler-tests} (see Appendix, Section~\ref{appendix:utilities}), specify \code{type-check-C0} as the type checker to use after \code{explicate-control}. In the process of assigning variables to stack locations, it is convenient for you to compute and store the size of the frame (in bytes) in the $\itm{info}$ field of the \key{Program} node, with the key \code{stack-space}, which is needed later to generate the conclusion of the \code{main} procedure. The x86-64 standard requires the frame size to be a multiple of 16 bytes. \index{frame} \begin{exercise} \normalfont Implement the \key{assign-homes} pass and test it on all of the example programs that you created for the previous passes pass. We recommend that \key{assign-homes} take an extra parameter that is a mapping of variable names to homes (stack locations for now). Use the \key{interp-tests} function (Appendix~\ref{appendix:utilities}) from \key{utilities.rkt} to test your passes on the example programs. \end{exercise} \section{Patch Instructions} \label{sec:patch-s0} The \code{patch-instructions} pass compiles $\text{x86}^{*}_0$ programs to $\text{x86}_0$ programs by making sure that each instruction adheres to the restrictions of the x86 assembly language. In particular, at most one argument of an instruction may be a memory reference. We return to the following running example. % s0_20.rkt \begin{lstlisting} (let ([a 42]) (let ([b a]) b)) \end{lstlisting} After the \key{assign-homes} pass, the above program has been translated to the following. \\ \begin{minipage}{0.5\textwidth} \begin{lstlisting} stack-space: 16 start: movq $42, -8(%rbp) movq -8(%rbp), -16(%rbp) movq -16(%rbp), %rax jmp conclusion \end{lstlisting} \end{minipage}\\ The second \key{movq} instruction is problematic because both arguments are stack locations. We suggest fixing this problem by moving from the source location to the register \key{rax} and then from \key{rax} to the destination location, as follows. \begin{lstlisting} movq -8(%rbp), %rax movq %rax, -16(%rbp) \end{lstlisting} \begin{exercise} \normalfont Implement the \key{patch-instructions} pass and test it on all of the example programs that you created for the previous passes and create three new example programs that are designed to exercise all of the interesting code in this pass. Use the \key{interp-tests} function (Appendix~\ref{appendix:utilities}) from \key{utilities.rkt} to test your passes on the example programs. \end{exercise} \section{Print x86} \label{sec:print-x86} The last step of the compiler from $R_1$ to x86 is to convert the $\text{x86}_0$ AST (defined in Figure~\ref{fig:x86-0-ast}) to the string representation (defined in Figure~\ref{fig:x86-0-concrete}). The Racket \key{format} and \key{string-append} functions are useful in this regard. The main work that this step needs to perform is to create the \key{main} function and the standard instructions for its prelude and conclusion, as shown in Figure~\ref{fig:p1-x86} of Section~\ref{sec:x86}. You need to know the number of stack-allocated variables, so we suggest computing it in the \key{assign-homes} pass (Section~\ref{sec:assign-r1}) and storing it in the $\itm{info}$ field of the \key{program} node. %% Your compiled code should print the result of the program's execution %% by using the \code{print\_int} function provided in %% \code{runtime.c}. If your compiler has been implemented correctly so %% far, this final result should be stored in the \key{rax} register. %% We'll talk more about how to perform function calls with arguments in %% general later on, but for now, place the following after the compiled %% code for the $R_1$ program but before the conclusion: %% \begin{lstlisting} %% movq %rax, %rdi %% callq print_int %% \end{lstlisting} %% These lines move the value in \key{rax} into the \key{rdi} register, which %% stores the first argument to be passed into \key{print\_int}. If you want your program to run on Mac OS X, your code needs to determine whether or not it is running on a Mac, and prefix underscores to labels like \key{main}. You can determine the platform with the Racket call \code{(system-type 'os)}, which returns \code{'macosx}, \code{'unix}, or \code{'windows}. %% In addition to %% placing underscores on \key{main}, you need to put them in front of %% \key{callq} labels (so \code{callq print\_int} becomes \code{callq %% \_print\_int}). \begin{exercise} \normalfont Implement the \key{print-x86} pass and test it on all of the example programs that you created for the previous passes. Use the \key{compiler-tests} function (Appendix~\ref{appendix:utilities}) from \key{utilities.rkt} to test your complete compiler on the example programs. See the \key{run-tests.rkt} script in the student support code for an example of how to use \key{compiler-tests}. Also, remember to compile the provided \key{runtime.c} file to \key{runtime.o} using \key{gcc}. \end{exercise} \section{Challenge: Partial Evaluator for $R_1$} \label{sec:pe-R1} \index{partial evaluation} This section describes optional challenge exercises that involve adapting and improving the partial evaluator for $R_0$ that was introduced in Section~\ref{sec:partial-evaluation}. \begin{exercise}\label{ex:pe-R1} \normalfont Adapt the partial evaluator from Section~\ref{sec:partial-evaluation} (Figure~\ref{fig:pe-arith}) so that it applies to $R_1$ programs instead of $R_0$ programs. Recall that $R_1$ adds \key{let} binding and variables to the $R_0$ language, so you will need to add cases for them in the \code{pe-exp} function. Also, note that the \key{program} form changes slightly to include an $\itm{info}$ field. Once complete, add the partial evaluation pass to the front of your compiler and make sure that your compiler still passes all of the tests. \end{exercise} The next exercise builds on Exercise~\ref{ex:pe-R1}. \begin{exercise} \normalfont Improve on the partial evaluator by replacing the \code{pe-neg} and \code{pe-add} auxiliary functions with functions that know more about arithmetic. For example, your partial evaluator should translate \begin{lstlisting} (+ 1 (+ (read) 1)) \end{lstlisting} into \begin{lstlisting} (+ 2 (read)) \end{lstlisting} To accomplish this, the \code{pe-exp} function should produce output in the form of the $\itm{residual}$ non-terminal of the following grammar. \[ \begin{array}{lcl} \itm{inert} &::=& \Var \mid (\key{read}) \mid (\key{-} \;(\key{read})) \mid (\key{+} \; \itm{inert} \; \itm{inert})\\ \itm{residual} &::=& \Int \mid (\key{+}\; \Int\; \itm{inert}) \mid \itm{inert} \end{array} \] The \code{pe-add} and \code{pe-neg} functions may therefore assume that their inputs are $\itm{residual}$ expressions and they should return $\itm{residual}$ expressions. Once the improvements are complete, make sure that your compiler still passes all of the tests. After all, fast code is useless if it produces incorrect results! \end{exercise} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Register Allocation} \label{ch:register-allocation-r1} \index{register allocation} In Chapter~\ref{ch:int-exp} we placed all variables on the stack to make our life easier. However, we can improve the performance of the generated code if we instead place some variables into registers. The CPU can access a register in a single cycle, whereas accessing the stack takes many cycles if the relevant data is in cache or many more to access main memory if the data is not in cache. Figure~\ref{fig:reg-eg} shows a program with four variables that serves as a running example. We show the source program and also the output of instruction selection. At that point the program is almost x86 assembly but not quite; it still contains variables instead of stack locations or registers. \begin{figure} \begin{minipage}{0.45\textwidth} Example $R_1$ program: % s0_28.rkt \begin{lstlisting} (let ([v 1]) (let ([w 42]) (let ([x (+ v 7)]) (let ([y x]) (let ([z (+ x w)]) (+ z (- y))))))) \end{lstlisting} \end{minipage} \begin{minipage}{0.45\textwidth} After instruction selection: \begin{lstlisting} locals-types: x : Integer, y : Integer, z : Integer, t : Integer, v : Integer, w : Integer start: movq $1, v movq $42, w movq v, x addq $7, x movq x, y movq x, z addq w, z movq y, t negq t movq z, %rax addq t, %rax jmp conclusion \end{lstlisting} \end{minipage} \caption{A running example program for register allocation.} \label{fig:reg-eg} \end{figure} The goal of register allocation is to fit as many variables into registers as possible. A program sometimes has more variables than registers, so we cannot always map each variable to a different register. Fortunately, it is common for different variables to be needed during different periods of time during program execution, and in such cases several variables can be mapped to the same register. Consider variables \code{x} and \code{y} in Figure~\ref{fig:reg-eg}. After the variable \code{x} is moved to \code{z} it is no longer needed. Variable \code{y}, on the other hand, is used only after this point, so \code{x} and \code{y} could share the same register. The topic of Section~\ref{sec:liveness-analysis-r1} is how to compute where a variable is needed. Once we have that information, we compute which variables are needed at the same time, i.e., which ones \emph{interfere} with each other, and represent this relation as an undirected graph whose vertices are variables and edges indicate when two variables interfere (Section~\ref{sec:build-interference}). We then model register allocation as a graph coloring problem, which we discuss in Section~\ref{sec:graph-coloring}. If we run out of registers despite these efforts, we place the remaining variables on the stack, similar to what we did in Chapter~\ref{ch:int-exp}. It is common to use the verb \emph{spill} for assigning a variable to a stack location. The decision to spill a variable is handled as part of the graph coloring process described in Section~\ref{sec:graph-coloring}. We make the simplifying assumption that each variable is assigned to one location (a register or stack address). A more sophisticated approach is to assign a variable to one or more locations in different regions of the program. For example, if a variable is used many times in short sequence and then only used again after many other instructions, it could be more efficient to assign the variable to a register during the initial sequence and then move it to the stack for the rest of its lifetime. We refer the interested reader to \citet{Cooper:1998ly} and \citet{Cooper:2011aa} for more information about that approach. % discuss prioritizing variables based on how much they are used. \section{Registers and Calling Conventions} \label{sec:calling-conventions} \index{calling conventions} As we perform register allocation, we need to be aware of the \emph{calling conventions} \index{calling conventions} that govern how functions calls are performed in x86. Function calls require coordination between the caller and the callee, which is often assembly code written by different programmers or generated by different compilers. Here we follow the System V calling conventions that are used by the \code{gcc} compiler on Linux and MacOS~\citep{Bryant:2005aa,Matz:2013aa}. % Even though $R_1$ does not include programmer-defined functions, our generated code will 1) include a \code{main} function that the operating system will call to initiate execution, and 2) make calls to the \code{read\_int} function in our runtime system. The calling conventions include rules about how functions share the use of registers. In particular, the caller is responsible for freeing up some registers prior to the function call for use by the callee. These are called the \emph{caller-saved registers} \index{caller-saved registers} and they are \begin{lstlisting} rax rcx rdx rsi rdi r8 r9 r10 r11 \end{lstlisting} On the other hand, the callee is responsible for preserving the values of the \emph{callee-saved registers}, \index{callee-saved registers} which are \begin{lstlisting} rsp rbp rbx r12 r13 r14 r15 \end{lstlisting} We can think about this caller/callee convention from two points of view, the caller view and the callee view: \begin{itemize} \item The caller should assume that all the caller-saved registers get overwritten with arbitrary values by the callee. On the other hand, the caller can safely assume that all the callee-saved registers contain the same values after the call that they did before the call. \item The callee can freely use any of the caller-saved registers. However, if the callee wants to use a callee-saved register, the callee must arrange to put the original value back in the register prior to returning to the caller, which is usually accomplished by saving the value to the stack in the prelude of the function and restoring the value in the conclusion of the function. \end{itemize} In x86, registers are also used for passing arguments to a function and for the return value. In particular, the first six arguments of a function are passed in the following six registers, in the order given. \begin{lstlisting} rdi rsi rdx rcx r8 r9 \end{lstlisting} If there are more than six arguments, then the convention is to use space on the frame of the caller for the rest of the arguments. However, in Chapter~\ref{ch:functions} we arrange to never need more than six arguments. For now, the only function we care about is \code{read\_int} and it takes zero argument. % The register \code{rax} is for the return value of a function. The next question is how these calling conventions impact register allocation. Consider the $R_1$ program in Figure~\ref{fig:example-calling-conventions}. We first analyze this example from the caller point of view and then from the callee point of view. The program makes two calls to the \code{read} function. Also, the variable \code{x} is in-use during the second call to \code{read}, so we need to make sure that the value in \code{x} does not get accidentally wiped out by the call to \code{read}. One obvious approach is to save all the values in caller-saved registers to the stack prior to each function call, and restore them after each call. That way, if the register allocator chooses to assign \code{x} to a caller-saved register, its value will be preserved across the call to \code{read}. However, the disadvantage of this approach is that saving and restoring to the stack is relatively slow. If \code{x} is not used many times, it may be better to assign \code{x} to a stack location in the first place. Or better yet, if we can arrange for \code{x} to be placed in a callee-saved register, then it won't need to be saved and restored during function calls. The approach that we recommend for variables that are in-use during a function call is to either assign them to callee-saved registers or to spill them to the stack. On the other hand, for variables that are not in-use during a function call, we try the following alternatives in order 1) look for an available caller-saved register (to leave room for other variables in the callee-saved register), 2) look for a callee-saved register, and 3) spill the variable to the stack. It is straightforward to implement this approach in a graph coloring register allocator. First, we know which variables are in-use during every function call because we compute that information for every instruction (Section~\ref{sec:liveness-analysis-r1}). Second, when we build the interference graph (Section~\ref{sec:build-interference}), we can place an edge between each of these variables and the caller-saved registers in the interference graph. This will prevent the graph coloring algorithm from assigning those variables to caller-saved registers. Returning to the example in Figure~\ref{fig:example-calling-conventions}, let us analyze the generated x86 code on the right-hand side, focusing on the \code{start} block. Notice that variable \code{x} is assigned to \code{rbx}, a callee-saved register. Thus, it is already in a safe place during the second call to \code{read\_int}. Next, notice that variable \code{y} is assigned to \code{rcx}, a caller-saved register, because there are no function calls in the remainder of the block. Next we analyze the example from the callee point of view, focusing on the prelude and conclusion of the \code{main} function. As usual the prelude begins with saving the \code{rbp} register to the stack and setting the \code{rbp} to the current stack pointer. We now know why it is necessary to save the \code{rbp}: it is a callee-saved register. The prelude then pushes \code{rbx} to the stack because 1) \code{rbx} is also a callee-saved register and 2) \code{rbx} is assigned to a variable (\code{x}). There are several more callee-saved register that are not saved in the prelude because they were not assigned to variables. The prelude subtracts 8 bytes from the \code{rsp} to make it 16-byte aligned and then jumps to the \code{start} block. Shifting attention to the \code{conclusion}, we see that \code{rbx} is restored from the stack with a \code{popq} instruction. \index{prelude}\index{conclusion} \begin{figure}[tp] \begin{minipage}{0.45\textwidth} Example $R_1$ program: %s0_14.rkt \begin{lstlisting} (let ([x (read)]) (let ([y (read)]) (+ (+ x y) 42))) \end{lstlisting} \end{minipage} \begin{minipage}{0.45\textwidth} Generated x86 assembly: \begin{lstlisting} start: callq read_int movq %rax, %rbx callq read_int movq %rax, %rcx addq %rcx, %rbx movq %rbx, %rax addq $42, %rax jmp _conclusion .globl main main: pushq %rbp movq %rsp, %rbp pushq %rbx subq $8, %rsp jmp start conclusion: addq $8, %rsp popq %rbx popq %rbp retq \end{lstlisting} \end{minipage} \caption{An example with function calls.} \label{fig:example-calling-conventions} \end{figure} \section{Liveness Analysis} \label{sec:liveness-analysis-r1} \index{liveness analysis} A variable or register is \emph{live} at a program point if its current value is used at some later point in the program. We refer to variables and registers collectively as \emph{locations}. % Consider the following code fragment in which there are two writes to \code{b}. Are \code{a} and \code{b} both live at the same time? \begin{lstlisting}[numbers=left,numberstyle=\tiny] movq $5, a movq $30, b movq a, c movq $10, b addq b, c \end{lstlisting} The answer is no because the integer \code{30} written to \code{b} on line 2 is never used. The variable \code{b} is read on line 5 and there is an intervening write to \code{b} on line 4, so the read on line 5 receives the value written on line 4, not line 2. \begin{wrapfigure}[19]{l}[1.0in]{0.6\textwidth} \small \begin{tcolorbox}[title=\href{https://docs.racket-lang.org/reference/sets.html}{The Racket Set Package}] A \emph{set} is an unordered collection of elements without duplicates. \index{set} \begin{description} \item[$\LP\code{set}\,v\,\ldots\RP$] constructs a set containing the specified elements. \item[$\LP\code{set-union}\,set_1\,set_2\RP$] returns the union of the two sets. \item[$\LP\code{set-subtract}\,set_1\,set_2\RP$] returns the difference of the two sets. \item[$\LP\code{set-member?}\,set\,v\RP$] is element $v$ in $set$? \item[$\LP\code{set-count}\,set\RP$] how many unique elements are in $set$? \item[$\LP\code{set->list}\,set\RP$] converts the set to a list. \end{description} \end{tcolorbox} \end{wrapfigure} The live locations can be computed by traversing the instruction sequence back to front (i.e., backwards in execution order). Let $I_1,\ldots, I_n$ be the instruction sequence. We write $L_{\mathsf{after}}(k)$ for the set of live locations after instruction $I_k$ and $L_{\mathsf{before}}(k)$ for the set of live locations before instruction $I_k$. The live locations after an instruction are always the same as the live locations before the next instruction. \index{live-after} \index{live-before} \begin{equation} \label{eq:live-after-before-next} L_{\mathsf{after}}(k) = L_{\mathsf{before}}(k+1) \end{equation} To start things off, there are no live locations after the last instruction\footnote{Technically, the \code{rax} register is live but we do not use it for register allocation.}, so \begin{equation}\label{eq:live-last-empty} L_{\mathsf{after}}(n) = \emptyset \end{equation} We then apply the following rule repeatedly, traversing the instruction sequence back to front. \begin{equation}\label{eq:live-before-after-minus-writes-plus-reads} L_{\mathtt{before}}(k) = (L_{\mathtt{after}}(k) - W(k)) \cup R(k), \end{equation} where $W(k)$ are the locations written to by instruction $I_k$ and $R(k)$ are the locations read by instruction $I_k$. There is a special case for \code{jmp} instructions. The locations that are live before a \code{jmp} should be the locations that are live before the instruction that follows the target label. So we recommend maintaining an alist, perhaps called \code{label->live}, that maps each label to a set of such locations. Recall that for now, the only \code{jmp} in a pseudo-x86 program is the one at the end, to the \code{conclusion}. (For example, see Figure~\ref{fig:reg-eg}.) So the alist should map \code{conclusion} to the set $\{\ttm{rax},\ttm{rsp}\}$. Let us walk through the above example, applying these formulas starting with the instruction on line 5. We collect the answers in the below listing. The $L_{\mathsf{after}}$ for the \code{addq b, c} instruction is $\emptyset$ because it is the last instruction (formula~\ref{eq:live-last-empty}). The $L_{\mathsf{before}}$ for this instruction is $\{\ttm{b},\ttm{c}\}$ because it reads from variables \code{b} and \code{c} (formula~\ref{eq:live-before-after-minus-writes-plus-reads}), that is \[ L_{\mathsf{before}}(5) = (\emptyset - \{\ttm{c}\}) \cup \{ \ttm{b}, \ttm{c} \} = \{ \ttm{b}, \ttm{c} \} \] Moving on the the instruction \code{movq \$10, b} at line 4, we copy the live-before set from line 5 to be the live-after set for this instruction (formula~\ref{eq:live-after-before-next}). \[ L_{\mathsf{after}}(4) = \{ \ttm{b}, \ttm{c} \} \] This move instruction writes to \code{b} and does not read from any variables, so we have the following live-before set (formula~\ref{eq:live-before-after-minus-writes-plus-reads}). \[ L_{\mathsf{before}}(4) = (\{\ttm{b},\ttm{c}\} - \{\ttm{b}\}) \cup \emptyset = \{ \ttm{c} \} \] The live-before for instruction \code{movq a, c} is $\{\ttm{a}\}$ because it writes to $\{\ttm{c}\}$ and reads from $\{\ttm{a}\}$ (formula~\ref{eq:live-before-after-minus-writes-plus-reads}). The live-before for \code{movq \$30, b} is $\{\ttm{a}\}$ because it writes to a variable that is not live and does not read from a variable. Finally, the live-before for \code{movq \$5, a} is $\emptyset$ because it writes to variable \code{a}. \begin{center} \begin{minipage}{0.45\textwidth} \begin{lstlisting}[numbers=left,numberstyle=\tiny] movq $5, a movq $30, b movq a, c movq $10, b addq b, c \end{lstlisting} \end{minipage} \vrule\hspace{10pt} \begin{minipage}{0.45\textwidth} \begin{align*} L_{\mathsf{before}}(1)= \emptyset, L_{\mathsf{after}}(1)= \{\ttm{a}\}\\ L_{\mathsf{before}}(2)= \{\ttm{a}\}, L_{\mathsf{after}}(2)= \{\ttm{a}\}\\ L_{\mathsf{before}}(3)= \{\ttm{a}\}, L_{\mathsf{after}}(2)= \{\ttm{c}\}\\ L_{\mathsf{before}}(4)= \{\ttm{c}\}, L_{\mathsf{after}}(4)= \{\ttm{b},\ttm{c}\}\\ L_{\mathsf{before}}(5)= \{\ttm{b},\ttm{c}\}, L_{\mathsf{after}}(5)= \emptyset \end{align*} \end{minipage} \end{center} Figure~\ref{fig:live-eg} shows the results of liveness analysis for the running example program, with the live-before and live-after sets shown between each instruction to make the figure easy to read. \begin{figure}[tp] \hspace{20pt} \begin{minipage}{0.45\textwidth} \begin{lstlisting} |$\{\ttm{rsp}\}$| movq $1, v |$\{\ttm{v},\ttm{rsp}\}$| movq $42, w |$\{\ttm{v},\ttm{w},\ttm{rsp}\}$| movq v, x |$\{\ttm{w},\ttm{x},\ttm{rsp}\}$| addq $7, x |$\{\ttm{w},\ttm{x},\ttm{rsp}\}$| movq x, y |$\{\ttm{w},\ttm{x},\ttm{y},\ttm{rsp}\}$| movq x, z |$\{\ttm{w},\ttm{y},\ttm{z},\ttm{rsp}\}$| addq w, z |$\{\ttm{y},\ttm{z},\ttm{rsp}\}$| movq y, t |$\{\ttm{t},\ttm{z},\ttm{rsp}\}$| negq t |$\{\ttm{t},\ttm{z},\ttm{rsp}\}$| movq z, %rax |$\{\ttm{rax},\ttm{t},\ttm{rsp}\}$| addq t, %rax |$\{\ttm{rax},\ttm{rsp}\}$| jmp conclusion \end{lstlisting} \end{minipage} \caption{The running example annotated with live-after sets.} \label{fig:live-eg} \end{figure} \begin{exercise}\normalfont Implement the compiler pass named \code{uncover-live} that computes the live-after sets. We recommend storing the live-after sets (a list of a set of variables) in the $\itm{info}$ field of the \code{Block} structure. % We recommend organizing your code to use a helper function that takes a list of instructions and an initial live-after set (typically empty) and returns the list of live-after sets. % We recommend creating helper functions to 1) compute the set of locations that appear in an argument (of an instruction), 2) compute the locations read by an instruction which corresponds to the $R$ function discussed above, and 3) the locations written by an instruction which corresponds to $W$. The \code{callq} instruction should include all of the caller-saved registers in its write-set $W$ because the calling convention says that those registers may be written to during the function call. Likewise, the \code{callq} instruction should include the appropriate number of argument passing registers in its read-set $R$, depending on the arity of the function being called. (This is why the abstract syntax for \code{callq} includes the arity.) \end{exercise} \section{Building the Interference Graph} \label{sec:build-interference} \begin{wrapfigure}[27]{r}[1.0in]{0.6\textwidth} \small \begin{tcolorbox}[title=\href{https://docs.racket-lang.org/graph/index.html}{The Racket Graph Library}] A \emph{graph} is a collection of vertices and edges where each edge connects two vertices. A graph is \emph{directed} if each edge points from a source to a target. Otherwise the graph is \emph{undirected}. \index{graph}\index{directed graph}\index{undirected graph} \begin{description} \item[$\LP\code{directed-graph}\,\itm{edges}\RP$] constructs a directed graph from a list of edges. Each edge is a list containing the source and target vertex. \item[$\LP\code{undirected-graph}\,\itm{edges}\RP$] constructs a undirected graph from a list of edges. Each edge is represented by a list containing two vertices. \item[$\LP\code{add-vertex!}\,\itm{graph}\,\itm{vertex}\RP$] inserts a vertex into the graph. \item[$\LP\code{add-edge!}\,\itm{graph}\,\itm{source}\,\itm{target}\RP$] inserts an edge between the two vertices into the graph. \item[$\LP\code{in-neighbors}\,\itm{graph}\,\itm{vertex}\RP$] returns a sequence of all the neighbors of the given vertex. \item[$\LP\code{in-vertices}\,\itm{graph}\RP$] returns a sequence of all the vertices in the graph. \end{description} \end{tcolorbox} \end{wrapfigure} Based on the liveness analysis, we know where each location is used (read from). However, during register allocation, we need to answer questions of the specific form: are locations $u$ and $v$ live at the same time? (And therefore cannot be assigned to the same register.) To make this question easier to answer, we create an explicit data structure, an \emph{interference graph}\index{interference graph}. An interference graph is an undirected graph that has an edge between two locations if they are live at the same time, that is, if they interfere with each other. The most obvious way to compute the interference graph is to look at the set of live location between each statement in the program and add an edge to the graph for every pair of variables in the same set. This approach is less than ideal for two reasons. First, it can be expensive because it takes $O(n^2)$ time to look at every pair in a set of $n$ live locations. Second, there is a special case in which two locations that are live at the same time do not actually interfere with each other: when they both contain the same value because we have assigned one to the other. A better way to compute the interference graph is to focus on the writes~\cite{Appel:2003fk}. We do not want the writes performed by an instruction to overwrite something in a live location. So for each instruction, we create an edge between the locations being written to and all the other live locations. (Except that one should not create self edges.) Recall that for a \key{callq} instruction, we consider all of the caller-saved registers as being written to, so an edge will be added between every live variable and every caller-saved register. For \key{movq}, we deal with the above-mentioned special case by not adding an edge between a live variable $v$ and destination $d$ if $v$ matches the source of the move. So we have the following two rules. \begin{enumerate} \item If instruction $I_k$ is a move such as \key{movq} $s$\key{,} $d$, then add the edge $(d,v)$ for every $v \in L_{\mathsf{after}}(k)$ unless $v = d$ or $v = s$. \item For any other instruction $I_k$, for every $d \in W(k)$ add an edge $(d,v)$ for every $v \in L_{\mathsf{after}}(k)$ unless $v = d$. %% \item If instruction $I_k$ is an arithmetic instruction such as %% \code{addq} $s$\key{,} $d$, then add the edge $(d,v)$ for every $v \in %% L_{\mathsf{after}}(k)$ unless $v = d$. %% \item If instruction $I_k$ is of the form \key{callq} %% $\mathit{label}$, then add an edge $(r,v)$ for every caller-saved %% register $r$ and every variable $v \in L_{\mathsf{after}}(k)$. \end{enumerate} Working from the top to bottom of Figure~\ref{fig:live-eg}, we apply the above rules to each instruction. We highlight a few of the instructions and then refer the reader to Figure~\ref{fig:interference-results} for all the interference results. The first instruction is \lstinline{movq $1, v}, so rule 3 applies, and the live-after set is $\{\ttm{v}\}$. We do not add any interference edges because the one live variable \code{v} is also the destination of this instruction. % For the second instruction, \lstinline{movq $42, w}, so rule 3 applies again, and the live-after set is $\{\ttm{v},\ttm{w}\}$. So the target $\ttm{w}$ of \key{movq} interferes with $\ttm{v}$. % Next we skip forward to the instruction \lstinline{movq x, y}. \begin{figure}[tbp] \begin{quote} \begin{tabular}{ll} \lstinline!movq $1, v!& \ttm{v} interferes with \ttm{rsp},\\ \lstinline!movq $42, w!& \ttm{w} interferes with \ttm{v} and \ttm{rsp},\\ \lstinline!movq v, x!& \ttm{x} interferes with \ttm{w} and \ttm{rsp},\\ \lstinline!addq $7, x!& \ttm{x} interferes with \ttm{w} and \ttm{rsp},\\ \lstinline!movq x, y!& \ttm{y} interferes with \ttm{w} and \ttm{rsp} but not \ttm{x},\\ \lstinline!movq x, z!& \ttm{z} interferes with \ttm{w}, \ttm{y}, and \ttm{rsp},\\ \lstinline!addq w, z!& \ttm{z} interferes with \ttm{y} and \ttm{rsp}, \\ \lstinline!movq y, t!& \ttm{t} interferes with \ttm{z} and \ttm{rsp}, \\ \lstinline!negq t!& \ttm{t} interferes with \ttm{z} and \ttm{rsp}, \\ \lstinline!movq z, %rax! & \ttm{rax} interferes with \ttm{t} and \ttm{rsp}, \\ \lstinline!addq t, %rax! & \ttm{rax} interferes with \ttm{rsp}. \\ \lstinline!jmp conclusion!& no interference. \end{tabular} \end{quote} \caption{Interference results for the running example.} \label{fig:interference-results} \end{figure} The resulting interference graph is shown in Figure~\ref{fig:interfere}. \begin{figure}[tbp] \large \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}$}; \node (rsp) at (9,2) {$\ttm{rsp}$}; \node (t1) at (0,2) {$\ttm{t}$}; \node (z) at (3,2) {$\ttm{z}$}; \node (x) at (6,2) {$\ttm{x}$}; \node (y) at (3,0) {$\ttm{y}$}; \node (w) at (6,0) {$\ttm{w}$}; \node (v) at (9,0) {$\ttm{v}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] \caption{The interference graph of the example program.} \label{fig:interfere} \end{figure} %% Our next concern is to choose a data structure for representing the %% interference graph. There are many choices for how to represent a %% graph, for example, \emph{adjacency matrix}, \emph{adjacency list}, %% and \emph{edge set}~\citep{Cormen:2001uq}. The right way to choose a %% data structure is to study the algorithm that uses the data structure, %% determine what operations need to be performed, and then choose the %% data structure that provide the most efficient implementations of %% those operations. Often times the choice of data structure can have an %% effect on the time complexity of the algorithm, as it does here. If %% you skim the next section, you will see that the register allocation %% algorithm needs to ask the graph for all of its vertices and, given a %% vertex, it needs to known all of the adjacent vertices. Thus, the %% correct choice of graph representation is that of an adjacency %% list. There are helper functions in \code{utilities.rkt} for %% representing graphs using the adjacency list representation: %% \code{make-graph}, \code{add-edge}, and \code{adjacent} %% (Appendix~\ref{appendix:utilities}). %% % %% \margincomment{\footnotesize To do: change to use the %% Racket graph library. \\ --Jeremy} %% % %% In particular, those functions use a hash table to map each vertex to %% the set of adjacent vertices, and the sets are represented using %% Racket's \key{set}, which is also a hash table. \begin{exercise}\normalfont Implement the compiler pass named \code{build-interference} according to the algorithm suggested above. We recommend using the \code{graph} package to create and inspect the interference graph. The output graph of this pass should be stored in the $\itm{info}$ field of the program, under the key \code{conflicts}. \end{exercise} \section{Graph Coloring via Sudoku} \label{sec:graph-coloring} \index{graph coloring} \index{Sudoku} \index{color} We come to the main event, mapping variables to registers (or to stack locations in the event that we run out of registers). We need to make sure that two variables do not get mapped to the same register if the two variables interfere with each other. Thinking about the interference graph, this means that adjacent vertices must be mapped to different registers. If we think of registers as colors, the register allocation problem becomes the widely-studied graph coloring problem~\citep{Balakrishnan:1996ve,Rosen:2002bh}. The reader may be more familiar with the graph coloring problem than he or she realizes; the popular game of Sudoku is an instance of the graph coloring problem. The following describes how to build a graph out of an initial Sudoku board. \begin{itemize} \item There is one vertex in the graph for each Sudoku square. \item There is an edge between two vertices if the corresponding squares are in the same row, in the same column, or if the squares are in the same $3\times 3$ region. \item Choose nine colors to correspond to the numbers $1$ to $9$. \item Based on the initial assignment of numbers to squares in the Sudoku board, assign the corresponding colors to the corresponding vertices in the graph. \end{itemize} If you can color the remaining vertices in the graph with the nine colors, then you have also solved the corresponding game of Sudoku. Figure~\ref{fig:sudoku-graph} shows an initial Sudoku game board and the corresponding graph with colored vertices. We map the Sudoku number 1 to blue, 2 to yellow, and 3 to red. We only show edges for a sampling of the vertices (the colored ones) because showing edges for all of the vertices would make the graph unreadable. \begin{figure}[tbp] \includegraphics[width=0.45\textwidth]{figs/sudoku} \includegraphics[width=0.5\textwidth]{figs/sudoku-graph} \caption{A Sudoku game board and the corresponding colored graph.} \label{fig:sudoku-graph} \end{figure} Given that Sudoku is an instance of graph coloring, one can use Sudoku strategies to come up with an algorithm for allocating registers. For example, one of the basic techniques for Sudoku is called Pencil Marks. The idea is to use a process of elimination to determine what numbers no longer make sense for a square and write down those numbers in the square (writing very small). For example, if the number $1$ is assigned to a square, then by process of elimination, you can write the pencil mark $1$ in all the squares in the same row, column, and region. Many Sudoku computer games provide automatic support for Pencil Marks. % The Pencil Marks technique corresponds to the notion of \emph{saturation}\index{saturation} due to \cite{Brelaz:1979eu}. The saturation of a vertex, in Sudoku terms, is the set of numbers that are no longer available. In graph terminology, we have the following definition: \begin{equation*} \mathrm{saturation}(u) = \{ c \;|\; \exists v. v \in \mathrm{neighbors}(u) \text{ and } \mathrm{color}(v) = c \} \end{equation*} where $\mathrm{neighbors}(u)$ is the set of vertices that share an edge with $u$. Using the Pencil Marks technique leads to a simple strategy for filling in numbers: if there is a square with only one possible number left, then choose that number! But what if there are no squares with only one possibility left? One brute-force approach is to try them all: choose the first and if it ultimately leads to a solution, great. If not, backtrack and choose the next possibility. One good thing about Pencil Marks is that it reduces the degree of branching in the search tree. Nevertheless, backtracking can be horribly time consuming. One way to reduce the amount of backtracking is to use the most-constrained-first heuristic. That is, when choosing a square, always choose one with the fewest possibilities left (the vertex with the highest saturation). The idea is that choosing highly constrained squares earlier rather than later is better because later on there may not be any possibilities left for those squares. However, register allocation is easier than Sudoku because the register allocator can map variables to stack locations when the registers run out. Thus, it makes sense to drop backtracking in favor of greedy search, that is, make the best choice at the time and keep going. We still wish to minimize the number of colors needed, so keeping the most-constrained-first heuristic is a good idea. Figure~\ref{fig:satur-algo} gives the pseudo-code for a simple greedy algorithm for register allocation based on saturation and the most-constrained-first heuristic. It is roughly equivalent to the DSATUR algorithm of \cite{Brelaz:1979eu} (also known as saturation degree ordering~\citep{Gebremedhin:1999fk,Omari:2006uq}). Just as in Sudoku, the algorithm represents colors with integers. The integers $0$ through $k-1$ correspond to the $k$ registers that we use for register allocation. The integers $k$ and larger correspond to stack locations. The registers that are not used for register allocation, such as \code{rax}, are assigned to negative integers. In particular, we assign $-1$ to \code{rax} and $-2$ to \code{rsp}. One might wonder why we include registers at all in the liveness analysis and interference graph, for example, we never allocate a variable to \code{rax} and \code{rsp}, so it would be harmless to leave them out. As we see in Chapter~\ref{ch:tuples}, when we begin to use register for passing arguments to functions, it will be necessary for those registers to appear in the interference graph because those registers will also be assigned to variables, and we don't want those two uses to encroach on each other. Regarding registers such as \code{rax} and \code{rsp} that are not used for variables, we could omit them from the interference graph but that would require adding special cases to our algorithm, which would complicate the logic for little gain. \begin{figure}[btp] \centering \begin{lstlisting}[basicstyle=\rmfamily,deletekeywords={for,from,with,is,not,in,find},morekeywords={while},columns=fullflexible] Algorithm: DSATUR Input: a graph |$G$| Output: an assignment |$\mathrm{color}[v]$| for each vertex |$v \in G$| |$W \gets \mathrm{vertices}(G)$| while |$W \neq \emptyset$| do pick a vertex |$u$| from |$W$| with the highest saturation, breaking ties randomly find the lowest color |$c$| that is not in |$\{ \mathrm{color}[v] \;:\; v \in \mathrm{adjacent}(u)\}$| |$\mathrm{color}[u] \gets c$| |$W \gets W - \{u\}$| \end{lstlisting} \caption{The saturation-based greedy graph coloring algorithm.} \label{fig:satur-algo} \end{figure} With the DSATUR algorithm in hand, let us return to the running example and consider how to color the interference graph in Figure~\ref{fig:interfere}. % We color the vertices for registers with their own color. For example, \code{rax} is assigned the color $-1$ and \code{rsp} is assigned $-2$. The vertices for variables are not yet colored, so they annotated with a dash. We then update the saturation for vertices that are adjacent to a register. For example, the saturation for \code{t} is $\{-1,-2\}$ because it interferes with both \code{rax} and \code{rsp}. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1\}$}; \node (t1) at (0,2) {$\ttm{t}:-,\{-1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:-,\{-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{-2\}$}; \node (y) at (3,0) {$\ttm{y}:-,\{-2\}$}; \node (w) at (6,0) {$\ttm{w}:-,\{-2\}$}; \node (v) at (10,0) {$\ttm{v}:-,\{-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] The algorithm says to select a maximally saturated vertex. So we pick $\ttm{t}$ and color it with the first available integer, which is $0$. We mark $0$ as no longer available for $\ttm{z}$, $\ttm{rax}$, and \ttm{rsp} because they interfere with $\ttm{t}$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1,0\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{-1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:-,\{0,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{-2\}$}; \node (y) at (3,0) {$\ttm{y}:-,\{-2\}$}; \node (w) at (6,0) {$\ttm{w}:-,\{-2\}$}; \node (v) at (10,0) {$\ttm{v}:-,\{-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] We repeat the process, selecting another maximally saturated vertex, which is \code{z}, and color it with the first available number, which is $1$. We add $1$ to the saturation for the neighboring vertices \code{t}, \code{y}, \code{w}, and \code{rsp}. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1,0,1\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{-1,1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{-2\}$}; \node (y) at (3,0) {$\ttm{y}:-,\{1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:-,\{1,-2\}$}; \node (v) at (10,0) {$\ttm{v}:-,\{-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] The most saturated vertices are now \code{w} and \code{y}. We color \code{w} with the first available color, which is $0$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1,0,1\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{-1,1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{0,-2\}$}; \node (y) at (3,0) {$\ttm{y}:-,\{0,1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:0,\{1,-2\}$}; \node (v) at (10,0) {$\ttm{v}:-,\{0,-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] Vertex \code{y} is now the most highly saturated, so we color \code{y} with $2$. We cannot choose $0$ or $1$ because those numbers are in \code{y}'s saturation set. Indeed, \code{y} interferes with \code{w} and \code{z}, whose colors are $0$ and $1$ respectively. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{-1,1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,2,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{0,-2\}$}; \node (y) at (3,0) {$\ttm{y}:2,\{0,1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:0,\{1,2,-2\}$}; \node (v) at (10,0) {$\ttm{v}:-,\{0,-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] Now \code{x} and \code{v} are the most saturated, so we color \code{v} with $1$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{-1,1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,2,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{0,-2\}$}; \node (y) at (3,0) {$\ttm{y}:2,\{0,1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:0,\{1,2,-2\}$}; \node (v) at (10,0) {$\ttm{v}:1,\{0,-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] In the last step of the algorithm, we color \code{x} with $1$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (10,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{-1,1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,2,-2\}$}; \node (x) at (6,2) {$\ttm{x}:1,\{0,-2\}$}; \node (y) at (3,0) {$\ttm{y}:2,\{0,1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:0,\{1,2,-2\}$}; \node (v) at (10,0) {$\ttm{v}:1,\{0,-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] With the coloring complete, we finalize the assignment of variables to registers and stack locations. Recall that if we have $k$ registers to use for allocation, we map the first $k$ colors to registers and the rest to stack locations. Suppose for the moment that we have just one register to use for register allocation, \key{rcx}. Then the following maps of colors to registers and stack allocations. \[ \{ 0 \mapsto \key{\%rcx}, \; 1 \mapsto \key{-8(\%rbp)}, \; 2 \mapsto \key{-16(\%rbp)} \} \] Putting this mapping together with the above coloring of the variables, we arrive at the following assignment. \begin{gather*} \{ \ttm{v} \mapsto \key{\%rcx}, \, \ttm{w} \mapsto \key{\%rcx}, \, \ttm{x} \mapsto \key{-8(\%rbp)}, \, \ttm{y} \mapsto \key{-16(\%rbp)}, \\ \ttm{z} \mapsto \key{-8(\%rbp)}, \, \ttm{t} \mapsto \key{\%rcx} \} \end{gather*} Applying this assignment to our running example, on the left, yields the program on the right. % why frame size of 32? -JGS \begin{center} \begin{minipage}{0.3\textwidth} \begin{lstlisting} movq $1, v movq $42, w movq v, x addq $7, x movq x, y movq x, z addq w, z movq y, t negq t movq z, %rax addq t, %rax jmp conclusion \end{lstlisting} \end{minipage} $\Rightarrow\qquad$ \begin{minipage}{0.45\textwidth} \begin{lstlisting} movq $1, %rcx movq $42, %rcx movq %rcx, -8(%rbp) addq $7, -8(%rbp) movq -8(%rbp), -16(%rbp) movq -8(%rbp), -8(%rbp) addq %rcx, -8(%rbp) movq -16(%rbp), %rcx negq %rcx movq -8(%rbp), %rax addq %rcx, %rax jmp conclusion \end{lstlisting} \end{minipage} \end{center} The resulting program is almost an x86 program. The remaining step is the patch instructions pass. In this example, the trivial move of \code{-8(\%rbp)} to itself is deleted and the addition of \code{-8(\%rbp)} to \key{-16(\%rbp)} is fixed by going through \code{rax} as follows. \begin{lstlisting} movq -8(%rbp), %rax addq %rax, -16(%rbp) \end{lstlisting} We recommend creating a helper function named \code{color-graph} that takes an interference graph and a list of all the variables in the program. This function should return a mapping of variables to their colors (represented as natural numbers). By creating this helper function, you will be able to reuse it in Chapter~\ref{ch:functions} when you add support for functions. To prioritize the processing of highly saturated nodes inside your \code{color-graph} function, we recommend using the priority queue data structure (see the side bar on the right). Note that you will also need to maintain a mapping from variables to their ``handles'' in the priority queue so that you can notify the priority queue when their saturation changes. \begin{wrapfigure}[23]{r}[1.0in]{0.6\textwidth} \small \begin{tcolorbox}[title=Priority Queue] A \emph{priority queue} is a collection of items in which the removal of items is governed by priority. In a ``min'' queue, lower priority items are removed first. An implementation is in \code{priority\_queue.rkt} of the support code. \index{priority queue} \index{minimum priority queue} \begin{description} \item[$\LP\code{make-pqueue}\,\itm{cmp}\RP$] constructs an empty priority queue that uses the $\itm{cmp}$ predicate to determine whether its first argument has lower or equal priority to its second argument. \item[$\LP\code{pqueue-count}\,\itm{queue}\RP$] returns the number of items in the queue. \item[$\LP\code{pqueue-push!