(*<*)
theory Paper
imports
"../Lexer"
"../Simplifying"
"../Positions"
"../SizeBound4"
"HOL-Library.LaTeXsugar"
begin
declare [[show_question_marks = false]]
notation (latex output)
If ("(\<^latex>\<open>\\textrm{\<close>if\<^latex>\<open>}\<close> (_)/ \<^latex>\<open>\\textrm{\<close>then\<^latex>\<open>}\<close> (_)/ \<^latex>\<open>\\textrm{\<close>else\<^latex>\<open>}\<close> (_))" 10) and
Cons ("_\<^latex>\<open>\\mbox{$\\,$}\<close>::\<^latex>\<open>\\mbox{$\\,$}\<close>_" [75,73] 73)
abbreviation
"der_syn r c \<equiv> der c r"
abbreviation
"ders_syn r s \<equiv> ders s r"
abbreviation
"bder_syn r c \<equiv> bder c r"
notation (latex output)
der_syn ("_\\_" [79, 1000] 76) and
ders_syn ("_\\_" [79, 1000] 76) and
bder_syn ("_\\_" [79, 1000] 76) and
bders ("_\\_" [79, 1000] 76) and
bders_simp ("_\\\<^sub>b\<^sub>s\<^sub>i\<^sub>m\<^sub>p _" [79, 1000] 76) and
ZERO ("\<^bold>0" 81) and
ONE ("\<^bold>1" 81) and
CH ("_" [1000] 80) and
ALT ("_ + _" [77,77] 78) and
SEQ ("_ \<cdot> _" [77,77] 78) and
STAR ("_\<^sup>*" [79] 78) and
val.Void ("Empty" 78) and
val.Char ("Char _" [1000] 78) and
val.Left ("Left _" [79] 78) and
val.Right ("Right _" [1000] 78) and
val.Seq ("Seq _ _" [79,79] 78) and
val.Stars ("Stars _" [79] 78) and
Prf ("\<turnstile> _ : _" [75,75] 75) and
Posix ("'(_, _') \<rightarrow> _" [63,75,75] 75) and
flat ("|_|" [75] 74) and
flats ("|_|" [72] 74) and
injval ("inj _ _ _" [79,77,79] 76) and
mkeps ("mkeps _" [79] 76) and
length ("len _" [73] 73) and
set ("_" [73] 73) and
AZERO ("ZERO" 81) and
AONE ("ONE _" [79] 78) and
ACHAR ("CHAR _ _" [79, 79] 80) and
AALTs ("ALTs _ _" [77,77] 78) and
ASEQ ("SEQ _ _ _" [79, 79,79] 78) and
ASTAR ("STAR _ _" [79, 79] 78) and
code ("code _" [79] 74) and
intern ("_\<^latex>\<open>\\mbox{$^\\uparrow$}\<close>" [900] 80) and
erase ("_\<^latex>\<open>\\mbox{$^\\downarrow$}\<close>" [1000] 74) and
bnullable ("bnullable _" [1000] 80) and
bsimp_AALTs ("bsimpALT _ _" [10,1000] 80) and
bsimp_ASEQ ("bsimpSEQ _ _ _" [10,1000,1000] 80) and
bmkeps ("bmkeps _" [1000] 80) and
srewrite ("_\<^latex>\<open>\\mbox{$\\,\\stackrel{s}{\\leadsto}$}\<close> _" [71, 71] 80) and
rrewrites ("_ \<^latex>\<open>\\mbox{$\\,\\leadsto^*$}\<close> _" [71, 71] 80) and
srewrites ("_ \<^latex>\<open>\\mbox{$\\,\\stackrel{s}{\\leadsto}^*$}\<close> _" [71, 71] 80) and
blexer_simp ("blexer\<^sup>+" 1000)
lemma better_retrieve:
shows "rs \<noteq> Nil ==> retrieve (AALTs bs (r#rs)) (Left v) = bs @ retrieve r v"
and "rs \<noteq> Nil ==> retrieve (AALTs bs (r#rs)) (Right v) = bs @ retrieve (AALTs [] rs) v"
apply (metis list.exhaust retrieve.simps(4))
by (metis list.exhaust retrieve.simps(5))
(*>*)
section {* Introduction *}
text {*
In the last fifteen or so years, Brzozowski's derivatives of regular
expressions have sparked quite a bit of interest in the functional
programming and theorem prover communities. The beauty of
Brzozowski's derivatives \cite{Brzozowski1964} is that they are neatly
expressible in any functional language, and easily definable and
reasoned about in theorem provers---the definitions just consist of
inductive datatypes and simple recursive functions. Derivatives of a
regular expression, written @{term "der c r"}, give a simple solution
to the problem of matching a string @{term s} with a regular
expression @{term r}: if the derivative of @{term r} w.r.t.\ (in
succession) all the characters of the string matches the empty string,
then @{term r} matches @{term s} (and {\em vice versa}). We are aware
of a mechanised correctness proof of Brzozowski's derivative-based matcher in HOL4 by
Owens and Slind~\cite{Owens2008}. Another one in Isabelle/HOL is part
of the work by Krauss and Nipkow~\cite{Krauss2011}. And another one
in Coq is given by Coquand and Siles \cite{Coquand2012}.
