| 677 |      1 | % !TEX program = xelatex
 | 
| 539 |      2 | \documentclass{article}
 | 
|  |      3 | \usepackage{../style}
 | 
|  |      4 | \usepackage{../langs}
 | 
|  |      5 | \usepackage{../graphics}
 | 
|  |      6 | \usepackage{../grammar}
 | 
| 677 |      7 | %%\usepackage{multicol}
 | 
| 539 |      8 | 
 | 
| 677 |      9 | %%\newcommand{\dn}{\stackrel{\mbox{\scriptsize def}}{=}}
 | 
| 539 |     10 | 
 | 
|  |     11 | \begin{document}
 | 
| 677 |     12 | \fnote{\copyright{} Christian Urban, King's College London, 2019}
 | 
| 539 |     13 | 
 | 
|  |     14 | 
 | 
| 677 |     15 | \section*{Handout 9 (LLVM, SSA and CPS)}
 | 
| 539 |     16 | 
 | 
| 678 |     17 | Reflecting on our tiny compiler targetting the JVM, the code generation
 | 
|  |     18 | part was actually not so hard, no? Pretty much just some post-traversal
 | 
| 680 |     19 | of the abstract syntax tree, yes? One of the reasons for this ease is
 | 
|  |     20 | that the JVM is a stack-based virtual machine and it is therefore not
 | 
|  |     21 | hard to translate deeply-nested arithmetic expressions into a sequence
 | 
|  |     22 | of instructions manipulating the stack. The problem is that ``real''
 | 
|  |     23 | CPUs, although supporting stack operations, are not really designed to
 | 
|  |     24 | be \emph{stack machines}.  The design of CPUs is more like, here is a
 | 
|  |     25 | chunk of memory---compiler, or better compiler writers, do something
 | 
|  |     26 | with it. Consequently, modern compilers need to go the extra mile in
 | 
|  |     27 | order to generate code that is much easier and faster to process by
 | 
|  |     28 | CPUs. To make this all tractable for this module, we target the LLVM
 | 
|  |     29 | Intermediate Language. In this way we can take advantage of the tools
 | 
|  |     30 | coming with LLVM. For example we do not have to worry about things like
 | 
|  |     31 | register allocations.\bigskip 
 | 
| 539 |     32 | 
 | 
| 678 |     33 | \noindent LLVM\footnote{\url{http://llvm.org}} is a beautiful example
 | 
|  |     34 | that projects from Academia can make a difference in the world. LLVM
 | 
|  |     35 | started in 2000 as a project by two researchers at the  University of
 | 
|  |     36 | Illinois at Urbana-Champaign. At the time the behemoth of compilers was
 | 
| 680 |     37 | gcc with its myriad of front-ends for other languages (C++, Fortran,
 | 
| 678 |     38 | Ada, Go, Objective-C, Pascal etc). The problem was that gcc morphed over
 | 
|  |     39 | time into a monolithic gigantic piece of m\ldots ehm software, which you
 | 
|  |     40 | could not mess about in an afternoon. In contrast, LLVM is designed to
 | 
|  |     41 | be a modular suite of tools with which you could play around easily and
 | 
|  |     42 | try out something new. LLVM became a big player once Apple hired one of
 | 
|  |     43 | the original developers (I cannot remember the reason why Apple did not
 | 
|  |     44 | want to use gcc, but maybe they were also just disgusted by its big
 | 
|  |     45 | monolithic codebase). Anyway, LLVM is now the big player and gcc is more
 | 
|  |     46 | or less legacy. This does not mean that programming languages like C and
 | 
|  |     47 | C++ are dying out any time soon---they are nicely supported by LLVM.
 | 
| 539 |     48 | 
 | 
| 680 |     49 | We will target the LLVM Intermediate Language, or Intermediate
 | 
|  |     50 | Representation (short LLVM-IR). As a result we can benefit from the
 | 
| 678 |     51 | modular structure of the LLVM compiler and let for example the compiler
 | 
| 680 |     52 | generate code for different CPUs, like X86 or ARM. That means we can be
 | 
|  |     53 | agnostic about where our code actually runs. We can also be ignorant
 | 
|  |     54 | about optimising code and allocating memory efficiently. However, what
 | 
|  |     55 | we have to do is to generate code in \emph{Static Single-Assignment}
 | 
|  |     56 | format (short SSA), because that is what the LLVM-IR expects from us. 