}\,\itm{queue}\,\itm{item}\RP$] inserts the item into the queue and returns a handle for the item in the queue. \item[$\LP\code{pqueue-pop!}\,\itm{queue}\RP$] returns the item with the lowest priority. \item[$\LP\code{pqueue-decrease-key!}\,\itm{queue}\,\itm{handle}\RP$] notifies the queue that the priority has decreased for the item associated with the given handle. \end{description} \end{tcolorbox} \end{wrapfigure} Once you have obtained the coloring from \code{color-graph}, you can assign the variables to registers or stack locations and then reuse code from the \code{assign-homes} pass from Section~\ref{sec:assign-r1} to replace the variables with their assigned location. \begin{exercise}\normalfont Implement the compiler pass \code{allocate-registers}, which should come after the \code{build-interference} pass. The three new passes described in this chapter replace the \code{assign-homes} pass of Section~\ref{sec:assign-r1}. % Test your updated compiler by creating new example programs that exercise all of the register allocation algorithm, such as forcing variables to be spilled to the stack. \end{exercise} \section{Print x86} \label{sec:print-x86-reg-alloc} \index{calling conventions} \index{prelude}\index{conclusion} Recall that the \code{print-x86} pass generates the prelude and conclusion instructions for the \code{main} function. % The prelude saved the values in \code{rbp} and \code{rsp} and the conclusion returned those values to \code{rbp} and \code{rsp}. The reason for this is that our \code{main} function must adhere to the x86 calling conventions that we described in Section~\ref{sec:calling-conventions}. Furthermore, if your register allocator assigned variables to other callee-saved registers (e.g. \code{rbx}, \code{r12}, etc.), then those variables must also be saved to the stack in the prelude and restored in the conclusion. The simplest approach is to save and restore all of the callee-saved registers. The more efficient approach is to keep track of which callee-saved registers were used and only save and restore them. Either way, make sure to take this use of stack space into account when you are calculating the size of the frame and adjusting the \code{rsp} in the prelude. Also, don't forget that the size of the frame needs to be a multiple of 16 bytes! An overview of all of the passes involved in register allocation is shown in Figure~\ref{fig:reg-alloc-passes}. \begin{figure}[tbp] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R1) at (0,2) {\large $R_1$}; \node (R1-2) at (3,2) {\large $R_1$}; \node (R1-3) at (6,2) {\large $R_1$}; \node (C0-1) at (3,0) {\large $C_0$}; \node (x86-2) at (3,-2) {\large $\text{x86}^{*}$}; \node (x86-3) at (6,-2) {\large $\text{x86}^{*}$}; \node (x86-4) at (9,-2) {\large $\text{x86}$}; \node (x86-5) at (9,-4) {\large $\text{x86}^{\dagger}$}; \node (x86-2-1) at (3,-4) {\large $\text{x86}^{*}$}; \node (x86-2-2) at (6,-4) {\large $\text{x86}^{*}$}; \path[->,bend left=15] (R1) edge [above] node {\ttfamily\footnotesize uniquify} (R1-2); \path[->,bend left=15] (R1-2) edge [above] node {\ttfamily\footnotesize remove-complex.} (R1-3); \path[->,bend left=15] (R1-3) edge [right] node {\ttfamily\footnotesize explicate-control} (C0-1); \path[->,bend right=15] (C0-1) edge [left] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend left=15] (x86-2) edge [right] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [right] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [right] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_1$ with register allocation.} \label{fig:reg-alloc-passes} \end{figure} \section{Challenge: Move Biasing} \label{sec:move-biasing} \index{move biasing} This section describes an optional enhancement to register allocation for those students who are looking for an extra challenge or who have a deeper interest in register allocation. We return to the running example, but we remove the supposition that we only have one register to use. So we have the following mapping of color numbers to registers. \[ \{ 0 \mapsto \key{\%rbx}, \; 1 \mapsto \key{\%rcx}, \; 2 \mapsto \key{\%rdx} \} \] Using the same assignment of variables to color numbers that was produced by the register allocator described in the last section, we get the following program. \begin{minipage}{0.3\textwidth} \begin{lstlisting} movq $1, v movq $42, w movq v, x addq $7, x movq x, y movq x, z addq w, z movq y, t negq t movq z, %rax addq t, %rax jmp conclusion \end{lstlisting} \end{minipage} $\Rightarrow\qquad$ \begin{minipage}{0.45\textwidth} \begin{lstlisting} movq $1, %rcx movq $42, $rbx movq %rcx, %rcx addq $7, %rcx movq %rcx, %rdx movq %rcx, %rcx addq %rbx, %rcx movq %rdx, %rbx negq %rbx movq %rcx, %rax addq %rbx, %rax jmp conclusion \end{lstlisting} \end{minipage} In the above output code there are two \key{movq} instructions that can be removed because their source and target are the same. However, if we had put \key{t}, \key{v}, \key{x}, and \key{y} into the same register, we could instead remove three \key{movq} instructions. We can accomplish this by taking into account which variables appear in \key{movq} instructions with which other variables. We say that two variables $p$ and $q$ are \emph{move related}\index{move related} if they participate together in a \key{movq} instruction, that is, \key{movq} $p$\key{,} $q$ or \key{movq} $q$\key{,} $p$. When the register allocator chooses a color for a variable, it should prefer a color that has already been used for a move-related variable (assuming that they do not interfere). Of course, this preference should not override the preference for registers over stack locations. This preference should be used as a tie breaker when choosing between registers or when choosing between stack locations. We recommend representing the move relationships in a graph, similar to how we represented interference. The following is the \emph{move graph} for our running example. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}$}; \node (rsp) at (9,2) {$\ttm{rsp}$}; \node (t) at (0,2) {$\ttm{t}$}; \node (z) at (3,2) {$\ttm{z}$}; \node (x) at (6,2) {$\ttm{x}$}; \node (y) at (3,0) {$\ttm{y}$}; \node (w) at (6,0) {$\ttm{w}$}; \node (v) at (9,0) {$\ttm{v}$}; \draw (v) to (x); \draw (x) to (y); \draw (x) to (z); \draw (y) to (t); \end{tikzpicture} \] Now we replay the graph coloring, pausing to see the coloring of \code{y}. Recall the following configuration. The most saturated vertices were \code{w} and \code{y}. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (9,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{-2\}$}; \node (y) at (3,0) {$\ttm{y}:-,\{1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:-,\{1,-2\}$}; \node (v) at (9,0) {$\ttm{v}:-,\{-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] % Last time we chose to color \code{w} with $0$. But this time we see that \code{w} is not move related to any vertex, but \code{y} is move related to \code{t}. So we choose to color \code{y} the same color as \code{t}, $0$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (9,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{-2\}$}; \node (y) at (3,0) {$\ttm{y}:0,\{1,-2\}$}; \node (w) at (6,0) {$\ttm{w}:-,\{0,1,-2\}$}; \node (v) at (9,0) {$\ttm{v}:-,\{-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] Now \code{w} is the most saturated, so we color it $2$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (9,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t1) at (0,2) {$\ttm{t}:0,\{1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,2,-2\}$}; \node (x) at (6,2) {$\ttm{x}:-,\{2,-2\}$}; \node (y) at (3,0) {$\ttm{y}:0,\{1,2,-2\}$}; \node (w) at (6,0) {$\ttm{w}:2,\{0,1,-2\}$}; \node (v) at (9,0) {$\ttm{v}:-,\{2,-2\}$}; \draw (t1) to (rax); \draw (t1) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] At this point, vertices \code{x} and \code{v} are most saturated, but \code{x} is move related to \code{y} and \code{z}, so we color \code{x} to $0$ to match \code{y}. Finally, we color \code{v} to $0$. \[ \begin{tikzpicture}[baseline=(current bounding box.center)] \node (rax) at (0,0) {$\ttm{rax}:-1,\{0,-2\}$}; \node (rsp) at (9,2) {$\ttm{rsp}:-2,\{-1,0,1,2\}$}; \node (t) at (0,2) {$\ttm{t}:0,\{1,-2\}$}; \node (z) at (3,2) {$\ttm{z}:1,\{0,2,-2\}$}; \node (x) at (6,2) {$\ttm{x}:0,\{2,-2\}$}; \node (y) at (3,0) {$\ttm{y}:0,\{1,2,-2\}$}; \node (w) at (6,0) {$\ttm{w}:2,\{0,1,-2\}$}; \node (v) at (9,0) {$\ttm{v}:0,\{2,-2\}$}; \draw (t1) to (rax); \draw (t) to (z); \draw (z) to (y); \draw (z) to (w); \draw (x) to (w); \draw (y) to (w); \draw (v) to (w); \draw (v) to (rsp); \draw (w) to (rsp); \draw (x) to (rsp); \draw (y) to (rsp); \path[-.,bend left=15] (z) edge node {} (rsp); \path[-.,bend left=10] (t1) edge node {} (rsp); \draw (rax) to (rsp); \end{tikzpicture} \] So we have the following assignment of variables to registers. \begin{gather*} \{ \ttm{v} \mapsto \key{\%rbx}, \, \ttm{w} \mapsto \key{\%rdx}, \, \ttm{x} \mapsto \key{\%rbx}, \, \ttm{y} \mapsto \key{\%rbx}, \, \ttm{z} \mapsto \key{\%rcx}, \, \ttm{t} \mapsto \key{\%rbx} \} \end{gather*} We apply this register assignment to the running example, on the left, to obtain the code in the middle. The \code{patch-instructions} then removes the three trivial moves from \key{rbx} to \key{rbx} to obtain the code on the right. \begin{minipage}{0.25\textwidth} \begin{lstlisting} movq $1, v movq $42, w movq v, x addq $7, x movq x, y movq x, z addq w, z movq y, t negq t movq z, %rax addq t, %rax jmp conclusion \end{lstlisting} \end{minipage} $\Rightarrow\qquad$ \begin{minipage}{0.25\textwidth} \begin{lstlisting} movq $1, %rbx movq $42, %rdx movq %rbx, %rbx addq $7, %rbx movq %rbx, %rbx movq %rbx, %rcx addq %rdx, %rcx movq %rbx, %rbx negq %rbx movq %rcx, %rax addq %rbx, %rax jmp conclusion \end{lstlisting} \end{minipage} $\Rightarrow\qquad$ \begin{minipage}{0.25\textwidth} \begin{lstlisting} movq $1, %rbx movq $42, %rdx addq $7, %rbx movq %rbx, %rcx addq %rdx, %rcx negq %rbx movq %rcx, %rax addq %rbx, %rax jmp conclusion \end{lstlisting} \end{minipage} \begin{exercise}\normalfont Change your implementation of \code{allocate-registers} to take move biasing into account. Make sure that your compiler still passes all of the previous tests. Create two new tests that include at least one opportunity for move biasing and visually inspect the output x86 programs to make sure that your move biasing is working properly. \end{exercise} \margincomment{\footnotesize To do: another neat challenge would be to do live range splitting~\citep{Cooper:1998ly}. \\ --Jeremy} \section{Output of the Running Example} \label{sec:reg-alloc-output} \index{prelude}\index{conclusion} Figure~\ref{fig:running-example-x86} shows the x86 code generated for the running example (Figure~\ref{fig:reg-eg}) with register allocation and move biasing. To demonstrate both the use of registers and the stack, we have limited the register allocator to use just two registers: \code{rbx} and \code{rcx}. In the prelude of the \code{main} function, we push \code{rbx} onto the stack because it is a callee-saved register and it was assigned to variable by the register allocator. We subtract \code{8} from the \code{rsp} at the end of the prelude to reserve space for the one spilled variable. After that subtraction, the \code{rsp} is aligned to 16 bytes. Moving on the the \code{start} block, we see how the registers were allocated. Variables \code{v}, \code{x}, and \code{y} were assigned to \code{rbx} and variable \code{z} was assigned to \code{rcx}. Variable \code{w} was spilled to the stack location \code{-16(\%rbp)}. Recall that the prelude saved the callee-save register \code{rbx} onto the stack. The spilled variables must be placed lower on the stack than the saved callee-save registers, so in this case \code{w} is placed at \code{-16(\%rbp)}. In the \code{conclusion}, we undo the work that was done in the prelude. We move the stack pointer up by \code{8} bytes (the room for spilled variables), then we pop the old values of \code{rbx} and \code{rbp} (callee-saved registers), and finish with \code{retq} to return control to the operating system. \begin{figure}[tbp] % s0_28.rkt % (use-minimal-set-of-registers! #t) % and only rbx rcx % tmp 0 rbx % z 1 rcx % y 0 rbx % w 2 16(%rbp) % v 0 rbx % x 0 rbx \begin{lstlisting} start: movq $1, %rbx movq $42, -16(%rbp) addq $7, %rbx movq %rbx, %rcx addq -16(%rbp), %rcx negq %rbx movq %rcx, %rax addq %rbx, %rax jmp conclusion .globl main main: pushq %rbp movq %rsp, %rbp pushq %rbx subq $8, %rsp jmp start conclusion: addq $8, %rsp popq %rbx popq %rbp retq \end{lstlisting} \caption{The x86 output from the running example (Figure~\ref{fig:reg-eg}).} \label{fig:running-example-x86} \end{figure} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Booleans and Control Flow} \label{ch:bool-types} \index{Boolean} \index{control flow} \index{conditional expression} The $R_0$ and $R_1$ languages only have a single kind of value, the integers. In this chapter we add a second kind of value, the Booleans, to create the $R_2$ language. The Boolean values \emph{true} and \emph{false} are written \key{\#t} and \key{\#f} respectively in Racket. The $R_2$ language includes several operations that involve Booleans (\key{and}, \key{not}, \key{eq?}, \key{<}, etc.) and the conditional \key{if} expression. With the addition of \key{if} expressions, programs can have non-trivial control flow which which significantly impacts the \code{explicate-control} and the liveness analysis for register allocation. Also, because we now have two kinds of values, we need to handle programs that apply an operation to the wrong kind of value, such as \code{(not 1)}. There are two language design options for such situations. One option is to signal an error and the other is to provide a wider interpretation of the operation. The Racket language uses a mixture of these two options, depending on the operation and the kind of value. For example, the result of \code{(not 1)} in Racket is \code{\#f} because Racket treats non-zero integers as if they were \code{\#t}. On the other hand, \code{(car 1)} results in a run-time error in Racket stating that \code{car} expects a pair. The Typed Racket language makes similar design choices as Racket, except much of the error detection happens at compile time instead of run time. Like Racket, Typed Racket accepts and runs \code{(not 1)}, producing \code{\#f}. But in the case of \code{(car 1)}, Typed Racket reports a compile-time error because Typed Racket expects the type of the argument to be of the form \code{(Listof T)} or \code{(Pairof T1 T2)}. For the $R_2$ language we choose to be more like Typed Racket in that we perform type checking during compilation. In Chapter~\ref{ch:type-dynamic} we study the alternative choice, that is, how to compile a dynamically typed language like Racket. The $R_2$ language is a subset of Typed Racket but by no means includes all of Typed Racket. For many operations we take a narrower interpretation than Typed Racket, for example, rejecting \code{(not 1)}. This chapter is organized as follows. We begin by defining the syntax and interpreter for the $R_2$ language (Section~\ref{sec:r2-lang}). We then introduce the idea of type checking and build a type checker for $R_2$ (Section~\ref{sec:type-check-r2}). To compile $R_2$ we need to enlarge the intermediate language $C_0$ into $C_1$, which we do in Section~\ref{sec:c1}. The remaining sections of this chapter discuss how our compiler passes need to change to accommodate Booleans and conditional control flow. \section{The $R_2$ Language} \label{sec:r2-lang} The concrete syntax of the $R_2$ language is defined in Figure~\ref{fig:r2-concrete-syntax} and the abstract syntax is defined in Figure~\ref{fig:r2-syntax}. The $R_2$ language includes all of $R_1$ (shown in gray), the Boolean literals \code{\#t} and \code{\#f}, and the conditional \code{if} expression. Also, we expand the operators to include \begin{enumerate} \item subtraction on integers, \item the logical operators \key{and}, \key{or} and \key{not}, \item the \key{eq?} operation for comparing two integers or two Booleans, and \item the \key{<}, \key{<=}, \key{>}, and \key{>=} operations for comparing integers. \end{enumerate} We reorganize the abstract syntax for the primitive operations in Figure~\ref{fig:r2-syntax}, using only one grammar rule for all of them. This means that the grammar no longer checks whether the arity of an operators matches the number of arguments. That responsibility is moved to the type checker for $R_2$, which we introduce in Section~\ref{sec:type-check-r2}. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \itm{bool} &::=& \key{\#t} \mid \key{\#f} \\ \itm{cmp} &::= & \key{eq?} \mid \key{<} \mid \key{<=} \mid \key{>} \mid \key{>=} \\ \Exp &::=& \gray{ \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} } \mid \CSUB{\Exp}{\Exp} \\ &\mid& \gray{ \Var \mid \CLET{\Var}{\Exp}{\Exp} } \\ &\mid& \itm{bool} \mid (\key{and}\;\Exp\;\Exp) \mid (\key{or}\;\Exp\;\Exp) \mid (\key{not}\;\Exp) \\ &\mid& (\itm{cmp}\;\Exp\;\Exp) \mid \CIF{\Exp}{\Exp}{\Exp} \\ R_2 &::=& \Exp \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_2$, extending $R_1$ (Figure~\ref{fig:r1-concrete-syntax}) with Booleans and conditionals.} \label{fig:r2-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \itm{bool} &::=& \code{\#t} \mid \code{\#f} \\ \itm{cmp} &::= & \code{eq?} \mid \code{<} \mid \code{<=} \mid \code{>} \mid \code{>=} \\ \itm{op} &::= & \itm{cmp} \mid \code{read} \mid \code{+} \mid \code{-} \mid \code{and} \mid \code{or} \mid \code{not} \\ \Exp &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \PRIM{\itm{op}}{\Exp\ldots}\\ &\mid& \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} \\ R_2 &::=& \PROGRAM{\code{'()}}{\Exp} \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_2$.} \label{fig:r2-syntax} \end{figure} Figure~\ref{fig:interp-R2} defines the interpreter for $R_2$, inheriting from the interpreter for $R_1$ (Figure~\ref{fig:interp-R1}). The literals \code{\#t} and \code{\#f} evaluate to the corresponding Boolean values. The conditional expression $(\key{if}\, \itm{cnd}\,\itm{thn}\,\itm{els})$ evaluates the Boolean expression \itm{cnd} and then either evaluates \itm{thn} or \itm{els} depending on whether \itm{cnd} produced \code{\#t} or \code{\#f}. The logical operations \code{not} and \code{and} behave as you might expect, but note that the \code{and} operation is short-circuiting. That is, given the expression $(\key{and}\,e_1\,e_2)$, the expression $e_2$ is not evaluated if $e_1$ evaluates to \code{\#f}. With the increase in the number of primitive operations, the interpreter code for them could become repetitive without some care. We factor out the different parts of the code for primitive operations into the \code{interp-op} method shown in in Figure~\ref{fig:interp-op-R2}. The match clause for \code{Prim} makes the recursive calls to interpret the arguments and then passes the resulting values to \code{interp-op}. We do not use \code{interp-op} for the \code{and} operation because of its short-circuiting behavior. \begin{figure}[tbp] \begin{lstlisting} (define interp-R2-class (class interp-R1-class (super-new) (define/public (interp-op op) ...) (define/override ((interp-exp env) e) (define recur (interp-exp env)) (match e [(Bool b) b] [(If cnd thn els) (define b (recur cnd)) (match b [#t (recur thn)] [#f (recur els)])] [(Prim 'and (list e1 e2)) (define v1 (recur e1)) (match v1 [#t (match (recur e2) [#t #t] [#f #f])] [#f #f])] [(Prim op args) (apply (interp-op op) (for/list ([e args]) (recur e)))] [else ((super interp-exp env) e)] )) )) (define (interp-R2 p) (send (new interp-R2-class) interp-program p)) \end{lstlisting} \caption{Interpreter for the $R_2$ language. (See Figure~\ref{fig:interp-op-R2} for \code{interp-op}.)} \label{fig:interp-R2} \end{figure} \begin{figure}[tbp] \begin{lstlisting} (define/public (interp-op op) (match op ['+ fx+] ['- fx-] ['read read-fixnum] ['not (lambda (v) (match v [#t #f] [#f #t]))] ['or (lambda (v1 v2) (cond [(and (boolean? v1) (boolean? v2)) (or v1 v2)]))] ['eq? (lambda (v1 v2) (cond [(or (and (fixnum? v1) (fixnum? v2)) (and (boolean? v1) (boolean? v2)) (and (vector? v1) (vector? v2))) (eq? v1 v2)]))] ['< (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (< v1 v2)]))] ['<= (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (<= v1 v2)]))] ['> (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (> v1 v2)]))] ['>= (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (>= v1 v2)]))] [else (error 'interp-op "unknown operator")] )) \end{lstlisting} \caption{Interpreter for the primitive operators in the $R_2$ language.} \label{fig:interp-op-R2} \end{figure} \section{Type Checking $R_2$ Programs} \label{sec:type-check-r2} \index{type checking} \index{semantic analysis} It is helpful to think about type checking in two complementary ways. A type checker predicts the type of value that will be produced by each expression in the program. For $R_2$, we have just two types, \key{Integer} and \key{Boolean}. So a type checker should predict that \begin{lstlisting} (+ 10 (- (+ 12 20))) \end{lstlisting} produces an \key{Integer} while \begin{lstlisting} (and (not #f) #t) \end{lstlisting} produces a \key{Boolean}. Another way to think about type checking is that it enforces a set of rules about which operators can be applied to which kinds of values. For example, our type checker for $R_2$ will signal an error for the below expression because, as we have seen above, the expression \code{(+ 10 ...)} has type \key{Integer} but the type checker enforces the rule that the argument of \code{not} must be a \key{Boolean}. \begin{lstlisting} (not (+ 10 (- (+ 12 20)))) \end{lstlisting} We implement type checking using classes and method overriding for the same reason that we use them to implement the interpreters. We separate the type checker for the $R_1$ fragment into its own class, shown in Figure~\ref{fig:type-check-R1}. The type checker for $R_2$ is shown in Figure~\ref{fig:type-check-R2}; inherits from the one for $R_1$. The code for these type checkers are in the files \code{type-check-R1.rkt} and \code{type-check-R2.rkt} of the support code. % Each type checker is a structurally recursive function over the AST. Given an input expression \code{e}, the type checker either signals an error or returns an expression and its type (\key{Integer} or \key{Boolean}). There are situations in which we want to change or update the expression. % The type of an integer literal is \code{Integer} and the type of a Boolean literal is \code{Boolean}. To handle variables, the type checker uses the environment \code{env} to map variables to types. Consider the clause for \key{let}. We type check the initializing expression to obtain its type \key{T} and then associate type \code{T} with the variable \code{x} in the environment used to type check the body of the \key{let}. Thus, when the type checker encounters a use of variable \code{x}, it can find its type in the environment. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define type-check-R1-class (class object% (super-new) (define/public (operator-types) '((+ . ((Integer Integer) . Integer)) (- . ((Integer) . Integer)) (read . (() . Integer)))) (define/public (type-equal? t1 t2) (equal? t1 t2)) (define/public (check-type-equal? t1 t2 e) (unless (type-equal? t1 t2) (error 'type-check "~a != ~a\nin ~v" t1 t2 e))) (define/public (type-check-op op arg-types e) (match (dict-ref (operator-types) op) [`(,param-types . ,return-type) (for ([at arg-types] [pt param-types]) (check-type-equal? at pt e)) return-type] [else (error 'type-check-op "unrecognized ~a" op)])) (define/public (type-check-exp env) (lambda (e) (debug 'type-check-exp "R1" e) (match e [(Var x) (values (Var x) (dict-ref env x))] [(Int n) (values (Int n) 'Integer)] [(Let x e body) (define-values (e^ Te) ((type-check-exp env) e)) (define-values (b Tb) ((type-check-exp (dict-set env x Te)) body)) (values (Let x e^ b) Tb)] [(Prim op es) (define-values (new-es ts) (for/lists (exprs types) ([e es]) ((type-check-exp env) e))) (values (Prim op new-es) (type-check-op op ts e))] [else (error 'type-check-exp "couldn't match" e)]))) (define/public (type-check-program e) (match e [(Program info body) (define-values (body^ Tb) ((type-check-exp '()) body)) (check-type-equal? Tb 'Integer body) (Program info body^)] [else (error 'type-check-R1 "couldn't match ~a" e)])) )) (define (type-check-R1 p) (send (new type-check-R1-class) type-check-program p)) \end{lstlisting} \caption{Type checker for the $R_1$ fragment of $R_2$.} \label{fig:type-check-R1} \end{figure} \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define type-check-R2-class (class type-check-R1-class (super-new) (inherit check-type-equal?) (define/override (operator-types) (append '((- . ((Integer Integer) . Integer)) (and . ((Boolean Boolean) . Boolean)) (or . ((Boolean Boolean) . Boolean)) (< . ((Integer Integer) . Boolean)) (<= . ((Integer Integer) . Boolean)) (> . ((Integer Integer) . Boolean)) (>= . ((Integer Integer) . Boolean)) (not . ((Boolean) . Boolean)) ) (super operator-types))) (define/override (type-check-exp env) (lambda (e) (match e [(Bool b) (values (Bool b) 'Boolean)] [(If cnd thn els) (define-values (cnd^ Tc) ((type-check-exp env) cnd)) (define-values (thn^ Tt) ((type-check-exp env) thn)) (define-values (els^ Te) ((type-check-exp env) els)) (check-type-equal? Tc 'Boolean e) (check-type-equal? Tt Te e) (values (If cnd^ thn^ els^) Te)] [(Prim 'eq? (list e1 e2)) (define-values (e1^ T1) ((type-check-exp env) e1)) (define-values (e2^ T2) ((type-check-exp env) e2)) (check-type-equal? T1 T2 e) (values (Prim 'eq? (list e1^ e2^)) 'Boolean)] [else ((super type-check-exp env) e)]))) )) (define (type-check-R2 p) (send (new type-check-R2-class) type-check-program p)) \end{lstlisting} \caption{Type checker for the $R_2$ language.} \label{fig:type-check-R2} \end{figure} Three auxiliary methods are used in the type checker. The method \code{operator-types} defines a dictionary that maps the operator names to their parameter and return types. The \code{type-equal?} method determines whether two types are equal, which for now simply dispatches to \code{equal?} (deep equality). The \code{type-check-op} method looks up the operator in the \code{operator-types} dictionary and then checks whether the argument types are equal to the parameter types. The result is the return type of the operator. \begin{exercise}\normalfont Create 10 new example programs in $R_2$. Half of the example programs should have a type error. For those programs, to signal that a type error is expected, create an empty file with the same base name but with file extension \code{.tyerr}. For example, if the test \code{r2\_14.rkt} is expected to error, then create an empty file named \code{r2\_14.tyerr}. The other half of the example programs should not have type errors. Note that if the type checker does not signal an error for a program, then interpreting that program should not encounter an error. \end{exercise} \section{Shrink the $R_2$ Language} \label{sec:shrink-r2} The $R_2$ language includes several operators that are easily expressible in terms of other operators. For example, subtraction is expressible in terms of addition and negation. \[ \key{(-}\; e_1 \; e_2\key{)} \quad \Rightarrow \quad \LP\key{+} \; e_1 \; \LP\key{-} \; e_2\RP\RP \] Several of the comparison operations are expressible in terms of less-than and logical negation. \[ \LP\key{<=}\; e_1 \; e_2\RP \quad \Rightarrow \quad \LP\key{let}~\LP\LS\key{tmp.1}~e_1\RS\RP~\LP\key{not}\;\LP\key{<}\;e_2\;\key{tmp.1})\RP\RP \] The \key{let} is needed in the above translation to ensure that expression $e_1$ is evaluated before $e_2$. By performing these translations near the front-end of the compiler, the later passes of the compiler do not need to deal with these constructs, making those passes shorter. On the other hand, sometimes these translations make it more difficult to generate the most efficient code with respect to the number of instructions. However, these differences typically do not affect the number of accesses to memory, which is the primary factor that determines execution time on modern computer architectures. \begin{exercise}\normalfont Implement the pass \code{shrink} that removes subtraction, \key{and}, \key{or}, \key{<=}, \key{>}, and \key{>=} from the language by translating them to other constructs in $R_2$. Create tests to make sure that the behavior of all of these constructs stays the same after translation. \end{exercise} \section{The x86$_1$ Language} \label{sec:x86-1} \index{x86} To implement the new logical operations, the comparison operations, and the \key{if} expression, we need to delve further into the x86 language. Figures~\ref{fig:x86-1-concrete} and \ref{fig:x86-1} define the concrete and abstract syntax for a larger subset of x86 that includes instructions for logical operations, comparisons, and conditional jumps. One small challenge is that x86 does not provide an instruction that directly implements logical negation (\code{not} in $R_2$ and $C_1$). However, the \code{xorq} instruction can be used to encode \code{not}. The \key{xorq} instruction takes two arguments, performs a pairwise exclusive-or ($\mathrm{XOR}$) operation on each bit of its arguments, and writes the results into its second argument. Recall the truth table for exclusive-or: \begin{center} \begin{tabular}{l|cc} & 0 & 1 \\ \hline 0 & 0 & 1 \\ 1 & 1 & 0 \end{tabular} \end{center} For example, applying $\mathrm{XOR}$ to each bit of the binary numbers $0011$ and $0101$ yields $0110$. Notice that in the row of the table for the bit $1$, the result is the opposite of the second bit. Thus, the \code{not} operation can be implemented by \code{xorq} with $1$ as the first argument: \[ \Var~ \key{=}~ \LP\key{not}~\Arg\RP\key{;} \qquad\Rightarrow\qquad \begin{array}{l} \key{movq}~ \Arg\key{,} \Var\\ \key{xorq}~ \key{\$1,} \Var \end{array} \] \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \itm{bytereg} &::=& \key{ah} \mid \key{al} \mid \key{bh} \mid \key{bl} \mid \key{ch} \mid \key{cl} \mid \key{dh} \mid \key{dl} \\ \Arg &::=& \gray{ \key{\$}\Int \mid \key{\%}\Reg \mid \Int\key{(}\key{\%}\Reg\key{)} } \mid \key{\%}\itm{bytereg}\\ \itm{cc} & ::= & \key{e} \mid \key{l} \mid \key{le} \mid \key{g} \mid \key{ge} \\ \Instr &::=& \gray{ \key{addq} \; \Arg\key{,} \Arg \mid \key{subq} \; \Arg\key{,} \Arg \mid \key{negq} \; \Arg \mid \key{movq} \; \Arg\key{,} \Arg \mid } \\ && \gray{ \key{callq} \; \itm{label} \mid \key{pushq}\;\Arg \mid \key{popq}\;\Arg \mid \key{retq} \mid \key{jmp}\,\itm{label} } \\ && \gray{ \itm{label}\key{:}\; \Instr } \mid \key{xorq}~\Arg\key{,}~\Arg \mid \key{cmpq}~\Arg\key{,}~\Arg \mid \\ && \key{set}cc~\Arg \mid \key{movzbq}~\Arg\key{,}~\Arg \mid \key{j}cc~\itm{label} \\ x86_1 &::= & \gray{ \key{.globl main} }\\ & & \gray{ \key{main:} \; \Instr\ldots } \end{array} \] \end{minipage} } \caption{The concrete syntax of x86$_1$ (extends x86$_0$ of Figure~\ref{fig:x86-0-concrete}).} \label{fig:x86-1-concrete} \end{figure} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \itm{bytereg} &::=& \key{ah} \mid \key{al} \mid \key{bh} \mid \key{bl} \mid \key{ch} \mid \key{cl} \mid \key{dh} \mid \key{dl} \\ \Arg &::=& \gray{\IMM{\Int} \mid \REG{\Reg} \mid \DEREF{\Reg}{\Int}} \mid \BYTEREG{\itm{bytereg}} \\ \itm{cc} & ::= & \key{e} \mid \key{l} \mid \key{le} \mid \key{g} \mid \key{ge} \\ \Instr &::=& \gray{ \BININSTR{\code{'addq}}{\Arg}{\Arg} \mid \BININSTR{\code{'subq}}{\Arg}{\Arg} } \\ &\mid& \gray{ \BININSTR{\code{'movq}}{\Arg}{\Arg} \mid \UNIINSTR{\code{'negq}}{\Arg} } \\ &\mid& \gray{ \CALLQ{\itm{label}}{\itm{int}} \mid \RETQ{} \mid \PUSHQ{\Arg} \mid \POPQ{\Arg} \mid \JMP{\itm{label}} } \\ &\mid& \BININSTR{\code{'xorq}}{\Arg}{\Arg} \mid \BININSTR{\code{'cmpq}}{\Arg}{\Arg}\\ &\mid& \BININSTR{\code{'set}}{\itm{cc}}{\Arg} \mid \BININSTR{\code{'movzbq}}{\Arg}{\Arg}\\ &\mid& \JMPIF{\itm{cc}}{\itm{label}} \\ \Block &::= & \gray{\BLOCK{\itm{info}}{\Instr\ldots}} \\ x86_1 &::= & \gray{\PROGRAM{\itm{info}}{\CFG{\key{(}\itm{label} \,\key{.}\, \Block \key{)}\ldots}}} \end{array} \] \end{minipage} } \caption{The abstract syntax of x86$_1$ (extends x86$_0$ of Figure~\ref{fig:x86-0-ast}).} \label{fig:x86-1} \end{figure} Next we consider the x86 instructions that are relevant for compiling the comparison operations. The \key{cmpq} instruction compares its two arguments to determine whether one argument is less than, equal, or greater than the other argument. The \key{cmpq} instruction is unusual regarding the order of its arguments and where the result is placed. The argument order is backwards: if you want to test whether $x < y$, then write \code{cmpq} $y$\code{,} $x$. The result of \key{cmpq} is placed in the special EFLAGS register. This register cannot be accessed directly but it can be queried by a number of instructions, including the \key{set} instruction. The \key{set} instruction puts a \key{1} or \key{0} into its destination depending on whether the comparison came out according to the condition code \itm{cc} (\key{e} for equal, \key{l} for less, \key{le} for less-or-equal, \key{g} for greater, \key{ge} for greater-or-equal). The \key{set} instruction has an annoying quirk in that its destination argument must be single byte register, such as \code{al} (L for lower bits) or \code{ah} (H for higher bits), which are part of the \code{rax} register. Thankfully, the \key{movzbq} instruction can then be used to move from a single byte register to a normal 64-bit register. The x86 instruction for conditional jump are relevant to the compilation of \key{if} expressions. The \key{JmpIf} instruction updates the program counter to point to the instruction after the indicated label depending on whether the result in the EFLAGS register matches the condition code \itm{cc}, otherwise the \key{JmpIf} instruction falls through to the next instruction. The abstract syntax for \key{JmpIf} differs from the concrete syntax for x86 in that it separates the instruction name from the condition code. For example, \code{(JmpIf le foo)} corresponds to \code{jle foo}. Because the \key{JmpIf} instruction relies on the EFLAGS register, it is common for the \key{JmpIf} to be immediately preceded by a \key{cmpq} instruction to set the EFLAGS register. \section{The $C_1$ Intermediate Language} \label{sec:c1} As with $R_1$, we compile $R_2$ to a C-like intermediate language, but we need to grow that intermediate language to handle the new features in $R_2$: Booleans and conditional expressions. Figure~\ref{fig:c1-syntax} defines the abstract syntax of $C_1$. (The concrete syntax is in the Appendix, Figure~\ref{fig:c1-concrete-syntax}.) The $C_1$ language adds logical and comparison operators to the $\Exp$ non-terminal and the literals \key{\#t} and \key{\#f} to the $\Arg$ non-terminal. Regarding control flow, $C_1$ differs considerably from $R_2$. Instead of \key{if} expressions, $C_1$ has \key{goto} and conditional \key{goto} in the grammar for $\Tail$. This means that a sequence of statements may now end with a \code{goto} or a conditional \code{goto}. The conditional \code{goto} jumps to one of two labels depending on the outcome of the comparison. In Section~\ref{sec:explicate-control-r2} we discuss how to translate from $R_2$ to $C_1$, bridging this gap between \key{if} expressions and \key{goto}'s. \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Atm &::=& \gray{\INT{\Int} \mid \VAR{\Var}} \mid \BOOL{\itm{bool}} \\ \itm{cmp} &::= & \key{eq?} \mid \key{<} \\ \Exp &::= & \gray{ \Atm \mid \READ{} }\\ &\mid& \gray{ \NEG{\Atm} \mid \ADD{\Atm}{\Atm} } \\ &\mid& \UNIOP{\key{'not}}{\Atm} \mid \BINOP{\key{'}\itm{cmp}}{\Atm}{\Atm} \\ \Stmt &::=& \gray{ \ASSIGN{\VAR{\Var}}{\Exp} } \\ \Tail &::= & \gray{\RETURN{\Exp} \mid \SEQ{\Stmt}{\Tail} } \mid \GOTO{\itm{label}} \\ &\mid& \IFSTMT{\BINOP{\itm{cmp}}{\Atm}{\Atm}}{\GOTO{\itm{label}}}{\GOTO{\itm{label}}} \\ C_1 & ::= & \gray{\PROGRAM{\itm{info}}{\CFG{\key{(}\itm{label}\,\key{.}\,\Tail\key{)}\ldots}}} \end{array} \] \end{minipage} } \caption{The abstract syntax of $C_1$, an extension of $C_0$ (Figure~\ref{fig:c0-syntax}).} \label{fig:c1-syntax} \end{figure} \clearpage \section{Remove Complex Operands} \label{sec:remove-complex-opera-R2} Add cases for \code{Bool} and \code{If} to the \code{rco-exp} and \code{rco-atom} functions according to the definition of the output language for this pass, $R_2^{\dagger}$, the administrative normal form of $R_2$, which is defined in Figure~\ref{fig:r2-anf-syntax}. The \code{Bool} form is an atomic expressions but \code{If} is not. All three sub-expressions of an \code{If} are allowed to be complex expressions in the output of \code{remove-complex-opera*}, but the operands of \code{not} and the comparisons must be atoms. Regarding the \code{If} form, it is particularly important to \textbf{not} replace its condition with a temporary variable because that would interfere with the generation of high-quality output in the \code{explicate-control} pass. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{rcl} \Atm &::=& \gray{ \INT{\Int} \mid \VAR{\Var} } \mid \BOOL{\itm{bool}}\\ \Exp &::=& \gray{ \Atm \mid \READ{} } \\ &\mid& \gray{ \NEG{\Atm} \mid \ADD{\Atm}{\Atm} } \\ &\mid& \gray{ \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \UNIOP{\key{'not}}{\Atm} \\ &\mid& \BINOP{\itm{cmp}}{\Atm}{\Atm} \mid \IF{\Exp}{\Exp}{\Exp} \\ R^{\dagger}_2 &::=& \PROGRAM{\code{'()}}{\Exp} \end{array} \] \end{minipage} } \caption{$R_2^{\dagger}$ is $R_2$ in administrative normal form (ANF).} \label{fig:r2-anf-syntax} \end{figure} \section{Explicate Control} \label{sec:explicate-control-r2} Recall that the purpose of \code{explicate-control} is to make the order of evaluation explicit in the syntax of the program. With the addition of \key{if} in $R_2$ this get more interesting. As a motivating example, consider the following program that has an \key{if} expression nested in the predicate of another \key{if}. % s1_41.rkt \begin{center} \begin{minipage}{0.96\textwidth} \begin{lstlisting} (let ([x (read)]) (let ([y (read)]) (if (if (< x 1) (eq? x 0) (eq? x 2)) (+ y 2) (+ y 10)))) \end{lstlisting} \end{minipage} \end{center} % The naive way to compile \key{if} and the comparison would be to handle each of them in isolation, regardless of their context. Each comparison would be translated into a \key{cmpq} instruction followed by a couple instructions to move the result from the EFLAGS register into a general purpose register or stack location. Each \key{if} would be translated into the combination of a \key{cmpq} and a conditional jump. The generated code for the inner \key{if} in the above example would be as follows. \begin{center} \begin{minipage}{0.96\textwidth} \begin{lstlisting} ... cmpq $1, x ;; (< x 1) setl %al movzbq %al, tmp cmpq $1, tmp ;; (if (< x 1) ...) je then_branch_1 jmp else_branch_1 ... \end{lstlisting} \end{minipage} \end{center} However, if we take context into account we can do better and reduce the use of \key{cmpq} and EFLAG-accessing instructions. One idea is to try and reorganize the code at the level of $R_2$, pushing the outer \key{if} inside the inner one. This would yield the following code. \begin{center} \begin{minipage}{0.96\textwidth} \begin{lstlisting} (let ([x (read)]) (let ([y (read)]) (if (< x 1) (if (eq? x 0) (+ y 2) (+ y 10)) (if (eq? x 2) (+ y 2) (+ y 10))))) \end{lstlisting} \end{minipage} \end{center} Unfortunately, this approach duplicates the two branches, and a compiler must never duplicate code! We need a way to perform the above transformation, but without duplicating code. That is, we need a way for different parts of a program to refer to the same piece of code, that is, to \emph{share} code. At the level of x86 assembly this is straightforward because we can label the code for each of the branches and insert jumps in all the places that need to execute the branches. At the higher level of our intermediate languages, we need to move away from abstract syntax \emph{trees} and instead use \emph{graphs}. In particular, we use a standard program representation called a \emph{control flow graph} (CFG), due to Frances Elizabeth \citet{Allen:1970uq}. \index{control-flow graph} Each vertex is a labeled sequence of code, called a \emph{basic block}, and each edge represents a jump to another block. The \key{Program} construct of $C_0$ and $C_1$ contains a control flow graph represented as an alist mapping labels to basic blocks. Each basic block is represented by the $\Tail$ non-terminal. Figure~\ref{fig:explicate-control-s1-38} shows the output of the \code{remove-complex-opera*} pass and then the \code{explicate-control} pass on the example program. We walk through the output program and then discuss the algorithm. % Following the order of evaluation in the output of \code{remove-complex-opera*}, we first have two calls to \code{(read)} and then the less-than-comparison to \code{1} in the predicate of the inner \key{if}. In the output of \code{explicate-control}, in the block labeled \code{start}, this becomes two assignment statements followed by a conditional \key{goto} to label \code{block40} or \code{block41}. The blocks associated with those labels contain the translations of the code \code{(eq? x 0)} and \code{(eq? x 2)}, respectively. Regarding the block labeled with \code{block40}, we start with the comparison to \code{0} and then have a conditional goto, either to label \code{block38} or label \code{block39}, which are the two branches of the outer \key{if}, i.e., \code{(+ y 2)} and \code{(+ y 10)}. The story for the block labeled \code{block41} is similar. \begin{figure}[tbp] \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} % s1_41.rkt \begin{lstlisting} (let ([x (read)]) (let ([y (read)]) (if (if (< x 1) (eq? x 0) (eq? x 2)) (+ y 2) (+ y 10)))) \end{lstlisting} \hspace{40pt}$\Downarrow$ \begin{lstlisting} (let ([x (read)]) (let ([y (read)]) (if (if (< x 1) (eq? x 0) (eq? x 2)) (+ y 2) (+ y 10)))) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.55\textwidth} \begin{lstlisting} start: x = (read); y = (read); if (< x 1) goto block40; else goto block41; block40: if (eq? x 0) goto block38; else goto block39; block41: if (eq? x 2) goto block38; else goto block39; block38: return (+ y 2); block39: return (+ y 10); \end{lstlisting} \end{minipage} \end{tabular} \caption{Translation from $R_2$ to $C_1$ via the \code{explicate-control}.