Also Ribeiro and Du Bois give one in Agda~\cite{RibeiroAgda2017}.
However, there are two difficulties with derivative-based matchers:
First, Brzozowski's original matcher only generates a yes/no answer
for whether a regular expression matches a string or not. This is too
little information in the context of lexing where separate tokens must
be identified and also classified (for example as keywords
or identifiers). Sulzmann and Lu~\cite{Sulzmann2014} overcome this
difficulty by cleverly extending Brzozowski's matching
algorithm. Their extended version generates additional information on
\emph{how} a regular expression matches a string following the POSIX
rules for regular expression matching. They achieve this by adding a
second ``phase'' to Brzozowski's algorithm involving an injection
function. In our own earlier work we provided the formal
specification of what POSIX matching means and proved in Isabelle/HOL
the correctness
of Sulzmann and Lu's extended algorithm accordingly
\cite{AusafDyckhoffUrban2016}.
The second difficulty is that Brzozowski's derivatives can
grow to arbitrarily big sizes. For example if we start with the
regular expression \mbox{@{text "(a + aa)\<^sup>*"}} and take
successive derivatives according to the character $a$, we end up with
a sequence of ever-growing derivatives like
\def\ll{\stackrel{\_\backslash{} a}{\longrightarrow}}
\begin{center}
\begin{tabular}{rll}
$(a + aa)^*$ & $\ll$ & $(\ONE + \ONE{}a) \cdot (a + aa)^*$\\
& $\ll$ & $(\ZERO + \ZERO{}a + \ONE) \cdot (a + aa)^* \;+\; (\ONE + \ONE{}a) \cdot (a + aa)^*$\\
& $\ll$ & $(\ZERO + \ZERO{}a + \ZERO) \cdot (a + aa)^* + (\ONE + \ONE{}a) \cdot (a + aa)^* \;+\; $\\
& & $\qquad(\ZERO + \ZERO{}a + \ONE) \cdot (a + aa)^* + (\ONE + \ONE{}a) \cdot (a + aa)^*$\\
& $\ll$ & \ldots \hspace{15mm}(regular expressions of sizes 98, 169, 283, 468, 767, \ldots)
\end{tabular}
\end{center}
\noindent where after around 35 steps we run out of memory on a
typical computer (we shall define shortly the precise details of our
regular expressions and the derivative operation). Clearly, the
notation involving $\ZERO$s and $\ONE$s already suggests
simplification rules that can be applied to regular regular
expressions, for example $\ZERO{}\,r \Rightarrow \ZERO$, $\ONE{}\,r
\Rightarrow r$, $\ZERO{} + r \Rightarrow r$ and $r + r \Rightarrow
r$. While such simple-minded simplifications have been proved in our
earlier work to preserve the correctness of Sulzmann and Lu's
algorithm \cite{AusafDyckhoffUrban2016}, they unfortunately do
\emph{not} help with limiting the growth of the derivatives shown
above: the growth is slowed, but the derivatives can still grow rather
quickly beyond any finite bound.