 | 
| 539 |     57 | 
 | 
| 678 |     58 | The idea behind the SSA format is to use very simple variable
 | 
|  |     59 | assignments where every variable is assigned only once. The assignments
 | 
|  |     60 | also need to be primitive in the sense that they can be just simple
 | 
|  |     61 | operations like addition, multiplication, jumps, comparisons and so on.
 | 
| 680 |     62 | Say, we have an expression $((1 + a) + (3 + (b * 5)))$, then the
 | 
|  |     63 | SSA is 
 | 
|  |     64 |  
 | 
|  |     65 | \begin{lstlisting}[language=LLVMIR,numbers=left]
 | 
|  |     66 | let tmp0 = add 1 a in   
 | 
|  |     67 | let tmp1 = mul b 5 in 
 | 
|  |     68 | let tmp2 = add 3 tmp1 in 
 | 
|  |     69 | let tmp3 = add tmp0 tmp2 in
 | 
|  |     70 |   tmp3 
 | 
| 677 |     71 | \end{lstlisting}
 | 
| 539 |     72 | 
 | 
| 678 |     73 | \noindent where every variable is used only once (we could not write
 | 
| 680 |     74 | \texttt{tmp1 = add 3 tmp1} in Line 3 for example).  There are
 | 
| 678 |     75 | sophisticated algorithms for imperative languages, like C, that
 | 
|  |     76 | efficiently transform a high-level program into SSA format. But we can
 | 
|  |     77 | ignore them here. We want to compile a functional language and there
 | 
|  |     78 | things get much more interesting than just sophisticated. We will need
 | 
|  |     79 | to have a look at CPS translations, where the CPS stands for
 | 
|  |     80 | Continuation-Passing-Style---basically black programming art or
 | 
|  |     81 | abracadabra programming. So sit tight.
 | 
| 539 |     82 | 
 | 
| 678 |     83 | \subsection*{LLVM-IR}
 | 
| 539 |     84 | 
 | 
| 678 |     85 | Before we start, lets first have a look at the \emph{LLVM Intermediate
 | 
| 680 |     86 | Representation} in more detail. What is good about our toy Fun language
 | 
|  |     87 | is that it basically only contains expressions (be they arithmetic
 | 
|  |     88 | expressions or boolean expressions or if-expressions). The exception are
 | 
|  |     89 | function definitions. Luckily, for them we can use the mechanism of
 | 
|  |     90 | defining functions in LLVM-IR. For example the simple Fun program 
 | 
| 539 |     91 | 
 | 
|  |     92 | 
 | 
| 677 |     93 | \begin{lstlisting}[language=Scala,numbers=none]
 | 
| 678 |     94 | def sqr(x) = x * x
 | 
| 677 |     95 | \end{lstlisting}
 | 
| 539 |     96 | 
 | 
| 677 |     97 | \noindent
 | 
|  |     98 | can be compiled into the following LLVM-IR function:
 | 
| 539 |     99 | 
 | 
| 677 |    100 | \begin{lstlisting}[language=LLVM]
 | 
| 678 |    101 | define i32 @sqr(i32 %x) {
 | 
|  |    102 |    %tmp = mul i32 %x, %x
 | 
| 677 |    103 |    ret i32 %tmp
 | 
|  |    104 | }    
 | 
|  |    105 | \end{lstlisting}
 | 
| 539 |    106 | 
 | 
| 680 |    107 | \noindent First notice that all variable names in the LLVM-IR are
 | 
|  |    108 | prefixed by \texttt{\%}; function names need to be prefixed with
 | 
|  |    109 | @-symbol. Also, the LLVM-IR is a fully typed language. The \texttt{i32}
 | 
|  |    110 | type stands for 32-bit integers. There are also types for 64-bit
 | 
|  |    111 | integers (\texttt{i64}), chars (\texttt{i8}), floats, arrays and even
 | 
|  |    112 | pointer types. In the code above, \texttt{sqr} takes an argument of type
 | 
|  |    113 | \texttt{i32} and produces a result of type \texttt{i32} (the result type
 | 
|  |    114 | is before the function name, like in C). Each arithmetic operation, like
 | 
|  |    115 | addition or multiplication, are also prefixed with the type they operate
 | 
|  |    116 | on. Obviously these types need to match up\ldots{} but since we have in
 | 
|  |    117 | our programs only integers, \texttt{i32} everywhere will do. 