} \label{fig:explicate-control-s1-38} \end{figure} %% The nice thing about the output of \code{explicate-control} is that %% there are no unnecessary comparisons and every comparison is part of a %% conditional jump. %% The down-side of this output is that it includes %% trivial blocks, such as the blocks labeled \code{block92} through %% \code{block95}, that only jump to another block. We discuss a solution %% to this problem in Section~\ref{sec:opt-jumps}. Recall that in Section~\ref{sec:explicate-control-r1} we implement \code{explicate-control} for $R_1$ using two mutually recursive functions, \code{explicate-tail} and \code{explicate-assign}. The former function translates expressions in tail position whereas the later function translates expressions on the right-hand-side of a \key{let}. With the addition of \key{if} expression in $R_2$ we have a new kind of context to deal with: the predicate position of the \key{if}. We need another function, \code{explicate-pred}, that takes an $R_2$ expression and two blocks for the then-branch and else-branch. The output of \code{explicate-pred} is a block. % %% Note that the three explicate functions need to construct a %% control-flow graph, which we recommend they do via updates to a global %% variable. % In the following paragraphs we discuss specific cases in the \code{explicate-pred} function as well as the additions to the \code{explicate-tail} and \code{explicate-assign} functions. The function \code{explicate-pred} will need a case for every expression that can have type \code{Boolean}. We detail a few cases here and leave the rest for the reader. The input to this function is an expression and two blocks, $B_1$ and $B_2$, for the two branches of the enclosing \key{if}, though some care will be needed regarding how we represent the blocks. Suppose the expression is the Boolean \code{\#t}. Then we can perform a kind of partial evaluation \index{partial evaluation} and translate it to the ``then'' branch $B_1$. Likewise, we translate \code{\#f} to the ``else`` branch $B_2$. \[ \key{\#t} \quad\Rightarrow\quad B_1, \qquad\qquad\qquad \key{\#f} \quad\Rightarrow\quad B_2 \] These two cases demonstrate that we sometimes discard one of the blocks that are input to \code{explicate-pred}. We will need to arrange for the blocks that we actually use to appear in the resulting control-flow graph, but not the discarded blocks. The case for \key{if} in \code{explicate-pred} is particularly illuminating as it deals with the challenges that we discussed above regarding the example of the nested \key{if} expressions. The ``then'' and ``else'' branches of the current \key{if} inherit their context from the current one, that is, predicate context. So we recursively apply \code{explicate-pred} to the ``then'' and ``else'' branches. For both of those recursive calls, we shall pass the blocks $B_1$ and $B_2$. Thus, $B_1$ may get used twice, once inside each recursive call, and likewise for $B_2$. As discussed above, to avoid duplicating code, we need to add these blocks to the control-flow graph so that we can instead refer to them by name and execute them with a \key{goto}. However, as we saw in the cases above for \key{\#t} and \key{\#f}, the blocks $B_1$ or $B_2$ may not get used at all and we don't want to prematurely add them to the control-flow graph if they end up being discarded. The solution to this conundrum is to use \emph{lazy evaluation} to delay adding the blocks to the control-flow graph until the points where we know they will be used~\citep{Friedman:1976aa}.\index{lazy evaluation} Racket provides support for lazy evaluation with the \href{https://docs.racket-lang.org/reference/Delayed_Evaluation.html}{\code{racket/promise}} package. The expression \key{(delay} $e_1 \ldots e_n$\key{)} \index{delay} creates a \emph{promise}\index{promise} in which the evaluation of the expressions is postponed. When \key{(force} $p$\key{)}\index{force} is applied to a promise $p$ for the first time, the expressions $e_1 \ldots e_n$ are evaluated and the result of $e_n$ is cached in the promise and returned. If \code{force} is applied again to the same promise, then the cached result is returned. We use lazy evaluation for the input and output blocks of the functions \code{explicate-pred} and \code{explicate-assign} and for the output block of \code{explicate-tail}. So instead of taking and returning blocks, they take and return promised blocks. Furthermore, when we come to a situation in which we a block might be used more than once, as in the case for \code{if} above, we transform the promise into a new promise that will add the block to the control-flow graph and return a \code{goto}. The following auxiliary function accomplishes this task. It begins with \code{delay} to create a promise. When forced, this promise will force the input block. If that block is already a \code{goto} (because it was already added to the control-flow graph), then we return that \code{goto}. Otherwise we add the block to the control-flow graph with another auxiliary function named \code{add-node} that returns the new label, and then return the \code{goto}. \begin{lstlisting} (define (block->goto block) (delay (define b (force block)) (match b [(Goto label) (Goto label)] [else (Goto (add-node b))] ))) \end{lstlisting} Getting back to the case for \code{if} in \code{explicate-pred}, we make the recursive calls to \code{explicate-pred} on the ``then'' and ``else'' branches with the arguments \code{(block->goto} $B_1$\code{)} and \code{(block->goto} $B_2$\code{)}. Let $B_3$ and $B_4$ be the results from the two recursive calls. We complete the case for \code{if} by recursively apply \code{explicate-pred} to the condition of the \code{if} with the promised blocks $B_3$ and $B_4$ to obtain the result $B_5$. \[ (\key{if}\; \itm{cnd}\; \itm{thn}\; \itm{els}) \quad\Rightarrow\quad B_5 \] Next, consider the case for a less-than comparison in \code{explicate-pred}. We translate it to an \code{if} statement, whose two branches are required to be \code{goto}'s. So we apply \code{block->goto} to $B_1$ and $B_2$ to obtain two promised goto's, which we can \code{force} to obtain the two actual goto's $G_1$ and $G_2$. The translation of the less-than comparison is as follows. \[ (\key{<}~e_1~e_2) \quad\Rightarrow\quad \begin{array}{l} \key{if}~(\key{<}~e_1~e_2) \; G_1\\ \key{else} \; G_2 \end{array} \] The \code{explicate-tail} function needs to be updated to use lazy evaluation and it needs an additional case for \key{if}. Each of the cases that return an AST node need use \code{delay} to instead return a promise of an AST node. Recall that \code{explicate-tail} has an accumulator parameter that is a block, which now becomes a promise of a block, which we refer to as $B_0$. In the case for \code{if} in \code{explicate-tail}, the two branches inherit the current context, so they are in tail position. Thus, the recursive calls on the ``then'' and ``else'' branch should be calls to \code{explicate-tail}. % We need to pass $B_0$ as the accumulator argument for both of these recursive calls, but we need to be careful not to duplicate $B_0$. Thus, we first apply \code{block->goto} to $B_0$ so that it gets added to the control-flow graph and obtain a promised goto $G_0$. % Let $B_1$ be the result of \code{explicate-tail} on the ``then'' branch and $G_0$ and let $B_2$ be the result of \code{explicate-tail} on the ``else'' branch and $G_0$. Let $B_3$ be the result of applying \code{explicate-pred} to the condition of the \key{if}, $B_1$, and $B_2$. Then the \key{if} as a whole translates to promise $B_3$. \[ (\key{if}\; \itm{cnd}\; \itm{thn}\; \itm{els}) \quad\Rightarrow\quad B_3 \] %% In the above discussion, we use the metavariables $B_1$, $B_2$, and %% $B_3$ to refer to blocks for the purposes of our discussion, but they %% should not be confused with the labels for the blocks that appear in %% the generated code. We initially construct unlabeled blocks; we only %% attach labels to blocks when we add them to the control-flow graph, as %% we see in the next case. Next consider the case for \key{if} in the \code{explicate-assign} function. The context of the \key{if} is an assignment to some variable $x$ and then the control continues to some promised block $B_1$. The code that we generate for both the ``then'' and ``else'' branches needs to continue to $B_1$, so to avoid duplicating $B_1$ we apply \code{block->goto} to it and obtain a promised goto $G_1$. The branches of the \key{if} inherit the current context, so they are in assignment positions. Let $B_2$ be the result of applying \code{explicate-assign} to the ``then'' branch, variable $x$, and $G_1$. Let $B_3$ be the result of applying \code{explicate-assign} to the ``else'' branch, variable $x$, and $G_1$. Finally, let $B_4$ be the result of applying \code{explicate-pred} to the predicate $\itm{cnd}$ and the promises $B_2$ and $B_3$. The \key{if} as a whole translates to the promise $B_4$. \[ (\key{if}\; \itm{cnd}\; \itm{thn}\; \itm{els}) \quad\Rightarrow\quad B_4 \] This completes the description of \code{explicate-control} for $R_2$. The way in which the \code{shrink} pass transforms logical operations such as \code{and} and \code{or} can impact the quality of code generated by \code{explicate-control}. For example, consider the following program. % s1_21.rkt \begin{lstlisting} (if (and (eq? (read) 0) (eq? (read) 1)) 0 42) \end{lstlisting} The \code{and} operation should transform into something that the \code{explicate-pred} function can still analyze and descend through to reach the underlying \code{eq?} conditions. Ideally, your \code{explicate-control} pass should generate code similar to the following for the above program. \begin{center} \begin{lstlisting} start: tmp1 = (read); if (eq? tmp1 0) goto block40; else goto block39; block40: tmp2 = (read); if (eq? tmp2 1) goto block38; else goto block39; block38: return 0; block39: return 42; \end{lstlisting} \end{center} \begin{exercise}\normalfont Implement the pass \code{explicate-control} by adding the cases for \key{if} to the functions for tail and assignment contexts, and implement \code{explicate-pred} for predicate contexts. Create test cases that exercise all of the new cases in the code for this pass. \end{exercise} \section{Select Instructions} \label{sec:select-r2} \index{instruction selection} Recall that the \code{select-instructions} pass lowers from our $C$-like intermediate representation to the pseudo-x86 language, which is suitable for conducting register allocation. The pass is implemented using three auxiliary functions, one for each of the non-terminals $\Atm$, $\Stmt$, and $\Tail$. For $\Atm$, we have new cases for the Booleans. We take the usual approach of encoding them as integers, with true as 1 and false as 0. \[ \key{\#t} \Rightarrow \key{1} \qquad \key{\#f} \Rightarrow \key{0} \] For $\Stmt$, we discuss a couple cases. The \code{not} operation can be implemented in terms of \code{xorq} as we discussed at the beginning of this section. Given an assignment $\itm{var}$ \key{=} \key{(not} $\Atm$\key{);}, if the left-hand side $\itm{var}$ is the same as $\Atm$, then just the \code{xorq} suffices. \[ \Var~\key{=}~ \key{(not}\; \Var\key{);} \quad\Rightarrow\quad \key{xorq}~\key{\$}1\key{,}~\Var \] Otherwise, a \key{movq} is needed to adapt to the update-in-place semantics of x86. Let $\Arg$ be the result of translating $\Atm$ to x86. Then we have \[ \Var~\key{=}~ \key{(not}\; \Atm\key{);} \quad\Rightarrow\quad \begin{array}{l} \key{movq}~\Arg\key{,}~\Var\\ \key{xorq}~\key{\$}1\key{,}~\Var \end{array} \] Next consider the cases for \code{eq?} and less-than comparison. Translating these operations to x86 is slightly involved due to the unusual nature of the \key{cmpq} instruction discussed above. We recommend translating an assignment from \code{eq?} into the following sequence of three instructions. \\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} |$\Var$| = (eq? |$\Atm_1$| |$\Atm_2$|); \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} cmpq |$\Arg_2$|, |$\Arg_1$| sete %al movzbq %al, |$\Var$| \end{lstlisting} \end{minipage} \end{tabular} \\ Regarding the $\Tail$ non-terminal, we have two new cases: \key{goto} and conditional \key{goto}. Both are straightforward to handle. A \key{goto} becomes a jump instruction. \[ \key{goto}\; \ell\key{;} \quad \Rightarrow \quad \key{jmp}\;\ell \] A conditional \key{goto} becomes a compare instruction followed by a conditional jump (for ``then'') and the fall-through is to a regular jump (for ``else'').\\ \begin{tabular}{lll} \begin{minipage}{0.4\textwidth} \begin{lstlisting} if (eq? |$\Atm_1$| |$\Atm_2$|) goto |$\ell_1$|; else goto |$\ell_2$|; \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} cmpq |$\Arg_2$|, |$\Arg_1$| je |$\ell_1$| jmp |$\ell_2$| \end{lstlisting} \end{minipage} \end{tabular} \\ \begin{exercise}\normalfont Expand your \code{select-instructions} pass to handle the new features of the $R_2$ language. Test the pass on all the examples you have created and make sure that you have some test programs that use the \code{eq?} and \code{<} operators, creating some if necessary. Test the output using the \code{interp-x86} interpreter (Appendix~\ref{appendix:interp}). \end{exercise} \section{Register Allocation} \label{sec:register-allocation-r2} \index{register allocation} The changes required for $R_2$ affect liveness analysis, building the interference graph, and assigning homes, but the graph coloring algorithm itself does not change. \subsection{Liveness Analysis} \label{sec:liveness-analysis-r2} \index{liveness analysis} Recall that for $R_1$ we implemented liveness analysis for a single basic block (Section~\ref{sec:liveness-analysis-r1}). With the addition of \key{if} expressions to $R_2$, \code{explicate-control} produces many basic blocks arranged in a control-flow graph. We recommend that you create a new auxiliary function named \code{uncover-live-CFG} that applies liveness analysis to a control-flow graph. The first question we need to consider is: what order should we process the basic blocks in the control-flow graph? To perform liveness analysis on a basic block, we need to know its live-after set. If a basic block has no successor blocks (i.e. no out-edges in the control flow graph), then it has an empty live-after set and we can immediately apply liveness analysis to it. If a basic block has some successors, then we need to complete liveness analysis on those blocks first. Thankfully, the control flow graph does not contain any cycles because $R_2$ does not include loops. (In Chapter~\ref{ch:loop} we add loops and study how to handle cycles in the control-flow graph.) % Returning to the question of what order should we process the basic blocks, the answer is reverse topological order. We recommend using the \code{tsort} (topological sort) and \code{transpose} functions of the Racket \code{graph} package to obtain this ordering. \index{topological order} \index{topological sort} The next question is how to analyze the jump instructions. In Section~\ref{sec:liveness-analysis-r1} we recommended that you maintain an alist named \code{label->live} that maps each label to the set of live locations at the beginning of the associated block. Now that we have many basic blocks, the alist needs to be extended as we process the blocks. In particular, after performing liveness analysis on a block, we can take the live-before set for its first instruction and associate that with the block's label in the alist. % As discussed in Section~\ref{sec:liveness-analysis-r1}, the live-before set for a $\JMP{\itm{label}}$ instruction is given by the mapping for $\itm{label}$ in \code{label->live}. Now for $x86_1$ we also have the conditional jump $\JMPIF{\itm{cc}}{\itm{label}}$ to deal with. This one is particularly interesting because during compilation we do not know, in general, which way a conditional jump will go, so we do not know whether to use the live-before set for the following instruction or the live-before set for $\itm{label}$. The solution to this challenge is based on the observation that there is no harm to the correctness of the compiler if we classify more locations as live than the ones that are truly live during a particular execution of the instruction. Thus, we can take the union of the live-before sets from the following instruction and from the mapping fro $\itm{label}$ in \code{label->live}. The helper functions for computing the variables in an instruction's argument and for computing the variables read-from ($R$) or written-to ($W$) by an instruction need to be updated to handle the new kinds of arguments and instructions in x86$_1$. \subsection{Build Interference} \label{sec:build-interference-r2} Many of the new instructions in x86$_1$ can be handled in the same way as the instructions in x86$_0$. Thus, if your code was already quite general, it will not need to be changed to handle the new instructions. If you code is not general enough, I recommend that you change your code to be more general. For example, you can factor out the computing of the the read and write sets for each kind of instruction into two auxiliary functions. Note that the \key{movzbq} instruction requires some special care, just like the \key{movq} instruction. See rule number 3 in Section~\ref{sec:build-interference}. %% \subsection{Assign Homes} %% \label{sec:assign-homes-r2} %% The \code{assign-homes} function (Section~\ref{sec:assign-r1}) needs %% to be updated to handle the \key{if} statement, simply by recursively %% processing the child nodes. Hopefully your code already handles the %% other new instructions, but if not, you can generalize your code. \begin{exercise}\normalfont Update the \code{register-allocation} pass so that it works for $R_2$ and test your compiler using your previously created programs on the \code{interp-x86} interpreter (Appendix~\ref{appendix:interp}). \end{exercise} \section{Patch Instructions} The second argument of the \key{cmpq} instruction must not be an immediate value (such as an integer). So if you are comparing two immediates, we recommend inserting a \key{movq} instruction to put the second argument in \key{rax}. Also, recall that instructions may have at most one memory reference. % The second argument of the \key{movzbq} must be a register. % There are no special restrictions on the x86 instructions \key{JmpIf} and \key{Jmp}. \begin{exercise}\normalfont Update \code{patch-instructions} to handle the new x86 instructions. Test your compiler using your previously created programs on the \code{interp-x86} interpreter (Appendix~\ref{appendix:interp}). \end{exercise} \begin{figure}[tbp] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R2) at (0,2) {\large $R_2$}; \node (R2-2) at (3,2) {\large $R_2$}; \node (R2-3) at (6,2) {\large $R_2$}; \node (R2-4) at (9,2) {\large $R_2$}; \node (R2-5) at (12,2) {\large $R_2$}; \node (C1-1) at (3,0) {\large $C_1$}; \node (x86-2) at (3,-2) {\large $\text{x86}^{*}_1$}; \node (x86-3) at (6,-2) {\large $\text{x86}^{*}_1$}; \node (x86-4) at (9,-2) {\large $\text{x86}^{*}_1$}; \node (x86-5) at (9,-4) {\large $\text{x86}^{\dagger}_1$}; \node (x86-2-1) at (3,-4) {\large $\text{x86}^{*}_1$}; \node (x86-2-2) at (6,-4) {\large $\text{x86}^{*}_1$}; \path[->,bend left=15] (R2) edge [above] node {\ttfamily\footnotesize type-check} (R2-2); \path[->,bend left=15] (R2-2) edge [above] node {\ttfamily\footnotesize shrink} (R2-3); \path[->,bend left=15] (R2-3) edge [above] node {\ttfamily\footnotesize uniquify} (R2-4); \path[->,bend left=15] (R2-4) edge [above] node {\ttfamily\footnotesize remove-complex.} (R2-5); \path[->,bend left=15] (R2-5) edge [left] node {\ttfamily\footnotesize explicate-control} (C1-1); \path[->,bend right=15] (C1-1) edge [left] node {\ttfamily\footnotesize select-instructions} (x86-2); \path[->,bend left=15] (x86-2) edge [right] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [right] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [right] node {\ttfamily\footnotesize print-x86 } (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_2$, a language with conditionals.} \label{fig:R2-passes} \end{figure} Figure~\ref{fig:R2-passes} lists all the passes needed for the compilation of $R_2$. \section{An Example Translation} Figure~\ref{fig:if-example-x86} shows a simple example program in $R_2$ translated to x86, showing the results of \code{explicate-control}, \code{select-instructions}, and the final x86 assembly code. \begin{figure}[tbp] \begin{tabular}{lll} \begin{minipage}{0.5\textwidth} % s1_20.rkt \begin{lstlisting} (if (eq? (read) 1) 42 0) \end{lstlisting} $\Downarrow$ \begin{lstlisting} start: tmp7951 = (read); if (eq? tmp7951 1) then goto block7952; else goto block7953; block7952: return 42; block7953: return 0; \end{lstlisting} $\Downarrow$ \begin{lstlisting} start: callq read_int movq %rax, tmp7951 cmpq $1, tmp7951 je block7952 jmp block7953 block7953: movq $0, %rax jmp conclusion block7952: movq $42, %rax jmp conclusion \end{lstlisting} \end{minipage} & $\Rightarrow\qquad$ \begin{minipage}{0.4\textwidth} \begin{lstlisting} start: callq read_int movq %rax, %rcx cmpq $1, %rcx je block7952 jmp block7953 block7953: movq $0, %rax jmp conclusion block7952: movq $42, %rax jmp conclusion .globl main main: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx pushq %r14 subq $0, %rsp jmp start conclusion: addq $0, %rsp popq %r14 popq %rbx popq %r12 popq %r13 popq %rbp retq \end{lstlisting} \end{minipage} \end{tabular} \caption{Example compilation of an \key{if} expression to x86.} \label{fig:if-example-x86} \end{figure} \section{Challenge: Remove Jumps} \label{sec:opt-jumps} %% Recall that in the example output of \code{explicate-control} in %% Figure~\ref{fig:explicate-control-s1-38}, \code{block57} through %% \code{block60} are trivial blocks, they do nothing but jump to another %% block. The first goal of this challenge assignment is to remove those %% blocks. Figure~\ref{fig:optimize-jumps} repeats the result of %% \code{explicate-control} on the left and shows the result of bypassing %% the trivial blocks on the right. Let us focus on \code{block61}. The %% \code{then} branch jumps to \code{block57}, which in turn jumps to %% \code{block55}. The optimized code on the right of %% Figure~\ref{fig:optimize-jumps} bypasses \code{block57}, with the %% \code{then} branch jumping directly to \code{block55}. The story is %% similar for the \code{else} branch, as well as for the two branches in %% \code{block62}. After the jumps in \code{block61} and \code{block62} %% have been optimized in this way, there are no longer any jumps to %% blocks \code{block57} through \code{block60}, so they can be removed. %% \begin{figure}[tbp] %% \begin{tabular}{lll} %% \begin{minipage}{0.4\textwidth} %% \begin{lstlisting} %% block62: %% tmp54 = (read); %% if (eq? tmp54 2) then %% goto block59; %% else %% goto block60; %% block61: %% tmp53 = (read); %% if (eq? tmp53 0) then %% goto block57; %% else %% goto block58; %% block60: %% goto block56; %% block59: %% goto block55; %% block58: %% goto block56; %% block57: %% goto block55; %% block56: %% return (+ 700 77); %% block55: %% return (+ 10 32); %% start: %% tmp52 = (read); %% if (eq? tmp52 1) then %% goto block61; %% else %% goto block62; %% \end{lstlisting} %% \end{minipage} %% & %% $\Rightarrow$ %% & %% \begin{minipage}{0.55\textwidth} %% \begin{lstlisting} %% block62: %% tmp54 = (read); %% if (eq? tmp54 2) then %% goto block55; %% else %% goto block56; %% block61: %% tmp53 = (read); %% if (eq? tmp53 0) then %% goto block55; %% else %% goto block56; %% block56: %% return (+ 700 77); %% block55: %% return (+ 10 32); %% start: %% tmp52 = (read); %% if (eq? tmp52 1) then %% goto block61; %% else %% goto block62; %% \end{lstlisting} %% \end{minipage} %% \end{tabular} %% \caption{Optimize jumps by removing trivial blocks.} %% \label{fig:optimize-jumps} %% \end{figure} %% The name of this pass is \code{optimize-jumps}. We recommend %% implementing this pass in two phases. The first phrase builds a hash %% table that maps labels to possibly improved labels. The second phase %% changes the target of each \code{goto} to use the improved label. If %% the label is for a trivial block, then the hash table should map the %% label to the first non-trivial block that can be reached from this %% label by jumping through trivial blocks. If the label is for a %% non-trivial block, then the hash table should map the label to itself; %% we do not want to change jumps to non-trivial blocks. %% The first phase can be accomplished by constructing an empty hash %% table, call it \code{short-cut}, and then iterating over the control %% flow graph. Each time you encouter a block that is just a \code{goto}, %% then update the hash table, mapping the block's source to the target %% of the \code{goto}. Also, the hash table may already have mapped some %% labels to the block's source, to you must iterate through the hash %% table and update all of those so that they instead map to the target %% of the \code{goto}. %% For the second phase, we recommend iterating through the $\Tail$ of %% each block in the program, updating the target of every \code{goto} %% according to the mapping in \code{short-cut}. %% \begin{exercise}\normalfont %% Implement the \code{optimize-jumps} pass as a transformation from %% $C_1$ to $C_1$, coming after the \code{explicate-control} pass. %% Check that \code{optimize-jumps} removes trivial blocks in a few %% example programs. Then check that your compiler still passes all of %% your tests. %% \end{exercise} There is an opportunity for optimizing jumps that is apparent in the example of Figure~\ref{fig:if-example-x86}. The \code{start} block end with a jump to \code{block7953} and there are no other jumps to \code{block7953} in the rest of the program. In this situation we can avoid the runtime overhead of this jump by merging \code{block7953} into the preceding block, in this case the \code{start} block. Figure~\ref{fig:remove-jumps} shows the output of \code{select-instructions} on the left and the result of this optimization on the right. \begin{figure}[tbp] \begin{tabular}{lll} \begin{minipage}{0.5\textwidth} % s1_20.rkt \begin{lstlisting} start: callq read_int movq %rax, tmp7951 cmpq $1, tmp7951 je block7952 jmp block7953 block7953: movq $0, %rax jmp conclusion block7952: movq $42, %rax jmp conclusion \end{lstlisting} \end{minipage} & $\Rightarrow\qquad$ \begin{minipage}{0.4\textwidth} \begin{lstlisting} start: callq read_int movq %rax, tmp7951 cmpq $1, tmp7951 je block7952 movq $0, %rax jmp conclusion block7952: movq $42, %rax jmp conclusion \end{lstlisting} \end{minipage} \end{tabular} \caption{Merging basic blocks by removing unnecessary jumps.} \label{fig:remove-jumps} \end{figure} \begin{exercise}\normalfont Implement a pass named \code{remove-jumps} that merges basic blocks into their preceding basic block, when there is only one preceding block. The pass should translate from pseudo $x86_1$ to pseudo $x86_1$ and it should come immediately after \code{select-instructions}. Check that \code{remove-jumps} accomplishes the goal of merging basic blocks on several test programs and check that your compiler passes all of your tests. \end{exercise} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Tuples and Garbage Collection} \label{ch:tuples} \index{tuple} \index{vector} \margincomment{\scriptsize To do: challenge assignments: mark-and-sweep, add simple structures. \\ --Jeremy} \margincomment{\scriptsize To do: look through Andre's code comments for extra things to discuss in this chapter. \\ --Jeremy} \margincomment{\scriptsize To do: Flesh out this chapter, e.g., make sure all the IR grammars are spelled out! \\ --Jeremy} \margincomment{\scriptsize Introduce has-type, but after flatten, remove it, but keep type annotations on vector creation and local variables, function parameters, etc. \\ --Jeremy} \margincomment{\scriptsize Be more explicit about how to deal with the root stack. \\ --Jeremy} In this chapter we study the implementation of mutable tuples (called ``vectors'' in Racket). This language feature is the first to use the computer's \emph{heap}\index{heap} because the lifetime of a Racket tuple is indefinite, that is, a tuple lives forever from the programmer's viewpoint. Of course, from an implementer's viewpoint, it is important to reclaim the space associated with a tuple when it is no longer needed, which is why we also study \emph{garbage collection} \emph{garbage collection} techniques in this chapter. Section~\ref{sec:r3} introduces the $R_3$ language including its interpreter and type checker. The $R_3$ language extends the $R_2$ language of Chapter~\ref{ch:bool-types} with vectors and Racket's \code{void} value. The reason for including the later is that the \code{vector-set!} operation returns a value of type \code{Void}\footnote{Racket's \code{Void} type corresponds to what is called the \code{Unit} type in the programming languages literature. Racket's \code{Void} type is inhabited by a single value \code{void} which corresponds to \code{unit} or \code{()} in the literature~\citep{Pierce:2002hj}.}. Section~\ref{sec:GC} describes a garbage collection algorithm based on copying live objects back and forth between two halves of the heap. The garbage collector requires coordination with the compiler so that it can see all of the \emph{root} pointers, that is, pointers in registers or on the procedure call stack. Sections~\ref{sec:expose-allocation} through \ref{sec:print-x86-gc} discuss all the necessary changes and additions to the compiler passes, including a new compiler pass named \code{expose-allocation}. \section{The $R_3$ Language} \label{sec:r3} Figure~\ref{fig:r3-concrete-syntax} defines the concrete syntax for $R_3$ and Figure~\ref{fig:r3-syntax} defines the abstract syntax. The $R_3$ language includes three new forms: \code{vector} for creating a tuple, \code{vector-ref} for reading an element of a tuple, and \code{vector-set!} for writing to an element of a tuple. The program in Figure~\ref{fig:vector-eg} shows the usage of tuples in Racket. We create a 3-tuple \code{t} and a 1-tuple that is stored at index $2$ of the 3-tuple, demonstrating that tuples are first-class values. The element at index $1$ of \code{t} is \code{\#t}, so the ``then'' branch of the \key{if} is taken. The element at index $0$ of \code{t} is \code{40}, to which we add \code{2}, the element at index $0$ of the 1-tuple. So the result of the program is \code{42}. \begin{figure}[tbp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Type &::=& \gray{\key{Integer} \mid \key{Boolean}} \mid \LP\key{Vector}\;\Type\ldots\RP \mid \key{Void}\\ \Exp &::=& \gray{ \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} \mid \CSUB{\Exp}{\Exp} } \\ &\mid& \gray{ \Var \mid \CLET{\Var}{\Exp}{\Exp} }\\ &\mid& \gray{ \key{\#t} \mid \key{\#f} \mid \LP\key{and}\;\Exp\;\Exp\RP \mid \LP\key{or}\;\Exp\;\Exp\RP \mid \LP\key{not}\;\Exp\RP } \\ &\mid& \gray{ \LP\itm{cmp}\;\Exp\;\Exp\RP \mid \CIF{\Exp}{\Exp}{\Exp} } \\ &\mid& \LP\key{vector}\;\Exp\ldots\RP \mid \LP\key{vector-length}\;\Exp\RP \\ &\mid& \LP\key{vector-ref}\;\Exp\;\Int\RP \mid \LP\key{vector-set!}\;\Exp\;\Int\;\Exp\RP \\ &\mid& \LP\key{void}\RP \mid \LP\key{has-type}~\Exp~\Type\RP\\ R_3 &::=& \Exp \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_3$, extending $R_2$ (Figure~\ref{fig:r2-concrete-syntax}).} \label{fig:r3-concrete-syntax} \end{figure} \begin{figure}[tbp] \begin{lstlisting} (let ([t (vector 40 #t (vector 2))]) (if (vector-ref t 1) (+ (vector-ref t 0) (vector-ref (vector-ref t 2) 0)) 44)) \end{lstlisting} \caption{Example program that creates tuples and reads from them.} \label{fig:vector-eg} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \itm{op} &::=& \ldots \mid \code{vector} \mid \code{vector-length} \\ \Exp &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \PRIM{\itm{op}}{\Exp\ldots} \mid \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} } \\ &\mid& \VECREF{\Exp}{\INT{\Int}}\\ &\mid& \VECSET{\Exp}{\INT{\Int}}{\Exp} \\ &\mid& \VOID{} \mid \LP\key{HasType}~\Exp~\Type \RP \\ R_3 &::=& \PROGRAM{\key{'()}}{\Exp} \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_3$.} \label{fig:r3-syntax} \end{figure} \index{allocate} \index{heap allocate} Tuples are our first encounter with heap-allocated data, which raises several interesting issues. First, variable binding performs a shallow-copy when dealing with tuples, which means that different variables can refer to the same tuple, that is, different variables can be \emph{aliases} for the same entity. Consider the following example in which both \code{t1} and \code{t2} refer to the same tuple. Thus, the mutation through \code{t2} is visible when referencing the tuple from \code{t1}, so the result of this program is \code{42}. \index{alias}\index{mutation} \begin{center} \begin{minipage}{0.96\textwidth} \begin{lstlisting} (let ([t1 (vector 3 7)]) (let ([t2 t1]) (let ([_ (vector-set! t2 0 42)]) (vector-ref t1 0)))) \end{lstlisting} \end{minipage} \end{center} The next issue concerns the lifetime of tuples. Of course, they are created by the \code{vector} form, but when does their lifetime end? Notice that $R_3$ does not include an operation for deleting tuples. Furthermore, the lifetime of a tuple is not tied to any notion of static scoping. For example, the following program returns \code{42} even though the variable \code{w} goes out of scope prior to the \code{vector-ref} that reads from the vector it was bound to. \begin{center} \begin{minipage}{0.96\textwidth} \begin{lstlisting} (let ([v (vector (vector 44))]) (let ([x (let ([w (vector 42)]) (let ([_ (vector-set! v 0 w)]) 0))]) (+ x (vector-ref (vector-ref v 0) 0)))) \end{lstlisting} \end{minipage} \end{center} From the perspective of programmer-observable behavior, tuples live forever. Of course, if they really lived forever, then many programs would run out of memory.\footnote{The $R_3$ language does not have looping or recursive functions, so it is nigh impossible to write a program in $R_3$ that will run out of memory. However, we add recursive functions in the next Chapter!} A Racket implementation must therefore perform automatic garbage collection. Figure~\ref{fig:interp-R3} shows the definitional interpreter for the $R_3$ language. We define the \code{vector}, \code{vector-length}, \code{vector-ref}, and \code{vector-set!} operations for $R_3$ in terms of the corresponding operations in Racket. One subtle point is that the \code{vector-set!} operation returns the \code{\#} value. The \code{\#} value can be passed around just like other values inside an $R_3$ program and a \code{\#} value can be compared for equality with another \code{\#} value. However, there are no other operations specific to the the \code{\#} value in $R_3$. In contrast, Racket defines the \code{void?} predicate that returns \code{\#t} when applied to \code{\#} and \code{\#f} otherwise. \begin{figure}[tbp] \begin{lstlisting} (define interp-R3-class (class interp-R2-class (super-new) (define/override (interp-op op) (match op ['eq? (lambda (v1 v2) (cond [(or (and (fixnum? v1) (fixnum? v2)) (and (boolean? v1) (boolean? v2)) (and (vector? v1) (vector? v2)) (and (void? v1) (void? v2))) (eq? v1 v2)]))] ['vector vector] ['vector-length vector-length] ['vector-ref vector-ref] ['vector-set! vector-set!] [else (super interp-op op)] )) (define/override ((interp-exp env) e) (define recur (interp-exp env)) (match e [(HasType e t) (recur e)] [(Void) (void)] [else ((super interp-exp env) e)] )) )) (define (interp-R3 p) (send (new interp-R3-class) interp-program p)) \end{lstlisting} \caption{Interpreter for the $R_3$ language.} \label{fig:interp-R3} \end{figure} Figure~\ref{fig:type-check-R3} shows the type checker for $R_3$, which deserves some explanation. When allocating a vector, we need to know which elements of the vector are pointers (i.e. are also vectors). We can obtain this information during type checking. The type checker in Figure~\ref{fig:type-check-R3} not only computes the type of an expression, it also wraps every \key{vector} creation with the form $(\key{HasType}~e~T)$, where $T$ is the vector's type. % To create the s-expression for the \code{Vector} type in Figure~\ref{fig:type-check-R3}, we use the \href{https://docs.racket-lang.org/reference/quasiquote.html}{unquote-splicing operator} \code{,@} to insert the list \code{t*} without its usual start and end parentheses. \index{unquote-slicing} \begin{figure}[tp] \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] (define type-check-R3-class (class type-check-R2-class (super-new) (inherit check-type-equal?) (define/override (type-check-exp env) (lambda (e) (define recur (type-check-exp env)) (match e [(Void) (values (Void) 'Void)] [(Prim 'vector es) (define-values (e* t*) (for/lists (e* t*) ([e es]) (recur e))) (define t `(Vector ,@t*)) (values (HasType (Prim 'vector e*) t) t)] [(Prim 'vector-ref (list e1 (Int i))) (define-values (e1^ t) (recur e1)) (match t [`(Vector ,ts ...) (unless (and (0 . <= . i) (i . < . (length ts))) (error 'type-check "index ~a out of bounds\nin ~v" i e)) (values (Prim 'vector-ref (list e1^ (Int i))) (list-ref ts i))] [else (error 'type-check "expect Vector, not ~a\nin ~v" t e)])] [(Prim 'vector-set! (list e1 (Int i) arg) ) (define-values (e-vec t-vec) (recur e1)) (define-values (e-arg^ t-arg) (recur arg)) (match t-vec [`(Vector ,ts ...) (unless (and (0 . <= . i) (i . < . (length ts))) (error 'type-check "index ~a out of bounds\nin ~v" i e)) (check-type-equal? (list-ref ts i) t-arg e) (values (Prim 'vector-set! (list e-vec (Int i) e-arg^)) 'Void)] [else (error 'type-check "expect Vector, not ~a\nin ~v" t-vec e)])] [(Prim 'vector-length (list e)) (define-values (e^ t) (recur e)) (match t [`(Vector ,ts ...) (values (Prim 'vector-length (list e^)) 'Integer)] [else (error 'type-check "expect Vector, not ~a\nin ~v" t e)])] [(Prim 'eq? (list arg1 arg2)) (define-values (e1 t1) (recur arg1)) (define-values (e2 t2) (recur arg2)) (match* (t1 t2) [(`(Vector ,ts1 ...) `(Vector ,ts2 ...)) (void)] [(other wise) (check-type-equal? t1 t2 e)]) (values (Prim 'eq? (list e1 e2)) 'Boolean)] [(HasType (Prim 'vector es) t) ((type-check-exp env) (Prim 'vector es))] [(HasType e1 t) (define-values (e1^ t^) (recur e1)) (check-type-equal? t t^ e) (values (HasType e1^ t) t)] [else ((super type-check-exp env) e)] ))) )) (define (type-check-R3 p) (send (new type-check-R3-class) type-check-program p)) \end{lstlisting} \caption{Type checker for the $R_3$ language.} \label{fig:type-check-R3} \end{figure} \section{Garbage Collection} \label{sec:GC} Here we study a relatively simple algorithm for garbage collection that is the basis of state-of-the-art garbage collectors~\citep{Lieberman:1983aa,Ungar:1984aa,Jones:1996aa,Detlefs:2004aa,Dybvig:2006aa,Tene:2011kx}. In particular, we describe a two-space copying collector~\citep{Wilson:1992fk} that uses Cheney's algorithm to perform the copy~\citep{Cheney:1970aa}. \index{copying collector} \index{two-space copying collector} Figure~\ref{fig:copying-collector} gives a coarse-grained depiction of what happens in a two-space collector, showing two time steps, prior to garbage collection (on the top) and after garbage collection (on the bottom). In a two-space collector, the heap is divided into two parts named the FromSpace and the ToSpace. Initially, all allocations go to the FromSpace until there is not enough room for the next allocation request. At that point, the garbage collector goes to work to make more room. \index{ToSpace} \index{FromSpace} The garbage collector must be careful not to reclaim tuples that will be used by the program in the future. Of course, it is impossible in general to predict what a program will do, but we can over approximate the will-be-used tuples by preserving all tuples that could be accessed by \emph{any} program given the current computer state. A program could access any tuple whose address is in a register or on the procedure call stack. These addresses are called the \emph{root set}\index{root set}. In addition, a program could access any tuple that is transitively reachable from the root set. Thus, it is safe for the garbage collector to reclaim the tuples that are not reachable in this way. So the goal of the garbage collector is twofold: \begin{enumerate} \item preserve all tuple that are reachable from the root set via a path of pointers, that is, the \emph{live} tuples, and \item reclaim the memory of everything else, that is, the \emph{garbage}. \end{enumerate} A copying collector accomplishes this by copying all of the live objects from the FromSpace into the ToSpace and then performs a slight of hand, treating the ToSpace as the new FromSpace and the old FromSpace as the new ToSpace. In the example of Figure~\ref{fig:copying-collector}, there are three pointers in the root set, one in a register and two on the stack. All of the live objects have been copied to the ToSpace (the right-hand side of Figure~\ref{fig:copying-collector}) in a way that preserves the pointer relationships. For example, the pointer in the register still points to a 2-tuple whose first element is a 3-tuple and whose second element is a 2-tuple. There are four tuples that are not reachable from the root set and therefore do not get copied into the ToSpace. The exact situation in Figure~\ref{fig:copying-collector} cannot be created by a well-typed program in $R_3$ because it contains a cycle. However, creating cycles will be possible once we get to $R_6$. We design the garbage collector to deal with cycles to begin with so we will not need to revisit this issue. \begin{figure}[tbp] \centering \includegraphics[width=\textwidth]{figs/copy-collect-1} \\[5ex] \includegraphics[width=\textwidth]{figs/copy-collect-2} \caption{A copying collector in action.} \label{fig:copying-collector} \end{figure} There are many alternatives to copying collectors (and their bigger siblings, the generational collectors) when its comes to garbage collection, such as mark-and-sweep~\citep{McCarthy:1960dz} and reference counting~\citep{Collins:1960aa}. The strengths of copying collectors are that allocation is fast (just a comparison and pointer increment), there is no fragmentation, cyclic garbage is collected, and the time complexity of collection only depends on the amount of live data, and not on the amount of garbage~\citep{Wilson:1992fk}. The main disadvantages of a two-space copying collector is that it uses a lot of space and takes a long time to perform the copy, though these problems are ameliorated in generational collectors. Racket and Scheme programs tend to allocate many small objects and generate a lot of garbage, so copying and generational collectors are a good fit. Garbage collection is an active research topic, especially concurrent garbage collection~\citep{Tene:2011kx}. Researchers are continuously developing new techniques and revisiting old trade-offs~\citep{Blackburn:2004aa,Jones:2011aa,Shahriyar:2013aa,Cutler:2015aa,Shidal:2015aa,Osterlund:2016aa,Jacek:2019aa,Gamari:2020aa}. Researchers meet every year at the International Symposium on Memory Management to present these findings. \subsection{Graph Copying via Cheney's Algorithm} \label{sec:cheney} \index{Cheney's algorithm} Let us take a closer look at the copying of the live objects. The allocated objects and pointers can be viewed as a graph and we need to copy the part of the graph that is reachable from the root set. To make sure we copy all of the reachable vertices in the graph, we need an exhaustive graph traversal algorithm, such as depth-first search or breadth-first search~\citep{Moore:1959aa,Cormen:2001uq}. Recall that such algorithms take into account the possibility of cycles by marking which vertices have already been visited, so as to ensure termination of the algorithm. These search algorithms also use a data structure such as a stack or queue as a to-do list to keep track of the vertices that need to be visited. We use breadth-first search and a trick due to \citet{Cheney:1970aa} for simultaneously representing the queue and copying tuples into the ToSpace. Figure~\ref{fig:cheney} shows several snapshots of the ToSpace as the copy progresses. The queue is represented by a chunk of contiguous memory at the beginning of the ToSpace, using two pointers to track the front and the back of the queue. The algorithm starts by copying all tuples that are immediately reachable from the root set into the ToSpace to form the initial queue. When we copy a tuple, we mark the old tuple to indicate that it has been visited. We discuss how this marking is accomplish in Section~\ref{sec:data-rep-gc}. Note that any pointers inside the copied tuples in the queue still point back to the FromSpace. Once the initial queue has been created, the algorithm enters a loop in which it repeatedly processes the tuple at the front of the queue and pops it off the queue. To process a tuple, the algorithm copies all the tuple that are directly reachable from it to the ToSpace, placing them at the back of the queue. The algorithm then updates the pointers in the popped tuple so they point to the newly copied tuples. \begin{figure}[tbp] \centering \includegraphics[width=0.9\textwidth]{figs/cheney} \caption{Depiction of the Cheney algorithm copying the live tuples.} \label{fig:cheney} \end{figure} Getting back to Figure~\ref{fig:cheney}, in the first step we copy the tuple whose second element is $42$ to the back of the queue. The other pointer goes to a tuple that has already been copied, so we do not need to copy it again, but we do need to update the pointer to the new location. This can be accomplished by storing a \emph{forwarding pointer} to the new location in the old tuple, back when we initially copied the tuple into the ToSpace. This completes one step of the algorithm. The algorithm continues in this way until the front of the queue is empty, that is, until the front catches up with the back. \subsection{Data Representation} \label{sec:data-rep-gc} The garbage collector places some requirements on the data representations used by our compiler. First, the garbage collector needs to distinguish between pointers and other kinds of data. There are several ways to accomplish this. \begin{enumerate} \item Attached a tag to each object that identifies what type of object it is~\citep{McCarthy:1960dz}. \item Store different types of objects in different regions~\citep{Steele:1977ab}. \item Use type information from the program to either generate type-specific code for collecting or to generate tables that can guide the collector~\citep{Appel:1989aa,Goldberg:1991aa,Diwan:1992aa}. \end{enumerate} Dynamically typed languages, such as Lisp, need to tag objects anyways, so option 1 is a natural choice for those languages. However, $R_3$ is a statically typed language, so it would be unfortunate to require tags on every object, especially small and pervasive objects like integers and Booleans. Option 3 is the best-performing choice for statically typed languages, but comes with a relatively high implementation complexity. To keep this chapter within a 2-week time budget, we recommend a combination of options 1 and 2, using separate strategies for the stack and the heap. Regarding the stack, we recommend using a separate stack for pointers, which we call a \emph{root stack}\index{root stack} (a.k.a. ``shadow stack'')~\citep{Siebert:2001aa,Henderson:2002aa,Baker:2009aa}. That is, when a local variable needs to be spilled and is of type \code{(Vector $\Type_1 \ldots \Type_n$)}, then we put it on the root stack instead of the normal procedure call stack. Furthermore, we always spill vector-typed variables if they are live during a call to the collector, thereby ensuring that no pointers are in registers during a collection. Figure~\ref{fig:shadow-stack} reproduces the example from Figure~\ref{fig:copying-collector} and contrasts it with the data layout using a root stack. The root stack contains the two pointers from the regular stack and also the pointer in the second register. \begin{figure}[tbp] \centering \includegraphics[width=0.60\textwidth]{figs/root-stack} \caption{Maintaining a root stack to facilitate garbage collection.