Sulzmann and Lu overcome this ``growth problem'' in a second algorithm
\cite{Sulzmann2014} where they introduce bitcoded
regular expressions. In this version, POSIX values are
represented as bitsequences and such sequences are incrementally generated
when derivatives are calculated. The compact representation
of bitsequences and regular expressions allows them to define a more
``aggressive'' simplification method that keeps the size of the
derivatives finitely bounded no matter what the length of the string is.
They make some informal claims about the correctness and linear behaviour
of this version, but do not provide any supporting proof arguments, not
even ``pencil-and-paper'' arguments. They write about their bitcoded
\emph{incremental parsing method} (that is the algorithm to be formalised
in this paper):
\begin{quote}\it
``Correctness Claim: We further claim that the incremental parsing
method [..] in combination with the simplification steps [..]
yields POSIX parse trees. We have tested this claim
extensively [..] but yet
have to work out all proof details.'' \cite[Page 14]{Sulzmann2014}
\end{quote}
\noindent{}\textbf{Contributions:} We have implemented in Isabelle/HOL
the derivative-based lexing algorithm of Sulzmann and Lu
\cite{Sulzmann2014} where regular expressions are annotated with
bitsequences. We define the crucial simplification function as a
recursive function, without the need of a fix-point operation. One objective of
the simplification function is to remove duplicates of regular
expressions. For this Sulzmann and Lu use in their paper the standard
@{text nub} function from Haskell's list library, but this function
does not achieve the intended objective with bitcoded regular expressions. The
reason is that in the bitcoded setting, each copy generally has a
different bitcode annotation---so @{text nub} would never ``fire''.
Inspired by Scala's library for lists, we shall instead use a @{text
distinctBy} function that finds duplicates under an erasing function
which deletes bitcodes.
We shall also introduce our own argument and definitions for
establishing the correctness of the bitcoded algorithm when
simplifications are included.\medskip
\noindent In this paper, we shall first briefly introduce the basic notions
of regular expressions and describe the definition
of POSIX lexing from our earlier work \cite{AusafDyckhoffUrban2016}. This serves
as a reference point for what correctness means in our Isabelle/HOL proofs. We shall then prove
the correctness for the bitcoded algorithm without simplification, and
after that extend the proof to include simplification.
*}
section {* Background *}
text {*
In our Isabelle/HOL formalisation strings are lists of characters with
the empty string being represented by the empty list, written $[]$,
and list-cons being written as $\_\!::\!\_\,$; string
concatenation is $\_ \,@\, \_\,$. We often use the usual
bracket notation for lists also for strings; for example a string
consisting of just a single character $c$ is written $[c]$.
Our regular expressions are defined as usual as the elements of the following inductive
datatype:
\begin{center}
@{text "r ::="} \;
@{const "ZERO"} $\mid$
@{const "ONE"} $\mid$
@{term "CH c"} $\mid$
@{term "ALT r\<^sub>1 r\<^sub>2"} $\mid$
@{term "SEQ r\<^sub>1 r\<^sub>2"} $\mid$
@{term "STAR r"}
\end{center}
\noindent where @{const ZERO} stands for the regular expression that does
not match any string, @{const ONE} for the regular expression that matches
only the empty string and @{term c} for matching a character literal.
The constructors $+$ and $\cdot$ represent alternatives and sequences, respectively.
We sometimes omit the $\cdot$ in a sequence regular expression for brevity.
The
\emph{language} of a regular expression, written $L(r)$, is defined as usual
and we omit giving the definition here (see for example \cite{AusafDyckhoffUrban2016}).
Central to Brzozowski's regular expression matcher are two functions
called @{text nullable} and \emph{derivative}. The latter is written
$r\backslash c$ for the derivative of the regular expression $r$
w.r.t.~the character $c$. Both functions are defined by recursion over
regular expressions.