 | 
| 539 |    118 |  
 | 
| 679 |    119 | Conveniently, you can use the program \texttt{lli}, which comes with
 | 
|  |    120 | LLVM, to interpret programs written in the LLVM-IR. So you can easily
 | 
|  |    121 | check whether the code you produced actually works. To get a running
 | 
|  |    122 | program that does something interesting you need to add some boilerplate
 | 
|  |    123 | about printing out numbers and a main-function that is the entrypoint
 | 
| 680 |    124 | for the program (see Figure~\ref{lli} for a complete listing). You can
 | 
|  |    125 | generate a binary for this program using \texttt{llc}-compiler and
 | 
|  |    126 | \texttt{gcc}---\texttt{llc} can generate an object file and then you can 
 | 
|  |    127 | use \texttt{gcc} (that is clang) for generating an executable  binary:
 | 
| 678 |    128 | 
 | 
| 679 |    129 | \begin{lstlisting}[language=bash,numbers=none]
 | 
|  |    130 | llc -filetype=obj sqr.ll
 | 
|  |    131 | gcc sqr.o -o a.out
 | 
|  |    132 | ./a.out
 | 
| 680 |    133 | > 25
 | 
| 679 |    134 | \end{lstlisting}
 | 
|  |    135 | 
 | 
|  |    136 | \begin{figure}[t]\small 
 | 
|  |    137 | \lstinputlisting[language=LLVM,numbers=left]{../progs/sqr.ll}
 | 
|  |    138 | \caption{An LLVM-IR program for calculating the square function. The 
 | 
|  |    139 | interesting function is \texttt{sqr} in Lines 13 -- 16. The main
 | 
|  |    140 | function calls \texttt{sqr} and prints out the result. The other
 | 
|  |    141 | code is boilerplate for printing out integers.\label{lli}}
 | 
| 678 |    142 | \end{figure}   
 | 
|  |    143 | 
 | 
| 679 |    144 | 
 | 
|  |    145 |     
 | 
|  |    146 | \subsection*{Our Own Intermediate Language}
 | 
|  |    147 | 
 | 
|  |    148 | Remember compilers have to solve the problem of bridging the gap between
 | 
| 680 |    149 | ``high-level'' programs and ``low-level'' hardware. If the gap is too
 | 
|  |    150 | wide for one step, then a good strategy is to lay a stepping stone
 | 
|  |    151 | somewhere in between. The LLVM-IR itself is such a stepping stone to
 | 
|  |    152 | make the task of generating and optimising code easier. Like a real
 | 
|  |    153 | compiler we will use our own stepping stone which I call the
 | 
|  |    154 | \emph{K-language}. For this remember expressions (and boolean
 | 
|  |    155 | expressions) in the Fun language. For convenience the Scala code is
 | 
|  |    156 | shown on top of Figure~\ref{absfun}. Below is the code for the
 | 
|  |    157 | K-language. There are two kinds of syntactic entities, namely
 | 
|  |    158 | \emph{K-values} and \emph{K-expressions}. The central point of the
 | 
|  |    159 | K-language is the \texttt{KLet}-constructor. For this recall that 
 | 
|  |    160 | arithmetic expressions such as $((1 + a) + (3 + (b * 5)))$ need
 | 
|  |    161 | to be broken up into smaller ``atomic'' steps, like so
 | 
|  |    162 |  
 | 
|  |    163 | \begin{lstlisting}[language=LLVMIR,numbers=none]
 | 
|  |    164 | let tmp0 = add 1 a in   
 | 
|  |    165 | let tmp1 = mul b 5 in 
 | 
|  |    166 | let tmp2 = add 3 tmp1 in 
 | 
|  |    167 | let tmp3 = add tmp0 tmp2 in
 | 
|  |    168 |   tmp3 
 | 
|  |    169 | \end{lstlisting}
 | 
|  |    170 | 
 | 
|  |    171 | \noindent
 | 
|  |    172 | Here \texttt{tmp3} will contain the result of what the expression stands
 | 
|  |    173 | for. In each step we can only perform an ``atomic'' operation, like
 | 
|  |    174 | addition or multiplication. We could not for example have an
 | 
|  |    175 | if-condition on the right-hand side of an equals. These constraints
 | 
|  |    176 | enforced upon us because how the SSA format works in the LLVM-IR. By
 | 
|  |    177 | having in \texttt{KLet},  first a string (standing for an intermediate
 | 
|  |    178 | result) and second a value, we can fulfil this constraint---there is no
 | 
|  |    179 | way we could write anything else than a value. To sum up, K-values are
 | 
|  |    180 | the atomic operations that can be on the right-hand side of equal-signs.