} \label{fig:shadow-stack} \end{figure} The problem of distinguishing between pointers and other kinds of data also arises inside of each tuple on the heap. We solve this problem by attaching a tag, an extra 64-bits, to each tuple. Figure~\ref{fig:tuple-rep} zooms in on the tags for two of the tuples in the example from Figure~\ref{fig:copying-collector}. Note that we have drawn the bits in a big-endian way, from right-to-left, with bit location 0 (the least significant bit) on the far right, which corresponds to the direction of the x86 shifting instructions \key{salq} (shift left) and \key{sarq} (shift right). Part of each tag is dedicated to specifying which elements of the tuple are pointers, the part labeled ``pointer mask''. Within the pointer mask, a 1 bit indicates there is a pointer and a 0 bit indicates some other kind of data. The pointer mask starts at bit location 7. We have limited tuples to a maximum size of 50 elements, so we just need 50 bits for the pointer mask. The tag also contains two other pieces of information. The length of the tuple (number of elements) is stored in bits location 1 through 6. Finally, the bit at location 0 indicates whether the tuple has yet to be copied to the ToSpace. If the bit has value 1, then this tuple has not yet been copied. If the bit has value 0 then the entire tag is a forwarding pointer. (The lower 3 bits of a pointer are always zero anyways because our tuples are 8-byte aligned.) \begin{figure}[tbp] \centering \includegraphics[width=0.8\textwidth]{figs/tuple-rep} \caption{Representation of tuples in the heap.} \label{fig:tuple-rep} \end{figure} \subsection{Implementation of the Garbage Collector} \label{sec:organize-gz} \index{prelude} An implementation of the copying collector is provided in the \code{runtime.c} file. Figure~\ref{fig:gc-header} defines the interface to the garbage collector that is used by the compiler. The \code{initialize} function creates the FromSpace, ToSpace, and root stack and should be called in the prelude of the \code{main} function. The arguments of \code{initialize} are the root stack size and the heap size. Both need to be multiples of $64$ and $16384$ is a good choice for both. The \code{initialize} function puts the address of the beginning of the FromSpace into the global variable \code{free\_ptr}. The global variable \code{fromspace\_end} points to the address that is 1-past the last element of the FromSpace. (We use half-open intervals to represent chunks of memory~\citep{Dijkstra:1982aa}.) The \code{rootstack\_begin} variable points to the first element of the root stack. As long as there is room left in the FromSpace, your generated code can allocate tuples simply by moving the \code{free\_ptr} forward. % The amount of room left in FromSpace is the difference between the \code{fromspace\_end} and the \code{free\_ptr}. The \code{collect} function should be called when there is not enough room left in the FromSpace for the next allocation. The \code{collect} function takes a pointer to the current top of the root stack (one past the last item that was pushed) and the number of bytes that need to be allocated. The \code{collect} function performs the copying collection and leaves the heap in a state such that the next allocation will succeed. \begin{figure}[tbp] \begin{lstlisting} void initialize(uint64_t rootstack_size, uint64_t heap_size); void collect(int64_t** rootstack_ptr, uint64_t bytes_requested); int64_t* free_ptr; int64_t* fromspace_begin; int64_t* fromspace_end; int64_t** rootstack_begin; \end{lstlisting} \caption{The compiler's interface to the garbage collector.} \label{fig:gc-header} \end{figure} %% \begin{exercise} %% In the file \code{runtime.c} you will find the implementation of %% \code{initialize} and a partial implementation of \code{collect}. %% The \code{collect} function calls another function, \code{cheney}, %% to perform the actual copy, and that function is left to the reader %% to implement. The following is the prototype for \code{cheney}. %% \begin{lstlisting} %% static void cheney(int64_t** rootstack_ptr); %% \end{lstlisting} %% The parameter \code{rootstack\_ptr} is a pointer to the top of the %% rootstack (which is an array of pointers). The \code{cheney} function %% also communicates with \code{collect} through the global %% variables \code{fromspace\_begin} and \code{fromspace\_end} %% mentioned in Figure~\ref{fig:gc-header} as well as the pointers for %% the ToSpace: %% \begin{lstlisting} %% static int64_t* tospace_begin; %% static int64_t* tospace_end; %% \end{lstlisting} %% The job of the \code{cheney} function is to copy all the live %% objects (reachable from the root stack) into the ToSpace, update %% \code{free\_ptr} to point to the next unused spot in the ToSpace, %% update the root stack so that it points to the objects in the %% ToSpace, and finally to swap the global pointers for the FromSpace %% and ToSpace. %% \end{exercise} %% \section{Compiler Passes} %% \label{sec:code-generation-gc} The introduction of garbage collection has a non-trivial impact on our compiler passes. We introduce a new compiler pass named \code{expose-allocation}. We make significant changes to \code{select-instructions}, \code{build-interference}, \code{allocate-registers}, and \code{print-x86} and make minor changes in several more passes. The following program will serve as our running example. It creates two tuples, one nested inside the other. Both tuples have length one. The program accesses the element in the inner tuple tuple via two vector references. % tests/s2_17.rkt \begin{lstlisting} (vector-ref (vector-ref (vector (vector 42)) 0) 0) \end{lstlisting} \section{Shrink} \label{sec:shrink-R3} Recall that the \code{shrink} pass translates the primitives operators into a smaller set of primitives. Because this pass comes after type checking, but before the passes that require the type information in the \code{HasType} AST nodes, the \code{shrink} pass must be modified to wrap \code{HasType} around each AST node that it generates. \section{Expose Allocation} \label{sec:expose-allocation} The pass \code{expose-allocation} lowers the \code{vector} creation form into a conditional call to the collector followed by the allocation. We choose to place the \code{expose-allocation} pass before \code{remove-complex-opera*} because the code generated by \code{expose-allocation} contains complex operands. We also place \code{expose-allocation} before \code{explicate-control} because \code{expose-allocation} introduces new variables using \code{let}, but \code{let} is gone after \code{explicate-control}. The output of \code{expose-allocation} is a language $R'_3$ that extends $R_3$ with the three new forms that we use in the translation of the \code{vector} form. \[ \begin{array}{lcl} \Exp &::=& \cdots \mid (\key{collect} \,\itm{int}) \mid (\key{allocate} \,\itm{int}\,\itm{type}) \mid (\key{global-value} \,\itm{name}) \end{array} \] The $(\key{collect}\,n)$ form runs the garbage collector, requesting $n$ bytes. It will become a call to the \code{collect} function in \code{runtime.c} in \code{select-instructions}. The $(\key{allocate}\,n\,T)$ form creates an tuple of $n$ elements. \index{allocate} The $T$ parameter is the type of the tuple: \code{(Vector $\Type_1 \ldots \Type_n$)} where $\Type_i$ is the type of the $i$th element in the tuple. The $(\key{global-value}\,\itm{name})$ form reads the value of a global variable, such as \code{free\_ptr}. In the following, we show the transformation for the \code{vector} form into 1) a sequence of let-bindings for the initializing expressions, 2) a conditional call to \code{collect}, 3) a call to \code{allocate}, and 4) the initialization of the vector. In the following, \itm{len} refers to the length of the vector and \itm{bytes} is how many total bytes need to be allocated for the vector, which is 8 for the tag plus \itm{len} times 8. \begin{lstlisting} (has-type (vector |$e_0 \ldots e_{n-1}$|) |\itm{type}|) |$\Longrightarrow$| (let ([|$x_0$| |$e_0$|]) ... (let ([|$x_{n-1}$| |$e_{n-1}$|]) (let ([_ (if (< (+ (global-value free_ptr) |\itm{bytes}|) (global-value fromspace_end)) (void) (collect |\itm{bytes}|))]) (let ([|$v$| (allocate |\itm{len}| |\itm{type}|)]) (let ([_ (vector-set! |$v$| |$0$| |$x_0$|)]) ... (let ([_ (vector-set! |$v$| |$n-1$| |$x_{n-1}$|)]) |$v$|) ... )))) ...) \end{lstlisting} In the above, we suppressed all of the \code{has-type} forms in the output for the sake of readability. The placement of the initializing expressions $e_0,\ldots,e_{n-1}$ prior to the \code{allocate} and the sequence of \code{vector-set!} is important, as those expressions may trigger garbage collection and we cannot have an allocated but uninitialized tuple on the heap during a collection. Figure~\ref{fig:expose-alloc-output} shows the output of the \code{expose-allocation} pass on our running example. \begin{figure}[tbp] % tests/s2_17.rkt \begin{lstlisting} (vector-ref (vector-ref (let ([vecinit7976 (let ([vecinit7972 42]) (let ([collectret7974 (if (< (+ (global-value free_ptr) 16) (global-value fromspace_end)) (void) (collect 16) )]) (let ([alloc7971 (allocate 1 (Vector Integer))]) (let ([initret7973 (vector-set! alloc7971 0 vecinit7972)]) alloc7971) ) ) ) ]) (let ([collectret7978 (if (< (+ (global-value free_ptr) 16) (global-value fromspace_end)) (void) (collect 16) )]) (let ([alloc7975 (allocate 1 (Vector (Vector Integer)))]) (let ([initret7977 (vector-set! alloc7975 0 vecinit7976)]) alloc7975) ) ) ) 0) 0) \end{lstlisting} \caption{Output of the \code{expose-allocation} pass, minus all of the \code{has-type} forms.} \label{fig:expose-alloc-output} \end{figure} \section{Remove Complex Operands} \label{sec:remove-complex-opera-R3} The new forms \code{collect}, \code{allocate}, and \code{global-value} should all be treated as complex operands. %% A new case for %% \code{HasType} is needed and the case for \code{Prim} needs to be %% handled carefully to prevent the \code{Prim} node from being separated %% from its enclosing \code{HasType}. Figure~\ref{fig:r3-anf-syntax} shows the grammar for the output language $R_3^{\dagger}$ of this pass, which is $R_3$ in administrative normal form. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{rcl} \Atm &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \BOOL{\itm{bool}} } \mid \VOID{} \\ \Exp &::=& \gray{ \Atm \mid \READ{} } \\ &\mid& \gray{ \NEG{\Atm} \mid \ADD{\Atm}{\Atm} } \\ &\mid& \gray{ \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \UNIOP{\key{'not}}{\Atm} } \\ &\mid& \gray{ \BINOP{\itm{cmp}}{\Atm}{\Atm} \mid \IF{\Exp}{\Exp}{\Exp} }\\ &\mid& \LP\key{Collect}~\Int\RP \mid \LP\key{Allocate}~\Int~\Type\RP \mid \LP\key{GlobalValue}~\Var\RP\\ % &\mid& \LP\key{HasType}~\Exp~\Type\RP \\ R^{\dagger}_3 &::=& \gray{ \PROGRAM{\code{'()}}{\Exp} } \end{array} \] \end{minipage} } \caption{$R_3^{\dagger}$ is $R_3$ in administrative normal form (ANF).} \label{fig:r3-anf-syntax} \end{figure} \section{Explicate Control and the $C_2$ language} \label{sec:explicate-control-r3} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Atm &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \BOOL{\itm{bool}} }\\ \itm{cmp} &::= & \gray{ \key{eq?} \mid \key{<} } \\ \Exp &::= & \gray{ \Atm \mid \READ{} } \\ &\mid& \gray{ \NEG{\Atm} \mid \ADD{\Atm}{\Atm} }\\ &\mid& \gray{ \UNIOP{\key{not}}{\Atm} \mid \BINOP{\itm{cmp}}{\Atm}{\Atm} } \\ &\mid& (\key{Allocate} \,\itm{int}\,\itm{type}) \\ &\mid& \BINOP{\key{'vector-ref}}{\Atm}{\INT{\Int}} \\ &\mid& (\key{Prim}~\key{'vector-set!}\,(\key{list}\,\Atm\,\INT{\Int}\,\Atm))\\ &\mid& (\key{GlobalValue} \,\Var) \mid (\key{Void})\\ \Stmt &::=& \gray{ \ASSIGN{\VAR{\Var}}{\Exp} } \mid (\key{Collect} \,\itm{int}) \\ \Tail &::= & \gray{ \RETURN{\Exp} \mid \SEQ{\Stmt}{\Tail} \mid \GOTO{\itm{label}} } \\ &\mid& \gray{ \IFSTMT{\BINOP{\itm{cmp}}{\Atm}{\Atm}}{\GOTO{\itm{label}}}{\GOTO{\itm{label}}} }\\ C_2 & ::= & \gray{ \PROGRAM{\itm{info}}{\CFG{(\itm{label}\,\key{.}\,\Tail)\ldots}} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $C_2$, extending $C_1$ (Figure~\ref{fig:c1-syntax}).} \label{fig:c2-syntax} \end{figure} The output of \code{explicate-control} is a program in the intermediate language $C_2$, whose abstract syntax is defined in Figure~\ref{fig:c2-syntax}. (The concrete syntax is defined in Figure~\ref{fig:c2-concrete-syntax} of the Appendix.) The new forms of $C_2$ include the \key{allocate}, \key{vector-ref}, and \key{vector-set!}, and \key{global-value} expressions and the \code{collect} statement. The \code{explicate-control} pass can treat these new forms much like the other expression forms that we've already encoutered. \section{Select Instructions and the x86$_2$ Language} \label{sec:select-instructions-gc} \index{instruction selection} %% void (rep as zero) %% allocate %% collect (callq collect) %% vector-ref %% vector-set! %% global (postpone) In this pass we generate x86 code for most of the new operations that were needed to compile tuples, including \code{Allocate}, \code{Collect}, \code{vector-ref}, \code{vector-set!}, and \code{void}. We compile \code{GlobalValue} to \code{Global} because the later has a different concrete syntax (see Figures~\ref{fig:x86-2-concrete} and \ref{fig:x86-2}). \index{x86} The \code{vector-ref} and \code{vector-set!} forms translate into \code{movq} instructions. (The plus one in the offset is to get past the tag at the beginning of the tuple representation.) \begin{lstlisting} |$\itm{lhs}$| = (vector-ref |$\itm{vec}$| |$n$|); |$\Longrightarrow$| movq |$\itm{vec}'$|, %r11 movq |$8(n+1)$|(%r11), |$\itm{lhs'}$| |$\itm{lhs}$| = (vector-set! |$\itm{vec}$| |$n$| |$\itm{arg}$|); |$\Longrightarrow$| movq |$\itm{vec}'$|, %r11 movq |$\itm{arg}'$|, |$8(n+1)$|(%r11) movq $0, |$\itm{lhs'}$| \end{lstlisting} The $\itm{lhs}'$, $\itm{vec}'$, and $\itm{arg}'$ are obtained by translating $\itm{vec}$ and $\itm{arg}$ to x86. The move of $\itm{vec}'$ to register \code{r11} ensures that offset expression \code{$-8(n+1)$(\%r11)} contains a register operand. This requires removing \code{r11} from consideration by the register allocating. Why not use \code{rax} instead of \code{r11}? Suppose we instead used \code{rax}. Then the generated code for \code{vector-set!} would be \begin{lstlisting} movq |$\itm{vec}'$|, %rax movq |$\itm{arg}'$|, |$8(n+1)$|(%rax) movq $0, |$\itm{lhs}'$| \end{lstlisting} Next, suppose that $\itm{arg}'$ ends up as a stack location, so \code{patch-instructions} would insert a move through \code{rax} as follows. \begin{lstlisting} movq |$\itm{vec}'$|, %rax movq |$\itm{arg}'$|, %rax movq %rax, |$8(n+1)$|(%rax) movq $0, |$\itm{lhs}'$| \end{lstlisting} But the above sequence of instructions does not work because we're trying to use \code{rax} for two different values ($\itm{vec}'$ and $\itm{arg}'$) at the same time! We compile the \code{allocate} form to operations on the \code{free\_ptr}, as shown below. The address in the \code{free\_ptr} is the next free address in the FromSpace, so we copy it into \code{r11} and then move it forward by enough space for the tuple being allocated, which is $8(\itm{len}+1)$ bytes because each element is 8 bytes (64 bits) and we use 8 bytes for the tag. We then initialize the \itm{tag} and finally copy the address in \code{r11} to the left-hand-side. Refer to Figure~\ref{fig:tuple-rep} to see how the tag is organized. We recommend using the Racket operations \code{bitwise-ior} and \code{arithmetic-shift} to compute the tag during compilation. The type annotation in the \code{vector} form is used to determine the pointer mask region of the tag. \begin{lstlisting} |$\itm{lhs}$| = (allocate |$\itm{len}$| (Vector |$\itm{type} \ldots$|)); |$\Longrightarrow$| movq free_ptr(%rip), %r11 addq |$8(\itm{len}+1)$|, free_ptr(%rip) movq $|$\itm{tag}$|, 0(%r11) movq %r11, |$\itm{lhs}'$| \end{lstlisting} The \code{collect} form is compiled to a call to the \code{collect} function in the runtime. The arguments to \code{collect} are 1) the top of the root stack and 2) the number of bytes that need to be allocated. We use another dedicated register, \code{r15}, to store the pointer to the top of the root stack. So \code{r15} is not available for use by the register allocator. \begin{lstlisting} (collect |$\itm{bytes}$|) |$\Longrightarrow$| movq %r15, %rdi movq $|\itm{bytes}|, %rsi callq collect \end{lstlisting} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Arg &::=& \gray{ \key{\$}\Int \mid \key{\%}\Reg \mid \Int\key{(}\key{\%}\Reg\key{)} \mid \key{\%}\itm{bytereg} } \mid \Var \key{(\%rip)} \\ x86_1 &::= & \gray{ \key{.globl main} }\\ & & \gray{ \key{main:} \; \Instr\ldots } \end{array} \] \end{minipage} } \caption{The concrete syntax of x86$_2$ (extends x86$_1$ of Figure~\ref{fig:x86-1-concrete}).} \label{fig:x86-2-concrete} \end{figure} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Arg &::=& \gray{ \INT{\Int} \mid \REG{\Reg} \mid \DEREF{\Reg}{\Int} \mid \BYTEREG{\Reg}} \\ &\mid& (\key{Global}~\Var) \\ x86_2 &::= & \gray{ \PROGRAM{\itm{info}}{\CFG{\key{(}\itm{label} \,\key{.}\, \Block \key{)}\ldots}} } \end{array} \] \end{minipage} } \caption{The abstract syntax of x86$_2$ (extends x86$_1$ of Figure~\ref{fig:x86-1}).} \label{fig:x86-2} \end{figure} The concrete and abstract syntax of the $x86_2$ language is defined in Figures~\ref{fig:x86-2-concrete} and \ref{fig:x86-2}. It differs from x86$_1$ just in the addition of the form for global variables. % Figure~\ref{fig:select-instr-output-gc} shows the output of the \code{select-instructions} pass on the running example. \begin{figure}[tbp] \centering % tests/s2_17.rkt \begin{minipage}[t]{0.5\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] block35: movq free_ptr(%rip), alloc9024 addq $16, free_ptr(%rip) movq alloc9024, %r11 movq $131, 0(%r11) movq alloc9024, %r11 movq vecinit9025, 8(%r11) movq $0, initret9026 movq alloc9024, %r11 movq 8(%r11), tmp9034 movq tmp9034, %r11 movq 8(%r11), %rax jmp conclusion block36: movq $0, collectret9027 jmp block35 block38: movq free_ptr(%rip), alloc9020 addq $16, free_ptr(%rip) movq alloc9020, %r11 movq $3, 0(%r11) movq alloc9020, %r11 movq vecinit9021, 8(%r11) movq $0, initret9022 movq alloc9020, vecinit9025 movq free_ptr(%rip), tmp9031 movq tmp9031, tmp9032 addq $16, tmp9032 movq fromspace_end(%rip), tmp9033 cmpq tmp9033, tmp9032 jl block36 jmp block37 block37: movq %r15, %rdi movq $16, %rsi callq 'collect jmp block35 block39: movq $0, collectret9023 jmp block38 \end{lstlisting} \end{minipage} \begin{minipage}[t]{0.45\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] start: movq $42, vecinit9021 movq free_ptr(%rip), tmp9028 movq tmp9028, tmp9029 addq $16, tmp9029 movq fromspace_end(%rip), tmp9030 cmpq tmp9030, tmp9029 jl block39 jmp block40 block40: movq %r15, %rdi movq $16, %rsi callq 'collect jmp block38 \end{lstlisting} \end{minipage} \caption{Output of the \code{select-instructions} pass.} \label{fig:select-instr-output-gc} \end{figure} \clearpage \section{Register Allocation} \label{sec:reg-alloc-gc} \index{register allocation} As discussed earlier in this chapter, the garbage collector needs to access all the pointers in the root set, that is, all variables that are vectors. It will be the responsibility of the register allocator to make sure that: \begin{enumerate} \item the root stack is used for spilling vector-typed variables, and \item if a vector-typed variable is live during a call to the collector, it must be spilled to ensure it is visible to the collector. \end{enumerate} The later responsibility can be handled during construction of the interference graph, by adding interference edges between the call-live vector-typed variables and all the callee-saved registers. (They already interfere with the caller-saved registers.) The type information for variables is in the \code{Program} form, so we recommend adding another parameter to the \code{build-interference} function to communicate this alist. The spilling of vector-typed variables to the root stack can be handled after graph coloring, when choosing how to assign the colors (integers) to registers and stack locations. The \code{Program} output of this pass changes to also record the number of spills to the root stack. % build-interference % % callq % extra parameter for var->type assoc. list % update 'program' and 'if' % allocate-registers % allocate spilled vectors to the rootstack % don't change color-graph \section{Print x86} \label{sec:print-x86-gc} \index{prelude}\index{conclusion} Figure~\ref{fig:print-x86-output-gc} shows the output of the \code{print-x86} pass on the running example. In the prelude and conclusion of the \code{main} function, we treat the root stack very much like the regular stack in that we move the root stack pointer (\code{r15}) to make room for the spills to the root stack, except that the root stack grows up instead of down. For the running example, there was just one spill so we increment \code{r15} by 8 bytes. In the conclusion we decrement \code{r15} by 8 bytes. One issue that deserves special care is that there may be a call to \code{collect} prior to the initializing assignments for all the variables in the root stack. We do not want the garbage collector to accidentally think that some uninitialized variable is a pointer that needs to be followed. Thus, we zero-out all locations on the root stack in the prelude of \code{main}. In Figure~\ref{fig:print-x86-output-gc}, the instruction % \lstinline{movq $0, (%r15)} % accomplishes this task. The garbage collector tests each root to see if it is null prior to dereferencing it. \begin{figure}[htbp] \begin{minipage}[t]{0.5\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] block35: movq free_ptr(%rip), %rcx addq $16, free_ptr(%rip) movq %rcx, %r11 movq $131, 0(%r11) movq %rcx, %r11 movq -8(%r15), %rax movq %rax, 8(%r11) movq $0, %rdx movq %rcx, %r11 movq 8(%r11), %rcx movq %rcx, %r11 movq 8(%r11), %rax jmp conclusion block36: movq $0, %rcx jmp block35 block38: movq free_ptr(%rip), %rcx addq $16, free_ptr(%rip) movq %rcx, %r11 movq $3, 0(%r11) movq %rcx, %r11 movq %rbx, 8(%r11) movq $0, %rdx movq %rcx, -8(%r15) movq free_ptr(%rip), %rcx addq $16, %rcx movq fromspace_end(%rip), %rdx cmpq %rdx, %rcx jl block36 movq %r15, %rdi movq $16, %rsi callq collect jmp block35 block39: movq $0, %rcx jmp block38 \end{lstlisting} \end{minipage} \begin{minipage}[t]{0.45\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] start: movq $42, %rbx movq free_ptr(%rip), %rdx addq $16, %rdx movq fromspace_end(%rip), %rcx cmpq %rcx, %rdx jl block39 movq %r15, %rdi movq $16, %rsi callq collect jmp block38 .globl main main: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx pushq %r14 subq $0, %rsp movq $16384, %rdi movq $16384, %rsi callq initialize movq rootstack_begin(%rip), %r15 movq $0, (%r15) addq $8, %r15 jmp start conclusion: subq $8, %r15 addq $0, %rsp popq %r14 popq %rbx popq %r12 popq %r13 popq %rbp retq \end{lstlisting} \end{minipage} \caption{Output of the \code{print-x86} pass.} \label{fig:print-x86-output-gc} \end{figure} \begin{figure}[p] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R3) at (0,2) {\large $R_3$}; \node (R3-2) at (3,2) {\large $R_3$}; \node (R3-3) at (6,2) {\large $R_3$}; \node (R3-4) at (9,2) {\large $R_3$}; \node (R3-5) at (12,2) {\large $R'_3$}; \node (C2-4) at (3,0) {\large $C_2$}; \node (x86-2) at (3,-2) {\large $\text{x86}^{*}_2$}; \node (x86-3) at (6,-2) {\large $\text{x86}^{*}_2$}; \node (x86-4) at (9,-2) {\large $\text{x86}^{*}_2$}; \node (x86-5) at (9,-4) {\large $\text{x86}^{\dagger}_2$}; \node (x86-2-1) at (3,-4) {\large $\text{x86}^{*}_2$}; \node (x86-2-2) at (6,-4) {\large $\text{x86}^{*}_2$}; %\path[->,bend left=15] (R3) edge [above] node {\ttfamily\footnotesize type-check} (R3-2); \path[->,bend left=15] (R3) edge [above] node {\ttfamily\footnotesize shrink} (R3-2); \path[->,bend left=15] (R3-2) edge [above] node {\ttfamily\footnotesize uniquify} (R3-3); \path[->,bend left=15] (R3-3) edge [above] node {\ttfamily\footnotesize expose-alloc.} (R3-4); \path[->,bend left=15] (R3-4) edge [above] node {\ttfamily\footnotesize remove-complex.} (R3-5); \path[->,bend left=20] (R3-5) edge [left] node {\ttfamily\footnotesize explicate-control} (C2-4); \path[->,bend left=15] (C2-4) edge [right] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend right=15] (x86-2) edge [left] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [right] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [right] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_3$, a language with tuples.} \label{fig:R3-passes} \end{figure} Figure~\ref{fig:R3-passes} gives an overview of all the passes needed for the compilation of $R_3$. \section{Challenge: Simple Structures} \label{sec:simple-structures} \index{struct} \index{structure} Figure~\ref{fig:r3s-concrete-syntax} defines the concrete syntax for $R^s_3$, which extends $R^3$ with support for simple structures. Recall that a \code{struct} in Typed Racket is a user-defined data type that contains named fields and that is heap allocated, similar to a vector. The following is an example of a structure definition, in this case the definition of a \code{point} type. \begin{lstlisting} (struct point ([x : Integer] [y : Integer]) #:mutable) \end{lstlisting} \begin{figure}[tbp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Type &::=& \gray{\key{Integer} \mid \key{Boolean} \mid (\key{Vector}\;\Type \ldots) \mid \key{Void} } \mid \Var \\ \itm{cmp} &::= & \gray{ \key{eq?} \mid \key{<} \mid \key{<=} \mid \key{>} \mid \key{>=} } \\ \Exp &::=& \gray{ \Int \mid (\key{read}) \mid (\key{-}\;\Exp) \mid (\key{+} \; \Exp\;\Exp) \mid (\key{-}\;\Exp\;\Exp) } \\ &\mid& \gray{ \Var \mid (\key{let}~([\Var~\Exp])~\Exp) }\\ &\mid& \gray{ \key{\#t} \mid \key{\#f} \mid (\key{and}\;\Exp\;\Exp) \mid (\key{or}\;\Exp\;\Exp) \mid (\key{not}\;\Exp) } \\ &\mid& \gray{ (\itm{cmp}\;\Exp\;\Exp) \mid (\key{if}~\Exp~\Exp~\Exp) } \\ &\mid& \gray{ (\key{vector}\;\Exp \ldots) \mid (\key{vector-ref}\;\Exp\;\Int) } \\ &\mid& \gray{ (\key{vector-set!}\;\Exp\;\Int\;\Exp) }\\ &\mid& \gray{ (\key{void}) } \mid (\Var\;\Exp \ldots)\\ \Def &::=& (\key{struct}\; \Var \; ([\Var \,\key{:}\, \Type] \ldots)\; \code{\#:mutable})\\ R_3 &::=& \Def \ldots \; \Exp \end{array} \] \end{minipage} } \caption{The concrete syntax of $R^s_3$, extending $R_3$ (Figure~\ref{fig:r3-concrete-syntax}).} \label{fig:r3s-concrete-syntax} \end{figure} An instance of a structure is created using function call syntax, with the name of the structure in the function position: \begin{lstlisting} (point 7 12) \end{lstlisting} Function-call syntax is also used to read the value in a field of a structure. The function name is formed by the structure name, a dash, and the field name. The following example uses \code{point-x} and \code{point-y} to access the \code{x} and \code{y} fields of two point instances. \begin{center} \begin{lstlisting} (let ([pt1 (point 7 12)]) (let ([pt2 (point 4 3)]) (+ (- (point-x pt1) (point-x pt2)) (- (point-y pt1) (point-y pt2))))) \end{lstlisting} \end{center} Similarly, to write to a field of a structure, use its set function, whose name starts with \code{set-}, followed by the structure name, then a dash, then the field name, and concluded with an exclamation mark. The following example uses \code{set-point-x!} to change the \code{x} field from \code{7} to \code{42}. \begin{center} \begin{lstlisting} (let ([pt (point 7 12)]) (let ([_ (set-point-x! pt 42)]) (point-x pt))) \end{lstlisting} \end{center} \begin{exercise}\normalfont Extend your compiler with support for simple structures, compiling $R^s_3$ to x86 assembly code. Create five new test cases that use structures and test your compiler. \end{exercise} \section{Challenge: Generational Collection} The copying collector described in Section~\ref{sec:GC} can incur significant runtime overhead because the call to \code{collect} takes time proportional to all of the live data. One way to reduce this overhead is to reduce how much data is inspected in each call to \code{collect}. In particular, researchers have observed that recently allocated data is more likely to become garbage then data that has survived one or more previous calls to \code{collect}. This insight motivated the creation of \emph{generational garbage collectors} \index{generational garbage collector} that 1) segregates data according to its age into two or more generations, 2) allocates less space for younger generations, so collecting them is faster, and more space for the older generations, and 3) performs collection on the younger generations more frequently then for older generations~\citep{Wilson:1992fk}. For this challenge assignment, the goal is to adapt the copying collector implemented in \code{runtime.c} to use two generations, one for young data and one for old data. Each generation consists of a FromSpace and a ToSpace. The following is a sketch of how to adapt the \code{collect} function to use the two generations. \begin{enumerate} \item Copy the young generation's FromSpace to its ToSpace then switch the role of the ToSpace and FromSpace \item If there is enough space for the requested number of bytes in the young FromSpace, then return from \code{collect}. \item If there is not enough space in the young FromSpace for the requested bytes, then move the data from the young generation to the old one with the following steps: \begin{enumerate} \item If there is enough room in the old FromSpace, copy the young FromSpace to the old FromSpace and then return. \item If there is not enough room in the old FromSpace, then collect the old generation by copying the old FromSpace to the old ToSpace and swap the roles of the old FromSpace and ToSpace. \item If there is enough room now, copy the young FromSpace to the old FromSpace and return. Otherwise, allocate a larger FromSpace and ToSpace for the old generation. Copy the young FromSpace and the old FromSpace into the larger FromSpace for the old generation and then return. \end{enumerate} \end{enumerate} We recommend that you generalize the \code{cheney} function so that it can be used for all the copies mentioned above: between the young FromSpace and ToSpace, between the old FromSpace and ToSpace, and between the young FromSpace and old FromSpace. This can be accomplished by adding parameters to \code{cheney} that replace its use of the global variables \code{fromspace\_begin}, \code{fromspace\_end}, \code{tospace\_begin}, and \code{tospace\_end}. Note that the collection of the young generation does not traverse the old generation. This introduces a potential problem: there may be young data that is only reachable through pointers in the old generation. If these pointers are not taken into account, the collector could throw away young data that is live! One solution, called \emph{pointer recording}, is to maintain a set of all the pointers from the old generation into the new generation and consider this set as part of the root set. To maintain this set, the compiler must insert extra instructions around every \code{vector-set!}. If the vector being modified is in the old generation, and if the value being written is a pointer into the new generation, than that pointer must be added to the set. Also, if the value being overwritten was a pointer into the new generation, then that pointer should be removed from the set. \begin{exercise}\normalfont Adapt the \code{collect} function in \code{runtime.c} to implement generational garbage collection, as outlined in this section. Update the code generation for \code{vector-set!} to implement pointer recording. Make sure that your new compiler and runtime passes your test suite. \end{exercise} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Functions} \label{ch:functions} \index{function} This chapter studies the compilation of functions similar to those found in the C language. This corresponds to a subset of Typed Racket in which only top-level function definitions are allowed. This kind of function is an important stepping stone to implementing lexically-scoped functions, that is, \key{lambda} abstractions, which is the topic of Chapter~\ref{ch:lambdas}. \section{The $R_4$ Language} The concrete and abstract syntax for function definitions and function application is shown in Figures~\ref{fig:r4-concrete-syntax} and \ref{fig:r4-syntax}, where we define the $R_4$ language. Programs in $R_4$ begin with zero or more function definitions. The function names from these definitions are in-scope for the entire program, including all other function definitions (so the ordering of function definitions does not matter). The concrete syntax for function application\index{function application} is $(\Exp \; \Exp \ldots)$ where the first expression must evaluate to a function and the rest are the arguments. The abstract syntax for function application is $\APPLY{\Exp}{\Exp\ldots}$. %% The syntax for function application does not include an explicit %% keyword, which is error prone when using \code{match}. To alleviate %% this problem, we translate the syntax from $(\Exp \; \Exp \ldots)$ to %% $(\key{app}\; \Exp \; \Exp \ldots)$ during type checking. Functions are first-class in the sense that a function pointer \index{function pointer} is data and can be stored in memory or passed as a parameter to another function. Thus, we introduce a function type, written \begin{lstlisting} (|$\Type_1$| |$\cdots$| |$\Type_n$| -> |$\Type_r$|) \end{lstlisting} for a function whose $n$ parameters have the types $\Type_1$ through $\Type_n$ and whose return type is $\Type_r$. The main limitation of these functions (with respect to Racket functions) is that they are not lexically scoped. That is, the only external entities that can be referenced from inside a function body are other globally-defined functions. The syntax of $R_4$ prevents functions from being nested inside each other. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Type &::=& \gray{ \key{Integer} \mid \key{Boolean} \mid (\key{Vector}\;\Type\ldots) \mid \key{Void} } \mid (\Type \ldots \; \key{->}\; \Type) \\ \itm{cmp} &::= & \gray{ \key{eq?} \mid \key{<} \mid \key{<=} \mid \key{>} \mid \key{>=} } \\ \Exp &::=& \gray{ \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} \mid \CSUB{\Exp}{\Exp} } \\ &\mid& \gray{ \Var \mid \CLET{\Var}{\Exp}{\Exp} }\\ &\mid& \gray{ \key{\#t} \mid \key{\#f} \mid (\key{and}\;\Exp\;\Exp) \mid (\key{or}\;\Exp\;\Exp) \mid (\key{not}\;\Exp)} \\ &\mid& \gray{(\itm{cmp}\;\Exp\;\Exp) \mid \CIF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{(\key{vector}\;\Exp\ldots) \mid (\key{vector-ref}\;\Exp\;\Int)} \\ &\mid& \gray{(\key{vector-set!}\;\Exp\;\Int\;\Exp)\mid (\key{void}) \mid \LP\key{has-type}~\Exp~\Type\RP } \\ &\mid& \LP\Exp \; \Exp \ldots\RP \\ \Def &::=& \CDEF{\Var}{\LS\Var \key{:} \Type\RS \ldots}{\Type}{\Exp} \\ R_4 &::=& \Def \ldots \; \Exp \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_4$, extending $R_3$ (Figure~\ref{fig:r3-concrete-syntax}).} \label{fig:r4-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::=& \gray{ \INT{\Int} \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \PRIM{\itm{op}}{\Exp\ldots} }\\ &\mid& \gray{ \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ \VOID{} \mid \LP\key{HasType}~\Exp~\Type \RP } \mid \APPLY{\Exp}{\Exp\ldots}\\ \Def &::=& \FUNDEF{\Var}{\LP[\Var \code{:} \Type]\ldots\RP}{\Type}{\code{'()}}{\Exp}\\ R_4 &::=& \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP)}{\Exp} \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_4$, extending $R_3$ (Figure~\ref{fig:r3-syntax}).} \label{fig:r4-syntax} \end{figure} The program in Figure~\ref{fig:r4-function-example} is a representative example of defining and using functions in $R_4$. We define a function \code{map-vec} that applies some other function \code{f} to both elements of a vector and returns a new vector containing the results. We also define a function \code{add1}. The program applies \code{map-vec} to \code{add1} and \code{(vector 0 41)}. The result is \code{(vector 1 42)}, from which we return the \code{42}. \begin{figure}[tbp] \begin{lstlisting} (define (map-vec [f : (Integer -> Integer)] [v : (Vector Integer Integer)]) : (Vector Integer Integer) (vector (f (vector-ref v 0)) (f (vector-ref v 1)))) (define (add1 [x : Integer]) : Integer (+ x 1)) (vector-ref (map-vec add1 (vector 0 41)) 1) \end{lstlisting} \caption{Example of using functions in $R_4$.} \label{fig:r4-function-example} \end{figure} The definitional interpreter for $R_4$ is in Figure~\ref{fig:interp-R4}. The case for the \code{ProgramDefsExp} form is responsible for setting up the mutual recursion between the top-level function definitions. We use the classic back-patching \index{back-patching} approach that uses mutable variables and makes two passes over the function definitions~\citep{Kelsey:1998di}. In the first pass we set up the top-level environment using a mutable cons cell for each function definition. Note that the \code{lambda} value for each function is incomplete; it does not yet include the environment. Once the top-level environment is constructed, we then iterate over it and update the \code{lambda} values to use the top-level environment. \begin{figure}[tp] \begin{lstlisting} (define interp-R4-class (class interp-R3-class (super-new) (define/override ((interp-exp env) e) (define recur (interp-exp env)) (match e [(Var x) (unbox (dict-ref env x))] [(Let x e body) (define new-env (dict-set env x (box (recur e)))) ((interp-exp new-env) body)] [(Apply fun args) (define fun-val (recur fun)) (define arg-vals (for/list ([e args]) (recur e))) (match fun-val [`(function (,xs ...) ,body ,fun-env) (define params-args (for/list ([x xs] [arg arg-vals]) (cons x (box arg)))) (define new-env (append params-args fun-env)) ((interp-exp new-env) body)] [else (error 'interp-exp "expected function, not ~a" fun-val)])] [else ((super interp-exp env) e)] )) (define/public (interp-def d) (match d [(Def f (list `[,xs : ,ps] ...) rt _ body) (cons f (box `(function ,xs ,body ())))])) (define/override (interp-program p) (match p [(ProgramDefsExp info ds body) (let ([top-level (for/list ([d ds]) (interp-def d))]) (for/list ([f (in-dict-values top-level)]) (set-box! f (match (unbox f) [`(function ,xs ,body ()) `(function ,xs ,body ,top-level)]))) ((interp-exp top-level) body))])) )) (define (interp-R4 p) (send (new interp-R4-class) interp-program p)) \end{lstlisting} \caption{Interpreter for the $R_4$ language.} \label{fig:interp-R4} \end{figure} \margincomment{TODO: explain type checker} The type checker for $R_4$ is is in Figure~\ref{fig:type-check-R4}. \begin{figure}[tp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define type-check-R4-class (class type-check-R3-class (super-new) (inherit check-type-equal?) (define/public (type-check-apply env e es) (define-values (e^ ty) ((type-check-exp env) e)) (define-values (e* ty*) (for/lists (e* ty*) ([e (in-list es)]) ((type-check-exp env) e))) (match ty [`(,ty^* ... -> ,rt) (for ([arg-ty ty*] [param-ty ty^*]) (check-type-equal? arg-ty param-ty (Apply e es))) (values e^ e* rt)])) (define/override (type-check-exp env) (lambda (e) (match e [(FunRef f) (values (FunRef f) (dict-ref env f))] [(Apply e es) (define-values (e^ es^ rt) (type-check-apply env e es)) (values (Apply e^ es^) rt)] [(Call e es) (define-values (e^ es^ rt) (type-check-apply env e es)) (values (Call e^ es^) rt)] [else ((super type-check-exp env) e)]))) (define/public (type-check-def env) (lambda (e) (match e [(Def f (and p:t* (list `[,xs : ,ps] ...)) rt info body) (define new-env (append (map cons xs ps) env)) (define-values (body^ ty^) ((type-check-exp new-env) body)) (check-type-equal? ty^ rt body) (Def f p:t* rt info body^)]))) (define/public (fun-def-type d) (match d [(Def f (list `[,xs : ,ps] ...) rt info body) `(,@ps -> ,rt)])) (define/override (type-check-program e) (match e [(ProgramDefsExp info ds body) (define new-env (for/list ([d ds]) (cons (Def-name d) (fun-def-type d)))) (define ds^ (for/list ([d ds]) ((type-check-def new-env) d))) (define-values (body^ ty) ((type-check-exp new-env) body)) (check-type-equal? ty 'Integer body) (ProgramDefsExp info ds^ body^)])))) (define (type-check-R4 p) (send (new type-check-R4-class) type-check-program p)) \end{lstlisting} \caption{Type checker for the $R_4$ language.} \label{fig:type-check-R4} \end{figure} \section{Functions in x86} \label{sec:fun-x86} \margincomment{\tiny Make sure callee-saved registers are discussed in enough depth, especially updating Fig 6.4 \\ --Jeremy } \margincomment{\tiny Talk about the return address on the stack and what callq and retq does.\\ --Jeremy } The x86 architecture provides a few features to support the implementation of functions. We have already seen that x86 provides labels so that one can refer to the location of an instruction, as is needed for jump instructions. Labels can also be used to mark the beginning of the instructions for a function. Going further, we can obtain the address of a label by using the \key{leaq} instruction and PC-relative addressing. For example, the following puts the address of the \code{add1} label into the \code{rbx} register. \begin{lstlisting} leaq add1(%rip), %rbx \end{lstlisting} The instruction pointer register \key{rip} (aka. the program counter \index{program counter}) always points to the next instruction to be executed. When combined with an label, as in \code{add1(\%rip)}, the linker computes the distance $d$ between the address of \code{add1} and where the \code{rip} would be at that moment and then changes \code{add1(\%rip)} to \code{$d$(\%rip)}, which at runtime will compute the address of \code{add1}. In Section~\ref{sec:x86} we used of the \code{callq} instruction to jump to a function whose location is given by a label. To support function calls in this chapter we instead will be jumping to a function whose location is given by an address in a register, that is, we need to make an \emph{indirect function call}. The x86 syntax for this is a \code{callq} instruction but with an asterisk before the register name.\index{indirect function call} \begin{lstlisting} callq *%rbx \end{lstlisting} \subsection{Calling Conventions} \index{calling conventions} The \code{callq} instruction provides partial support for implementing functions: it pushes the return address on the stack and it jumps to the target. However, \code{callq} does not handle \begin{enumerate} \item parameter passing, \item pushing frames on the procedure call stack and popping them off, or \item determining how registers are shared by different functions. \end{enumerate} Regarding (1) parameter passing, recall that the following six registers are used to pass arguments to a function, in this order. \begin{lstlisting} rdi rsi rdx rcx r8 r9 \end{lstlisting} If there are more than six arguments, then the convention is to use space on the frame of the caller for the rest of the arguments. However, to ease the implementation of efficient tail calls (Section~\ref{sec:tail-call}), we arrange to never need more than six arguments. % Also recall that the register \code{rax} is for the return value of the function. \index{prelude}\index{conclusion} Regarding (2) frames \index{frame} and the procedure call stack, \index{procedure call stack} recall from Section~\ref{sec:x86} that the stack grows down, with each function call using a chunk of space called a frame. The caller sets the stack pointer, register \code{rsp}, to the last data item in its frame. The callee must not change anything in the caller's frame, that is, anything that is at or above the stack pointer. The callee is free to use locations that are below the stack pointer. Recall that we are storing variables of vector type on the root stack. So the prelude needs to move the root stack pointer \code{r15} up and the conclusion needs to move the root stack pointer back down. Also, the prelude must initialize to \code{0} this frame's slots in the root stack to signal to the garbage collector that those slots do not yet contain a pointer to a vector. Otherwise the garbage collector will interpret the garbage bits in those slots as memory addresses and try to traverse them, causing serious mayhem! Regarding (3) the sharing of registers between different functions, recall from Section~\ref{sec:calling-conventions} that the registers are divided into two groups, the caller-saved registers and the callee-saved registers. The caller should assume that all the caller-saved registers get overwritten with arbitrary values by the callee. That is why we recommend in Section~\ref{sec:calling-conventions} that variables that are live during a function call should not be assigned to caller-saved registers. On the flip side, if the callee wants to use a callee-saved register, the callee must save the contents of those registers on their stack frame and then put them back prior to returning to the caller. That is why we recommended in Section~\ref{sec:calling-conventions} that if the register allocator assigns a variable to a callee-saved register, then the prelude of the \code{main} function must save that register to the stack and the conclusion of \code{main} must restore it. This recommendation now generalizes to all functions. Also recall that the base pointer, register \code{rbp}, is used as a point-of-reference within a frame, so that each local variable can be accessed at a fixed offset from the base pointer (Section~\ref{sec:x86}). % Figure~\ref{fig:call-frames} shows the general layout of the caller and callee frames. \begin{figure}[tbp] \centering \begin{tabular}{r|r|l|l} \hline Caller View & Callee View & Contents & Frame \\ \hline 8(\key{\%rbp}) & & return address & \multirow{5}{*}{Caller}\\ 0(\key{\%rbp}) & & old \key{rbp} \\ -8(\key{\%rbp}) & & callee-saved $1$ \\ \ldots & & \ldots \\ $-8j$(\key{\%rbp}) & & callee-saved $j$ \\ $-8(j+1)$(\key{\%rbp}) & & local variable $1$ \\ \ldots & & \ldots \\ $-8(j+k)$(\key{\%rbp}) & & local variable $k$ \\ %% & & \\ %% $8n-8$\key{(\%rsp)} & $8n+8$(\key{\%rbp})& argument $n$ \\ %% & \ldots & \ldots \\ %% 0\key{(\%rsp)} & 16(\key{\%rbp}) & argument $1$ & \\ \hline & 8(\key{\%rbp}) & return address & \multirow{5}{*}{Callee}\\ & 0(\key{\%rbp}) & old \key{rbp} \\ & -8(\key{\%rbp}) & callee-saved $1$ \\ & \ldots & \ldots \\ & $-8n$(\key{\%rbp}) & callee-saved $n$ \\ & $-8(n+1)$(\key{\%rbp}) & local variable $1$ \\ & \ldots & \ldots \\ & $-8(n+m)$(\key{\%rsp}) & local variable $m$\\ \hline \end{tabular} \caption{Memory layout of caller and callee frames.