\begin{center}
\begin{tabular}{cc}
\begin{tabular}{r@ {\hspace{2mm}}c@ {\hspace{2mm}}l}
@{thm (lhs) der.simps(1)} & $\dn$ & @{thm (rhs) der.simps(1)}\\
@{thm (lhs) der.simps(2)} & $\dn$ & @{thm (rhs) der.simps(2)}\\
@{thm (lhs) der.simps(3)} & $\dn$ & @{thm (rhs) der.simps(3)}\\
@{thm (lhs) der.simps(4)[of c "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) der.simps(4)[of c "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) der.simps(5)[of c "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{text "if"} @{term "nullable(r\<^sub>1)"}\\
& & @{text "then"} @{term "ALT (SEQ (der c r\<^sub>1) r\<^sub>2) (der c r\<^sub>2)"}\\
& & @{text "else"} @{term "SEQ (der c r\<^sub>1) r\<^sub>2"}\\
% & & @{thm (rhs) der.simps(5)[of c "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) der.simps(6)} & $\dn$ & @{thm (rhs) der.simps(6)}
\end{tabular}
&
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) nullable.simps(1)} & $\dn$ & @{thm (rhs) nullable.simps(1)}\\
@{thm (lhs) nullable.simps(2)} & $\dn$ & @{thm (rhs) nullable.simps(2)}\\
@{thm (lhs) nullable.simps(3)} & $\dn$ & @{thm (rhs) nullable.simps(3)}\\
@{thm (lhs) nullable.simps(4)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) nullable.simps(4)[of "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) nullable.simps(5)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) nullable.simps(5)[of "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) nullable.simps(6)} & $\dn$ & @{thm (rhs) nullable.simps(6)}\medskip\\
\end{tabular}
\end{tabular}
\end{center}
\noindent
We can extend this definition to give derivatives w.r.t.~strings:
\begin{center}
\begin{tabular}{cc}
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) ders.simps(1)} & $\dn$ & @{thm (rhs) ders.simps(1)}
\end{tabular}
&
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) ders.simps(2)} & $\dn$ & @{thm (rhs) ders.simps(2)}
\end{tabular}
\end{tabular}
\end{center}
\noindent
Using @{text nullable} and the derivative operation, we can
define the following simple regular expression matcher:
%
\[
@{text "match s r"} \;\dn\; @{term nullable}(r\backslash s)
\]
\noindent This is essentially Brzozowski's algorithm from 1964. Its
main virtue is that the algorithm can be easily implemented as a
functional program (either in a functional programming language or in
a theorem prover). The correctness proof for @{text match} amounts to
establishing the property
%
\begin{proposition}\label{matchcorr}
@{text "match s r"} \;\;\text{if and only if}\;\; $s \in L(r)$
\end{proposition}
\noindent
It is a fun exercise to formally prove this property in a theorem prover.
The novel idea of Sulzmann and Lu is to extend this algorithm for
lexing, where it is important to find out which part of the string
is matched by which part of the regular expression.
For this Sulzmann and Lu presented two lexing algorithms in their paper
\cite{Sulzmann2014}. The first algorithm consists of two phases: first a
matching phase (which is Brzozowski's algorithm) and then a value
construction phase. The values encode \emph{how} a regular expression
matches a string. \emph{Values} are defined as the inductive datatype
\begin{center}
@{text "v :="}
@{const "Void"} $\mid$
@{term "val.Char c"} $\mid$
@{term "Left v"} $\mid$
@{term "Right v"} $\mid$
@{term "Seq v\<^sub>1 v\<^sub>2"} $\mid$
@{term "Stars vs"}
\end{center}
\noindent where we use @{term vs} to stand for a list of values. The
string underlying a value can be calculated by a @{const flat}
function, written @{term "flat DUMMY"}. It traverses a value and
collects the characters contained in it. Sulzmann and Lu also define inductively an
inhabitation relation that associates values to regular expressions:
\begin{center}
\begin{tabular}{c}
\\[-8mm]
@{thm[mode=Axiom] Prf.intros(4)} \qquad
@{thm[mode=Axiom] Prf.intros(5)[of "c"]}\\[4mm]
@{thm[mode=Rule] Prf.intros(2)[of "v\<^sub>1" "r\<^sub>1" "r\<^sub>2"]} \qquad
@{thm[mode=Rule] Prf.intros(3)[of "v\<^sub>2" "r\<^sub>2" "r\<^sub>1"]}\\[4mm]
@{thm[mode=Rule] Prf.intros(1)[of "v\<^sub>1" "r\<^sub>1" "v\<^sub>2" "r\<^sub>2"]} \qquad
@{thm[mode=Rule] Prf.intros(6)[of "vs" "r"]}
\end{tabular}
\end{center}
\noindent Note that no values are associated with the regular expression
@{term ZERO}, since it cannot match any string.