 | 
|  |    181 | The K-language is restricted such that it is easy to generate the SSA format 
 | 
|  |    182 | for the LLVM-IR.
 | 
|  |    183 | 
 | 
| 679 |    184 | 
 | 
|  |    185 | 
 | 
|  |    186 | \begin{figure}[p]\small
 | 
| 678 |    187 | \begin{lstlisting}[language=Scala,numbers=none]
 | 
| 679 |    188 | // Fun-language (expressions)
 | 
|  |    189 | abstract class Exp 
 | 
|  |    190 | abstract class BExp 
 | 
| 678 |    191 | 
 | 
|  |    192 | case class Call(name: String, args: List[Exp]) extends Exp
 | 
|  |    193 | case class If(a: BExp, e1: Exp, e2: Exp) extends Exp
 | 
|  |    194 | case class Write(e: Exp) extends Exp
 | 
|  |    195 | case class Var(s: String) extends Exp
 | 
|  |    196 | case class Num(i: Int) extends Exp
 | 
|  |    197 | case class Aop(o: String, a1: Exp, a2: Exp) extends Exp
 | 
|  |    198 | case class Sequence(e1: Exp, e2: Exp) extends Exp
 | 
| 679 |    199 | case class Bop(o: String, a1: Exp, a2: Exp) extends BExp  
 | 
|  |    200 | 
 | 
|  |    201 | 
 | 
|  |    202 | 
 | 
|  |    203 | // K-language (K-expressions, K-values)
 | 
|  |    204 | abstract class KExp
 | 
|  |    205 | abstract class KVal
 | 
|  |    206 | 
 | 
|  |    207 | case class KVar(s: String) extends KVal
 | 
|  |    208 | case class KNum(i: Int) extends KVal
 | 
|  |    209 | case class Kop(o: String, v1: KVal, v2: KVal) extends KVal
 | 
|  |    210 | case class KCall(o: String, vrs: List[KVal]) extends KVal
 | 
|  |    211 | case class KWrite(v: KVal) extends KVal
 | 
|  |    212 | 
 | 
|  |    213 | case class KIf(x1: String, e1: KExp, e2: KExp) extends KExp
 | 
| 680 |    214 | case class KLet(x: String, v: KVal, e: KExp) extends KExp
 | 
| 679 |    215 | case class KReturn(v: KVal) extends KExp
 | 
| 678 |    216 | \end{lstlisting}
 | 
|  |    217 | \caption{Abstract syntax trees for the Fun language.\label{absfun}}
 | 
|  |    218 | \end{figure}
 | 
| 679 |    219 | 
 | 
| 678 |    220 | 
 | 
|  |    221 | 
 | 
|  |    222 | \subsection*{CPS-Translations}
 | 
|  |    223 | 
 | 
|  |    224 | 
 | 
| 679 |    225 | 
 | 
|  |    226 | 
 | 
|  |    227 | 
 | 
|  |    228 | Another reason why it makes sense to go the extra mile is that stack
 | 
|  |    229 | instructions are very difficult to optimise---you cannot just re-arrange
 | 
|  |    230 | instructions without messing about with what is calculated on the stack.
 | 
|  |    231 | Also it is hard to find out if all the calculations on the stack are
 | 
|  |    232 | actually necessary and not by chance dead code. The JVM has for all this
 | 
|  |    233 | sophisticated machinery to make such ``high-level'' code still run fast,
 | 
|  |    234 | but let's say that for the sake of argument we do not want to rely on
 | 
|  |    235 | it. We want to generate fast code ourselves. This means we have to work
 | 
|  |    236 | around the intricacies of what instructions CPUs can actually process.
 | 
|  |    237 | 
 | 
| 539 |    238 | \end{document}
 | 
|  |    239 | 
 | 
|  |    240 | 
 | 
|  |    241 | %%% Local Variables: 
 | 
|  |    242 | %%% mode: latex
 | 
|  |    243 | %%% TeX-master: t
 | 
|  |    244 | %%% End: 
 |