} \label{fig:call-frames} \end{figure} %% Recall from Section~\ref{sec:x86} that the stack is also used for %% local variables and for storing the values of callee-saved registers %% (we shall refer to all of these collectively as ``locals''), and that %% at the beginning of a function we move the stack pointer \code{rsp} %% down to make room for them. %% We recommend storing the local variables %% first and then the callee-saved registers, so that the local variables %% can be accessed using \code{rbp} the same as before the addition of %% functions. %% To make additional room for passing arguments, we shall %% move the stack pointer even further down. We count how many stack %% arguments are needed for each function call that occurs inside the %% body of the function and find their maximum. Adding this number to the %% number of locals gives us how much the \code{rsp} should be moved at %% the beginning of the function. In preparation for a function call, we %% offset from \code{rsp} to set up the stack arguments. We put the first %% stack argument in \code{0(\%rsp)}, the second in \code{8(\%rsp)}, and %% so on. %% Upon calling the function, the stack arguments are retrieved by the %% callee using the base pointer \code{rbp}. The address \code{16(\%rbp)} %% is the location of the first stack argument, \code{24(\%rbp)} is the %% address of the second, and so on. Figure~\ref{fig:call-frames} shows %% the layout of the caller and callee frames. Notice how important it is %% that we correctly compute the maximum number of arguments needed for %% function calls; if that number is too small then the arguments and %% local variables will smash into each other! \subsection{Efficient Tail Calls} \label{sec:tail-call} In general, the amount of stack space used by a program is determined by the longest chain of nested function calls. That is, if function $f_1$ calls $f_2$, $f_2$ calls $f_3$, $\ldots$, and $f_{n-1}$ calls $f_n$, then the amount of stack space is bounded by $O(n)$. The depth $n$ can grow quite large in the case of recursive or mutually recursive functions. However, in some cases we can arrange to use only constant space, i.e. $O(1)$, instead of $O(n)$. If a function call is the last action in a function body, then that call is said to be a \emph{tail call}\index{tail call}. For example, in the following program, the recursive call to \code{tail-sum} is a tail call. \begin{center} \begin{lstlisting} (define (tail-sum [n : Integer] [r : Integer]) : Integer (if (eq? n 0) r (tail-sum (- n 1) (+ n r)))) (+ (tail-sum 5 0) 27) \end{lstlisting} \end{center} At a tail call, the frame of the caller is no longer needed, so we can pop the caller's frame before making the tail call. With this approach, a recursive function that only makes tail calls will only use $O(1)$ stack space. Functional languages like Racket typically rely heavily on recursive functions, so they typically guarantee that all tail calls will be optimized in this way. \index{frame} However, some care is needed with regards to argument passing in tail calls. As mentioned above, for arguments beyond the sixth, the convention is to use space in the caller's frame for passing arguments. But for a tail call we pop the caller's frame and can no longer use it. Another alternative is to use space in the callee's frame for passing arguments. However, this option is also problematic because the caller and callee's frame overlap in memory. As we begin to copy the arguments from their sources in the caller's frame, the target locations in the callee's frame might overlap with the sources for later arguments! We solve this problem by not using the stack for passing more than six arguments but instead using the heap, as we describe in the Section~\ref{sec:limit-functions-r4}. As mentioned above, for a tail call we pop the caller's frame prior to making the tail call. The instructions for popping a frame are the instructions that we usually place in the conclusion of a function. Thus, we also need to place such code immediately before each tail call. These instructions include restoring the callee-saved registers, so it is good that the argument passing registers are all caller-saved registers. One last note regarding which instruction to use to make the tail call. When the callee is finished, it should not return to the current function, but it should return to the function that called the current one. Thus, the return address that is already on the stack is the right one, and we should not use \key{callq} to make the tail call, as that would unnecessarily overwrite the return address. Instead we can simply use the \key{jmp} instruction. Like the indirect function call, we write an \emph{indirect jump}\index{indirect jump} with a register prefixed with an asterisk. We recommend using \code{rax} to hold the jump target because the preceding conclusion overwrites just about everything else. \begin{lstlisting} jmp *%rax \end{lstlisting} \section{Shrink $R_4$} \label{sec:shrink-r4} The \code{shrink} pass performs a minor modification to ease the later passes. This pass introduces an explicit \code{main} function and changes the top \code{ProgramDefsExp} form to \code{ProgramDefs} as follows. \begin{lstlisting} (ProgramDefsExp |$\itm{info}$| (|$\Def\ldots$|) |$\Exp$|) |$\Rightarrow$| (ProgramDefs |$\itm{info}$| (|$\Def\ldots$| |$\itm{mainDef}$|)) \end{lstlisting} where $\itm{mainDef}$ is \begin{lstlisting} (Def 'main '() 'Integer '() |$\Exp'$|) \end{lstlisting} \section{Reveal Functions and the $F_1$ language} \label{sec:reveal-functions-r4} The syntax of $R_4$ is inconvenient for purposes of compilation in one respect: it conflates the use of function names and local variables. This is a problem because we need to compile the use of a function name differently than the use of a local variable; we need to use \code{leaq} to convert the function name (a label in x86) to an address in a register. Thus, it is a good idea to create a new pass that changes function references from just a symbol $f$ to $\FUNREF{f}$. This pass is named \code{reveal-functions} and the output language, $F_1$, is defined in Figure~\ref{fig:f1-syntax}. The concrete syntax for a function reference is $\CFUNREF{f}$. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Exp &::=& \ldots \mid \FUNREF{\Var}\\ \Def &::=& \gray{ \FUNDEF{\Var}{([\Var \code{:} \Type]\ldots)}{\Type}{\code{'()}}{\Exp} }\\ F_1 &::=& \PROGRAMDEFS{\code{'()}}{\LP \Def\ldots \RP} \end{array} \] \end{minipage} } \caption{The abstract syntax $F_1$, an extension of $R_4$ (Figure~\ref{fig:r4-syntax}).} \label{fig:f1-syntax} \end{figure} % TODO: rename $F_1$ to $R'_4$ %% Distinguishing between calls in tail position and non-tail position %% requires the pass to have some notion of context. We recommend using %% two mutually recursive functions, one for processing expressions in %% tail position and another for the rest. Placing this pass after \code{uniquify} will make sure that there are no local variables and functions that share the same name. On the other hand, \code{reveal-functions} needs to come before the \code{explicate-control} pass because that pass helps us compile \code{FunRef} forms into assignment statements. \section{Limit Functions} \label{sec:limit-functions-r4} Recall that we wish to limit the number of function parameters to six so that we do not need to use the stack for argument passing, which makes it easier to implement efficient tail calls. However, because the input language $R_4$ supports arbitrary numbers of function arguments, we have some work to do! This pass transforms functions and function calls that involve more than six arguments to pass the first five arguments as usual, but it packs the rest of the arguments into a vector and passes it as the sixth argument. Each function definition with too many parameters is transformed as follows. \begin{lstlisting} (Def |$f$| ([|$x_1$|:|$T_1$|] |$\ldots$| [|$x_n$|:|$T_n$|]) |$T_r$| |$\itm{info}$| |$\itm{body}$|) |$\Rightarrow$| (Def |$f$| ([|$x_1$|:|$T_1$|] |$\ldots$| [|$x_5$|:|$T_5$|] [vec : (Vector |$T_6 \ldots T_n$|)]) |$T_r$| |$\itm{info}$| |$\itm{body}'$|) \end{lstlisting} where the $\itm{body}$ is transformed into $\itm{body}'$ by replacing the occurrences of the later parameters with vector references. \begin{lstlisting} (Var |$x_i$|) |$\Rightarrow$| (Prim 'vector-ref (list vec (Int |$(i - 6)$|))) \end{lstlisting} For function calls with too many arguments, the \code{limit-functions} pass transforms them in the following way. \begin{tabular}{lll} \begin{minipage}{0.2\textwidth} \begin{lstlisting} (|$e_0$| |$e_1$| |$\ldots$| |$e_n$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.4\textwidth} \begin{lstlisting} (|$e_0$| |$e_1 \ldots e_5$| (vector |$e_6 \ldots e_n$|)) \end{lstlisting} \end{minipage} \end{tabular} \section{Remove Complex Operands} \label{sec:rco-r4} The primary decisions to make for this pass is whether to classify \code{FunRef} and \code{Apply} as either atomic or complex expressions. Recall that a simple expression will eventually end up as just an immediate argument of an x86 instruction. Function application will be translated to a sequence of instructions, so \code{Apply} must be classified as complex expression. On the other hand, the arguments of \code{Apply} should be atomic expressions. % Regarding \code{FunRef}, as discussed above, the function label needs to be converted to an address using the \code{leaq} instruction. Thus, even though \code{FunRef} seems rather simple, it needs to be classified as a complex expression so that we generate an assignment statement with a left-hand side that can serve as the target of the \code{leaq}. Figure~\ref{fig:r4-anf-syntax} defines the output language $R_4^{\dagger}$ of this pass. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{rcl} \Atm &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \BOOL{\itm{bool}} \mid \VOID{} } \\ \Exp &::=& \gray{ \Atm \mid \READ{} } \\ &\mid& \gray{ \NEG{\Atm} \mid \ADD{\Atm}{\Atm} } \\ &\mid& \gray{ \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \UNIOP{\key{'not}}{\Atm} } \\ &\mid& \gray{ \BINOP{\itm{cmp}}{\Atm}{\Atm} \mid \IF{\Exp}{\Exp}{\Exp} }\\ &\mid& \gray{ \LP\key{Collect}~\Int\RP \mid \LP\key{Allocate}~\Int~\Type\RP \mid \LP\key{GlobalValue}~\Var\RP }\\ &\mid& \FUNREF{\Var} \mid \APPLY{\Atm}{\Atm\ldots}\\ \Def &::=& \gray{ \FUNDEF{\Var}{([\Var \code{:} \Type]\ldots)}{\Type}{\code{'()}}{\Exp} }\\ R^{\dagger}_4 &::=& \gray{ \PROGRAMDEFS{\code{'()}}{\Def} } \end{array} \] \end{minipage} } \caption{$R_4^{\dagger}$ is $R_4$ in administrative normal form (ANF).} \label{fig:r4-anf-syntax} \end{figure} \section{Explicate Control and the $C_3$ language} \label{sec:explicate-control-r4} Figure~\ref{fig:c3-syntax} defines the abstract syntax for $C_3$, the output of \key{explicate-control}. (The concrete syntax is given in Figure~\ref{fig:c3-concrete-syntax} of the Appendix.) The auxiliary functions for assignment and tail contexts should be updated with cases for \code{Apply} and \code{FunRef} and the function for predicate context should be updated for \code{Apply} but not \code{FunRef}. (A \code{FunRef} can't be a Boolean.) In assignment and predicate contexts, \code{Apply} becomes \code{Call}, whereas in tail position \code{Apply} becomes \code{TailCall}. We recommend defining a new auxiliary function for processing function definitions. This code is similar to the case for \code{Program} in $R_3$. The top-level \code{explicate-control} function that handles the \code{ProgramDefs} form of $R_4$ can then apply this new function to all the function definitions. \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Atm &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \BOOL{\itm{bool}} }\\ \itm{cmp} &::= & \gray{ \key{eq?} \mid \key{<} } \\ \Exp &::= & \gray{ \Atm \mid \READ{} } \\ &\mid& \gray{ \NEG{\Atm} \mid \ADD{\Atm}{\Atm} }\\ &\mid& \gray{ \UNIOP{\key{not}}{\Atm} \mid \BINOP{\itm{cmp}}{\Atm}{\Atm} } \\ &\mid& \gray{ \LP\key{Allocate} \,\itm{int}\,\itm{type}\RP } \\ &\mid& \gray{ \BINOP{\key{'vector-ref}}{\Atm}{\INT{\Int}} }\\ &\mid& \gray{ \LP\key{Prim}~\key{'vector-set!}\,\LP\key{list}\,\Atm\,\INT{\Int}\,\Atm\RP\RP }\\ &\mid& \gray{ \LP\key{GlobalValue} \,\Var\RP \mid \LP\key{Void}\RP }\\ &\mid& \FUNREF{\itm{label}} \mid \CALL{\Atm}{\LP\Atm\ldots\RP} \\ \Stmt &::=& \gray{ \ASSIGN{\VAR{\Var}}{\Exp} \mid \LP\key{Collect} \,\itm{int}\RP } \\ \Tail &::= & \gray{ \RETURN{\Exp} \mid \SEQ{\Stmt}{\Tail} \mid \GOTO{\itm{label}} } \\ &\mid& \gray{ \IFSTMT{\BINOP{\itm{cmp}}{\Atm}{\Atm}}{\GOTO{\itm{label}}}{\GOTO{\itm{label}}} }\\ &\mid& \TAILCALL{\Atm}{\Atm\ldots} \\ \Def &::=& \DEF{\itm{label}}{\LP[\Var\key{:}\Type]\ldots\RP}{\Type}{\itm{info}}{\LP\LP\itm{label}\,\key{.}\,\Tail\RP\ldots\RP}\\ C_3 & ::= & \PROGRAMDEFS{\itm{info}}{\LP\Def\ldots\RP} \end{array} \] \end{minipage} } \caption{The abstract syntax of $C_3$, extending $C_2$ (Figure~\ref{fig:c2-syntax}).} \label{fig:c3-syntax} \end{figure} \section{Select Instructions and the x86$_3$ Language} \label{sec:select-r4} \index{instruction selection} The output of select instructions is a program in the x86$_3$ language, whose syntax is defined in Figure~\ref{fig:x86-3}. \index{x86} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Arg &::=& \gray{ \key{\$}\Int \mid \key{\%}\Reg \mid \Int\key{(}\key{\%}\Reg\key{)} \mid \key{\%}\itm{bytereg} } \mid \Var \key{(\%rip)} \mid \LP\key{fun-ref}\; \itm{label}\RP\\ \itm{cc} & ::= & \gray{ \key{e} \mid \key{l} \mid \key{le} \mid \key{g} \mid \key{ge} } \\ \Instr &::=& \ldots \mid \key{callq}\;\key{*}\Arg \mid \key{tailjmp}\;\Arg \mid \key{leaq}\;\Arg\key{,}\;\key{\%}\Reg \\ \Block &::= & \Instr\ldots \\ \Def &::= & \LP\key{define} \; \LP\itm{label}\RP \;\LP\LP\itm{label} \,\key{.}\, \Block\RP\ldots\RP\RP\\ x86_3 &::= & \Def\ldots \end{array} \] \end{minipage} } \caption{The concrete syntax of x86$_3$ (extends x86$_2$ of Figure~\ref{fig:x86-2-concrete}).} \label{fig:x86-3-concrete} \end{figure} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Arg &::=& \gray{ \INT{\Int} \mid \REG{\Reg} \mid \DEREF{\Reg}{\Int} \mid \BYTEREG{\Reg} } \\ &\mid& \gray{ (\key{Global}~\Var) } \mid \FUNREF{\itm{label}} \\ \Instr &::=& \ldots \mid \INDCALLQ{\Arg}{\itm{int}} \mid \TAILJMP{\Arg}{\itm{int}}\\ &\mid& \BININSTR{\code{'leaq}}{\Arg}{\REG{\Reg}}\\ \Block &::= & \BLOCK{\itm{info}}{\LP\Instr\ldots\RP}\\ \Def &::= & \DEF{\itm{label}}{\code{'()}}{\Type}{\itm{info}}{\LP\LP\itm{label}\,\key{.}\,\Block\RP\ldots\RP} \\ x86_3 &::= & \PROGRAMDEFS{\itm{info}}{\LP\Def\ldots\RP} \end{array} \] \end{minipage} } \caption{The abstract syntax of x86$_3$ (extends x86$_2$ of Figure~\ref{fig:x86-2}).} \label{fig:x86-3} \end{figure} An assignment of a function reference to a variable becomes a load-effective-address instruction as follows: \\ \begin{tabular}{lcl} \begin{minipage}{0.35\textwidth} \begin{lstlisting} |$\itm{lhs}$| = (fun-ref |$f$|); \end{lstlisting} \end{minipage} & $\Rightarrow$\qquad\qquad & \begin{minipage}{0.3\textwidth} \begin{lstlisting} leaq (fun-ref |$f$|), |$\itm{lhs}'$| \end{lstlisting} \end{minipage} \end{tabular} \\ Regarding function definitions, we need to remove the parameters and instead perform parameter passing using the conventions discussed in Section~\ref{sec:fun-x86}. That is, the arguments are passed in registers. We recommend turning the parameters into local variables and generating instructions at the beginning of the function to move from the argument passing registers to these local variables. \begin{lstlisting} (Def |$f$| '([|$x_1$| : |$T_1$|] [|$x_2$| : |$T_2$|] |$\ldots$| ) |$T_r$| |$\itm{info}$| |$G$|) |$\Rightarrow$| (Def |$f$| '() 'Integer |$\itm{info}'$| |$G'$|) \end{lstlisting} The $G'$ control-flow graph is the same as $G$ except that the \code{start} block is modified to add the instructions for moving from the argument registers to the parameter variables. So the \code{start} block of $G$ shown on the left is changed to the code on the right. \begin{center} \begin{minipage}{0.3\textwidth} \begin{lstlisting} start: |$\itm{instr}_1$| |$\vdots$| |$\itm{instr}_n$| \end{lstlisting} \end{minipage} $\Rightarrow$ \begin{minipage}{0.3\textwidth} \begin{lstlisting} start: movq %rdi, |$x_1$| movq %rsi, |$x_2$| |$\vdots$| |$\itm{instr}_1$| |$\vdots$| |$\itm{instr}_n$| \end{lstlisting} \end{minipage} \end{center} By changing the parameters to local variables, we are giving the register allocator control over which registers or stack locations to use for them. If you implemented the move-biasing challenge (Section~\ref{sec:move-biasing}), the register allocator will try to assign the parameter variables to the corresponding argument register, in which case the \code{patch-instructions} pass will remove the \code{movq} instruction. This happens in the example translation in Figure~\ref{fig:add-fun} of Section~\ref{sec:functions-example}, in the \code{add} function. % Also, note that the register allocator will perform liveness analysis on this sequence of move instructions and build the interference graph. So, for example, $x_1$ will be marked as interfering with \code{rsi} and that will prevent the assignment of $x_1$ to \code{rsi}, which is good, because that would overwrite the argument that needs to move into $x_2$. Next, consider the compilation of function calls. In the mirror image of handling the parameters of function definitions, the arguments need to be moved to the argument passing registers. The function call itself is performed with an indirect function call. The return value from the function is stored in \code{rax}, so it needs to be moved into the \itm{lhs}. \begin{lstlisting} |\itm{lhs}| = (call |\itm{fun}| |$\itm{arg}_1~\itm{arg}_2\ldots$|)); |$\Rightarrow$| movq |$\itm{arg}_1$|, %rdi movq |$\itm{arg}_2$|, %rsi |$\vdots$| callq *|\itm{fun}| movq %rax, |\itm{lhs}| \end{lstlisting} The \code{IndirectCallq} AST node includes an integer for the arity of the function, i.e., the number of parameters. That information is useful in the \code{uncover-live} pass for determining which argument-passing registers are potentially read during the call. For tail calls, the parameter passing is the same as non-tail calls: generate instructions to move the arguments into to the argument passing registers. After that we need to pop the frame from the procedure call stack. However, we do not yet know how big the frame is; that gets determined during register allocation. So instead of generating those instructions here, we invent a new instruction that means ``pop the frame and then do an indirect jump'', which we name \code{TailJmp}. The abstract syntax for this instruction includes an argument that specifies where to jump and an integer that represents the arity of the function being called. Recall that in Section~\ref{sec:explicate-control-r1} we recommended using the label \code{start} for the initial block of a program, and in Section~\ref{sec:select-r1} we recommended labeling the conclusion of the program with \code{conclusion}, so that $(\key{Return}\;\Arg)$ can be compiled to an assignment to \code{rax} followed by a jump to \code{conclusion}. With the addition of function definitions, we will have a starting block and conclusion for each function, but their labels need to be unique. We recommend prepending the function's name to \code{start} and \code{conclusion}, respectively, to obtain unique labels. (Alternatively, one could \code{gensym} labels for the start and conclusion and store them in the $\itm{info}$ field of the function definition.) \section{Register Allocation} \label{sec:register-allocation-r4} \subsection{Liveness Analysis} \label{sec:liveness-analysis-r4} \index{liveness analysis} %% The rest of the passes need only minor modifications to handle the new %% kinds of AST nodes: \code{fun-ref}, \code{indirect-callq}, and %% \code{leaq}. The \code{IndirectCallq} instruction should be treated like \code{Callq} regarding its written locations $W$, in that they should include all the caller-saved registers. Recall that the reason for that is to force call-live variables to be assigned to callee-saved registers or to be spilled to the stack. Regarding the set of read locations $R$ the arity field of \code{TailJmp} and \code{IndirectCallq} determines how many of the argument-passing registers should be considered as read by those instructions. \subsection{Build Interference Graph} \label{sec:build-interference-r4} With the addition of function definitions, we compute an interference graph for each function (not just one for the whole program). Recall that in Section~\ref{sec:reg-alloc-gc} we discussed the need to spill vector-typed variables that are live during a call to the \code{collect}. With the addition of functions to our language, we need to revisit this issue. Many functions perform allocation and therefore have calls to the collector inside of them. Thus, we should not only spill a vector-typed variable when it is live during a call to \code{collect}, but we should spill the variable if it is live during any function call. Thus, in the \code{build-interference} pass, we recommend adding interference edges between call-live vector-typed variables and the callee-saved registers (in addition to the usual addition of edges between call-live variables and the caller-saved registers). \subsection{Allocate Registers} The primary change to the \code{allocate-registers} pass is adding an auxiliary function for handling definitions (the \Def{} non-terminal in Figure~\ref{fig:x86-3}) with one case for function definitions. The logic is the same as described in Chapter~\ref{ch:register-allocation-r1}, except now register allocation is performed many times, once for each function definition, instead of just once for the whole program. \section{Patch Instructions} In \code{patch-instructions}, you should deal with the x86 idiosyncrasy that the destination argument of \code{leaq} must be a register. Additionally, you should ensure that the argument of \code{TailJmp} is \itm{rax}, our reserved register---this is to make code generation more convenient, because we trample many registers before the tail call (as explained in the next section). \section{Print x86} For the \code{print-x86} pass, the cases for \code{FunRef} and \code{IndirectCallq} are straightforward: output their concrete syntax. \begin{lstlisting} (FunRef |\itm{label}|) |$\Rightarrow$| |\itm{label}|(%rip) (IndirectCallq |\itm{arg}| |\itm{int}|) |$\Rightarrow$| callq *|\itm{arg}'| \end{lstlisting} The \code{TailJmp} node requires a bit work. A straightforward translation of \code{TailJmp} would be \code{jmp *$\itm{arg}$}, but before the jump we need to pop the current frame. This sequence of instructions is the same as the code for the conclusion of a function, except the \code{retq} is replaced with \code{jmp *$\itm{arg}$}. Regarding function definitions, you will need to generate a prelude and conclusion for each one. This code is similar to the prelude and conclusion that you generated for the \code{main} function in Chapter~\ref{ch:tuples}. To review, the prelude of every function should carry out the following steps. \begin{enumerate} \item Start with \code{.global} and \code{.align} directives followed by the label for the function. (See Figure~\ref{fig:add-fun} for an example.) \item Push \code{rbp} to the stack and set \code{rbp} to current stack pointer. \item Push to the stack all of the callee-saved registers that were used for register allocation. \item Move the stack pointer \code{rsp} down by the size of the stack frame for this function, which depends on the number of regular spills. (Aligned to 16 bytes.) \item Move the root stack pointer \code{r15} up by the size of the root-stack frame for this function, which depends on the number of spilled vectors. \label{root-stack-init} \item Initialize to zero all of the entries in the root-stack frame. \item Jump to the start block. \end{enumerate} The prelude of the \code{main} function has one additional task: call the \code{initialize} function to set up the garbage collector and move the value of the global \code{rootstack\_begin} in \code{r15}. This should happen before step \ref{root-stack-init} above, which depends on \code{r15}. The conclusion of every function should do the following. \begin{enumerate} \item Move the stack pointer back up by the size of the stack frame for this function. \item Restore the callee-saved registers by popping them from the stack. \item Move the root stack pointer back down by the size of the root-stack frame for this function. \item Restore \code{rbp} by popping it from the stack. \item Return to the caller with the \code{retq} instruction. \end{enumerate} \begin{exercise}\normalfont Expand your compiler to handle $R_4$ as outlined in this chapter. Create 5 new programs that use functions, including examples that pass functions and return functions from other functions, recursive functions, functions that create vectors, and functions that make tail calls. Test your compiler on these new programs and all of your previously created test programs. \end{exercise} \begin{figure}[tbp] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R4) at (0,2) {\large $R_4$}; \node (R4-1) at (3,2) {\large $R_4$}; \node (R4-2) at (6,2) {\large $R_4$}; \node (F1-1) at (12,0) {\large $F_1$}; \node (F1-2) at (9,0) {\large $F_1$}; \node (F1-3) at (6,0) {\large $F_1$}; \node (F1-4) at (3,0) {\large $F_1$}; \node (C3-2) at (3,-2) {\large $C_3$}; \node (x86-2) at (3,-4) {\large $\text{x86}^{*}_3$}; \node (x86-3) at (6,-4) {\large $\text{x86}^{*}_3$}; \node (x86-4) at (9,-4) {\large $\text{x86}_3$}; \node (x86-5) at (9,-6) {\large $\text{x86}^{\dagger}_3$}; \node (x86-2-1) at (3,-6) {\large $\text{x86}^{*}_3$}; \node (x86-2-2) at (6,-6) {\large $\text{x86}^{*}_3$}; \path[->,bend left=15] (R4) edge [above] node {\ttfamily\footnotesize shrink} (R4-1); \path[->,bend left=15] (R4-1) edge [above] node {\ttfamily\footnotesize uniquify} (R4-2); \path[->,bend left=15] (R4-2) edge [right] node {\ttfamily\footnotesize ~~reveal-functions} (F1-1); \path[->,bend left=15] (F1-1) edge [below] node {\ttfamily\footnotesize limit-functions} (F1-2); \path[->,bend right=15] (F1-2) edge [above] node {\ttfamily\footnotesize expose-alloc.} (F1-3); \path[->,bend right=15] (F1-3) edge [above] node {\ttfamily\footnotesize remove-complex.} (F1-4); \path[->,bend left=15] (F1-4) edge [right] node {\ttfamily\footnotesize explicate-control} (C3-2); \path[->,bend right=15] (C3-2) edge [left] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend left=15] (x86-2) edge [left] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [left] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend right=15] (x86-4) edge [left] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_4$, a language with functions.} \label{fig:R4-passes} \end{figure} Figure~\ref{fig:R4-passes} gives an overview of the passes for compiling $R_4$ to x86. \section{An Example Translation} \label{sec:functions-example} Figure~\ref{fig:add-fun} shows an example translation of a simple function in $R_4$ to x86. The figure also includes the results of the \code{explicate-control} and \code{select-instructions} passes. \begin{figure}[htbp] \begin{tabular}{ll} \begin{minipage}{0.5\textwidth} % s3_2.rkt \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] (define (add [x : Integer] [y : Integer]) : Integer (+ x y)) (add 40 2) \end{lstlisting} $\Downarrow$ \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] (define (add86 [x87 : Integer] [y88 : Integer]) : Integer add86start: return (+ x87 y88); ) (define (main) : Integer () mainstart: tmp89 = (fun-ref add86); (tail-call tmp89 40 2) ) \end{lstlisting} \end{minipage} & $\Rightarrow$ \begin{minipage}{0.5\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] (define (add86) : Integer add86start: movq %rdi, x87 movq %rsi, y88 movq x87, %rax addq y88, %rax jmp add11389conclusion ) (define (main) : Integer mainstart: leaq (fun-ref add86), tmp89 movq $40, %rdi movq $2, %rsi tail-jmp tmp89 ) \end{lstlisting} $\Downarrow$ \end{minipage} \end{tabular} \begin{tabular}{ll} \begin{minipage}{0.3\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] .globl add86 .align 16 add86: pushq %rbp movq %rsp, %rbp jmp add86start add86start: movq %rdi, %rax addq %rsi, %rax jmp add86conclusion add86conclusion: popq %rbp retq \end{lstlisting} \end{minipage} & \begin{minipage}{0.5\textwidth} \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] .globl main .align 16 main: pushq %rbp movq %rsp, %rbp movq $16384, %rdi movq $16384, %rsi callq initialize movq rootstack_begin(%rip), %r15 jmp mainstart mainstart: leaq add86(%rip), %rcx movq $40, %rdi movq $2, %rsi movq %rcx, %rax popq %rbp jmp *%rax mainconclusion: popq %rbp retq \end{lstlisting} \end{minipage} \end{tabular} \caption{Example compilation of a simple function to x86.} \label{fig:add-fun} \end{figure} % Challenge idea: inlining! (simple version) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Lexically Scoped Functions} \label{ch:lambdas} \index{lambda} \index{lexical scoping} This chapter studies lexically scoped functions as they appear in functional languages such as Racket. By lexical scoping we mean that a function's body may refer to variables whose binding site is outside of the function, in an enclosing scope. % Consider the example in Figure~\ref{fig:lexical-scoping} written in $R_5$, which extends $R_4$ with anonymous functions using the \key{lambda} form. The body of the \key{lambda}, refers to three variables: \code{x}, \code{y}, and \code{z}. The binding sites for \code{x} and \code{y} are outside of the \key{lambda}. Variable \code{y} is bound by the enclosing \key{let} and \code{x} is a parameter of function \code{f}. The \key{lambda} is returned from the function \code{f}. The main expression of the program includes two calls to \code{f} with different arguments for \code{x}, first \code{5} then \code{3}. The functions returned from \code{f} are bound to variables \code{g} and \code{h}. Even though these two functions were created by the same \code{lambda}, they are really different functions because they use different values for \code{x}. Applying \code{g} to \code{11} produces \code{20} whereas applying \code{h} to \code{15} produces \code{22}. The result of this program is \code{42}. \begin{figure}[btp] % s4_6.rkt \begin{lstlisting} (define (f [x : Integer]) : (Integer -> Integer) (let ([y 4]) (lambda: ([z : Integer]) : Integer (+ x (+ y z))))) (let ([g (f 5)]) (let ([h (f 3)]) (+ (g 11) (h 15)))) \end{lstlisting} \caption{Example of a lexically scoped function.} \label{fig:lexical-scoping} \end{figure} The approach that we take for implementing lexically scoped functions is to compile them into top-level function definitions, translating from $R_5$ into $R_4$. However, the compiler will need to provide special treatment for variable occurrences such as \code{x} and \code{y} in the body of the \code{lambda} of Figure~\ref{fig:lexical-scoping}. After all, an $R_4$ function may not refer to variables defined outside of it. To identify such variable occurrences, we review the standard notion of free variable. \begin{definition} A variable is \emph{free in expression} $e$ if the variable occurs inside $e$ but does not have an enclosing binding in $e$.\index{free variable} \end{definition} For example, in the expression \code{(+ x (+ y z))} the variables \code{x}, \code{y}, and \code{z} are all free. On the other hand, only \code{x} and \code{y} are free in the following expression because \code{z} is bound by the \code{lambda}. \begin{lstlisting} (lambda: ([z : Integer]) : Integer (+ x (+ y z))) \end{lstlisting} So the free variables of a \code{lambda} are the ones that will need special treatment. We need to arrange for some way to transport, at runtime, the values of those variables from the point where the \code{lambda} was created to the point where the \code{lambda} is applied. An efficient solution to the problem, due to \citet{Cardelli:1983aa}, is to bundle into a vector the values of the free variables together with the function pointer for the lambda's code, an arrangement called a \emph{flat closure} (which we shorten to just ``closure''). \index{closure}\index{flat closure} Fortunately, we have all the ingredients to make closures, Chapter~\ref{ch:tuples} gave us vectors and Chapter~\ref{ch:functions} gave us function pointers. The function pointer resides at index $0$ and the values for the free variables will fill in the rest of the vector. Let us revisit the example in Figure~\ref{fig:lexical-scoping} to see how closures work. It's a three-step dance. The program first calls function \code{f}, which creates a closure for the \code{lambda}. The closure is a vector whose first element is a pointer to the top-level function that we will generate for the \code{lambda}, the second element is the value of \code{x}, which is \code{5}, and the third element is \code{4}, the value of \code{y}. The closure does not contain an element for \code{z} because \code{z} is not a free variable of the \code{lambda}. Creating the closure is step 1 of the dance. The closure is returned from \code{f} and bound to \code{g}, as shown in Figure~\ref{fig:closures}. % The second call to \code{f} creates another closure, this time with \code{3} in the second slot (for \code{x}). This closure is also returned from \code{f} but bound to \code{h}, which is also shown in Figure~\ref{fig:closures}. \begin{figure}[tbp] \centering \includegraphics[width=0.6\textwidth]{figs/closures} \caption{Example closure representation for the \key{lambda}'s in Figure~\ref{fig:lexical-scoping}.} \label{fig:closures} \end{figure} Continuing with the example, consider the application of \code{g} to \code{11} in Figure~\ref{fig:lexical-scoping}. To apply a closure, we obtain the function pointer in the first element of the closure and call it, passing in the closure itself and then the regular arguments, in this case \code{11}. This technique for applying a closure is step 2 of the dance. % But doesn't this \code{lambda} only take 1 argument, for parameter \code{z}? The third and final step of the dance is generating a top-level function for a \code{lambda}. We add an additional parameter for the closure and we insert a \code{let} at the beginning of the function for each free variable, to bind those variables to the appropriate elements from the closure parameter. % This three-step dance is known as \emph{closure conversion}. We discuss the details of closure conversion in Section~\ref{sec:closure-conversion} and the code generated from the example in Section~\ref{sec:example-lambda}. But first we define the syntax and semantics of $R_5$ in Section~\ref{sec:r5}. \section{The $R_5$ Language} \label{sec:r5} The concrete and abstract syntax for $R_5$, a language with anonymous functions and lexical scoping, is defined in Figures~\ref{fig:r5-concrete-syntax} and ~\ref{fig:r5-syntax}. It adds the \key{lambda} form to the grammar for $R_4$, which already has syntax for function application. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Type &::=& \gray{\key{Integer} \mid \key{Boolean} \mid (\key{Vector}\;\Type\ldots) \mid \key{Void} \mid (\Type\ldots \; \key{->}\; \Type)} \\ \Exp &::=& \gray{ \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} \mid \CSUB{\Exp}{\Exp} } \\ &\mid& \gray{ \Var \mid \CLET{\Var}{\Exp}{\Exp} }\\ &\mid& \gray{\key{\#t} \mid \key{\#f} \mid (\key{and}\;\Exp\;\Exp) \mid (\key{or}\;\Exp\;\Exp) \mid (\key{not}\;\Exp) } \\ &\mid& \gray{ (\key{eq?}\;\Exp\;\Exp) \mid \CIF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ (\key{vector}\;\Exp\ldots) \mid (\key{vector-ref}\;\Exp\;\Int)} \\ &\mid& \gray{(\key{vector-set!}\;\Exp\;\Int\;\Exp)\mid (\key{void}) \mid (\Exp \; \Exp\ldots) } \\ &\mid& \LP \key{procedure-arity}~\Exp\RP \\ &\mid& \CLAMBDA{\LP\LS\Var \key{:} \Type\RS\ldots\RP}{\Type}{\Exp} \\ \Def &::=& \gray{ \CDEF{\Var}{\LS\Var \key{:} \Type\RS\ldots}{\Type}{\Exp} } \\ R_5 &::=& \gray{\Def\ldots \; \Exp} \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_5$, extending $R_4$ (Figure~\ref{fig:r4-concrete-syntax}) with \key{lambda}.} \label{fig:r5-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \itm{op} &::=& \ldots \mid \code{procedure-arity} \\ \Exp &::=& \gray{ \INT{\Int} \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \PRIM{\itm{op}}{\Exp\ldots} }\\ &\mid& \gray{ \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ \VOID{} \mid \LP\key{HasType}~\Exp~\Type \RP \mid \APPLY{\Exp}{\Exp\ldots} }\\ &\mid& \LAMBDA{\LP\LS\Var\code{:}\Type\RS\ldots\RP}{\Type}{\Exp}\\ \Def &::=& \gray{ \FUNDEF{\Var}{\LP\LS\Var \code{:} \Type\RS\ldots\RP}{\Type}{\code{'()}}{\Exp} }\\ R_5 &::=& \gray{ \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP}{\Exp} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_5$, extending $R_4$ (Figure~\ref{fig:r4-syntax}).} \label{fig:r5-syntax} \end{figure} \index{interpreter} \label{sec:interp-R5} Figure~\ref{fig:interp-R5} shows the definitional interpreter for $R_5$. The clause for \key{lambda} saves the current environment inside the returned \key{lambda}. Then the clause for \key{Apply} uses the environment from the \key{lambda}, the \code{lam-env}, when interpreting the body of the \key{lambda}. The \code{lam-env} environment is extended with the mapping of parameters to argument values. \begin{figure}[tbp] \begin{lstlisting} (define interp-R5-class (class interp-R4-class (super-new) (define/override (interp-op op) (match op ['procedure-arity (lambda (v) (match v [`(function (,xs ...) ,body ,lam-env) (length xs)] [else (error 'interp-op "expected a function, not ~a" v)]))] [else (super interp-op op)])) (define/override ((interp-exp env) e) (define recur (interp-exp env)) (match e [(Lambda (list `[,xs : ,Ts] ...) rT body) `(function ,xs ,body ,env)] [else ((super interp-exp env) e)])) )) (define (interp-R5 p) (send (new interp-R5-class) interp-program p)) \end{lstlisting} \caption{Interpreter for $R_5$.} \label{fig:interp-R5} \end{figure} \label{sec:type-check-r5} \index{type checking} Figure~\ref{fig:type-check-R5} shows how to type check the new \key{lambda} form. The body of the \key{lambda} is checked in an environment that includes the current environment (because it is lexically scoped) and also includes the \key{lambda}'s parameters. We require the body's type to match the declared return type. \begin{figure}[tbp] \begin{lstlisting} (define (type-check-R5 env) (lambda (e) (match e [(Lambda (and params `([,xs : ,Ts] ...)) rT body) (define-values (new-body bodyT) ((type-check-exp (append (map cons xs Ts) env)) body)) (define ty `(,@Ts -> ,rT)) (cond [(equal? rT bodyT) (values (HasType (Lambda params rT new-body) ty) ty)] [else (error "mismatch in return type" bodyT rT)])] ... ))) \end{lstlisting} \caption{Type checking the \key{lambda}'s in $R_5$.} \label{fig:type-check-R5} \end{figure} \section{Reveal Functions and the $F_2$ language} \label{sec:reveal-functions-r5} To support the \code{procedure-arity} operator we need to communicate the arity of a function to the point of closure creation. We can accomplish this by replacing the $\FUNREF{\Var}$ struct with one that has a second field for the arity: $\FUNREFARITY{\Var}{\Int}$. The output of this pass is the language $F_2$, whose syntax is defined in Figure~\ref{fig:f2-syntax}. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Exp &::=& \ldots \mid \FUNREFARITY{\Var}{\Int}\\ \Def &::=& \gray{ \FUNDEF{\Var}{([\Var \code{:} \Type]\ldots)}{\Type}{\code{'()}}{\Exp} }\\ F_2 &::=& \gray{\PROGRAMDEFS{\code{'()}}{\LP \Def\ldots \RP}} \end{array} \] \end{minipage} } \caption{The abstract syntax $F_2$, an extension of $R_5$ (Figure~\ref{fig:r5-syntax}).} \label{fig:f2-syntax} \end{figure} \section{Closure Conversion} \label{sec:closure-conversion} \index{closure conversion} The compiling of lexically-scoped functions into top-level function definitions is accomplished in the pass \code{convert-to-closures} that comes after \code{reveal-functions} and before \code{limit-functions}. As usual, we implement the pass as a recursive function over the AST. All of the action is in the clauses for \key{Lambda} and \key{Apply}. We transform a \key{Lambda} expression into an expression that creates a closure, that is, a vector whose first element is a function pointer and the rest of the elements are the free variables of the \key{Lambda}. We use the struct \code{Closure} here instead of using \code{vector} so that we can distinguish closures from vectors in Section~\ref{sec:optimize-closures} and to record the arity. In the generated code below, the \itm{name} is a unique symbol generated to identify the function and the \itm{arity} is the number of parameters (the length of \itm{ps}). \begin{lstlisting} (Lambda |\itm{ps}| |\itm{rt}| |\itm{body}|) |$\Rightarrow$| (Closure |\itm{arity}| (cons (FunRef |\itm{name}|) |\itm{fvs}|)) \end{lstlisting} In addition to transforming each \key{Lambda} into a \key{Closure}, we create a top-level function definition for each \key{Lambda}, as shown below.\\ \begin{minipage}{0.8\textwidth} \begin{lstlisting} (Def |\itm{name}| ([clos : (Vector _ |\itm{fvts}| ...)] |\itm{ps'}| ...) |\itm{rt'}| (Let |$\itm{fvs}_1$| (Prim 'vector-ref (list (Var clos) (Int 1))) ... (Let |$\itm{fvs}_n$| (Prim 'vector-ref (list (Var clos) (Int |$n$|))) |\itm{body'}|)...)) \end{lstlisting} \end{minipage}\\ The \code{clos} parameter refers to the closure. Translate the type annotations in \itm{ps} and the return type \itm{rt}, as discussed in the next paragraph, to obtain \itm{ps'} and \itm{rt'}. The types $\itm{fvts}$ are the types of the free variables in the lambda and the underscore \code{\_} is a dummy type that we use because it is rather difficult to give a type to the function in the closure's type.\footnote{To give an accurate type to a closure, we would need to add existential types to the type checker~\citep{Minamide:1996ys}.} The dummy type is considered to be equal to any other type during type checking. The sequence of \key{Let} forms bind the free variables to their values obtained from the closure. Closure conversion turns functions into vectors, so the type annotations in the program must also be translated. We recommend defining a auxiliary recursive function for this purpose. Function types should be translated as follows. \begin{lstlisting} (|$T_1, \ldots, T_n$| -> |$T_r$|) |$\Rightarrow$| (Vector ((Vector _) |$T'_1, \ldots, T'_n$| -> |$T'_r$|)) \end{lstlisting} The above type says that the first thing in the vector is a function pointer. The first parameter of the function pointer is a vector (a closure) and the rest of the parameters are the ones from the original function, with types $T'_1, \ldots, T'_n$. The \code{Vector} type for the closure omits the types of the free variables because 1) those types are not available in this context and 2) we do not need them in the code that is generated for function application. We transform function application into code that retrieves the function pointer from the closure and then calls the function, passing in the closure as the first argument. We bind $e'$ to a temporary variable to avoid code duplication. \begin{lstlisting} (Apply |$e$| |\itm{es}|) |$\Rightarrow$| (Let |\itm{tmp}| |$e'$| (Apply (Prim 'vector-ref (list (Var |\itm{tmp}|) (Int 0))) (cons |\itm{tmp}| |\itm{es'}|))) \end{lstlisting} There is also the question of what to do with references top-level function definitions. To maintain a uniform translation of function application, we turn function references into closures. \begin{tabular}{lll} \begin{minipage}{0.3\textwidth} \begin{lstlisting} (FunRefArity |$f$| |$n$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.5\textwidth} \begin{lstlisting} (Closure |$n$| (FunRef |$f$|) '()) \end{lstlisting} \end{minipage} \end{tabular} \\ % The top-level function definitions need to be updated as well to take an extra closure parameter. \section{An Example Translation} \label{sec:example-lambda} Figure~\ref{fig:lexical-functions-example} shows the result of \code{reveal-functions} and \code{convert-to-closures} for the example program demonstrating lexical scoping that we discussed at the beginning of this chapter. \begin{figure}[tbp] \begin{minipage}{0.8\textwidth} % tests/lambda_test_6.rkt \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define (f6 [x7 : Integer]) : (Integer -> Integer) (let ([y8 4]) (lambda: ([z9 : Integer]) : Integer (+ x7 (+ y8 z9))))) (define (main) : Integer (let ([g0 ((fun-ref-arity f6 1) 5)]) (let ([h1 ((fun-ref-arity f6 1) 3)]) (+ (g0 11) (h1 15))))) \end{lstlisting} $\Rightarrow$ \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define (f6 [fvs4 : _] [x7 : Integer]) : (Vector ((Vector _) Integer -> Integer)) (let ([y8 4]) (closure 1 (list (fun-ref lambda2) x7 y8)))) (define (lambda2 [fvs3 : (Vector _ Integer Integer)] [z9 : Integer]) : Integer (let ([x7 (vector-ref fvs3 1)]) (let ([y8 (vector-ref fvs3 2)]) (+ x7 (+ y8 z9))))) (define (main) : Integer (let ([g0 (let ([clos5 (closure 1 (list (fun-ref f6)))]) ((vector-ref clos5 0) clos5 5))]) (let ([h1 (let ([clos6 (closure 1 (list (fun-ref f6)))]) ((vector-ref clos6 0) clos6 3))]) (+ ((vector-ref g0 0) g0 11) ((vector-ref h1 0) h1 15))))) \end{lstlisting} \end{minipage} \caption{Example of closure conversion.