It is routine to establish how values ``inhabiting'' a regular
expression correspond to the language of a regular expression, namely
\begin{proposition}
@{thm L_flat_Prf}
\end{proposition}
In general there is more than one value inhabited by a regular
expression (meaning regular expressions can typically match more
than one string). But even when fixing a string from the language of the
regular expression, there are generally more than one way of how the
regular expression can match this string. POSIX lexing is about
identifying the unique value for a given regular expression and a
string that satisfies the informal POSIX rules (see
\cite{POSIX,Kuklewicz,OkuiSuzuki2010,Sulzmann2014,Vansummeren2006}).\footnote{POSIX
lexing acquired its name from the fact that the corresponding
rules were described as part of the POSIX specification for
Unix-like operating systems \cite{POSIX}.} Sometimes these
informal rules are called \emph{maximal munch rule} and \emph{rule priority}.
One contribution of our earlier paper is to give a convenient
specification for what POSIX values are (the inductive rules are shown in
Figure~\ref{POSIXrules}).
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
@{thm[mode=Axiom] Posix.intros(1)}\<open>P\<close>@{term "ONE"} \qquad
@{thm[mode=Axiom] Posix.intros(2)}\<open>P\<close>@{term "c"}\medskip\\
@{thm[mode=Rule] Posix.intros(3)[of "s" "r\<^sub>1" "v" "r\<^sub>2"]}\<open>P+L\<close>\qquad
@{thm[mode=Rule] Posix.intros(4)[of "s" "r\<^sub>2" "v" "r\<^sub>1"]}\<open>P+R\<close>\medskip\\
$\mprset{flushleft}
\inferrule
{@{thm (prem 1) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]} \qquad
@{thm (prem 2) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]} \\\\
@{thm (prem 3) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]}}
{@{thm (concl) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]}}$\<open>PS\<close>\medskip\smallskip\\
@{thm[mode=Axiom] Posix.intros(7)}\<open>P[]\<close>\qquad
$\mprset{flushleft}
\inferrule
{@{thm (prem 1) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]} \qquad
@{thm (prem 2) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]} \qquad
@{thm (prem 3) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]} \\\\
@{thm (prem 4) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]}}
{@{thm (concl) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]}}$\<open>P\<star>\<close>\\[-4mm]
\end{tabular}
\end{center}
\caption{The inductive definition of POSIX values taken from our earlier paper \cite{AusafDyckhoffUrban2016}. The ternary relation, written $(s, r) \rightarrow v$, formalises the notion
of given a string $s$ and a regular
expression $r$ what is the unique value $v$ that satisfies the informal POSIX constraints for
regular expression matching.}\label{POSIXrules}
\end{figure}
The clever idea by Sulzmann and Lu \cite{Sulzmann2014} in their first algorithm is to define
an injection function on values that mirrors (but inverts) the
construction of the derivative on regular expressions. Essentially it
injects back a character into a value.
For this they define two functions called @{text mkeps} and @{text inj}:
\begin{center}
\begin{tabular}{l}
\begin{tabular}{lcl}
@{thm (lhs) mkeps.simps(1)} & $\dn$ & @{thm (rhs) mkeps.simps(1)}\\
@{thm (lhs) mkeps.simps(2)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) mkeps.simps(2)[of "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) mkeps.simps(3)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) mkeps.simps(3)[of "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) mkeps.simps(4)} & $\dn$ & @{thm (rhs) mkeps.simps(4)}\\
\end{tabular}\smallskip\\
\begin{tabular}{lcl}
@{thm (lhs) injval.simps(1)} & $\dn$ & @{thm (rhs) injval.simps(1)}\\
@{thm (lhs) injval.simps(2)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1"]} & $\dn$ &
@{thm (rhs) injval.simps(2)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1"]}\\
@{thm (lhs) injval.simps(3)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]} & $\dn$ &
@{thm (rhs) injval.simps(3)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]}\\
@{thm (lhs) injval.simps(4)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]} & $\dn$
& @{thm (rhs) injval.simps(4)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]}\\
@{thm (lhs) injval.simps(5)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]} & $\dn$
& @{thm (rhs) injval.simps(5)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]}\\
@{thm (lhs) injval.simps(6)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]} & $\dn$
& @{thm (rhs) injval.simps(6)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]}\\
@{thm (lhs) injval.simps(7)[of "r" "c" "v" "vs"]} & $\dn$
& @{thm (rhs) injval.simps(7)[of "r" "c" "v" "vs"]}
\end{tabular}
\end{tabular}
\end{center}
\noindent
The function @{text mkeps} is run when the last derivative is nullable, that is
the string to be matched is in the language of the regular expression. It generates
a value for how the last derivative can match the empty string. The injection function
then calculates the corresponding value for each intermediate derivative until
a value for the original regular expression is generated.
Graphically the algorithm by
Sulzmann and Lu can be illustrated by the picture in Figure~\ref{Sulz}
where the path from the left to the right involving @{term derivatives}/@{const
nullable} is the first phase of the algorithm (calculating successive
\Brz's derivatives) and @{const mkeps}/@{text inj}, the path from right to
left, the second phase. The picture above shows the steps required when a
regular expression, say @{text "r\<^sub>1"}, matches the string @{term
"[a,b,c]"}. The first lexing algorithm by Sulzmann and Lu can be defined as:
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}[scale=2,node distance=1.3cm,
every node/.style={minimum size=6mm}]
\node (r1) {@{term "r\<^sub>1"}};
\node (r2) [right=of r1]{@{term "r\<^sub>2"}};
\draw[->,line width=1mm](r1)--(r2) node[above,midway] {@{term "der a DUMMY"}};
\node (r3) [right=of r2]{@{term "r\<^sub>3"}};
\draw[->,line width=1mm](r2)--(r3) node[above,midway] {@{term "der b DUMMY"}};
\node (r4) [right=of r3]{@{term "r\<^sub>4"}};
\draw[->,line width=1mm](r3)--(r4) node[above,midway] {@{term "der c DUMMY"}};
\draw (r4) node[anchor=west] {\;\raisebox{3mm}{@{term nullable}}};
\node (v4) [below=of r4]{@{term "v\<^sub>4"}};
\draw[->,line width=1mm](r4) -- (v4);
\node (v3) [left=of v4] {@{term "v\<^sub>3"}};
\draw[->,line width=1mm](v4)--(v3) node[below,midway] {\<open>inj r\<^sub>3 c\<close>};
\node (v2) [left=of v3]{@{term "v\<^sub>2"}};
\draw[->,line width=1mm](v3)--(v2) node[below,midway] {\<open>inj r\<^sub>2 b\<close>};
\node (v1) [left=of v2] {@{term "v\<^sub>1"}};
\draw[->,line width=1mm](v2)--(v1) node[below,midway] {\<open>inj r\<^sub>1 a\<close>};
\draw (r4) node[anchor=north west] {\;\raisebox{-8mm}{@{term "mkeps"}}};
\end{tikzpicture}
\end{center}
\mbox{}\\[-13mm]
\caption{The two phases of the first algorithm by Sulzmann \& Lu \cite{Sulzmann2014},
matching the string @{term "[a,b,c]"}. The first phase (the arrows from
left to right) is \Brz's matcher building successive derivatives. If the
last regular expression is @{term nullable}, then the functions of the
second phase are called (the top-down and right-to-left arrows): first
@{term mkeps} calculates a value @{term "v\<^sub>4"} witnessing
how the empty string has been recognised by @{term "r\<^sub>4"}. After
that the function @{term inj} ``injects back'' the characters of the string into
the values. The value @{term "v\<^sub>1"} is the result of the algorithm representing
the POSIX value for this string and
regular expression.