} \label{fig:lexical-functions-example} \end{figure} \begin{exercise}\normalfont Expand your compiler to handle $R_5$ as outlined in this chapter. Create 5 new programs that use \key{lambda} functions and make use of lexical scoping. Test your compiler on these new programs and all of your previously created test programs. \end{exercise} \section{Expose Allocation} \label{sec:expose-allocation-r5} Compile the $\CLOSURE{\itm{arity}}{\LP\Exp\ldots\RP}$ form into code that allocates and initializes a vector, similar to the translation of the \code{vector} operator in Section~\ref{sec:expose-allocation}. The only difference is replacing the use of \ALLOC{\itm{len}}{\itm{type}} with \ALLOCCLOS{\itm{len}}{\itm{type}}{\itm{arity}}. \section{Explicate Control and $C_4$} \label{sec:explicate-r5} The output language of \code{explicate-control} is $C_4$ whose abstract syntax is defined in Figure~\ref{fig:c4-syntax}. The only difference with respect to $C_3$ is the addition of the \code{AllocateClosure} form to the grammar for $\Exp$. The handling of \code{AllocateClosure} in the \code{explicate-control} pass is similar to the handling of other expressions such as primitive operators. \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::= & \ldots \mid \ALLOCCLOS{\Int}{\Type}{\Int} \\ \Stmt &::=& \gray{ \ASSIGN{\VAR{\Var}}{\Exp} \mid \LP\key{Collect} \,\itm{int}\RP } \\ \Tail &::= & \gray{ \RETURN{\Exp} \mid \SEQ{\Stmt}{\Tail} \mid \GOTO{\itm{label}} } \\ &\mid& \gray{ \IFSTMT{\BINOP{\itm{cmp}}{\Atm}{\Atm}}{\GOTO{\itm{label}}}{\GOTO{\itm{label}}} }\\ &\mid& \gray{ \TAILCALL{\Atm}{\Atm\ldots} } \\ \Def &::=& \gray{ \DEF{\itm{label}}{\LP[\Var\key{:}\Type]\ldots\RP}{\Type}{\itm{info}}{\LP\LP\itm{label}\,\key{.}\,\Tail\RP\ldots\RP} }\\ C_4 & ::= & \gray{ \PROGRAMDEFS{\itm{info}}{\LP\Def\ldots\RP} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $C_4$, extending $C_3$ (Figure~\ref{fig:c3-syntax}).} \label{fig:c4-syntax} \end{figure} \section{Select Instructions} \label{sec:select-instructions-R5} Compile \ALLOCCLOS{\itm{len}}{\itm{type}}{\itm{arity}} in almost the same way as the \ALLOC{\itm{len}}{\itm{type}} form (Section~\ref{sec:select-instructions-gc}). The only difference is that you should place the \itm{arity} in the tag that is stored at position $0$ of the vector. Recall that in Section~\ref{sec:select-instructions-gc} we used the first $56$ bits of the 64-bit tag, but that the rest were unused. So the arity goes into the tag in bit positions $57$ through $63$. Compile the \code{procedure-arity} operator into a sequence of instructions that access the tag from position $0$ of the vector and shift it by $57$ bits to the right. \begin{figure}[p] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R4) at (0,2) {\large $R_4$}; \node (R4-2) at (3,2) {\large $R_4$}; \node (R4-3) at (6,2) {\large $R_4$}; \node (F1-1) at (12,0) {\large $F_1$}; \node (F1-2) at (9,0) {\large $F_1$}; \node (F1-3) at (6,0) {\large $F_1$}; \node (F1-4) at (3,0) {\large $F_1$}; \node (F1-5) at (0,0) {\large $F_1$}; \node (C3-2) at (3,-2) {\large $C_3$}; \node (x86-2) at (3,-4) {\large $\text{x86}^{*}_3$}; \node (x86-3) at (6,-4) {\large $\text{x86}^{*}_3$}; \node (x86-4) at (9,-4) {\large $\text{x86}^{*}_3$}; \node (x86-5) at (9,-6) {\large $\text{x86}^{\dagger}_3$}; \node (x86-2-1) at (3,-6) {\large $\text{x86}^{*}_3$}; \node (x86-2-2) at (6,-6) {\large $\text{x86}^{*}_3$}; \path[->,bend left=15] (R4) edge [above] node {\ttfamily\footnotesize shrink} (R4-2); \path[->,bend left=15] (R4-2) edge [above] node {\ttfamily\footnotesize uniquify} (R4-3); \path[->,bend left=15] (R4-3) edge [right] node {\ttfamily\footnotesize reveal-functions} (F1-1); \path[->,bend left=15] (F1-1) edge [below] node {\ttfamily\footnotesize convert-to-clos.} (F1-2); \path[->,bend right=15] (F1-2) edge [above] node {\ttfamily\footnotesize limit-fun.} (F1-3); \path[->,bend right=15] (F1-3) edge [above] node {\ttfamily\footnotesize expose-alloc.} (F1-4); \path[->,bend right=15] (F1-4) edge [above] node {\ttfamily\footnotesize remove-complex.} (F1-5); \path[->,bend right=15] (F1-5) edge [right] node {\ttfamily\footnotesize explicate-control} (C3-2); \path[->,bend left=15] (C3-2) edge [left] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend right=15] (x86-2) edge [left] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [left] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [right] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_5$, a language with lexically-scoped functions.} \label{fig:R5-passes} \end{figure} Figure~\ref{fig:R5-passes} provides an overview of all the passes needed for the compilation of $R_5$. \clearpage \section{Challenge: Optimize Closures} \label{sec:optimize-closures} In this chapter we compiled lexically-scoped functions into a relatively efficient representation: flat closures. However, even this representation comes with some overhead. For example, consider the following program with a function \code{tail-sum} that does not have any free variables and where all the uses of \code{tail-sum} are in applications where we know that only \code{tail-sum} is being applied (and not any other functions). \begin{center} \begin{minipage}{0.95\textwidth} \begin{lstlisting} (define (tail-sum [n : Integer] [r : Integer]) : Integer (if (eq? n 0) r (tail-sum (- n 1) (+ n r)))) (+ (tail-sum 5 0) 27) \end{lstlisting} \end{minipage} \end{center} As described in this chapter, we uniformly apply closure conversion to all functions, obtaining the following output for this program. \begin{center} \begin{minipage}{0.95\textwidth} \begin{lstlisting} (define (tail_sum1 [fvs5 : _] [n2 : Integer] [r3 : Integer]) : Integer (if (eq? n2 0) r3 (let ([clos4 (closure (list (fun-ref tail_sum1)))]) ((vector-ref clos4 0) clos4 (+ n2 -1) (+ n2 r3))))) (define (main) : Integer (+ (let ([clos6 (closure (list (fun-ref tail_sum1)))]) ((vector-ref clos6 0) clos6 5 0)) 27)) \end{lstlisting} \end{minipage} \end{center} In the previous Chapter, there would be no allocation in the program and the calls to \code{tail-sum} would be direct calls. In contrast, the above program allocates memory for each \code{closure} and the calls to \code{tail-sum} are indirect. These two differences incur considerable overhead in a program such as this one, where the allocations and indirect calls occur inside a tight loop. One might think that this problem is trivial to solve: can't we just recognize calls of the form \code{((fun-ref $f$) $e_1 \ldots e_n$)} and compile them to direct calls \code{((fun-ref $f$) $e'_1 \ldots e'_n$)} instead of treating it like a call to a closure? We would also drop the \code{fvs5} parameter of \code{tail\_sum1}. % However, this problem is not so trivial because a global function may ``escape'' and become involved in applications that also involve closures. Consider the following example in which the application \code{(f 41)} needs to be compiled into a closure application, because the \code{lambda} may get bound to \code{f}, but the \code{add1} function might also get bound to \code{f}. \begin{lstlisting} (define (add1 [x : Integer]) : Integer (+ x 1)) (let ([y (read)]) (let ([f (if (eq? (read) 0) add1 (lambda: ([x : Integer]) : Integer (- x y)))]) (f 41))) \end{lstlisting} If a global function name is used in any way other than as the operator in a direct call, then we say that the function \emph{escapes}. If a global function does not escape, then we do not need to perform closure conversion on the function. \begin{exercise}\normalfont Implement an auxiliary function for detecting which global functions escape. Using that function, implement an improved version of closure conversion that does not apply closure conversion to global functions that do not escape but instead compiles them as regular functions. Create several new test cases that check whether you properly detect whether global functions escape or not. \end{exercise} So far we have reduced the overhead of calling global functions, but it would also be nice to reduce the overhead of calling a \code{lambda} when we can determine at compile time which \code{lambda} will be called. We refer to such calls as \emph{known calls}. Consider the following example in which a \code{lambda} is bound to \code{f} and then applied. \begin{lstlisting} (let ([y (read)]) (let ([f (lambda: ([x : Integer]) : Integer (+ x y))]) (f 21))) \end{lstlisting} Closure conversion compiles \code{(f 21)} into an indirect call: \begin{lstlisting} (define (lambda5 [fvs6 : (Vector _ Integer)] [x3 : Integer]) : Integer (let ([y2 (vector-ref fvs6 1)]) (+ x3 y2))) (define (main) : Integer (let ([y2 (read)]) (let ([f4 (Closure 1 (list (fun-ref lambda5) y2))]) ((vector-ref f4 0) f4 21)))) \end{lstlisting} but we can instead compile the application \code{(f 21)} into a direct call to \code{lambda5}: \begin{lstlisting} (define (main) : Integer (let ([y2 (read)]) (let ([f4 (Closure 1 (list (fun-ref lambda5) y2))]) ((fun-ref lambda5) f4 21)))) \end{lstlisting} The problem of determining which lambda will be called from a particular application is quite challenging in general and the topic of considerable research~\citep{Shivers:1988aa,Gilray:2016aa}. For the following exercise we recommend that you compile an application to a direct call when the operator is a variable and the variable is \code{let}-bound to a closure. This can be accomplished by maintaining an environment mapping \code{let}-bound variables to function names. Extend the environment whenever you encounter a closure on the right-hand side of a \code{let}, mapping the \code{let}-bound variable to the name of the global function for the closure. This pass should come after closure conversion. \begin{exercise}\normalfont Implement a compiler pass, named \code{optimize-known-calls}, that compiles known calls into direct calls. Verify that your compiler is successful in this regard on several example programs. \end{exercise} These exercises only scratches the surface of optimizing of closures. A good next step for the interested reader is to look at the work of \citet{Keep:2012ab}. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Dynamic Typing} \label{ch:type-dynamic} \index{dynamic typing} In this chapter we discuss the compilation of $R_7$, a dynamically typed language that is a subset of Racket. This is in contrast to the previous chapters, which have studied the compilation of Typed Racket. In dynamically typed languages such as $R_7$, a given expression may produce a value of a different type each time it is executed. Consider the following example with a conditional \code{if} expression that may return a Boolean or an integer depending on the input to the program. % part of dynamic_test_25.rkt \begin{lstlisting} (not (if (eq? (read) 1) #f 0)) \end{lstlisting} Languages that allow expressions to produce different kinds of values are called \emph{polymorphic}, a word composed of the Greek roots ``poly'', meaning ``many'', and ``morph'', meaning ``shape''. There are several kinds of polymorphism in programming languages, such as subtype polymorphism and parametric polymorphism~\citep{Cardelli:1985kx}. The kind of polymorphism we study in this chapter does not have a special name but it is the kind that arises in dynamically typed languages. Another characteristic of dynamically typed languages is that primitive operations, such as \code{not}, are often defined to operate on many different types of values. In fact, in Racket, the \code{not} operator produces a result for any kind of value: given \code{\#f} it returns \code{\#t} and given anything else it returns \code{\#f}. Furthermore, even when primitive operations restrict their inputs to values of a certain type, this restriction is enforced at runtime instead of during compilation. For example, the following vector reference results in a run-time contract violation because the index must be in integer, not a Boolean such as \code{\#t}. \begin{lstlisting} (vector-ref (vector 42) #t) \end{lstlisting} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.97\textwidth} \[ \begin{array}{rcl} \itm{cmp} &::= & \key{eq?} \mid \key{<} \mid \key{<=} \mid \key{>} \mid \key{>=} \\ \Exp &::=& \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} \mid \CSUB{\Exp}{\Exp} \\ &\mid& \Var \mid \CLET{\Var}{\Exp}{\Exp} \\ &\mid& \key{\#t} \mid \key{\#f} \mid \CBINOP{\key{and}}{\Exp}{\Exp} \mid \CBINOP{\key{or}}{\Exp}{\Exp} \mid \CUNIOP{\key{not}}{\Exp} \\ &\mid& \LP\itm{cmp}\;\Exp\;\Exp\RP \mid \CIF{\Exp}{\Exp}{\Exp} \\ &\mid& \LP\key{vector}\;\Exp\ldots\RP \mid \LP\key{vector-ref}\;\Exp\;\Exp\RP \\ &\mid& \LP\key{vector-set!}\;\Exp\;\Exp\;\Exp\RP \mid \LP\key{void}\RP \\ &\mid& \LP\Exp \; \Exp\ldots\RP \mid \LP\key{lambda}\;\LP\Var\ldots\RP\;\Exp\RP \\ & \mid & \LP\key{boolean?}\;\Exp\RP \mid \LP\key{integer?}\;\Exp\RP\\ & \mid & \LP\key{vector?}\;\Exp\RP \mid \LP\key{procedure?}\;\Exp\RP \mid \LP\key{void?}\;\Exp\RP \\ \Def &::=& \LP\key{define}\; \LP\Var \; \Var\ldots\RP \; \Exp\RP \\ R_7 &::=& \Def\ldots\; \Exp \end{array} \] \end{minipage} } \caption{Syntax of $R_7$, an untyped language (a subset of Racket).} \label{fig:r7-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::=& \INT{\Int} \mid \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} \\ &\mid& \PRIM{\itm{op}}{\Exp\ldots} \\ &\mid& \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} \\ &\mid& \VOID{} \mid \APPLY{\Exp}{\Exp\ldots} \\ &\mid& \LAMBDA{\LP\Var\ldots\RP}{\code{'Any}}{\Exp}\\ \Def &::=& \FUNDEF{\Var}{\LP\Var\ldots\RP}{\code{'Any}}{\code{'()}}{\Exp} \\ R_7 &::=& \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP}{\Exp} \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_7$.} \label{fig:r7-syntax} \end{figure} The concrete and abstract syntax of $R_7$, our subset of Racket, is defined in Figures~\ref{fig:r7-concrete-syntax} and \ref{fig:r7-syntax}. % There is no type checker for $R_7$ because it is not a statically typed language (it's dynamically typed!). The definitional interpreter for $R_7$ is presented in Figure~\ref{fig:interp-R7} and its auxiliary functions are defined in Figure~\ref{fig:interp-R7-aux}. Consider the match clause for \code{(Int n)}. Instead of simply returning the integer \code{n} (as in the interpreter for $R_1$ in Figure~\ref{fig:interp-R1}), the interpreter for $R_7$ creates a \emph{tagged value}\index{tagged value} that combines an underlying value with a tag that identifies what kind of value it is. We define the following struct to represented tagged values. \begin{lstlisting} (struct Tagged (value tag) #:transparent) \end{lstlisting} The tags are \code{Integer}, \code{Boolean}, \code{Void}, \code{Vector}, and \code{Procedure}. Tags are closely related to types but don't always capture all the information that a type does. For example, a vector of type \code{(Vector Any Any)} is tagged with \code{Vector} and a procedure of type \code{(Any Any -> Any)} is tagged with \code{Procedure}. Next consider the match clause for \code{vector-ref}. The \code{check-tag} auxiliary function (Figure~\ref{fig:interp-R7-aux}) is used to ensure that the first argument is a vector and the second is an integer. If they are not, a \code{trapped-error} is raised. Recall from Section~\ref{sec:interp-R0} that when a definition interpreter raises a \code{trapped-error} error, the compiled code must also signal an error by exiting with return code \code{255}. A \code{trapped-error} is also raised if the index is not less than length of the vector. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define ((interp-R7-exp env) ast) (define recur (interp-R7-exp env)) (match ast [(Var x) (lookup x env)] [(Int n) (Tagged n 'Integer)] [(Bool b) (Tagged b 'Boolean)] [(Lambda xs rt body) (Tagged `(function ,xs ,body ,env) 'Procedure)] [(Prim 'vector es) (Tagged (apply vector (for/list ([e es]) (recur e))) 'Vector)] [(Prim 'vector-ref (list e1 e2)) (define vec (recur e1)) (define i (recur e2)) (check-tag vec 'Vector ast) (check-tag i 'Integer ast) (unless (< (Tagged-value i) (vector-length (Tagged-value vec))) (error 'trapped-error "index ~a too big\nin ~v" (Tagged-value i) ast)) (vector-ref (Tagged-value vec) (Tagged-value i))] [(Prim 'vector-set! (list e1 e2 e3)) (define vec (recur e1)) (define i (recur e2)) (define arg (recur e3)) (check-tag vec 'Vector ast) (check-tag i 'Integer ast) (unless (< (Tagged-value i) (vector-length (Tagged-value vec))) (error 'trapped-error "index ~a too big\nin ~v" (Tagged-value i) ast)) (vector-set! (Tagged-value vec) (Tagged-value i) arg) (Tagged (void) 'Void)] [(Let x e body) ((interp-R7-exp (cons (cons x (recur e)) env)) body)] [(Prim 'and (list e1 e2)) (recur (If e1 e2 (Bool #f)))] [(Prim 'or (list e1 e2)) (define v1 (recur e1)) (match (Tagged-value v1) [#f (recur e2)] [else v1])] [(Prim 'eq? (list l r)) (Tagged (equal? (recur l) (recur r)) 'Boolean)] [(Prim op (list e1)) #:when (set-member? type-predicates op) (tag-value ((interp-op op) (Tagged-value (recur e1))))] [(Prim op es) (define args (map recur es)) (define tags (for/list ([arg args]) (Tagged-tag arg))) (unless (for/or ([expected-tags (op-tags op)]) (equal? expected-tags tags)) (error 'trapped-error "illegal argument tags ~a\nin ~v" tags ast)) (tag-value (apply (interp-op op) (for/list ([a args]) (Tagged-value a))))] [(If q t f) (match (Tagged-value (recur q)) [#f (recur f)] [else (recur t)])] [(Apply f es) (define new-f (recur f)) (define args (map recur es)) (check-tag new-f 'Procedure ast) (define f-val (Tagged-value new-f)) (match f-val [`(function ,xs ,body ,lam-env) (unless (eq? (length xs) (length args)) (error 'trapped-error "~a != ~a\nin ~v" (length args) (length xs) ast)) (define new-env (append (map cons xs args) lam-env)) ((interp-R7-exp new-env) body)] [else (error "interp-R7-exp, expected function, not" f-val)])])) \end{lstlisting} \caption{Interpreter for the $R_7$ language.} \label{fig:interp-R7} \end{figure} \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define (interp-op op) (match op ['+ fx+] ['- fx-] ['read read-fixnum] ['not (lambda (v) (match v [#t #f] [#f #t]))] ['< (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (< v1 v2)]))] ['<= (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (<= v1 v2)]))] ['> (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (> v1 v2)]))] ['>= (lambda (v1 v2) (cond [(and (fixnum? v1) (fixnum? v2)) (>= v1 v2)]))] ['boolean? boolean?] ['integer? fixnum?] ['void? void?] ['vector? vector?] ['vector-length vector-length] ['procedure? (match-lambda [`(functions ,xs ,body ,env) #t] [else #f])] [else (error 'interp-op "unknown operator" op)])) (define (op-tags op) (match op ['+ '((Integer Integer))] ['- '((Integer Integer) (Integer))] ['read '(())] ['not '((Boolean))] ['< '((Integer Integer))] ['<= '((Integer Integer))] ['> '((Integer Integer))] ['>= '((Integer Integer))] ['vector-length '((Vector))])) (define type-predicates (set 'boolean? 'integer? 'vector? 'procedure? 'void?)) (define (tag-value v) (cond [(boolean? v) (Tagged v 'Boolean)] [(fixnum? v) (Tagged v 'Integer)] [(procedure? v) (Tagged v 'Procedure)] [(vector? v) (Tagged v 'Vector)] [(void? v) (Tagged v 'Void)] [else (error 'tag-value "unidentified value ~a" v)])) (define (check-tag val expected ast) (define tag (Tagged-tag val)) (unless (eq? tag expected) (error 'trapped-error "expected ~a, not ~a\nin ~v" expected tag ast))) \end{lstlisting} \caption{Auxiliary functions for the $R_7$ interpreter.} \label{fig:interp-R7-aux} \end{figure} \clearpage \section{Representation of Tagged Values} The interpreter for $R_7$ introduced a new kind of value, a tagged value. To compile $R_7$ to x86 we must decide how to represent tagged values at the bit level. Because almost every operation in $R_7$ involves manipulating tagged values, the representation must be efficient. Recall that all of our values are 64 bits. We shall steal the 3 right-most bits to encode the tag. We use $001$ to identify integers, $100$ for Booleans, $010$ for vectors, $011$ for procedures, and $101$ for the void value. We define the following auxiliary function for mapping types to tag codes. \begin{align*} \itm{tagof}(\key{Integer}) &= 001 \\ \itm{tagof}(\key{Boolean}) &= 100 \\ \itm{tagof}((\key{Vector} \ldots)) &= 010 \\ \itm{tagof}((\ldots \key{->} \ldots)) &= 011 \\ \itm{tagof}(\key{Void}) &= 101 \end{align*} This stealing of 3 bits comes at some price: our integers are reduced to ranging from $-2^{60}$ to $2^{60}$. The stealing does not adversely affect vectors and procedures because those values are addresses, and our addresses are 8-byte aligned so the rightmost 3 bits are unused, they are always $000$. Thus, we do not lose information by overwriting the rightmost 3 bits with the tag and we can simply zero-out the tag to recover the original address. To make tagged values into first-class entities, we can give them a type, called \code{Any}, and define operations such as \code{Inject} and \code{Project} for creating and using them, yielding the $R_6$ intermediate language. We describe how to compile $R_7$ to $R_6$ in Section~\ref{sec:compile-r7} but first we describe the $R_6$ language in greater detail. \section{The $R_6$ Language} \label{sec:r6-lang} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \itm{op} &::= & \ldots \mid \code{any-vector-length} \mid \code{any-vector-ref} \mid \code{any-vector-set!}\\ &\mid& \code{boolean?} \mid \code{integer?} \mid \code{vector?} \mid \code{procedure?} \mid \code{void?} \\ \Exp &::=& \ldots \mid \gray{ \PRIM{\itm{op}}{\Exp\ldots} } \\ &\mid& \INJECT{\Exp}{\FType} \mid \PROJECT{\Exp}{\FType} \\ \Def &::=& \gray{ \FUNDEF{\Var}{\LP[\Var \code{:} \Type]\ldots\RP}{\Type}{\code{'()}}{\Exp} }\\ R_6 &::=& \gray{ \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP}{\Exp} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_6$, extending $R_5$ (Figure~\ref{fig:r5-syntax}).} \label{fig:r6-syntax} \end{figure} The abstract syntax of $R_6$ is defined in Figure~\ref{fig:r6-syntax}. (The concrete syntax of $R_6$ is in the Appendix, Figure~\ref{fig:r6-concrete-syntax}.) The $\INJECT{e}{T}$ form converts the value produced by expression $e$ of type $T$ into a tagged value. The $\PROJECT{e}{T}$ form converts the tagged value produced by expression $e$ into a value of type $T$ or else halts the program if the type tag is not equivalent to $T$. % Note that in both \code{Inject} and \code{Project}, the type $T$ is restricted to a flat type $\FType$, which simplifies the implementation and corresponds with what is needed for compiling $R_7$. The \code{any-vector} operators adapt the vector operations so that they can be applied to a value of type \code{Any}. They also generalize the vector operations in that the index is not restricted to be a literal integer in the grammar but is allowed to be any expression. The type predicates such as \key{boolean?} expect their argument to produce a tagged value; they return \key{\#t} if the tag corresponds to the predicate and they return \key{\#f} otherwise. The type checker for $R_6$ is shown in Figures~\ref{fig:type-check-R6-part-1} and \ref{fig:type-check-R6-part-2} and uses the auxiliary functions in Figure~\ref{fig:type-check-R6-aux}. % The interpreter for $R_6$ is in Figure~\ref{fig:interp-R6} and the auxiliary functions \code{apply-inject} and \code{apply-project} are in Figure~\ref{fig:apply-project}. \begin{figure}[btp] \begin{lstlisting}[basicstyle=\ttfamily\small] (define type-check-R6-class (class type-check-R5-class (super-new) (inherit check-type-equal?) (define/override (type-check-exp env) (lambda (e) (define recur (type-check-exp env)) (match e [(Inject e1 ty) (unless (flat-ty? ty) (error 'type-check "may only inject from flat type, not ~a" ty)) (define-values (new-e1 e-ty) (recur e1)) (check-type-equal? e-ty ty e) (values (Inject new-e1 ty) 'Any)] [(Project e1 ty) (unless (flat-ty? ty) (error 'type-check "may only project to flat type, not ~a" ty)) (define-values (new-e1 e-ty) (recur e1)) (check-type-equal? e-ty 'Any e) (values (Project new-e1 ty) ty)] [(Prim 'any-vector-length (list e1)) (define-values (e1^ t1) (recur e1)) (check-type-equal? t1 'Any e) (values (Prim 'any-vector-length (list e1^)) 'Integer)] [(Prim 'any-vector-ref (list e1 e2)) (define-values (e1^ t1) (recur e1)) (define-values (e2^ t2) (recur e2)) (check-type-equal? t1 'Any e) (check-type-equal? t2 'Integer e) (values (Prim 'any-vector-ref (list e1^ e2^)) 'Any)] [(Prim 'any-vector-set! (list e1 e2 e3)) (define-values (e1^ t1) (recur e1)) (define-values (e2^ t2) (recur e2)) (define-values (e3^ t3) (recur e3)) (check-type-equal? t1 'Any e) (check-type-equal? t2 'Integer e) (check-type-equal? t3 'Any e) (values (Prim 'any-vector-set! (list e1^ e2^ e3^)) 'Void)] \end{lstlisting} \caption{Type checker for the $R_6$ language, part 1.} \label{fig:type-check-R6-part-1} \end{figure} \begin{figure}[btp] \begin{lstlisting}[basicstyle=\ttfamily\small] [(ValueOf e ty) (define-values (new-e e-ty) (recur e)) (values (ValueOf new-e ty) ty)] [(Prim pred (list e1)) #:when (set-member? (type-predicates) pred) (define-values (new-e1 e-ty) (recur e1)) (check-type-equal? e-ty 'Any e) (values (Prim pred (list new-e1)) 'Boolean)] [(If cnd thn els) (define-values (cnd^ Tc) (recur cnd)) (define-values (thn^ Tt) (recur thn)) (define-values (els^ Te) (recur els)) (check-type-equal? Tc 'Boolean cnd) (check-type-equal? Tt Te e) (values (If cnd^ thn^ els^) (combine-types Tt Te))] [(Exit) (values (Exit) '_)] [(Prim 'eq? (list arg1 arg2)) (define-values (e1 t1) (recur arg1)) (define-values (e2 t2) (recur arg2)) (match* (t1 t2) [(`(Vector ,ts1 ...) `(Vector ,ts2 ...)) (void)] [(other wise) (check-type-equal? t1 t2 e)]) (values (Prim 'eq? (list e1 e2)) 'Boolean)] [else ((super type-check-exp env) e)]))) )) \end{lstlisting} \caption{Type checker for the $R_6$ language, part 2.} \label{fig:type-check-R6-part-2} \end{figure} \begin{figure}[tbp] \begin{lstlisting} (define/override (operator-types) (append '((integer? . ((Any) . Boolean)) (vector? . ((Any) . Boolean)) (procedure? . ((Any) . Boolean)) (void? . ((Any) . Boolean)) (tag-of-any . ((Any) . Integer)) (make-any . ((_ Integer) . Any)) ) (super operator-types))) (define/public (type-predicates) (set 'boolean? 'integer? 'vector? 'procedure? 'void?)) (define/public (combine-types t1 t2) (match (list t1 t2) [(list '_ t2) t2] [(list t1 '_) t1] [(list `(Vector ,ts1 ...) `(Vector ,ts2 ...)) `(Vector ,@(for/list ([t1 ts1] [t2 ts2]) (combine-types t1 t2)))] [(list `(,ts1 ... -> ,rt1) `(,ts2 ... -> ,rt2)) `(,@(for/list ([t1 ts1] [t2 ts2]) (combine-types t1 t2)) -> ,(combine-types rt1 rt2))] [else t1])) (define/public (flat-ty? ty) (match ty [(or `Integer `Boolean '_ `Void) #t] [`(Vector ,ts ...) (for/and ([t ts]) (eq? t 'Any))] [`(,ts ... -> ,rt) (and (eq? rt 'Any) (for/and ([t ts]) (eq? t 'Any)))] [else #f])) \end{lstlisting} \caption{Auxiliary methods for type checking $R_6$.} \label{fig:type-check-R6-aux} \end{figure} \begin{figure}[btp] \begin{lstlisting} (define interp-R6-class (class interp-R5-class (super-new) (define/override (interp-op op) (match op ['boolean? (match-lambda [`(tagged ,v1 ,tg) (equal? tg (any-tag 'Boolean))] [else #f])] ['integer? (match-lambda [`(tagged ,v1 ,tg) (equal? tg (any-tag 'Integer))] [else #f])] ['vector? (match-lambda [`(tagged ,v1 ,tg) (equal? tg (any-tag `(Vector Any)))] [else #f])] ['procedure? (match-lambda [`(tagged ,v1 ,tg) (equal? tg (any-tag `(Any -> Any)))] [else #f])] ['eq? (match-lambda* [`((tagged ,v1^ ,tg1) (tagged ,v2^ ,tg2)) (and (eq? v1^ v2^) (equal? tg1 tg2))] [ls (apply (super interp-op op) ls)])] ['any-vector-ref (lambda (v i) (match v [`(tagged ,v^ ,tg) (vector-ref v^ i)]))] ['any-vector-set! (lambda (v i a) (match v [`(tagged ,v^ ,tg) (vector-set! v^ i a)]))] ['any-vector-length (lambda (v) (match v [`(tagged ,v^ ,tg) (vector-length v^)]))] [else (super interp-op op)])) (define/override ((interp-exp env) e) (define recur (interp-exp env)) (match e [(Inject e ty) `(tagged ,(recur e) ,(any-tag ty))] [(Project e ty2) (apply-project (recur e) ty2)] [else ((super interp-exp env) e)])) )) (define (interp-R6 p) (send (new interp-R6-class) interp-program p)) \end{lstlisting} \caption{Interpreter for $R_6$.} \label{fig:interp-R6} \end{figure} \begin{figure}[tbp] \begin{lstlisting} (define/public (apply-inject v tg) (Tagged v tg)) (define/public (apply-project v ty2) (define tag2 (any-tag ty2)) (match v [(Tagged v1 tag1) (cond [(eq? tag1 tag2) (match ty2 [`(Vector ,ts ...) (define l1 ((interp-op 'vector-length) v1)) (cond [(eq? l1 (length ts)) v1] [else (error 'apply-project "vector length mismatch, ~a != ~a" l1 (length ts))])] [`(,ts ... -> ,rt) (match v1 [`(function ,xs ,body ,env) (cond [(eq? (length xs) (length ts)) v1] [else (error 'apply-project "arity mismatch ~a != ~a" (length xs) (length ts))])] [else (error 'apply-project "expected function not ~a" v1)])] [else v1])] [else (error 'apply-project "tag mismatch ~a != ~a" tag1 tag2)])] [else (error 'apply-project "expected tagged value, not ~a" v)])) \end{lstlisting} \caption{Auxiliary functions for injection and projection.} \label{fig:apply-project} \end{figure} \clearpage \section{Cast Insertion: Compiling $R_7$ to $R_6$} \label{sec:compile-r7} The \code{cast-insert} pass compiles from $R_7$ to $R_6$. Figure~\ref{fig:compile-r7-r6} shows the compilation of many of the $R_7$ forms into $R_6$. An important invariant of this pass is that given a subexpression $e$ in the $R_7$ program, the pass will produce an expression $e'$ in $R_6$ that has type \key{Any}. For example, the first row in Figure~\ref{fig:compile-r7-r6} shows the compilation of the Boolean \code{\#t}, which must be injected to produce an expression of type \key{Any}. % The second row of Figure~\ref{fig:compile-r7-r6}, the compilation of addition, is representative of compilation for many primitive operations: the arguments have type \key{Any} and must be projected to \key{Integer} before the addition can be performed. The compilation of \key{lambda} (third row of Figure~\ref{fig:compile-r7-r6}) shows what happens when we need to produce type annotations: we simply use \key{Any}. % The compilation of \code{if} and \code{eq?} demonstrate how this pass has to account for some differences in behavior between $R_7$ and $R_6$. The $R_7$ language is more permissive than $R_6$ regarding what kind of values can be used in various places. For example, the condition of an \key{if} does not have to be a Boolean. For \key{eq?}, the arguments need not be of the same type (in that case the result is \code{\#f}). \begin{figure}[btp] \centering \begin{tabular}{|lll|} \hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} #t \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (inject #t Boolean) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (+ |$e_1$| |$e_2$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (inject (+ (project |$e'_1$| Integer) (project |$e'_2$| Integer)) Integer) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (lambda (|$x_1 \ldots x_n$|) |$e$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (inject (lambda: ([|$x_1$|:Any]|$\ldots$|[|$x_n$|:Any]):Any |$e'$|) (Any|$\ldots$|Any -> Any)) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (|$e_0$| |$e_1 \ldots e_n$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} ((project |$e'_0$| (Any|$\ldots$|Any -> Any)) |$e'_1 \ldots e'_n$|) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (vector-ref |$e_1$| |$e_2$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (any-vector-ref |$e_1'$| |$e_2'$|) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (if |$e_1$| |$e_2$| |$e_3$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (if (eq? |$e'_1$| (inject #f Boolean)) |$e'_3$| |$e'_2$|) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (eq? |$e_1$| |$e_2$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (inject (eq? |$e'_1$| |$e'_2$|) Boolean) \end{lstlisting} \end{minipage} \\[2ex]\hline \begin{minipage}{0.27\textwidth} \begin{lstlisting} (not |$e_1$|) \end{lstlisting} \end{minipage} & $\Rightarrow$ & \begin{minipage}{0.65\textwidth} \begin{lstlisting} (if (eq? |$e'_1$| (inject #f Boolean)) (inject #t Boolean) (inject #f Boolean)) \end{lstlisting} \end{minipage} \\[2ex]\hline \end{tabular} \caption{Cast Insertion} \label{fig:compile-r7-r6} \end{figure} \section{Reveal Casts} \label{sec:reveal-casts-r6} % TODO: define R'_6 In the \code{reveal-casts} pass we recommend compiling \code{project} into an \code{if} expression that checks whether the value's tag matches the target type; if it does, the value is converted to a value of the target type by removing the tag; if it does not, the program exits. To perform these actions we need a new primitive operation, \code{tag-of-any}, and two new forms, \code{ValueOf} and \code{Exit}. The \code{tag-of-any} operation retrieves the type tag from a tagged value of type \code{Any}. The \code{ValueOf} form retrieves the underlying value from a tagged value. The \code{ValueOf} form includes the type for the underlying value which is used by the type checker. Finally, the \code{Exit} form ends the execution of the program. If the target type of the projection is \code{Boolean} or \code{Integer}, then \code{Project} can be translated as follows. \begin{center} \begin{minipage}{1.0\textwidth} \begin{lstlisting} (Project |$e$| |$\FType$|) |$\Rightarrow$| (Let |$\itm{tmp}$| |$e'$| (If (Prim 'eq? (list (Prim 'tag-of-any (list (Var |$\itm{tmp}$|))) (Int |$\itm{tagof}(\FType)$|))) (ValueOf |$\itm{tmp}$| |$\FType$|) (Exit))) \end{lstlisting} \end{minipage} \end{center} If the target type of the projection is a vector or function type, then there is a bit more work to do. For vectors, check that the length of the vector type matches the length of the vector (using the \code{vector-length} primitive). For functions, check that the number of parameters in the function type matches the function's arity (using \code{procedure-arity}). Regarding \code{inject}, we recommend compiling it to a slightly lower-level primitive operation named \code{make-any}. This operation takes a tag instead of a type. \begin{center} \begin{minipage}{1.0\textwidth} \begin{lstlisting} (Inject |$e$| |$\FType$|) |$\Rightarrow$| (Prim 'make-any (list |$e'$| (Int |$\itm{tagof}(\FType)$|))) \end{lstlisting} \end{minipage} \end{center} The type predicates (\code{boolean?}, etc.) can be translated into uses of \code{tag-of-any} and \code{eq?} in a similar way as in the translation of \code{Project}. The \code{any-vector-ref} and \code{any-vector-set!} operations combine the projection action with the vector operation. Also, the read and write operations allow arbitrary expressions for the index so the type checker for $R_6$ (Figure~\ref{fig:type-check-R6-part-1}) cannot guarantee that the index is within bounds. Thus, we insert code to perform bounds checking at runtime. The translation for \code{any-vector-ref} is as follows and the other two operations are translated in a similar way. \begin{lstlisting} (Prim 'any-vector-ref (list |$e_1$| |$e_2$|)) |$\Rightarrow$| (Let |$v$| |$e'_1$| (Let |$i$| |$e'_2$| (If (Prim 'eq? (list (Prim 'tag-of-any (list (Var |$v$|))) (Int 2))) (If (Prim '< (list (Var |$i$|) (Prim 'any-vector-length (list (Var |$v$|))))) (Prim 'any-vector-ref (list (Var |$v$|) (Var |$i$|))) (Exit)))) \end{lstlisting} \section{Remove Complex Operands} \label{sec:rco-r6} The \code{ValueOf} and \code{Exit} forms are both complex expressions. The subexpression of \code{ValueOf} must be atomic. \section{Explicate Control and $C_5$} \label{sec:explicate-r6} The output of \code{explicate-control} is the $C_5$ language whose syntax is defined in Figure~\ref{fig:c5-syntax}. The \code{ValueOf} form that we added to $R_6$ remains an expression and the \code{Exit} expression becomes a $\Tail$. Also, note that the index argument of \code{vector-ref} and \code{vector-set!} is an $\Atm$ instead of an integer, as in $C_2$ (Figure~\ref{fig:c2-syntax}). \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::= & \ldots \mid \BINOP{\key{'any-vector-ref}}{\Atm}{\Atm} \\ &\mid& (\key{Prim}~\key{'any-vector-set!}\,(\key{list}\,\Atm\,\Atm\,\Atm))\\ &\mid& \VALUEOF{\Exp}{\FType} \\ \Stmt &::=& \gray{ \ASSIGN{\VAR{\Var}}{\Exp} \mid \LP\key{Collect} \,\itm{int}\RP }\\ \Tail &::= & \gray{ \RETURN{\Exp} \mid \SEQ{\Stmt}{\Tail} \mid \GOTO{\itm{label}} } \\ &\mid& \gray{ \IFSTMT{\BINOP{\itm{cmp}}{\Atm}{\Atm}}{\GOTO{\itm{label}}}{\GOTO{\itm{label}}} }\\ &\mid& \gray{ \TAILCALL{\Atm}{\Atm\ldots} } \mid \LP\key{Exit}\RP \\ \Def &::=& \gray{ \DEF{\itm{label}}{\LP[\Var\key{:}\Type]\ldots\RP}{\Type}{\itm{info}}{\LP\LP\itm{label}\,\key{.}\,\Tail\RP\ldots\RP} }\\ C_4 & ::= & \gray{ \PROGRAMDEFS{\itm{info}}{\LP\Def\ldots\RP} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $C_5$, extending $C_4$ (Figure~\ref{fig:c4-syntax}).} \label{fig:c5-syntax} \end{figure} \section{Select Instructions} \label{sec:select-r6} In the \code{select-instructions} pass we translate the primitive operations on the \code{Any} type to x86 instructions that involve manipulating the 3 tag bits of the tagged value. \paragraph{Make-any} We recommend compiling the \key{make-any} primitive as follows if the tag is for \key{Integer} or \key{Boolean}. The \key{salq} instruction shifts the destination to the left by the number of bits specified its source argument (in this case $3$, the length of the tag) and it preserves the sign of the integer. We use the \key{orq} instruction to combine the tag and the value to form the tagged value. \\ \begin{lstlisting} (Assign |\itm{lhs}| (Prim 'make-any (list |$e$| (Int |$\itm{tag}$|)))) |$\Rightarrow$| movq |$e'$|, |\itm{lhs'}| salq $3, |\itm{lhs'}| orq $|$\itm{tag}$|, |\itm{lhs'}| \end{lstlisting} The instruction selection for vectors and procedures is different because their is no need to shift them to the left. The rightmost 3 bits are already zeros as described at the beginning of this chapter. So we just combine the value and the tag using \key{orq}. \\ \begin{lstlisting} (Assign |\itm{lhs}| (Prim 'make-any (list |$e$| (Int |$\itm{tag}$|)))) |$\Rightarrow$| movq |$e'$|, |\itm{lhs'}| orq $|$\itm{tag}$|, |\itm{lhs'}| \end{lstlisting} \paragraph{Tag-of-any} Recall that the \code{tag-of-any} operation extracts the type tag from a value of type \code{Any}. The type tag is the bottom three bits, so we obtain the tag by taking the bitwise-and of the value with $111$ ($7$ in decimal). \begin{lstlisting} (Assign |\itm{lhs}| (Prim 'tag-of-any (list |$e$|))) |$\Rightarrow$| movq |$e'$|, |\itm{lhs'}| andq $7, |\itm{lhs'}| \end{lstlisting} \paragraph{ValueOf} Like \key{make-any}, the instructions for \key{ValueOf} are different depending on whether the type $T$ is a pointer (vector or procedure) or not (Integer or Boolean). The following shows the instruction selection for Integer and Boolean. We produce an untagged value by shifting it to the right by 3 bits. \begin{lstlisting} (Assign |\itm{lhs}| (ValueOf |$e$| |$T$|)) |$\Rightarrow$| movq |$e'$|, |\itm{lhs'}| sarq $3, |\itm{lhs'}| \end{lstlisting} % In the case for vectors and procedures, there is no need to shift. Instead we just need to zero-out the rightmost 3 bits. We accomplish this by creating the bit pattern $\ldots 0111$ ($7$ in decimal) and apply \code{bitwise-not} to obtain $\ldots 11111000$ (-8 in decimal) which we \code{movq} into the destination $\itm{lhs}$. We then apply \code{andq} with the tagged value to get the desired result. \\ \begin{lstlisting} (Assign |\itm{lhs}| (ValueOf |$e$| |$T$|)) |$\Rightarrow$| movq $|$-8$|, |\itm{lhs'}| andq |$e'$|, |\itm{lhs'}| \end{lstlisting} %% \paragraph{Type Predicates} We leave it to the reader to %% devise a sequence of instructions to implement the type predicates %% \key{boolean?}, \key{integer?}, \key{vector?}, and \key{procedure?}. \paragraph{Any-vector-length} \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'any-vector-length (list |$a_1$|))) |$\Longrightarrow$| movq |$\neg 111$|, %r11 andq |$a_1'$|, %r11 movq 0(%r11), %r11 andq $126, %r11 sarq $1, %r11 movq %r11, |$\itm{lhs'}$| \end{lstlisting} \paragraph{Any-vector-ref} The index may be an arbitrary atom so instead of computing the offset at compile time, instructions need to be generated to compute the offset at runtime as follows. Note the use of the new instruction \code{imulq}. \begin{center} \begin{minipage}{0.96\textwidth} \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'any-vector-ref (list |$a_1$| |$a_2$|))) |$\Longrightarrow$| movq |$\neg 111$|, %r11 andq |$a_1'$|, %r11 movq |$a_2'$|, %rax addq $1, %rax imulq $8, %rax addq %rax, %r11 movq 0(%r11) |$\itm{lhs'}$| \end{lstlisting} \end{minipage} \end{center} \paragraph{Any-vector-set!} The code generation for \code{any-vector-set!} is similar to the other \code{any-vector} operations. \section{Register Allocation for $R_6$} \label{sec:register-allocation-r6} \index{register allocation} There is an interesting interaction between tagged values and garbage collection that has an impact on register allocation. A variable of type \code{Any} might refer to a vector and therefore it might be a root that needs to be inspected and copied during garbage collection. Thus, we need to treat variables of type \code{Any} in a similar way to variables of type \code{Vector} for purposes of register allocation. In particular, \begin{itemize} \item If a variable of type \code{Any} is live during a function call, then it must be spilled. This can be accomplished by changing \code{build-interference} to mark all variables of type \code{Any} that are live after a \code{callq} as interfering with all the registers. \item If a variable of type \code{Any} is spilled, it must be spilled to the root stack instead of the normal procedure call stack. \end{itemize} Another concern regarding the root stack is that the garbage collector needs to differentiate between (1) plain old pointers to tuples, (2) a tagged value that points to a tuple, and (3) a tagged value that is not a tuple. We enable this differentiation by choosing not to use the tag $000$ in the $\itm{tagof}$ function. Instead, that bit pattern is reserved for identifying plain old pointers to tuples. That way, if one of the first three bits is set, then we have a tagged value and inspecting the tag can differentiation between vectors ($010$) and the other kinds of values. \begin{exercise}\normalfont Expand your compiler to handle $R_6$ as discussed in the last few sections. Create 5 new programs that use the \code{Any} type and the new operations (\code{inject}, \code{project}, \code{boolean?}, etc.). Test your compiler on these new programs and all of your previously created test programs. \end{exercise} \begin{exercise}\normalfont Expand your compiler to handle $R_7$ as outlined in this chapter. Create tests for $R_7$ by adapting ten of your previous test programs by removing type annotations. Add 5 more tests programs that specifically rely on the language being dynamically typed. That is, they should not be legal programs in a statically typed language, but nevertheless, they should be valid $R_7$ programs that run to completion without error. \end{exercise} \begin{figure}[p] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R4) at (0,4) {\large $R_7$}; \node (R4-2) at (3,4) {\large $R_7$}; \node (R4-3) at (6,4) {\large $R_7$}; \node (R4-4) at (9,4) {\large $R'_7$}; \node (R4-5) at (9,2) {\large $R'_6$}; \node (R4-6) at (12,2) {\large $R'_6$}; \node (R4-7) at (12,0) {\large $R'_6$}; \node (F1-2) at (9,0) {\large $R'_6$}; \node (F1-3) at (6,0) {\large $R'_6$}; \node (F1-4) at (3,0) {\large $R'_6$}; \node (F1-5) at (0,0) {\large $R'_6$}; \node (C3-2) at (3,-2) {\large $C_3$}; \node (x86-2) at (3,-4) {\large $\text{x86}^{*}_3$}; \node (x86-3) at (6,-4) {\large $\text{x86}^{*}_3$}; \node (x86-4) at (9,-4) {\large $\text{x86}^{*}_3$}; \node (x86-5) at (9,-6) {\large $\text{x86}^{\dagger}_3$}; \node (x86-2-1) at (3,-6) {\large $\text{x86}^{*}_3$}; \node (x86-2-2) at (6,-6) {\large $\text{x86}^{*}_3$}; \path[->,bend left=15] (R4) edge [above] node {\ttfamily\footnotesize shrink} (R4-2); \path[->,bend left=15] (R4-2) edge [above] node {\ttfamily\footnotesize uniquify} (R4-3); \path[->,bend left=15] (R4-3) edge [above] node {\ttfamily\footnotesize reveal-functions} (R4-4); \path[->,bend right=15] (R4-4) edge [left] node {\ttfamily\footnotesize cast-insert} (R4-5); \path[->,bend left=15] (R4-5) edge [above] node {\ttfamily\footnotesize check-bounds} (R4-6); \path[->,bend left=15] (R4-6) edge [left] node {\ttfamily\footnotesize reveal-casts} (R4-7); \path[->,bend left=15] (R4-7) edge [below] node {\ttfamily\footnotesize convert-to-clos.} (F1-2); \path[->,bend right=15] (F1-2) edge [above] node {\ttfamily\footnotesize limit-fun.} (F1-3); \path[->,bend right=15] (F1-3) edge [above] node {\ttfamily\footnotesize expose-alloc.} (F1-4); \path[->,bend right=15] (F1-4) edge [above] node {\ttfamily\footnotesize remove-complex.} (F1-5); \path[->,bend right=15] (F1-5) edge [right] node {\ttfamily\footnotesize explicate-control} (C3-2); \path[->,bend left=15] (C3-2) edge [left] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend right=15] (x86-2) edge [left] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [left] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [right] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_7$, a dynamically typed language.} \label{fig:R7-passes} \end{figure} Figure~\ref{fig:R7-passes} provides an overview of all the passes needed for the compilation of $R_7$. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Loops and Assignment} \label{ch:loop} % todo: define R'_8 In this chapter we study two features that are the hallmarks of imperative programming languages: loops and assignments to local variables. The following example demonstrates these new features by computing the sum of the first five positive integers. % similar to loop_test_1.rkt \begin{lstlisting} (let ([sum 0]) (let ([i 5]) (begin (while (> i 0) (begin (set! sum (+ sum i)) (set! i (- i 1)))) sum))) \end{lstlisting} The \code{while} loop consists of a condition and a body. % The \code{set!} consists of a variable and a right-hand-side expression. % The primary purpose of both the \code{while} loop and \code{set!} is to cause side effects, so it is convenient to also include in a language feature for sequencing side effects: the \code{begin} expression. It consists of one or more subexpressions that are evaluated left-to-right. \section{The $R_8$ Language} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::=& \gray{ \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} \mid \CSUB{\Exp}{\Exp} } \\ &\mid& \gray{ \Var \mid \CLET{\Var}{\Exp}{\Exp} }\\ &\mid& \gray{\key{\#t} \mid \key{\#f} \mid (\key{and}\;\Exp\;\Exp) \mid (\key{or}\;\Exp\;\Exp) \mid (\key{not}\;\Exp) } \\ &\mid& \gray{ (\key{eq?}\;\Exp\;\Exp) \mid \CIF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ (\key{vector}\;\Exp\ldots) \mid (\key{vector-ref}\;\Exp\;\Int)} \\ &\mid& \gray{(\key{vector-set!}\;\Exp\;\Int\;\Exp)\mid (\key{void}) \mid (\Exp \; \Exp\ldots) } \\ &\mid& \gray{ \LP \key{procedure-arity}~\Exp\RP \mid \CLAMBDA{\LP\LS\Var \key{:} \Type\RS\ldots\RP}{\Type}{\Exp} } \\ &\mid& \CSETBANG{\Var}{\Exp} \mid \CBEGIN{\Exp\ldots}{\Exp} \mid \CWHILE{\Exp}{\Exp} \\ \Def &::=& \gray{ \CDEF{\Var}{\LS\Var \key{:} \Type\RS\ldots}{\Type}{\Exp} } \\ R_8 &::=& \gray{\Def\ldots \; \Exp} \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_8$, extending $R_6$ (Figure~\ref{fig:r6-concrete-syntax}).} \label{fig:r8-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::=& \gray{ \INT{\Int} \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \PRIM{\itm{op}}{\Exp\ldots} }\\ &\mid& \gray{ \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ \VOID{} \mid \LP\key{HasType}~\Exp~\Type \RP \mid \APPLY{\Exp}{\Exp\ldots} }\\ &\mid& \gray{ \LAMBDA{\LP\LS\Var\code{:}\Type\RS\ldots\RP}{\Type}{\Exp} }\\ &\mid& \SETBANG{\Var}{\Exp} \mid \BEGIN{\LP\Exp\ldots\RP}{\Exp} \mid \WHILE{\Exp}{\Exp} \\ \Def &::=& \gray{ \FUNDEF{\Var}{\LP\LS\Var \code{:} \Type\RS\ldots\RP}{\Type}{\code{'()}}{\Exp} }\\ R_8 &::=& \gray{ \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP}{\Exp} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_8$, extending $R_6$ (Figure~\ref{fig:r6-syntax}).} \label{fig:r8-syntax} \end{figure} The concrete syntax of $R_8$ is defined in Figure~\ref{fig:r8-concrete-syntax} and its abstract syntax is defined in Figure~\ref{fig:r8-syntax}. % The definitional interpreter for $R_8$ is shown in Figure~\ref{fig:interp-R8}. We add three new cases for \code{SetBang}, \code{WhileLoop}, and \code{Begin} and we make changes to the cases for \code{Var}, \code{Let}, and \code{Apply} regarding variables. To support assignment to variables and to make their lifetimes indefinite (see the second example in Section~\ref{sec:assignment-scoping}), we box the value that is bound to each variable (in \code{Let}) and function parameter (in \code{Apply}). The case for \code{Var} unboxes the value. % Now to discuss the new cases. For \code{SetBang}, we lookup the variable in the environment to obtain a boxed value and then we change it using \code{set-box!} to the result of evaluating the right-hand side. The result value of a \code{SetBang} is \code{void}. % For the \code{WhileLoop}, we repeatedly 1) evaluate the condition, and if the result is true, 2) evaluate the body. The result value of a \code{while} loop is also \code{void}. % Finally, the $\BEGIN{\itm{es}}{\itm{body}}$ expression evaluates the subexpressions \itm{es} for their effects and then evaluates and returns the result from \itm{body}. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define interp-R8-class (class interp-R6-class (super-new) (define/override ((interp-exp env) e) (define recur (interp-exp env)) (match e [(SetBang x rhs) (set-box! (lookup x env) (recur rhs))] [(WhileLoop cnd body) (define (loop) (cond [(recur cnd) (recur body) (loop)] [else (void)])) (loop)] [(Begin es body) (for ([e es]) (recur e)) (recur body)] [else ((super interp-exp env) e)])) )) (define (interp-R8 p) (send (new interp-R8-class) interp-program p)) \end{lstlisting} \caption{Interpreter for $R_8$.} \label{fig:interp-R8} \end{figure} The type checker for $R_8$ is define in Figure~\ref{fig:type-check-R8}. For \code{SetBang}, the type of the variable and the right-hand-side must agree. The result type is \code{Void}. For the \code{WhileLoop}, the condition must be a \code{Boolean}. The result type is also \code{Void}. For \code{Begin}, the result type is the type of its last subexpression. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define type-check-R8-class (class type-check-R6-class (super-new) (inherit check-type-equal?) (define/override (type-check-exp env) (lambda (e) (define recur (type-check-exp env)) (match e [(SetBang x rhs) (define-values (rhs^ rhsT) (recur rhs)) (define varT (dict-ref env x)) (check-type-equal? rhsT varT e) (values (SetBang x rhs^) 'Void)] [(WhileLoop cnd body) (define-values (cnd^ Tc) (recur cnd)) (check-type-equal? Tc 'Boolean e) (define-values (body^ Tbody) ((type-check-exp env) body)) (values (WhileLoop cnd^ body^) 'Void)] [(Begin es body) (define-values (es^ ts) (for/lists (l1 l2) ([e es]) (recur e))) (define-values (body^ Tbody) (recur body)) (values (Begin es^ body^) Tbody)] [else ((super type-check-exp env) e)]))) )) (define (type-check-R8 p) (send (new type-check-R8-class) type-check-program p)) \end{lstlisting} \caption{Type checking \key{SetBang}, \key{WhileLoop}, and \code{Begin} in $R_8$.} \label{fig:type-check-R8} \end{figure} At first glance, the translation of these language features to x86 seems straightforward because the $C_3$ intermediate language already supports all of the ingredients that we need: assignment, \code{goto}, conditional branching, and sequencing. However, there are two complications that arise which we discuss in the next two sections. After that we introduce one new compiler pass and the changes necessary to the existing passes. \section{Assignment and Lexically Scoped Functions} \label{sec:assignment-scoping} The addition of assignment raises a problem with our approach to implementing lexically-scoped functions. Consider the following example in which function \code{f} has a free variable \code{x} that is changed after \code{f} is created but before the call to \code{f}. % loop_test_11.rkt \begin{lstlisting} (let ([x 0]) (let ([y 0]) (let ([z 20]) (let ([f (lambda: ([a : Integer]) : Integer (+ a (+ x z)))]) (begin (set! x 10) (set! y 12) (f y)))))) \end{lstlisting} The correct output for this example is \code{42} because the call to \code{f} is required to use the current value of \code{x} (which is \code{10}). Unfortunately, the closure conversion pass (Section~\ref{sec:closure-conversion}) generates code for the \code{lambda} that copies the old value of \code{x} into a closure. Thus, if we naively add support for assignment to our current compiler, the output of this program would be \code{32}. A first attempt at solving this problem would be to save a pointer to \code{x} in the closure and change the occurrences of \code{x} inside the lambda to dereference the pointer. Of course, this would require assigning \code{x} to the stack and not to a register. However, the problem goes a bit deeper. Consider the following example in which we create a counter abstraction by creating a pair of functions that share the free variable \code{x}. % similar to loop_test_10.rkt \begin{lstlisting} (define (f [x : Integer]) : (Vector ( -> Integer) ( -> Void)) (vector (lambda: () : Integer x) (lambda: () : Void (set! x (+ 1 x))))) (let ([counter (f 0)]) (let ([get (vector-ref counter 0)]) (let ([inc (vector-ref counter 1)]) (begin (inc) (get))))) \end{lstlisting} In this example, the lifetime of \code{x} extends beyond the lifetime of the call to \code{f}. Thus, if we were to store \code{x} on the stack frame for the call to \code{f}, it would be gone by the time we call \code{inc} and \code{get}, leaving us with dangling pointers for \code{x}. This example demonstrates that when a variable occurs free inside a \code{lambda}, its lifetime becomes indefinite. Thus, the value of the variable needs to live on the heap. The verb ``box'' is often used for allocating a single value on the heap, producing a pointer, and ``unbox'' for dereferencing the pointer. We recommend solving these problems by ``boxing'' the local variables that are in the intersection of 1) variables that appear on the left-hand-side of a \code{set!} and 2) variables that occur free inside a \code{lambda}. We shall introduce a new pass named \code{convert-assignments} in Section~\ref{sec:convert-assignments} to perform this translation. But before diving into the compiler passes, we one more problem to discuss. \section{Cyclic Control Flow and Dataflow Analysis} \label{sec:dataflow-analysis} Up until this point the control-flow graphs generated in \code{explicate-control} were guaranteed to be acyclic. However, each \code{while} loop introduces a cycle in the control-flow graph. But does that matter? % Indeed it does. Recall that for register allocation, the compiler performs liveness analysis to determine which variables can share the same register. In Section~\ref{sec:liveness-analysis-r2} we analyze the control-flow graph in reverse topological order, but topological order is only well-defined for acyclic graphs. Let us return to the example of computing the sum of the first five positive integers. Here is the program after instruction selection but before register allocation. \begin{center} \begin{minipage}{0.45\textwidth} \begin{lstlisting} (define (main) : Integer mainstart: movq $0, sum1 movq $5, i2 jmp block5 block5: movq i2, tmp3 cmpq tmp3, $0 jl block7 jmp block8 \end{lstlisting} \end{minipage} \begin{minipage}{0.45\textwidth} \begin{lstlisting} block7: addq i2, sum1 movq $1, tmp4 negq tmp4 addq tmp4, i2 jmp block5 block8: movq $27, %rax addq sum1, %rax jmp mainconclusion ) \end{lstlisting} \end{minipage} \end{center} Recall that liveness analysis works backwards, starting at the end of each function. For this example we could start with \code{block8} because we know what is live at the beginning of the conclusion, just \code{rax} and \code{rsp}. So the live-before set for \code{block8} is $\{\ttm{rsp},\ttm{sum1}\}$. % Next we might try to analyze \code{block5} or \code{block7}, but \code{block5} jumps to \code{block7} and vice versa, so it seems that we are stuck. The way out of this impasse comes from the realization that one can perform liveness analysis starting with an empty live-after set to compute an under-approximation of the live-before set. By \emph{under-approximation}, we mean that the set only contains variables that are really live, but it may be missing some. Next, the under-approximations for each block can be improved by 1) updating the live-after set for each block using the approximate live-before sets from the other blocks and 2) perform liveness analysis again on each block. In fact, by iterating this process, the under-approximations eventually become the correct solutions! % This approach of iteratively analyzing a control-flow graph is applicable to many static analysis problems and goes by the name \emph{dataflow analysis}\index{dataflow analysis}. It was invented by \citet{Kildall:1973vn} in his Ph.D. thesis at the University of Washington. Let us apply this approach to the above example. We use the empty set for the initial live-before set for each block. Let $m_0$ be the following mapping from label names to sets of locations (variables and registers). \begin{center} \begin{lstlisting} mainstart: {} block5: {} block7: {} block8: {} \end{lstlisting} \end{center} Using the above live-before approximations, we determine the live-after for each block and then apply liveness analysis to each block. This produces our next approximation $m_1$ of the live-before sets. \begin{center} \begin{lstlisting} mainstart: {} block5: {i2} block7: {i2, sum1} block8: {rsp, sum1} \end{lstlisting} \end{center} For the second round, the live-after for \code{mainstart} is the current live-before for \code{block5}, which is \code{\{i2\}}. So the liveness analysis for \code{mainstart} computes the empty set. The live-after for \code{block5} is the union of the live-before sets for \code{block7} and \code{block8}, which is \code{\{i2 , rsp, sum1\}}. So the liveness analysis for \code{block5} computes \code{\{i2 , rsp, sum1\}}. The live-after for \code{block7} is the live-before for \code{block5} (from the previous iteration), which is \code{\{i2\}}. So the liveness analysis for \code{block7} remains \code{\{i2, sum1\}}. Together these yield the following approximation $m_2$ of the live-before sets. \begin{center} \begin{lstlisting} mainstart: {} block5: {i2, rsp, sum1} block7: {i2, sum1} block8: {rsp, sum1} \end{lstlisting} \end{center} In the preceding iteration, only \code{block5} changed, so we can limit our attention to \code{mainstart} and \code{block7}, the two blocks that jump to \code{block5}. As a result, the live-before sets for \code{mainstart} and \code{block7} are updated to include \code{rsp}, yielding the following approximation $m_3$. \begin{center} \begin{lstlisting} mainstart: {rsp} block5: {i2, rsp, sum1} block7: {i2, rsp, sum1} block8: {rsp, sum1} \end{lstlisting} \end{center} Because \code{block7} changed, we analyze \code{block5} once more, but its live-before set remains \code{\{ i2, rsp, sum1 \}}. At this point our approximations have converged, so $m_3$ is the solution. This iteration process is guaranteed to converge to a solution by the Kleene Fixed-Point Theorem, a general theorem about functions on lattices~\citep{Kleene:1952aa}. Roughly speaking, a \emph{lattice} is any collection that comes with a partial ordering $\sqsubseteq$ on its elements, a least element $\bot$ (pronounced bottom), and a join operator $\sqcup$.\index{lattice}\index{bottom}\index{partial ordering}\index{join}\footnote{Technically speaking, we will be working with join semi-lattices.} When two elements are ordered $m_i \sqsubseteq m_j$, it means that $m_j$ contains at least as much information as $m_i$, so we can think of $m_j$ as a better-or-equal approximation than $m_i$. The bottom element $\bot$ represents the complete lack of information, i.e., the worst approximation. The join operator takes two lattice elements and combines their information, i.e., it produces the least upper bound of the two.\index{least upper bound} A dataflow analysis typically involves two lattices: one lattice to represent abstract states and another lattice that aggregates the abstract states of all the blocks in the control-flow graph. For liveness analysis, an abstract state is a set of locations. We form the lattice $L$ by taking its elements to be sets of locations, the ordering to be set inclusion ($\subseteq$), the bottom to be the empty set, and the join operator to be set union. % We form a second lattice $M$ by taking its elements to be mappings from the block labels to sets of locations (elements of $L$). We order the mappings point-wise, using the ordering of $L$. So given any two mappings $m_i$ and $m_j$, $m_i \sqsubseteq_M m_j$ when $m_i(\ell) \subseteq m_j(\ell)$ for every block label $\ell$ in the program. The bottom element of $M$ is the mapping $\bot_M$ that sends every label to the empty set, i.e., $\bot_M(\ell) = \emptyset$. We can think of one iteration of liveness analysis as being a function $f$ on the lattice $M$. It takes a mapping as input and computes a new mapping. \[ f(m_i) = m_{i+1} \] Next let us think for a moment about what a final solution $m_s$ should look like. If we perform liveness analysis using the solution $m_s$ as input, we should get $m_s$ again as the output. That is, the solution should be a \emph{fixed point} of the function $f$.\index{fixed point} \[ f(m_s) = m_s \] Furthermore, the solution should only include locations that are forced to be there by performing liveness analysis on the program, so the solution should be the \emph{least} fixed point.\index{least fixed point} The Kleene Fixed-Point Theorem states that if a function $f$ is monotone (better inputs produce better outputs), then the least fixed point of $f$ is the least upper bound of the \emph{ascending Kleene chain} obtained by starting at $\bot$ and iterating $f$ as follows.\index{Kleene Fixed-Point Theorem} \[ \bot \sqsubseteq f(\bot) \sqsubseteq f(f(\bot)) \sqsubseteq \cdots \sqsubseteq f^n(\bot) \sqsubseteq \cdots \] When a lattice contains only finitely-long ascending chains, then every Kleene chain tops out at some fixed point after a number of iterations of $f$. So that fixed point is also a least upper bound of the chain. \[ \bot \sqsubseteq f(\bot) \sqsubseteq f(f(\bot)) \sqsubseteq \cdots \sqsubseteq f^k(\bot) = f^{k+1}(\bot) = m_s \] The liveness analysis is indeed a monotone function and the lattice $M$ only has finitely-long ascending chains because there are only a finite number of variables and blocks in the program. Thus we are guaranteed that iteratively applying liveness analysis to all blocks in the program will eventually produce the least fixed point solution. Next let us consider dataflow analysis in general and discuss the generic work list algorithm (Figure~\ref{fig:generic-dataflow}). % The algorithm has four parameters: the control-flow graph \code{G}, a function \code{transfer} that applies the analysis to one block, the \code{bottom} and \code{join} operator for the lattice of abstract states. The algorithm begins by creating the bottom mapping, represented by a hash table. It then pushes all of the nodes in the control-flow graph onto the work list (a queue). The algorithm repeats the \code{while} loop as long as there are items in the work list. In each iteration, a node is popped from the work list and processed. The \code{input} for the node is computed by taking the join of the abstract states of all the predecessor nodes. The \code{transfer} function is then applied to obtain the \code{output} abstract state. If the output differs from the previous state for this block, the mapping for this block is updated and its successor nodes are pushed onto the work list. \begin{figure}[tb] \begin{lstlisting} (define (analyze-dataflow G transfer bottom join) (define mapping (make-hash)) (for ([v (in-vertices G)]) (dict-set! mapping v bottom)) (define worklist (make-queue)) (for ([v (in-vertices G)]) (enqueue! worklist v)) (define trans-G (transpose G)) (while (not (queue-empty? worklist)) (define node (dequeue! worklist)) (define input (for/fold ([state bottom]) ([pred (in-neighbors trans-G node)]) (join state (dict-ref mapping pred)))) (define output (transfer node input)) (cond [(not (equal? output (dict-ref mapping node))) (dict-set! mapping node output) (for ([v (in-neighbors G node)]) (enqueue! worklist v))])) mapping) \end{lstlisting} \caption{Generic work list algorithm for dataflow analysis} \label{fig:generic-dataflow} \end{figure} Having discussed the two complications that arise from adding support for assignment and loops, we turn to discussing the one new compiler pass and the significant changes to existing passes. \section{Convert Assignments} \label{sec:convert-assignments} Recall that in Section~\ref{sec:assignment-scoping} we learned that the combination of assignments and lexically-scoped functions requires that we box those variables that are both assigned-to and that appear free inside a \code{lambda}. The purpose of the \code{convert-assignments} pass is to carry out that transformation. We recommend placing this pass after \code{uniquify} but before \code{reveal-functions}. Consider again the first example from Section~\ref{sec:assignment-scoping}: \begin{lstlisting} (let ([x 0]) (let ([y 0]) (let ([z 20]) (let ([f (lambda: ([a : Integer]) : Integer (+ a (+ x z)))]) (begin (set! x 10) (set! y 12) (f y)))))) \end{lstlisting} The variables \code{x} and \code{y} are assigned-to. The variables \code{x} and \code{z} occur free inside the \code{lambda}. Thus, variable \code{x} needs to be boxed but not \code{y} and \code{z}. The boxing of \code{x} consists of three transformations: initialize \code{x} with a vector, replace reads from \code{x} with \code{vector-ref}'s, and replace each \code{set!} on \code{x} with a \code{vector-set!}. The output of \code{convert-assignments} for this example is as follows. \begin{lstlisting} (define (main) : Integer (let ([x0 (vector 0)]) (let ([y1 0]) (let ([z2 20]) (let ([f4 (lambda: ([a3 : Integer]) : Integer (+ a3 (+ (vector-ref x0 0) z2)))]) (begin (vector-set! x0 0 10) (set! y1 12) (f4 y1))))))) \end{lstlisting} \paragraph{Assigned \& Free} We recommend defining an auxiliary function named \code{assigned\&free} that takes an expression and simultaneously computes 1) a set of assigned variables $A$, 2) a set $F$ of variables that occur free within lambda's, and 3) a new version of the expression that records which bound variables occurred in the intersection of $A$ and $F$. You can use the struct \code{AssignedFree} to do this. Consider the case for $\LET{x}{\itm{rhs}}{\itm{body}}$. Suppose the the recursive call on $\itm{rhs}$ produces $\itm{rhs}'$, $A_r$, and $F_r$ and the recursive call on the $\itm{body}$ produces $\itm{body}'$, $A_b$, and $F_b$. If $x$ is in $A_b\cap F_b$, then transforms the \code{Let} as follows. \begin{lstlisting} (Let |$x$| |$rhs$| |$body$|) |$\Rightarrow$| (Let (AssignedFree |$x$|) |$rhs'$| |$body'$|) \end{lstlisting} If $x$ is not in $A_b\cap F_b$ then omit the use of \code{AssignedFree}. The set of assigned variables for this \code{Let} is $A_r \cup (A_b - \{x\})$ and the set of variables free in lambda's is $F_r \cup (F_b - \{x\})$. The case for $\SETBANG{x}{\itm{rhs}}$ is straightforward but important. Recursively process \itm{rhs} to obtain \itm{rhs'}, $A_r$, and $F_r$. The result is $\SETBANG{x}{\itm{rhs'}}$, $\{x\} \cup A_r$, and $F_r$. The case for $\LAMBDA{\itm{params}}{T}{\itm{body}}$ is a bit more involved. Let \itm{body'}, $A_b$, and $F_b$ be the result of recursively processing \itm{body}. Wrap each of parameter that occurs in $A_b \cap F_b$ with \code{AssignedFree} to produce \itm{params'}. Let $P$ be the set of parameter names in \itm{params}. The result is $\LAMBDA{\itm{params'}}{T}{\itm{body'}}$, $A_b - P$, and $(F_b \cup \mathrm{FV}(\itm{body})) - P$, where $\mathrm{FV}$ computes the free variables of an expression (see Chapter~\ref{ch:lambdas}). \paragraph{Convert Assignments} Next we discuss the \code{convert-assignment} pass with its auxiliary functions for expressions and definitions. The function for expressions, \code{cnvt-assign-exp}, should take an expression and a set of assigned-and-free variables (obtained from the result of \code{assigned\&free}. In the case for $\VAR{x}$, if $x$ is assigned-and-free, then unbox it by translating $\VAR{x}$ to a \code{vector-ref}. \begin{lstlisting} (Var |$x$|) |$\Rightarrow$| (Prim 'vector-ref (list (Var |$x$|) (Int 0))) \end{lstlisting} % In the case for $\LET{\LP\code{AssignedFree}\, x\RP}{\itm{rhs}}{\itm{body}}$, recursively process \itm{rhs} to obtain \itm{rhs'}. Next, recursively process \itm{body} to obtain \itm{body'} but with $x$ added to the set of assigned-and-free variables. Translate the let-expression as follows to bind $x$ to a boxed value. \begin{lstlisting} (Let (AssignedFree |$x$|) |$rhs$| |$body$|) |$\Rightarrow$| (Let |$x$| (Prim 'vector (list |$rhs'$|)) |$body'$|) \end{lstlisting} % In the case for $\SETBANG{x}{\itm{rhs}}$, recursively process \itm{rhs} to obtain \itm{rhs'}. If $x$ is in the assigned-and-free variables, translate the \code{set!} into a \code{vector-set!} as follows. \begin{lstlisting} (SetBang |$x$| |$\itm{rhs}$|) |$\Rightarrow$| (Prim 'vector-set! (list (Var |$x$|) (Int 0) |$\itm{rhs'}$|)) \end{lstlisting} % The case for \code{Lambda} is non-trivial, but it is similar to the case for function definitions, which we discuss next. The auxiliary function for definitions, \code{cnvt-assign-def}, applies assignment conversion to function definitions. We translate a function definition as follows. \begin{lstlisting} (Def |$f$| |$\itm{params}$| |$T$| |$\itm{info}$| |$\itm{body_1}$|) |$\Rightarrow$| (Def |$f$| |$\itm{params'}$| |$T$| |$\itm{info}$| |$\itm{body_4}$|) \end{lstlisting} So it remains to explain \itm{params'} and $\itm{body}_4$. Let \itm{body_2}, $A_b$, and $F_b$ be the result of \code{assigned\&free} on $\itm{body_1}$. Let $P$ be the parameter names in \itm{params}. We then apply \code{cnvt-assign-exp} to $\itm{body_2}$ to obtain \itm{body_3}, passing $A_b \cap F_b \cap P$ as the set of assigned-and-free variables. Finally, we obtain \itm{body_4} by wrapping \itm{body_3} in a sequence of let-expressions that box the parameters that are in $A_b \cap F_b$. % Regarding \itm{params'}, change the names of the parameters that are in $A_b \cap F_b$ to maintain uniqueness (and so the let-bound variables can retain the original names). Recall the second example in Section~\ref{sec:assignment-scoping} involving a counter abstraction. The following is the output of assignment version for function \code{f}. \begin{lstlisting} (define (f0 [x1 : Integer]) : (Vector ( -> Integer) ( -> Void)) (vector (lambda: () : Integer x1) (lambda: () : Void (set! x1 (+ 1 x1))))) |$\Rightarrow$| (define (f0 [param_x1 : Integer]) : (Vector (-> Integer) (-> Void)) (let ([x1 (vector param_x1)]) (vector (lambda: () : Integer (vector-ref x1 0)) (lambda: () : Void (vector-set! x1 0 (+ 1 (vector-ref x1 0))))))) \end{lstlisting} \section{Remove Complex Operands} \label{sec:rco-loop} The three new language forms, \code{while}, \code{set!}, and \code{begin} are all complex expressions and their subexpressions are allowed to be complex. Figure~\ref{fig:r4-anf-syntax} defines the output language $R_4^{\dagger}$ of this pass. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{rcl} \Atm &::=& \gray{ \INT{\Int} \mid \VAR{\Var} \mid \BOOL{\itm{bool}} \mid \VOID{} } \\ \Exp &::=& \ldots \mid \gray{ \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \WHILE{\Exp}{\Exp} \mid \SETBANG{\Var}{\Exp} \mid \BEGIN{\LP\Exp\ldots\RP}{\Exp} \\ \Def &::=& \gray{ \FUNDEF{\Var}{([\Var \code{:} \Type]\ldots)}{\Type}{\code{'()}}{\Exp} }\\ R^{\dagger}_8 &::=& \gray{ \PROGRAMDEFS{\code{'()}}{\Def} } \end{array} \] \end{minipage} } \caption{$R_8^{\dagger}$ is $R_8$ in administrative normal form (ANF).} \label{fig:r8-anf-syntax} \end{figure} As usual, when a complex expression appears in a grammar position that needs to be atomic, such as the argument of a primitive operator, we must introduce a temporary variable and bind it to the complex expression. This approach applies, unchanged, to handle the new language forms. For example, in the following code there are two \code{begin} expressions appearing as arguments to \code{+}. The output of \code{rco-exp} is shown below, in which the \code{begin} expressions have been bound to temporary variables. Recall that \code{let} expressions in $R_8^{\dagger}$ are allowed to have arbitrary expressions in their right-hand-side expression, so it is fine to place \code{begin} there. \begin{lstlisting} (let ([x0 10]) (let ([y1 0]) (+ (+ (begin (set! y1 (read)) x0) (begin (set! x0 (read)) y1)) x0))) |$\Rightarrow$| (let ([x0 10]) (let ([y1 0]) (let ([tmp2 (begin (set! y1 (read)) x0)]) (let ([tmp3 (begin (set! x0 (read)) y1)]) (let ([tmp4 (+ tmp2 tmp3)]) (+ tmp4 x0)))))) \end{lstlisting} \section{Explicate Control and $C_7$} \label{sec:explicate-loop} Recall that in the \code{explicate-control} pass we define one helper function for each kind of position in the program. For the $R_1$ language of integers and variables we needed kinds of positions: assignment and tail. The \code{if} expressions of $R_2$ introduced predicate positions. For $R_8$, the \code{begin} expression introduces yet another kind of position: effect position. Except for the last subexpression, the subexpressions inside a \code{begin} are evaluated only for their effect. Their result values are discarded. We can generate better code by taking this fact into account. The output language of \code{explicate-control} is $C_7$ (Figure~\ref{fig:c7-syntax}), which is nearly identical to $C_4$. The only difference is that \code{Call}, \code{vector-set!}, and \code{read} may also appear as statements. \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Stmt &::=& \gray{ \ASSIGN{\VAR{\Var}}{\Exp} \mid \LP\key{Collect} \,\itm{int}\RP } \\ &\mid& \CALL{\Atm}{\LP\Atm\ldots\RP} \mid \READ{}\\ &\mid& \LP\key{Prim}~\key{'vector-set!}\,\LP\key{list}\,\Atm\,\INT{\Int}\,\Atm\RP\RP \\ \Def &::=& \DEF{\itm{label}}{\LP\LS\Var\key{:}\Type\RS\ldots\RP}{\Type}{\itm{info}}{\LP\LP\itm{label}\,\key{.}\,\Tail\RP\ldots\RP}\\ C_7 & ::= & \PROGRAMDEFS{\itm{info}}{\LP\Def\ldots\RP} \end{array} \] \end{minipage} } \caption{The abstract syntax of $C_7$, extending $C_4$ (Figure~\ref{fig:c4-syntax}).} \label{fig:c7-syntax} \end{figure} The new auxiliary function \code{explicate-effect} takes an expression (in an effect position) and a promise of a continuation block. The function returns a promise for a $\Tail$ that includes the generated code for the input expression followed by the continuation block. If the expression is obviously pure, that is, never causes side effects, then the expression can be removed, so the result is just the continuation block. % The $\WHILE{\itm{cnd}}{\itm{body}}$ expression is the most interesting case. First, you will need a fresh label $\itm{loop}$ for the top of the loop. Recursively process the \itm{body} (in effect position) with the a \code{goto} to $\itm{loop}$ as the continuation, producing \itm{body'}. Next, process the \itm{cnd} (in predicate position) with \itm{body'} as the then-branch and the continuation block as the else-branch. The result should be added to the control-flow graph with the label \itm{loop}. The result for the whole \code{while} loop is a \code{goto} to the \itm{loop} label. Note that the loop should only be added to the control-flow graph if the loop is indeed used, which can be accomplished using \code{delay}. The auxiliary functions for tail, assignment, and predicate positions need to be updated. The three new language forms, \code{while}, \code{set!}, and \code{begin}, can appear in assignment and tail positions. Only \code{begin} may appear in predicate positions; the other two have result type \code{Void}. \section{Select Instructions} \label{sec:select-instructions-loop} Only three small additions are needed in the \code{select-instructions} pass to handle the changes to $C_7$. That is, \code{Call}, \code{read}, and \code{vector-set!} may now appear as stand-alone statements instead of only appearing on the right-hand side of an assignment statement. The code generation is nearly identical; just leave off the instruction for moving the result into the left-hand side. \section{Register Allocation} \label{sec:register-allocation-loop} As discussed in Section~\ref{sec:dataflow-analysis}, the presence of loops in $R_8$ means that the control-flow graphs may contain cycles, which complicates the liveness analysis needed for register allocation. \subsection{Liveness Analysis} \label{sec:liveness-analysis-r8} We recommend using the generic \code{analyze-dataflow} function that was presented at the end of Section~\ref{sec:dataflow-analysis} to perform liveness analysis, replacing the code in \code{uncover-live-CFG} that processed the basic blocks in topological order (Section~\ref{sec:liveness-analysis-r2}). The \code{analyze-dataflow} function has four parameters. \begin{enumerate} \item The first parameter \code{G} should be a directed graph from the \code{racket/graph} package (see the sidebar in Section~\ref{sec:build-interference}) that represents the control-flow graph. \item The second parameter \code{transfer} is a function that applies liveness analysis to a basic block. It takes two parameters: the label for the block to analyze and the live-after set for that block. The transfer function should return the live-before set for the block. Also, as a side-effect, it should update the block's $\itm{info}$ with the liveness information for each instruction. To implement the \code{transfer} function, you should be able to reuse the code you already have for analyzing basic blocks. \item The third and fourth parameters of \code{analyze-dataflow} are \code{bottom} and \code{join} for the lattice of abstract states, i.e. sets of locations. The bottom of the lattice is the empty set \code{(set)} and the join operator is \code{set-union}. \end{enumerate} \begin{figure}[p] \begin{tikzpicture}[baseline=(current bounding box.center)] \node (R4) at (0,2) {\large $R_8$}; \node (R4-2) at (3,2) {\large $R_8$}; \node (R4-3) at (6,2) {\large $R_8$}; \node (R4-4) at (9,2) {\large $R'_8$}; \node (F1-1) at (12,0) {\large $R'_8$}; \node (F1-2) at (9,0) {\large $R'_8$}; \node (F1-3) at (6,0) {\large $R'_8$}; \node (F1-4) at (3,0) {\large $R'_8$}; \node (F1-5) at (0,0) {\large $R'_8$}; \node (C3-2) at (3,-2) {\large $C_3$}; \node (x86-2) at (3,-4) {\large $\text{x86}^{*}_3$}; \node (x86-3) at (6,-4) {\large $\text{x86}^{*}_3$}; \node (x86-4) at (9,-4) {\large $\text{x86}^{*}_3$}; \node (x86-5) at (9,-6) {\large $\text{x86}^{\dagger}_3$}; \node (x86-2-1) at (3,-6) {\large $\text{x86}^{*}_3$}; \node (x86-2-2) at (6,-6) {\large $\text{x86}^{*}_3$}; %% \path[->,bend left=15] (R4) edge [above] node %% {\ttfamily\footnotesize type-check} (R4-2); \path[->,bend left=15] (R4) edge [above] node {\ttfamily\footnotesize shrink} (R4-2); \path[->,bend left=15] (R4-2) edge [above] node {\ttfamily\footnotesize uniquify} (R4-3); \path[->,bend left=15] (R4-3) edge [above] node {\ttfamily\footnotesize reveal-functions} (R4-4); \path[->,bend left=15] (R4-4) edge [right] node {\ttfamily\footnotesize convert-assignments} (F1-1); \path[->,bend left=15] (F1-1) edge [below] node {\ttfamily\footnotesize convert-to-clos.} (F1-2); \path[->,bend right=15] (F1-2) edge [above] node {\ttfamily\footnotesize limit-fun.} (F1-3); \path[->,bend right=15] (F1-3) edge [above] node {\ttfamily\footnotesize expose-alloc.} (F1-4); \path[->,bend right=15] (F1-4) edge [above] node {\ttfamily\footnotesize remove-complex.} (F1-5); \path[->,bend right=15] (F1-5) edge [right] node {\ttfamily\footnotesize explicate-control} (C3-2); \path[->,bend left=15] (C3-2) edge [left] node {\ttfamily\footnotesize select-instr.} (x86-2); \path[->,bend right=15] (x86-2) edge [left] node {\ttfamily\footnotesize uncover-live} (x86-2-1); \path[->,bend right=15] (x86-2-1) edge [below] node {\ttfamily\footnotesize build-inter.} (x86-2-2); \path[->,bend right=15] (x86-2-2) edge [left] node {\ttfamily\footnotesize allocate-reg.} (x86-3); \path[->,bend left=15] (x86-3) edge [above] node {\ttfamily\footnotesize patch-instr.} (x86-4); \path[->,bend left=15] (x86-4) edge [right] node {\ttfamily\footnotesize print-x86} (x86-5); \end{tikzpicture} \caption{Diagram of the passes for $R_8$ (loops and assignment).} \label{fig:R8-passes} \end{figure} Figure~\ref{fig:R8-passes} provides an overview of all the passes needed for the compilation of $R_8$. % TODO: challenge assignment %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Gradual Typing} \label{ch:gradual-typing} \index{gradual typing} This chapter studies a language, $R_9$, in which the programmer can choose between static and dynamic type checking for different regions of a program, thereby mixing the statically typed $R_8$ language with the dynamically typed $R_7$. There are several approaches to mixing static and dynamic typing, including multi-language integration~\citep{Tobin-Hochstadt:2006fk,Matthews:2007zr} and hybrid type checking~\citep{Flanagan:2006mn,Gronski:2006uq}. In this chapter we focus on \emph{gradual typing}\index{gradual typing}, in which the programmer controls the amount of static versus dynamic checking by adding or removing type annotations on parameters and variables~\citep{Anderson:2002kd,Siek:2006bh}. % The concrete syntax of $R_9$ is defined in Figure~\ref{fig:r9-concrete-syntax} and its abstract syntax is defined in Figure~\ref{fig:r9-syntax}. The main syntactic difference between $R_8$ and $R_9$ is the additional \itm{param} and \itm{ret} non-terminals that make type annotations optional. The return types are not optional in the abstract syntax; the parser fills in \code{Any} when the return type is not specified in the concrete syntax. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \itm{param} &::=& \Var \mid \LS\Var \key{:} \Type\RS \\ \itm{ret} &::=& \epsilon \mid \key{:} \Type \\ \Exp &::=& \gray{ \Int \mid \CREAD{} \mid \CNEG{\Exp} \mid \CADD{\Exp}{\Exp} \mid \CSUB{\Exp}{\Exp} } \\ &\mid& \gray{ \Var \mid \CLET{\Var}{\Exp}{\Exp} }\\ &\mid& \gray{\key{\#t} \mid \key{\#f} \mid (\key{and}\;\Exp\;\Exp) \mid (\key{or}\;\Exp\;\Exp) \mid (\key{not}\;\Exp) } \\ &\mid& \gray{ (\key{eq?}\;\Exp\;\Exp) \mid \CIF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ (\key{vector}\;\Exp\ldots) \mid (\key{vector-ref}\;\Exp\;\Int)} \\ &\mid& \gray{(\key{vector-set!}\;\Exp\;\Int\;\Exp)\mid (\key{void}) \mid (\Exp \; \Exp\ldots) } \\ &\mid& \gray{ \LP \key{procedure-arity}~\Exp\RP } \mid \CGLAMBDA{\LP\itm{param}\ldots\RP}{\itm{ret}}{\Exp} \\ &\mid& \gray{ \CSETBANG{\Var}{\Exp} \mid \CBEGIN{\Exp\ldots}{\Exp} \mid \CWHILE{\Exp}{\Exp} } \\ \Def &::=& \CGDEF{\Var}{\itm{param}\ldots}{\itm{ret}}{\Exp} \\ R_9 &::=& \gray{\Def\ldots \; \Exp} \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_9$, extending $R_8$ (Figure~\ref{fig:r8-concrete-syntax}).} \label{fig:r9-concrete-syntax} \end{figure} \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \itm{param} &::=& \Var \mid \LS\Var \key{:} \Type\RS \\ \Exp &::=& \gray{ \INT{\Int} \VAR{\Var} \mid \LET{\Var}{\Exp}{\Exp} } \\ &\mid& \gray{ \PRIM{\itm{op}}{\Exp\ldots} }\\ &\mid& \gray{ \BOOL{\itm{bool}} \mid \IF{\Exp}{\Exp}{\Exp} } \\ &\mid& \gray{ \VOID{} \mid \LP\key{HasType}~\Exp~\Type \RP \mid \APPLY{\Exp}{\Exp\ldots} }\\ &\mid& \LAMBDA{\LP\itm{param}\ldots\RP}{\Type}{\Exp} \\ &\mid& \gray{ \SETBANG{\Var}{\Exp} \mid \BEGIN{\LP\Exp\ldots\RP}{\Exp} } \\ &\mid& \gray{ \WHILE{\Exp}{\Exp} } \\ \Def &::=& \FUNDEF{\Var}{\LP\itm{param}\ldots\RP}{\Type}{\code{'()}}{\Exp} \\ R_9 &::=& \gray{ \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP}{\Exp} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $R_9$, extending $R_8$ (Figure~\ref{fig:r8-syntax}).} \label{fig:r9-syntax} \end{figure} Both the type checker and the interpreter for $R_9$ require some interesting changes to enable gradual typing, which we discuss in the next two sections while revisiting the \code{map-vec} example from Chapter~\ref{ch:functions}. In Figure~\ref{fig:gradual-map-vec} we present the example again but this time we leave off the type annotations from the \code{add1} function. \begin{figure}[btp] % gradual_test_9.rkt \begin{lstlisting} (define (map-vec [f : (Integer -> Integer)] [v : (Vector Integer Integer)]) : (Vector Integer Integer) (vector (f (vector-ref v 0)) (f (vector-ref v 1)))) (define (add1 x) (+ x 1)) (vector-ref (map-vec add1 (vector 0 41)) 1) \end{lstlisting} \caption{A partially-typed version of the \code{map-vec} example.} \label{fig:gradual-map-vec} \end{figure} \section{Type Checking $R_9$, Casts, and $R'_9$} \label{sec:gradual-type-check} The type checker for $R_9$ uses the \code{Any} type for missing parameter and return types. For example, the \code{x} parameter of \code{add1} in Figure~\ref{fig:gradual-map-vec} is given the type \code{Any} and the return type of \code{add1} is \code{Any}. Next consider the \code{+} operator inside \code{add1}. It expects both arguments to have type \code{Integer}, but its first argument \code{x} has type \code{Any}. In a gradually typed language, such differences are allowed so long as the types are \emph{consistent}, that is, they are equal except in places where there is an \code{Any} type. The type \code{Any} is consistent with every other type. Figure~\ref{fig:consistent} defines the \code{consistent?} predicate. \begin{figure}[tbp] \begin{lstlisting} (define/public (consistent? t1 t2) (match* (t1 t2) [('Integer 'Integer) #t] [('Boolean 'Boolean) #t] [('Void 'Void) #t] [('Any t2) #t] [(t1 'Any) #t] [(`(Vector ,ts1 ...) `(Vector ,ts2 ...)) (for/and ([t1 ts1] [t2 ts2]) (consistent? t1 t2))] [(`(,ts1 ... -> ,rt1) `(,ts2 ... -> ,rt2)) (and (for/and ([t1 ts1] [t2 ts2]) (consistent? t1 t2)) (consistent? rt1 rt2))] [(other wise) #f])) \end{lstlisting} \caption{The consistency predicate on types, a method in \code{type-check-gradual-class}.} \label{fig:consistent} \end{figure} Returning to the \code{map-vec} example of Figure~\ref{fig:gradual-map-vec}, the \code{add1} function has type \code{(Any -> Any)} but parameter \code{f} of \code{map-vec} has type \code{(Integer -> Integer)}. The type checker for $R_9$ allows this because the two types are consistent. In particular, \code{->} is equal to \code{->} and because \code{Any} is consistent with \code{Integer}. Next consider a program with an error, such as applying the \code{map-vec} to a function that sometimes returns a Boolean, as shown in Figure~\ref{fig:map-vec-maybe-add1}. The type checker for $R_9$ accepts this program because the type of \code{maybe-add1} is consistent with the type of parameter \code{f} of \code{map-vec}, that is, \code{(Any -> Any)} is consistent with \code{(Integer -> Integer)}. One might say that a gradual type checker is optimistic in that it accepts programs that might execute without a runtime type error. % Unfortunately, running this program with input \code{1} triggers an error when the \code{maybe-add1} function returns \code{\#t}. $R_9$ performs checking at runtime to ensure the integrity of the static types, such as the \code{(Integer -> Integer)} annotation on parameter \code{f} of \code{map-vec}. This runtime checking is carried out by a new \code{Cast} form that is inserted by the type checker. Thus, the output of the type checker is a program in the $R'_9$ language, which adds \code{Cast} to $R_8$, as shown in Figure~\ref{fig:r9-prime-syntax}. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Exp &::=& \ldots \mid \CAST{\Exp}{\Type}{\Type} \\ R'_9 &::=& \gray{ \PROGRAMDEFSEXP{\code{'()}}{\LP\Def\ldots\RP}{\Exp} } \end{array} \] \end{minipage} } \caption{The abstract syntax of $R'_9$, extending $R_8$ (Figure~\ref{fig:r8-syntax}).} \label{fig:r9-prime-syntax} \end{figure} \begin{figure}[tbp] \begin{lstlisting} (define (map-vec [f : (Integer -> Integer)] [v : (Vector Integer Integer)]) : (Vector Integer Integer) (vector (f (vector-ref v 0)) (f (vector-ref v 1)))) (define (add1 x) (+ x 1)) (define (true) #t) (define (maybe-add1 x) (if (eq? 0 (read)) (add1 x) (true))) (vector-ref (map-vec maybe-add1 (vector 0 41)) 0) \end{lstlisting} \caption{A variant of the \code{map-vec} example with an error.} \label{fig:map-vec-maybe-add1} \end{figure} Figure~\ref{fig:map-vec-cast} shows the output of the type checker for \code{map-vec} and \code{maybe-add1}. The idea is that \code{Cast} is inserted every time the type checker sees two types that are consistent but not equal. In the \code{add1} function, \code{x} is cast to \code{Integer} and the result of the \code{+} is cast to \code{Any}. In the call to \code{map-vec}, the \code{add1} argument is cast from \code{(Any -> Any)} to \code{(Integer -> Integer)}. \begin{figure}[btp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define (map-vec [f : (Integer -> Integer)] [v : (Vector Integer Integer)]) : (Vector Integer Integer) (vector (f (vector-ref v 0)) (f (vector-ref v 1)))) (define (add1 [x : Any]) : Any (cast (+ (cast x Any Integer) 1) Integer Any)) (define (true) : Any (cast #t Boolean Any)) (define (maybe-add1 [x : Any]) : Any (if (eq? 0 (read)) (add1 x) (true))) (vector-ref (map-vec (cast maybe-add1 (Any -> Any) (Integer -> Integer)) (vector 0 41)) 0) \end{lstlisting} \caption{Output of type checking \code{map-vec} and \code{maybe-add1}.} \label{fig:map-vec-cast} \end{figure} The type checker for $R_9$ is defined in Figures~\ref{fig:type-check-R9-1}, \ref{fig:type-check-R9-2}, and \ref{fig:type-check-R9-3}. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] (define type-check-gradual-class (class type-check-R8-class (super-new) (inherit operator-types type-predicates) (define/override (type-check-exp env) (lambda (e) (define recur (type-check-exp env)) (match e [(Prim 'vector-length (list e1)) (define-values (e1^ t) (recur e1)) (match t [`(Vector ,ts ...) (values (Prim 'vector-length (list e1^)) 'Integer)] ['Any (values (Prim 'any-vector-length (list e1^)) 'Integer)])] [(Prim 'vector-ref (list e1 e2)) (define-values (e1^ t1) (recur e1)) (define-values (e2^ t2) (recur e2)) (check-consistent? t2 'Integer e) (match t1 [`(Vector ,ts ...) (match e2^ [(Int i) (unless (and (0 . <= . i) (i . < . (length ts))) (error 'type-check "invalid index ~a in ~a" i e)) (values (Prim 'vector-ref (list e1^ (Int i))) (list-ref ts i))] [else (define e1^^ (make-cast e1^ t1 'Any)) (define e2^^ (make-cast e2^ t2 'Integer)) (values (Prim 'any-vector-ref (list e1^^ e2^^)) 'Any)])] ['Any (define e2^^ (make-cast e2^ t2 'Integer)) (values (Prim 'any-vector-ref (list e1^ e2^^)) 'Any)] [else (error 'type-check "expected vector not ~a\nin ~v" t1 e)])] [(Prim 'vector-set! (list e1 e2 e3) ) (define-values (e1^ t1) (recur e1)) (define-values (e2^ t2) (recur e2)) (define-values (e3^ t3) (recur e3)) (check-consistent? t2 'Integer e) (match t1 [`(Vector ,ts ...) (match e2^ [(Int i) (unless (and (0 . <= . i) (i . < . (length ts))) (error 'type-check "invalid index ~a in ~a" i e)) (check-consistent? (list-ref ts i) t3 e) (define e3^^ (make-cast e3^ t3 (list-ref ts i))) (values (Prim 'vector-set! (list e1^ (Int i) e3^^)) 'Void)] [else (define e1^^ (make-cast e1^ t1 'Any)) (define e2^^ (make-cast e2^ t2 'Integer)) (define e3^^ (make-cast e3^ t3 'Any)) (values (Prim 'any-vector-set! (list e1^^ e2^^ e3^^)) 'Void)])] ['Any (define e2^^ (make-cast e2^ t2 'Integer)) (define e3^^ (make-cast e3^ t3 'Any)) (values (Prim 'any-vector-set! (list e1^ e2^^ e3^^)) 'Void)] [else (error 'type-check "expected vector not ~a\nin ~v" t1 e)])] \end{lstlisting} \caption{Type checker for the $R_9$ language, part 1.} \label{fig:type-check-R9-1} \end{figure} \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] [(Prim 'eq? (list e1 e2)) (define-values (e1^ t1) (recur e1)) (define-values (e2^ t2) (recur e2)) (check-consistent? t1 t2 e) (define T (meet t1 t2)) (values (Prim 'eq? (list (make-cast e1^ t1 T) (make-cast e2^ t2 T))) 'Boolean)] [(Prim 'not (list e1)) (recur (If (Prim 'eq? (list e1 (Inject (Bool #f) 'Boolean))) (Bool #t) (Bool #f)))] [(Prim 'and (list e1 e2)) (recur (If e1 e2 (Bool #f)))] [(Prim 'or (list e1 e2)) (define tmp (gensym 'tmp)) (recur (Let tmp e1 (If (Var tmp) (Var tmp) e2)))] [(Prim op es) #:when (not (set-member? explicit-prim-ops op)) (define-values (new-es ts) (for/lists (exprs types) ([e es]) (recur e))) (define-values (t-ret new-es^) (type-check-op op ts new-es e)) (values (Prim op new-es^) t-ret)] [(If e1 e2 e3) (define-values (e1^ T1) (recur e1)) (define-values (e2^ T2) (recur e2)) (define-values (e3^ T3) (recur e3)) (check-consistent? T2 T3 e) (match T1 ['Boolean (define Tif (join T2 T3)) (values (If e1^ (make-cast e2^ T2 Tif) (make-cast e3^ T3 Tif)) Tif)] ['Any (define Tif (meet T2 T3)) (values (If (Prim 'eq? (list e1^ (Inject (Bool #f) 'Boolean))) (make-cast e3^ T3 Tif) (make-cast e2^ T2 Tif)) Tif)] [else (error 'type-check "expected Boolean not ~a\nin ~v" T1 e)])] [(HasType e1 T) (define-values (e1^ T1) (recur e1)) (check-consistent? T1 T) (values (make-cast e1^ T1 T) T)] [(SetBang x e1) (define-values (e1^ T1) (recur e1)) (define varT (dict-ref env x)) (check-consistent? T1 varT e) (values (SetBang x (make-cast e1^ T1 varT)) 'Void)] [(WhileLoop e1 e2) (define-values (e1^ T1) (recur e1)) (check-consistent? T1 'Boolean e) (define-values (e2^ T2) ((type-check-exp env) e2)) (values (WhileLoop (make-cast e1^ T1 'Boolean) e2^) 'Void)] \end{lstlisting} \caption{Type checker for the $R_9$ language, part 2.} \label{fig:type-check-R9-2} \end{figure} \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] [(Apply e1 e2s) (define-values (e1^ T1) (recur e1)) (define-values (e2s^ T2s) (for/lists (e* ty*) ([e2 e2s]) (recur e2))) (match T1 [`(,T1ps ... -> ,T1rt) (for ([T2 T2s] [Tp T1ps]) (check-consistent? T2 Tp e)) (define e2s^^ (for/list ([e2 e2s^] [src T2s] [tgt T1ps]) (make-cast e2 src tgt))) (values (Apply e1^ e2s^^) T1rt)] [`Any (define e1^^ (make-cast e1^ 'Any `(,@(for/list ([e e2s]) 'Any) -> Any))) (define e2s^^ (for/list ([e2 e2s^] [src T2s]) (make-cast e2 src 'Any))) (values (Apply e1^^ e2s^^) 'Any)] [else (error 'type-check "expected function not ~a\nin ~v" T1 e)])] [(Lambda params Tr e1) (define-values (xs Ts) (for/lists (l1 l2) ([p params]) (match p [`[,x : ,T] (values x T)] [(? symbol? x) (values x 'Any)]))) (define-values (e1^ T1) ((type-check-exp (append (map cons xs Ts) env)) e1)) (check-consistent? Tr T1 e) (values (Lambda (for/list ([x xs] [T Ts]) `[,x : ,T]) Tr (make-cast e1^ T1 Tr)) `(,@Ts -> ,Tr))] [else ((super type-check-exp env) e)] ))) \end{lstlisting} \caption{Type checker for the $R_9$ language, part 3.} \label{fig:type-check-R9-3} \end{figure} \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\scriptsize] (define/public (join t1 t2) (match* (t1 t2) [('Integer 'Integer) 'Integer] [('Boolean 'Boolean) 'Boolean] [('Void 'Void) 'Void] [('Any t2) t2] [(t1 'Any) t1] [(`(Vector ,ts1 ...) `(Vector ,ts2 ...)) `(Vector ,@(for/list ([t1 ts1] [t2 ts2]) (join t1 t2)))] [(`(,ts1 ... -> ,rt1) `(,ts2 ... -> ,rt2)) `(,@(for/list ([t1 ts1] [t2 ts2]) (join t1 t2)) -> ,(join rt1 rt2))])) (define/public (meet t1 t2) (match* (t1 t2) [('Integer 'Integer) 'Integer] [('Boolean 'Boolean) 'Boolean] [('Void 'Void) 'Void] [('Any t2) 'Any] [(t1 'Any) 'Any] [(`(Vector ,ts1 ...) `(Vector ,ts2 ...)) `(Vector ,@(for/list ([t1 ts1] [t2 ts2]) (meet t1 t2)))] [(`(,ts1 ... -> ,rt1) `(,ts2 ... -> ,rt2)) `(,@(for/list ([t1 ts1] [t2 ts2]) (meet t1 t2)) -> ,(meet rt1 rt2))])) (define/public (make-cast e src tgt) (cond [(equal? src tgt) e] [else (Cast e src tgt)])) (define/public (check-consistent? t1 t2 e) (unless (consistent? t1 t2) (error 'type-check "~a is inconsistent with ~a\nin ~v" t1 t2 e))) (define/override (type-check-op op arg-types args e) (match (dict-ref (operator-types) op) [`(,param-types . ,return-type) (for ([at arg-types] [pt param-types]) (check-consistent? at pt e)) (values return-type (for/list ([e args] [s arg-types] [t param-types]) (make-cast e s t)))] [else (error 'type-check-op "unrecognized ~a" op)])) (define explicit-prim-ops (set-union (type-predicates) (set 'procedure-arity 'eq? 'vector 'vector-length 'vector-ref 'vector-set! 'any-vector-length 'any-vector-ref 'any-vector-set!))) (define/override (fun-def-type d) (match d [(Def f params rt info body) (define ps (for/list ([p params]) (match p [`[,x : ,T] T] [(? symbol?) 'Any] [else (error 'fun-def-type "unmatched parameter ~a" p)]))) `(,@ps -> ,rt)] [else (error 'fun-def-type "ill-formed function definition in ~a" d)])) \end{lstlisting} \caption{Auxiliary functions for type checking $R_9$.} \label{fig:type-check-R9-aux} \end{figure} \clearpage \section{Interpreting $R'_9$} \label{sec:interp-casts} The runtime behavior of first-order casts is straightforward, that is, casts involving simple types such as \code{Integer} and \code{Boolean}. For example, a cast from \code{Integer} to \code{Any} can be accomplished with the \code{Inject} operator of $R_6$, which puts the integer into a tagged value (Figure~\ref{fig:interp-R6}). Similarly, a cast from \code{Any} to \code{Integer} is accomplished with the \code{Project} operator, that is, by checking the value's tag and either retrieving the underlying integer or signaling an error if it the tag is not the one for integers (Figure~\ref{fig:apply-project}). % Things get more interesting for higher-order casts, that is, casts involving function or vector types. Consider the cast of the function \code{maybe-add1} from \code{(Any -> Any)} to \code{(Integer -> Integer)}. When a function flows through this cast at runtime, we can't know in general whether the function will always return an integer.\footnote{Predicting the return value of a function is equivalent to the halting problem, which is undecidable.} The $R'_9$ interpreter therefore delays the checking of the cast until the function is applied. This is accomplished by wrapping \code{maybe-add1} in a new function that casts its parameter from \code{Integer} to \code{Any}, applies \code{maybe-add1}, and then casts the return value from \code{Any} to \code{Integer}. Turning our attention to casts involving vector types, we consider the example in Figure~\ref{fig:map-vec-bang} that defines a partially-typed version of \code{map-vec} whose parameter \code{v} has type \code{(Vector Any Any)} and that updates \code{v} in place instead of returning a new vector. So we name this function \code{map-vec!}. We apply \code{map-vec!} to a vector of integers, so the type checker inserts a cast from \code{(Vector Integer Integer)} to \code{(Vector Any Any)}. A naive way for the $R'_9$ interpreter to cast between vector types would be a build a new vector whose elements are the result of casting each of the original elements to the appropriate target type. However, this approach is only valid for immutable vectors; and our vectors are mutable. In the example of Figure~\ref{fig:map-vec-bang}, if the cast created a new vector, then the updates inside of \code{map-vec!} would happen to the new vector and not the original one. \begin{figure}[tbp] % gradual_test_11.rkt \begin{lstlisting} (define (map-vec! [f : (Any -> Any)] [v : (Vector Any Any)]) : Void (begin (vector-set! v 0 (f (vector-ref v 0))) (vector-set! v 1 (f (vector-ref v 1))))) (define (add1 x) (+ x 1)) (let ([v (vector 0 41)]) (begin (map-vec! add1 v) (vector-ref v 1))) \end{lstlisting} \caption{An example involving casts on vectors.} \label{fig:map-vec-bang} \end{figure} Instead the interpreter needs to create a new kind of value, a \emph{vector proxy}, that intercepts every vector operation. On a read, the proxy reads from the underlying vector and then applies a cast to the resulting value. On a write, the proxy casts the argument value and then performs the write to the underlying vector. For the first \code{(vector-ref v 0)} in \code{map-vec!}, the proxy casts \code{0} from \code{Integer} to \code{Any}. For the first \code{vector-set!}, the proxy casts a tagged \code{1} from \code{Any} to \code{Integer}. The final category of cast that we need to consider are casts between the \code{Any} type and either a function or a vector type. Figure~\ref{fig:map-vec-any} shows a variant of \code{map-vec!} in which parameter \code{v} does not have a type annotation, so it is given type \code{Any}. In the call to \code{map-vec!}, the vector has type \code{(Vector Integer Integer)} so the type checker inserts a cast from \code{(Vector Integer Integer)} to \code{Any}. A first thought is to use \code{Inject}, but that doesn't work because \code{(Vector Integer Integer)} is not a flat type. Instead, we must first cast to \code{(Vector Any Any)} (which is flat) and then inject to \code{Any}. \begin{figure}[tbp] \begin{lstlisting} (define (map-vec! [f : (Any -> Any)] v) : Void (begin (vector-set! v 0 (f (vector-ref v 0))) (vector-set! v 1 (f (vector-ref v 1))))) (define (add1 x) (+ x 1)) (let ([v (vector 0 41)]) (begin (map-vec! add1 v) (vector-ref v 1))) \end{lstlisting} \caption{Casting a vector to \code{Any}.} \label{fig:map-vec-any} \end{figure} The $R'_9$ interpreter uses an auxiliary function named \code{apply-cast} to cast a value from a source type to a target type, shown in Figure~\ref{fig:apply-cast}. You'll find that it handles all of the kinds of casts that we've discussed in this section. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define/public (apply-cast v s t) (match* (s t) [(t1 t2) #:when (equal? t1 t2) v] [('Any t2) (match t2 [`(,ts ... -> ,rt) (define any->any `(,@(for/list ([t ts]) 'Any) -> Any)) (define v^ (apply-project v any->any)) (apply-cast v^ any->any `(,@ts -> ,rt))] [`(Vector ,ts ...) (define vec-any `(Vector ,@(for/list ([t ts]) 'Any))) (define v^ (apply-project v vec-any)) (apply-cast v^ vec-any `(Vector ,@ts))] [else (apply-project v t2)])] [(t1 'Any) (match t1 [`(,ts ... -> ,rt) (define any->any `(,@(for/list ([t ts]) 'Any) -> Any)) (define v^ (apply-cast v `(,@ts -> ,rt) any->any)) (apply-inject v^ (any-tag any->any))] [`(Vector ,ts ...) (define vec-any `(Vector ,@(for/list ([t ts]) 'Any))) (define v^ (apply-cast v `(Vector ,@ts) vec-any)) (apply-inject v^ (any-tag vec-any))] [else (apply-inject v (any-tag t1))])] [(`(Vector ,ts1 ...) `(Vector ,ts2 ...)) (define x (gensym 'x)) (define cast-reads (for/list ([t1 ts1] [t2 ts2]) `(function (,x) ,(Cast (Var x) t1 t2) ()))) (define cast-writes (for/list ([t1 ts1] [t2 ts2]) `(function (,x) ,(Cast (Var x) t2 t1) ()))) `(vector-proxy ,(vector v (apply vector cast-reads) (apply vector cast-writes)))] [(`(,ts1 ... -> ,rt1) `(,ts2 ... -> ,rt2)) (define xs (for/list ([t2 ts2]) (gensym 'x))) `(function ,xs ,(Cast (Apply (Value v) (for/list ([x xs][t1 ts1][t2 ts2]) (Cast (Var x) t2 t1))) rt1 rt2) ())] )) \end{lstlisting} \caption{The \code{apply-cast} auxiliary method.} \label{fig:apply-cast} \end{figure} The interpreter for $R'_9$ is defined in Figure~\ref{fig:interp-R9-prime}, with the case for \code{Cast} dispatching to \code{apply-cast}. To handle the addition of vector proxies, we update the vector primitives in \code{interp-op} using the functions in Figure~\ref{fig:guarded-vector}. \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define interp-R9-prime-class (class interp-R8-class (super-new) (inherit apply-fun apply-inject apply-project) (define/override (interp-op op) (match op ['vector-length guarded-vector-length] ['vector-ref guarded-vector-ref] ['vector-set! guarded-vector-set!] ['any-vector-ref (lambda (v i) (match v [`(tagged ,v^ ,tg) (guarded-vector-ref v^ i)]))] ['any-vector-set! (lambda (v i a) (match v [`(tagged ,v^ ,tg) (guarded-vector-set! v^ i a)]))] ['any-vector-length (lambda (v) (match v [`(tagged ,v^ ,tg) (guarded-vector-length v^)]))] [else (super interp-op op)] )) (define/override ((interp-exp env) e) (define (recur e) ((interp-exp env) e)) (match e [(Value v) v] [(Cast e src tgt) (apply-cast (recur e) src tgt)] [else ((super interp-exp env) e)])) )) (define (interp-R9-prime p) (send (new interp-R9-prime-class) interp-program p)) \end{lstlisting} \caption{The interpreter for $R'_9$.} \label{fig:interp-R9-prime} \end{figure} \begin{figure}[tbp] \begin{lstlisting}[basicstyle=\ttfamily\footnotesize] (define (guarded-vector-ref vec i) (match vec [`(vector-proxy ,proxy) (define val (guarded-vector-ref (vector-ref proxy 0) i)) (define rd (vector-ref (vector-ref proxy 1) i)) (apply-fun rd (list val) 'guarded-vector-ref)] [else (vector-ref vec i)])) (define (guarded-vector-set! vec i arg) (match vec [`(vector-proxy ,proxy) (define wr (vector-ref (vector-ref proxy 2) i)) (define arg^ (apply-fun wr (list arg) 'guarded-vector-set!)) (guarded-vector-set! (vector-ref proxy 0) i arg^)] [else (vector-set! vec i arg)])) (define (guarded-vector-length vec) (match vec [`(vector-proxy ,proxy) (guarded-vector-length (vector-ref proxy 0))] [else (vector-length vec)])) \end{lstlisting} \caption{The guarded-vector auxiliary functions.} \label{fig:guarded-vector} \end{figure} \section{Lower Casts} \label{sec:lower-casts} The next step in the journey towards x86 is the \code{lower-casts} pass that translates the casts in $R'_9$ to the lower-level \code{Inject} and \code{Project} operators and a new operator for creating vector proxies, extending the $R'_8$ language to create $R''_8$. We recommend creating an auxiliary function named \code{lower-cast} that takes an expression (in $R'_9$), a source type, and a target type, and translates it to expression in $R''_8$ that has the same behavior as casting the expression from the source to the target type in the interpreter. The \code{lower-cast} function can follow a code structure similar to the \code{apply-cast} function (Figure~\ref{fig:apply-cast}) used in the interpreter for $R'_9$ because it must handle the same cases as \code{apply-cast} and it needs to mimic the behavior of \code{apply-cast}. The most interesting cases are those concerning the casts between two vector types and between two function types. As mentioned in Section~\ref{sec:interp-casts}, a cast from one vector type to another vector type is accomplished by creating a proxy that intercepts the operations on the underlying vector. Here we make the creation of the proxy explicit with the \code{vector-proxy} primitive operation. It takes three arguments, the first is an expression for the vector, the second is a vector of functions for casting an element that is being read from the vector, and the third is a vector of functions for casting an element that is being written to the vector. You can create the functions using \code{Lambda}. Also, as we shall see in the next section, we need to differentiate these vectors from the user-created ones, so we recommend using a new primitive operator named \code{raw-vector} instead of \code{vector} to create these vectors of functions. Figure~\ref{fig:map-vec-bang-lower-cast} shows the output of \code{lower-casts} on the example in Figure~\ref{fig:map-vec-bang} that involved casting a vector of integers to a vector of \code{Any}. \begin{figure}[tbp] \begin{lstlisting} (define (map-vec! [f : (Any -> Any)] [v : (Vector Any Any)]) : Void (begin (vector-set! v 0 (f (vector-ref v 0))) (vector-set! v 1 (f (vector-ref v 1))))) (define (add1 [x : Any]) : Any (inject (+ (project x Integer) 1) Integer)) (let ([v (vector 0 41)]) (begin (map-vec! add1 (vector-proxy v (raw-vector (lambda: ([x9 : Integer]) : Any (inject x9 Integer)) (lambda: ([x9 : Integer]) : Any (inject x9 Integer))) (raw-vector (lambda: ([x9 : Any]) : Integer (project x9 Integer)) (lambda: ([x9 : Any]) : Integer (project x9 Integer))))) (vector-ref v 1))) \end{lstlisting} \caption{Output of \code{lower-casts} on the example in Figure~\ref{fig:map-vec-bang}.} \label{fig:map-vec-bang-lower-cast} \end{figure} A cast from one function type to another function type is accomplished by generating a \code{Lambda} whose parameter and return types match the target function type. The body of the \code{Lambda} should cast the parameters from the target type to the source type (yes, backwards! functions are contravariant\index{contravariant} in the parameters), then call the underlying function, and finally cast the result from the source return type to the target return type. Figure~\ref{fig:map-vec-lower-cast} shows the output of the \code{lower-casts} pass on the \code{map-vec} example in Figure~\ref{fig:gradual-map-vec}. Note that the \code{add1} argument in the call to \code{map-vec} is wrapped in a \code{lambda}. \begin{figure}[tbp] \begin{lstlisting} (define (map-vec [f : (Integer -> Integer)] [v : (Vector Integer Integer)]) : (Vector Integer Integer) (vector (f (vector-ref v 0)) (f (vector-ref v 1)))) (define (add1 [x : Any]) : Any (inject (+ (project x Integer) 1) Integer)) (vector-ref (map-vec (lambda: ([x9 : Integer]) : Integer (project (add1 (inject x9 Integer)) Integer)) (vector 0 41)) 1) \end{lstlisting} \caption{Output of \code{lower-casts} on the example in Figure~\ref{fig:gradual-map-vec}.} \label{fig:map-vec-lower-cast} \end{figure} \section{Differentiate Proxies} \label{sec:differentiate-proxies} So far the job of differentiating vectors and vector proxies has been the job of the interpreter. For example, the interpreter for $R'_9$ implements \code{vector-ref} using the \code{guarded-vector-ref} function in Figure~\ref{fig:guarded-vector}. In the \code{differentiate-proxies} pass we shift this responsibility to the generated code. We begin by designing the output language $R_{\mathrm{proxy}}$. In $R_9$ we used the type \code{Vector} for both real vectors and vector proxies. In $R_{\mathrm{proxy}}$ we return the \code{Vector} type to its original meaning, as the type of real vectors, and we introduce a new type, \code{GVector}, whose values can be either real vectors or vector proxies. This new type comes with a suite of new primitive operations for creating and using values of type \code{GVector}. We don't need to introduce a new type to represent vector proxies. A proxy is represented by a vector containing three things: 1) the underlying vector, 2) a vector of functions for casting elements that are read from the vector, and 3) a vector of functions for casting values to be written to the vector. So we define the following abbreviation for the type of a vector proxy: \[ \itm{Proxy} (T\ldots \Rightarrow T'\ldots) = (\ttm{Vector}~(\ttm{GVector}~ T\ldots) ~R~ W) \to (\key{GVector}~ T' \ldots) \] where $R = (\ttm{Vector}~(T\to T') \ldots)$ and $W = (\ttm{Vector}~(T'\to T) \ldots)$. % Next we describe each of the new primitive operations. \begin{description} \item[\code{inject-vector} : (\key{Vector} $T \ldots$) $\to$ (\key{GVector} $T \ldots$)]\ \\ % This operation brands a vector as a value of the \code{GVector} type. \item[\code{inject-proxy} : $\itm{Proxy}(T\ldots \Rightarrow T'\ldots)$ $\to$ (\key{GVector} $T' \ldots$)]\ \\ % This operation brands a vector proxy as value of the \code{GVector} type. \item[\code{proxy?} : (\key{GVector} $T \ldots$) $\to$ \code{Boolean}] \ \\ % returns true if the value is a vector proxy and false if it is a real vector. \item[\code{project-vector} : (\key{GVector} $T \ldots$) $\to$ (\key{Vector} $T \ldots$)]\ \\ % Assuming that the input is a vector (and not a proxy), this operation returns the vector. \item[\code{proxy-vector-length} : (\key{GVector} $T \ldots$) $\to$ \code{Boolean}]\ \\ % Given a vector proxy, this operation returns the length of the underlying vector. \item[\code{proxy-vector-ref} : (\key{GVector} $T \ldots$) $\to$ ($i$ : \code{Integer}) $\to$ $T_i$]\ \\ % Given a vector proxy, this operation returns the $i$th element of the underlying vector. \item[\code{proxy-vector-set!} : (\key{GVector} $T \ldots$) $\to$ ($i$ : \code{Integer}) $\to$ $T_i$ $\to$ \key{Void}]\ \\ Given a vector proxy, this operation writes a value to the $i$th element of the underlying vector. \end{description} Now to discuss the translation that differentiates vectors from proxies. First, every type annotation in the program must be translated (recursively) to replace \code{Vector} with \code{GVector}. Next, we must insert uses of \code{GVector} operations in the appropriate places. For example, we wrap every vector creation with an \code{inject-vector}. \begin{lstlisting} (vector |$e_1 \ldots e_n$|) |$\Rightarrow$| (inject-vector (vector |$e'_1 \ldots e'_n$|)) \end{lstlisting} The \code{raw-vector} operator that we introduced in the previous section does not get injected. \begin{lstlisting} (raw-vector |$e_1 \ldots e_n$|) |$\Rightarrow$| (vector |$e'_1 \ldots e'_n$|) \end{lstlisting} The \code{vector-proxy} primitive translates as follows. \begin{lstlisting} (vector-proxy |$e_1~e_2~e_3$|) |$\Rightarrow$| (inject-proxy (vector |$e'_1~e'_2~e'_3$|)) \end{lstlisting} We translate the vector operations into conditional expressions that check whether the value is a proxy and then dispatch to either the appropriate proxy vector operation or the regular vector operation. For example, the following is the translation for \code{vector-ref}. \begin{lstlisting} (vector-ref |$e_1$| |$i$|) |$\Rightarrow$| (let ([|$v~e_1$|]) (if (proxy? |$v$|) (proxy-vector-ref |$v$| |$i$|) (vector-ref (project-vector |$v$|) |$i$|) \end{lstlisting} Note in the case of a real vector, we must apply \code{project-vector} before the \code{vector-ref}. \section{Reveal Casts} \label{sec:reveal-casts-gradual} Recall that the \code{reveal-casts} pass (Section~\ref{sec:reveal-casts-r6}) is responsible for lowering \code{Inject} and \code{Project} into lower-level operations. In particular, \code{Project} turns into a conditional expression that inspects the tag and retrieves the underlying value. Here we need to augment the translation of \code{Project} to handle the situation when the target type is \code{GVector}. Instead of using \code{vector-length} we need to use \code{proxy-vector-length}. \begin{lstlisting} (project |$e$| (GVector Any|$_1$| |$\ldots$| Any|$_n$|)) |$\Rightarrow$| (let |$\itm{tmp}$| |$e'$| (if (eq? (tag-of-any |$\itm{tmp}$| 2)) (let |$\itm{vec}$| (value-of |$\itm{tmp}$| (GVector Any |$\ldots$| Any)) (if (eq? (proxy-vector-length |$\itm{vec}$|) |$n$|) |$\itm{vec}$| (exit))) (exit))) \end{lstlisting} \section{Closure Conversion} \label{sec:closure-conversion-gradual} The closure conversion pass only requires one minor adjustment. The auxiliary function that translates type annotations needs to be updated to handle the \code{GVector} type. \section{Explicate Control} \label{sec:explicate-control-gradual} Update the \code{explicate-control} pass to handle the new primitive operations on the \code{GVector} type. \section{Select Instructions} \label{sec:select-instructions-gradual} Recall that the \code{select-instructions} pass is responsible for lowering the primitive operations into x86 instructions. So we need to translate the new \code{GVector} operations to x86. To do so, the first question we need to answer is how will we differentiate the two kinds of values (vectors and proxies) that can inhabit \code{GVector}. We need just one bit to accomplish this, so we use the $57$th bit of the 64-bit tag at the front of every vector (see Figure~\ref{fig:tuple-rep}). So far, this bit has been set to $0$, so for \code{inject-vector} we leave it that way. \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'inject-vector (list |$e_1$|))) |$\Rightarrow$| movq |$e'_1$|, |$\itm{lhs'}$| \end{lstlisting} On the other hand, \code{inject-proxy} sets the $57$th bit to $1$. \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'inject-proxy (list |$e_1$|))) |$\Rightarrow$| movq |$e'_1$|, %r11 movq |$(1 << 57)$|, %rax orq 0(%r11), %rax movq %rax, 0(%r11) movq %r11, |$\itm{lhs'}$| \end{lstlisting} The \code{proxy?} operation consumes the information so carefully stashed away by \code{inject-vector} and \code{inject-proxy}. It isolates the $57$th bit to tell whether the value is a real vector or a proxy. \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'proxy? (list e))) |$\Rightarrow$| movq |$e_1'$|, %r11 movq 0(%r11), %rax sarq $57, %rax andq $1, %rax movq %rax, |$\itm{lhs'}$| \end{lstlisting} The \code{project-vector} operation is straightforward to translate, so we leave it up to the reader. Regarding the \code{proxy-vector} operations, the runtime provides procedures that implement them (they are recursive functions!) so here we simply need to translate these vector operations into the appropriate function call. For example, here is the translation for \code{proxy-vector-ref}. \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'proxy-vector-ref (list |$e_1$| |$e_2$|))) |$\Rightarrow$| movq |$e_1'$|, %rdi movq |$e_2'$|, %rsi callq proxy_vector_ref movq %rax, |$\itm{lhs'}$| \end{lstlisting} We have another batch of vector operations to deal with, those for the \code{Any} type. Recall that the type checker for $R_9$ generates an \code{any-vector-ref} when there is a \code{vector-ref} on something of type \code{Any}, and similarly for \code{any-vector-set!} and \code{any-vector-length} (Figure~\ref{fig:type-check-R9-1}). In Section~\ref{sec:select-r6} we selected instructions for these operations based on the idea that the underlying value was a real vector. But in the current setting, the underlying value is of type \code{GVector}. So \code{any-vector-ref} can be translates to pseudo-x86 as follows. We begin by projecting the underlying value out of the tagged value and then call the \code{proxy\_vector\_ref} procedure in the runtime. \begin{lstlisting} (Assign |$\itm{lhs}$| (Prim 'any-vector-ref (list |$e_1$| |$e_2$|))) movq |$\neg 111$|, %rdi andq |$e_1'$|, %rdi movq |$e_2'$|, %rsi callq proxy_vector_ref movq %rax, |$\itm{lhs'}$| \end{lstlisting} The \code{any-vector-set!} and \code{any-vector-length} operators can be translated in a similar way. \begin{exercise}\normalfont Implement a compiler for the gradually-typed $R_9$ language by extending and adapting your compiler for $R_8$. Create 10 new partially-typed test programs. In addition to testing with these new programs, also test your compiler on all the tests for $R_8$ and tests for $R_7$. Sometimes you may get a type checking error on the $R_7$ programs but you can adapt them by inserting a cast to the \code{Any} type around each subexpression causing a type error. While $R_7$ doesn't have explicit casts, you can induce one by wrapping the subexpression \code{e} with a call to an un-annotated identity function, like this: \code{((lambda (x) x) e)}. \end{exercise} \section{Further Reading} This chapter just scratches the surface of gradual typing. The basic approach described here is missing two key ingredients that one would want in a implementation of gradual typing: blame tracking~\citep{Tobin-Hochstadt:2006fk,Wadler:2009qv} and space-efficient casts~\citep{Herman:2006uq,Herman:2010aa}. The problem addressed by blame tracking is that when a cast on a higher-order value fails, it often does so at a point in the program that is far removed from the original cast. Blame tracking is a technique for propagating extra information through casts and proxies so that when a cast fails, the error message can point back to the original location of the cast in the source program. The problem addressed by space-efficient casts also relates to higher-order casts. It turns out that in partially typed programs, a function or vector can flow through very-many casts at runtime. With the approach described in this chapter, each cast adds another \code{lambda} wrapper or a vector proxy. Not only does this take up considerable space, but it also makes the function calls and vector operations slow. For example, a partially-typed version of quicksort could, in the worst case, build a chain of proxies of length $O(n)$ around the vector, changing the overall time complexity of the algorithm from $O(n^2)$ to $O(n^3)$! \citet{Herman:2006uq} suggested a solution to this problem by representing casts using the coercion calculus of \citet{Henglein:1994nz}, which prevents the creation of long chains of proxies by compressing them into a concise normal form. \citet{Siek:2015ab} give and algorithm for compressing coercions and \citet{Kuhlenschmidt:2019aa} show how to implement these ideas in the Grift compiler. \begin{center} \url{https://github.com/Gradual-Typing/Grift} \end{center} There are also interesting interactions between gradual typing and other language features, such as parametetric polymorphism, information-flow types, and type inference, to name a few. We recommend the reader to the online gradual typing bibliography: \begin{center} \url{http://samth.github.io/gradual-typing-bib/} \end{center} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Parametric Polymorphism} \label{ch:parametric-polymorphism} \index{parametric polymorphism} \index{generics} This chapter may be based on ideas from \citet{Cardelli:1984aa}, \citet{Leroy:1992qb}, \citet{Shao:1997uj}, or \citet{Harper:1995um}. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %% \chapter{High-level Optimization} %% \label{ch:high-level-optimization} %% This chapter will present a procedure inlining pass based on the %% algorithm of \citet{Waddell:1997fk}. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \chapter{Appendix} \section{Interpreters} \label{appendix:interp} \index{interpreter} We provide interpreters for each of the source languages $R_0$, $R_1$, $\ldots$ in the files \code{interp-R0.rkt}, \code{interp-R1.rkt}, etc. The interpreters for the intermediate languages $C_0$ and $C_1$ are in \code{interp-C0.rkt} and \code{interp-C1.rkt}. The interpreters for $C_2$, $C_3$, pseudo-x86, and x86 are in the \key{interp.rkt} file. \section{Utility Functions} \label{appendix:utilities} The utility functions described in this section are in the \key{utilities.rkt} file of the support code. \paragraph{\code{interp-tests}} The \key{interp-tests} function runs the compiler passes and the interpreters on each of the specified tests to check whether each pass is correct. The \key{interp-tests} function has the following parameters: \begin{description} \item[name (a string)] a name to identify the compiler, \item[typechecker] a function of exactly one argument that either raises an error using the \code{error} function when it encounters a type error, or returns \code{\#f} when it encounters a type error. If there is no type error, the type checker returns the program. \item[passes] a list with one entry per pass. An entry is a list with four things: \begin{enumerate} \item a string giving the name of the pass, \item the function that implements the pass (a translator from AST to AST), \item a function that implements the interpreter (a function from AST to result value) for the output language, \item and a type checker for the output language. Type checkers for the $R$ and $C$ languages are provided in the support code. For example, the type checkers for $R_1$ and $C_0$ are in \code{type-check-R1.rkt}. The type checker entry is optional. The support code does not provide type checkers for the x86 languages. \end{enumerate} \item[source-interp] an interpreter for the source language. The interpreters from Appendix~\ref{appendix:interp} make a good choice. \item[test-family (a string)] for example, \code{"r1"}, \code{"r2"}, etc. \item[tests] a list of test numbers that specifies which tests to run. (see below) \end{description} % The \key{interp-tests} function assumes that the subdirectory \key{tests} has a collection of Racket programs whose names all start with the family name, followed by an underscore and then the test number, ending with the file extension \key{.rkt}. Also, for each test program that calls \code{read} one or more times, there is a file with the same name except that the file extension is \key{.in} that provides the input for the Racket program. If the test program is expected to fail type checking, then there should be an empty file of the same name but with extension \key{.tyerr}. \paragraph{\code{compiler-tests}} runs the compiler passes to generate x86 (a \key{.s} file) and then runs the GNU C compiler (gcc) to generate machine code. It runs the machine code and checks that the output is $42$. The parameters to the \code{compiler-tests} function are similar to those of the \code{interp-tests} function, and consist of \begin{itemize} \item a compiler name (a string), \item a type checker, \item description of the passes, \item name of a test-family, and \item a list of test numbers. \end{itemize} \paragraph{\code{compile-file}} takes a description of the compiler passes (see the comment for \key{interp-tests}) and returns a function that, given a program file name (a string ending in \key{.rkt}), applies all of the passes and writes the output to a file whose name is the same as the program file name but with \key{.rkt} replaced with \key{.s}. \paragraph{\code{read-program}} takes a file path and parses that file (it must be a Racket program) into an abstract syntax tree. \paragraph{\code{parse-program}} takes an S-expression representation of an abstract syntax tree and converts it into the struct-based representation. \paragraph{\code{assert}} takes two parameters, a string (\code{msg}) and Boolean (\code{bool}), and displays the message \key{msg} if the Boolean \key{bool} is false. \paragraph{\code{lookup}} % remove discussion of lookup? -Jeremy takes a key and an alist, and returns the first value that is associated with the given key, if there is one. If not, an error is triggered. The alist may contain both immutable pairs (built with \key{cons}) and mutable pairs (built with \key{mcons}). %The \key{map2} function ... \section{x86 Instruction Set Quick-Reference} \label{sec:x86-quick-reference} \index{x86} Table~\ref{tab:x86-instr} lists some x86 instructions and what they do. We write $A \to B$ to mean that the value of $A$ is written into location $B$. Address offsets are given in bytes. The instruction arguments $A, B, C$ can be immediate constants (such as \code{\$4}), registers (such as \code{\%rax}), or memory references (such as \code{-4(\%ebp)}). Most x86 instructions only allow at most one memory reference per instruction. Other operands must be immediates or registers. \begin{table}[tbp] \centering \begin{tabular}{l|l} \textbf{Instruction} & \textbf{Operation} \\ \hline \texttt{addq} $A$, $B$ & $A + B \to B$\\ \texttt{negq} $A$ & $- A \to A$ \\ \texttt{subq} $A$, $B$ & $B - A \to B$\\ \texttt{callq} $L$ & Pushes the return address and jumps to label $L$ \\ \texttt{callq} \texttt{*}$A$ & Calls the function at the address $A$. \\ %\texttt{leave} & $\texttt{ebp} \to \texttt{esp};$ \texttt{popl \%ebp} \\ \texttt{retq} & Pops the return address and jumps to it \\ \texttt{popq} $A$ & $*\mathtt{rsp} \to A; \mathtt{rsp} + 8 \to \mathtt{rsp}$ \\ \texttt{pushq} $A$ & $\texttt{rsp} - 8 \to \texttt{rsp}; A \to *\texttt{rsp}$\\ \texttt{leaq} $A$,$B$ & $A \to B$ ($B$ must be a register) \\ \texttt{cmpq} $A$, $B$ & compare $A$ and $B$ and set the flag register ($B$ must not be an immediate) \\ \texttt{je} $L$ & \multirow{5}{3.7in}{Jump to label $L$ if the flag register matches the condition code of the instruction, otherwise go to the next instructions. The condition codes are \key{e} for ``equal'', \key{l} for ``less'', \key{le} for ``less or equal'', \key{g} for ``greater'', and \key{ge} for ``greater or equal''.} \\ \texttt{jl} $L$ & \\ \texttt{jle} $L$ & \\ \texttt{jg} $L$ & \\ \texttt{jge} $L$ & \\ \texttt{jmp} $L$ & Jump to label $L$ \\ \texttt{movq} $A$, $B$ & $A \to B$ \\ \texttt{movzbq} $A$, $B$ & \multirow{3}{3.7in}{$A \to B$, \text{where } $A$ is a single-byte register (e.g., \texttt{al} or \texttt{cl}), $B$ is a 8-byte register, and the extra bytes of $B$ are set to zero.} \\ & \\ & \\ \texttt{notq} $A$ & $\sim A \to A$ \qquad (bitwise complement)\\ \texttt{orq} $A$, $B$ & $A | B \to B$ \qquad (bitwise-or)\\ \texttt{andq} $A$, $B$ & $A \& B \to B$ \qquad (bitwise-and)\\ \texttt{salq} $A$, $B$ & $B$ \texttt{<<} $A \to B$ (arithmetic shift left, where $A$ is a constant)\\ \texttt{sarq} $A$, $B$ & $B$ \texttt{>>} $A \to B$ (arithmetic shift right, where $A$ is a constant)\\ \texttt{sete} $A$ & \multirow{5}{3.7in}{If the flag matches the condition code, then $1 \to A$, else $0 \to A$. Refer to \texttt{je} above for the description of the condition codes. $A$ must be a single byte register (e.g., \texttt{al} or \texttt{cl}).} \\ \texttt{setl} $A$ & \\ \texttt{setle} $A$ & \\ \texttt{setg} $A$ & \\ \texttt{setge} $A$ & \end{tabular} \vspace{5pt} \caption{Quick-reference for the x86 instructions used in this book.} \label{tab:x86-instr} \end{table} \cleardoublepage \section{Concrete Syntax for Intermediate Languages} The concrete syntax of $R_6$ is defined in Figure~\ref{fig:r6-concrete-syntax}. \begin{figure}[tp] \centering \fbox{ \begin{minipage}{0.97\textwidth}\small \[ \begin{array}{lcl} \Type &::=& \gray{\key{Integer} \mid \key{Boolean} \mid \LP\key{Vector}\;\Type\ldots\RP \mid \key{Void}} \\ &\mid& \gray{\LP\Type\ldots \; \key{->}\; \Type\RP} \mid \key{Any} \\ \FType &::=& \key{Integer} \mid \key{Boolean} \mid \key{Void} \mid \LP\key{Vector}\; \key{Any}\ldots\RP \\ &\mid& \LP\key{Any}\ldots \; \key{->}\; \key{Any}\RP\\ \Exp &::=& \ldots \CINJECT{\Exp}{\FType}\RP \mid \CPROJECT{\Exp}{\FType}\\ &\mid& \LP\key{any-vector-length}\;\Exp\RP \mid \LP\key{any-vector-ref}\;\Exp\;\Exp\RP \\ &\mid& \LP\key{any-vector-set!}\;\Exp\;\Exp\;\Exp\RP\\ &\mid& \LP\key{boolean?}\;\Exp\RP \mid \LP\key{integer?}\;\Exp\RP \mid \LP\key{void?}\;\Exp\RP \\ &\mid& \LP\key{vector?}\;\Exp\RP \mid \LP\key{procedure?}\;\Exp\RP \\ \Def &::=& \gray{ \CDEF{\Var}{\LS\Var \key{:} \Type\RS\ldots}{\Type}{\Exp} } \\ R_6 &::=& \gray{\Def\ldots \; \Exp} \end{array} \] \end{minipage} } \caption{The concrete syntax of $R_6$, extending $R_5$ (Figure~\ref{fig:r5-syntax}) with \key{Any}.} \label{fig:r6-concrete-syntax} \end{figure} The concrete syntax for $C_0$, $C_1$, $C_2$ and $C_3$ is defined in Figures~\ref{fig:c0-concrete-syntax}, \ref{fig:c1-concrete-syntax}, \ref{fig:c2-concrete-syntax}, and \ref{fig:c3-concrete-syntax}, respectively. \begin{figure}[tbp] \fbox{ \begin{minipage}{0.96\textwidth} \[ \begin{array}{lcl} \Atm &::=& \Int \mid \Var \\ \Exp &::=& \Atm \mid \key{(read)} \mid \key{(-}~\Atm\key{)} \mid \key{(+}~\Atm~\Atm\key{)}\\ \Stmt &::=& \Var~\key{=}~\Exp\key{;} \\ \Tail &::= & \key{return}~\Exp\key{;} \mid \Stmt~\Tail \\ C_0 & ::= & (\itm{label}\key{:}~ \Tail)\ldots \end{array} \] \end{minipage} } \caption{The concrete syntax of the $C_0$ intermediate language.} \label{fig:c0-concrete-syntax} \end{figure} \begin{figure}[tbp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Atm &::=& \gray{ \Int \mid \Var } \mid \itm{bool} \\ \itm{cmp} &::= & \key{eq?} \mid \key{<} \\ \Exp &::=& \gray{ \Atm \mid \key{(read)} \mid \key{(-}~\Atm\key{)} \mid \key{(+}~\Atm~\Atm\key{)} } \\ &\mid& \LP \key{not}~\Atm \RP \mid \LP \itm{cmp}~\Atm~\Atm\RP \\ \Stmt &::=& \gray{ \Var~\key{=}~\Exp\key{;} } \\ \Tail &::= & \gray{ \key{return}~\Exp\key{;} \mid \Stmt~\Tail } \mid \key{goto}~\itm{label}\key{;}\\ &\mid& \key{if}~\LP \itm{cmp}~\Atm~\Atm \RP~ \key{goto}~\itm{label}\key{;} ~\key{else}~\key{goto}~\itm{label}\key{;} \\ C_1 & ::= & \gray{ (\itm{label}\key{:}~ \Tail)\ldots } \end{array} \] \end{minipage} } \caption{The concrete syntax of the $C_1$ intermediate language.} \label{fig:c1-concrete-syntax} \end{figure} \begin{figure}[tbp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Atm &::=& \gray{ \Int \mid \Var \mid \itm{bool} } \\ \itm{cmp} &::= & \gray{ \key{eq?} \mid \key{<} } \\ \Exp &::=& \gray{ \Atm \mid \key{(read)} \mid \key{(-}~\Atm\key{)} \mid \key{(+}~\Atm~\Atm\key{)} } \\ &\mid& \gray{ \LP \key{not}~\Atm \RP \mid \LP \itm{cmp}~\Atm~\Atm\RP } \\ &\mid& \LP \key{allocate}~\Int~\Type \RP \\ &\mid& (\key{vector-ref}\;\Atm\;\Int) \mid (\key{vector-set!}\;\Atm\;\Int\;\Atm)\\ &\mid& \LP \key{global-value}~\Var \RP \mid \LP \key{void} \RP \\ \Stmt &::=& \gray{ \Var~\key{=}~\Exp\key{;} } \mid \LP\key{collect}~\Int \RP\\ \Tail &::= & \gray{ \key{return}~\Exp\key{;} \mid \Stmt~\Tail } \mid \gray{ \key{goto}~\itm{label}\key{;} }\\ &\mid& \gray{ \key{if}~\LP \itm{cmp}~\Atm~\Atm \RP~ \key{goto}~\itm{label}\key{;} ~\key{else}~\key{goto}~\itm{label}\key{;} } \\ C_2 & ::= & \gray{ (\itm{label}\key{:}~ \Tail)\ldots } \end{array} \] \end{minipage} } \caption{The concrete syntax of the $C_2$ intermediate language.} \label{fig:c2-concrete-syntax} \end{figure} \begin{figure}[tp] \fbox{ \begin{minipage}{0.96\textwidth} \small \[ \begin{array}{lcl} \Atm &::=& \gray{ \Int \mid \Var \mid \key{\#t} \mid \key{\#f} } \\ \itm{cmp} &::= & \gray{ \key{eq?} \mid \key{<} } \\ \Exp &::= & \gray{ \Atm \mid \LP\key{read}\RP \mid \LP\key{-}\;\Atm\RP \mid \LP\key{+} \; \Atm\;\Atm\RP \mid \LP\key{not}\;\Atm\RP \mid \LP\itm{cmp}\;\Atm\;\Atm\RP } \\ &\mid& \gray{ \LP\key{allocate}\,\Int\,\Type\RP \mid \LP\key{vector-ref}\, \Atm\, \Int\RP } \\ &\mid& \gray{ \LP\key{vector-set!}\,\Atm\,\Int\,\Atm\RP \mid \LP\key{global-value} \,\itm{name}\RP \mid \LP\key{void}\RP } \\ &\mid& \LP\key{fun-ref}~\itm{label}\RP \mid \LP\key{call} \,\Atm\,\Atm\ldots\RP \\ \Stmt &::=& \gray{ \ASSIGN{\Var}{\Exp} \mid \RETURN{\Exp} \mid \LP\key{collect} \,\itm{int}\RP }\\ \Tail &::= & \gray{\RETURN{\Exp} \mid \LP\key{seq}\;\Stmt\;\Tail\RP} \\ &\mid& \gray{\LP\key{goto}\,\itm{label}\RP \mid \IF{\LP\itm{cmp}\, \Atm\,\Atm\RP}{\LP\key{goto}\,\itm{label}\RP}{\LP\key{goto}\,\itm{label}\RP}} \\ &\mid& \LP\key{tail-call}\,\Atm\,\Atm\ldots\RP \\ \Def &::=& \LP\key{define}\; \LP\itm{label} \; [\Var \key{:} \Type]\ldots\RP \key{:} \Type \; \LP\LP\itm{label}\,\key{.}\,\Tail\RP\ldots\RP\RP \\ C_3 & ::= & \Def\ldots \end{array} \] \end{minipage} } \caption{The $C_3$ language, extending $C_2$ (Figure~\ref{fig:c2-concrete-syntax}) with functions.} \label{fig:c3-concrete-syntax} \end{figure} \cleardoublepage \addcontentsline{toc}{chapter}{Index} \printindex \cleardoublepage \bibliographystyle{plainnat} \bibliography{all} \addcontentsline{toc}{chapter}{Bibliography} \end{document} %% LocalWords: Dybvig Waddell Abdulaziz Ghuloum Dipanwita Sussman %% LocalWords: Sarkar lcl Matz aa representable Chez Ph Dan's nano %% LocalWords: fk bh Siek plt uq Felleisen Bor Yuh ASTs AST Naur eq %% LocalWords: BNF fixnum datatype arith prog backquote quasiquote %% LocalWords: ast Reynold's reynolds interp cond fx evaluator jane %% LocalWords: quasiquotes pe nullary unary rcl env lookup gcc rax %% LocalWords: addq movq callq rsp rbp rbx rcx rdx rsi rdi subq nx %% LocalWords: negq pushq popq retq globl Kernighan uniquify lll ve %% LocalWords: allocator gensym env subdirectory scm rkt tmp lhs Tt %% LocalWords: runtime Liveness liveness undirected Balakrishnan je %% LocalWords: Rosen DSATUR SDO Gebremedhin Omari morekeywords cnd %% LocalWords: fullflexible vertices Booleans Listof Pairof thn els %% LocalWords: boolean type-check notq cmpq sete movzbq jmp al xorq %% LocalWords: EFLAGS thns elss elselabel endlabel Tuples tuples os %% LocalWords: tuple args lexically leaq Polymorphism msg bool nums %% LocalWords: macosx unix Cormen vec callee xs maxStack numParams %% LocalWords: arg bitwise XOR'd thenlabel immediates optimizations %% LocalWords: deallocating Ungar Detlefs Tene kx FromSpace ToSpace %% LocalWords: Appel Diwan Siebert ptr fromspace rootstack typedef %% LocalWords: len prev rootlen heaplen setl lt Kohlbecker dk multi % LocalWords: Bloomington Wollowski definitional whitespace deref JM % LocalWords: subexpression subexpressions iteratively ANF Danvy rco % LocalWords: goto stmt JS ly cmp ty le ge jle goto's EFLAG CFG pred % LocalWords: acyclic worklist Aho qf tsort implementer's hj Shidal % LocalWords: nonnegative Shahriyar endian salq sarq uint cheney ior % LocalWords: tospace vecinit collectret alloc initret decrement jl % LocalWords: dereferencing GC di vals ps mcons ds mcdr callee's th % LocalWords: mainDef tailcall prepending mainstart num params rT qb % LocalWords: mainconclusion Cardelli bodyT fvs clos fvts subtype uj % LocalWords: polymorphism untyped elts tys tagof Vectorof tyeq orq % LocalWords: andq untagged Shao inlining ebp jge setle setg setge % LocalWords: struct symtab Friedman's MacOS Nystrom alist sam kate % LocalWords: alists arity github unordered pqueue exprs ret param % LocalWords: tyerr bytereg dh dl JmpIf HasType Osterlund Jacek TODO % LocalWords: Gamari GlobalValue ProgramDefsExp prm ProgramDefs vn % LocalWords: FunRef TailCall tailjmp IndirectCallq TailJmp Gilray % LocalWords: dereference unbox Dataflow versa dataflow Kildall rhs % LocalWords: Kleene enqueue dequeue AssignedFree FV cnvt SetBang tg % LocalWords: ValueOf typechecker