\label{Sulz}}
\end{figure}
\begin{center}
\begin{tabular}{lcl}
@{thm (lhs) lexer.simps(1)} & $\dn$ & @{thm (rhs) lexer.simps(1)}\\
@{thm (lhs) lexer.simps(2)} & $\dn$ & @{text "case"} @{term "lexer (der c r) s"} @{text of}\\
& & \phantom{$|$} @{term "None"} @{text "\<Rightarrow>"} @{term None}\\
& & $|$ @{term "Some v"} @{text "\<Rightarrow>"} @{term "Some (injval r c v)"}
\end{tabular}
\end{center}
We have shown in our earlier paper \cite{AusafDyckhoffUrban2016} that
this algorithm is correct, that is it generates POSIX values. The
central property we established relates the derivative operation to the
injection function.
\begin{proposition}\label{Posix2}
\textit{If} $(s,\; r\backslash c) \rightarrow v$ \textit{then} $(c :: s,\; r) \rightarrow$ \textit{inj} $r\; c\; v$.
\end{proposition}
\noindent
With this in place we were able to prove:
\begin{proposition}\mbox{}\smallskip\\\label{lexercorrect}
\begin{tabular}{ll}
(1) & @{thm (lhs) lexer_correct_None} if and only if @{thm (rhs) lexer_correct_None}\\
(2) & @{thm (lhs) lexer_correct_Some} if and only if @{thm (rhs) lexer_correct_Some}\\
\end{tabular}
\end{proposition}
\noindent
In fact we have shown that, in the success case, the generated POSIX value $v$ is
unique and in the failure case that there is no POSIX value $v$ that satisfies
$(s, r) \rightarrow v$. While the algorithm is correct, it is excruciatingly
slow in cases where the derivatives grow arbitrarily (recall the example from the
Introduction). However it can be used as a convenient reference point for the correctness
proof of the second algorithm by Sulzmann and Lu, which we shall describe next.
*}
section {* Bitcoded Regular Expressions and Derivatives *}
text {*
In the second part of their paper \cite{Sulzmann2014},
Sulzmann and Lu describe another algorithm that also generates POSIX
values but dispenses with the second phase where characters are
injected ``back'' into values. For this they annotate bitcodes to
regular expressions, which we define in Isabelle/HOL as the datatype
\begin{center}
\begin{tabular}{lcl}
@{term breg} & $::=$ & @{term "AZERO"} $\quad\mid\quad$ @{term "AONE bs"}\\
& $\mid$ & @{term "ACHAR bs c"}\\
& $\mid$ & @{term "AALTs bs rs"}\\
& $\mid$ & @{term "ASEQ bs r\<^sub>1 r\<^sub>2"}\\
& $\mid$ & @{term "ASTAR bs r"}
\end{tabular}
\end{center}
\noindent where @{text bs} stands for bitsequences; @{text r},
@{text "r\<^sub>1"} and @{text "r\<^sub>2"} for bitcoded regular
expressions; and @{text rs} for lists of bitcoded regular
expressions. The binary alternative @{text "ALT bs r\<^sub>1 r\<^sub>2"}
is just an abbreviation for \mbox{@{text "ALTs bs [r\<^sub>1, r\<^sub>2]"}}.
For bitsequences we use lists made up of the
constants @{text Z} and @{text S}. The idea with bitcoded regular
expressions is to incrementally generate the value information (for
example @{text Left} and @{text Right}) as bitsequences. For this
Sulzmann and Lu define a coding
function for how values can be coded into bitsequences.
\begin{center}
\begin{tabular}{cc}
\begin{tabular}{lcl}
@{thm (lhs) code.simps(1)} & $\dn$ & @{thm (rhs) code.simps(1)}\\
@{thm (lhs) code.simps(2)} & $\dn$ & @{thm (rhs) code.simps(2)}\\
@{thm (lhs) code.simps(3)} & $\dn$ & @{thm (rhs) code.simps(3)}\\
@{thm (lhs) code.simps(4)} & $\dn$ & @{thm (rhs) code.simps(4)}
\end{tabular}
&
\begin{tabular}{lcl}
@{thm (lhs) code.simps(5)[of "v\<^sub>1" "v\<^sub>2"]} & $\dn$ & @{thm (rhs) code.simps(5)[of "v\<^sub>1" "v\<^sub>2"]}\\
@{thm (lhs) code.simps(6)} & $\dn$ & @{thm (rhs) code.simps(6)}\\
@{thm (lhs) code.simps(7)} & $\dn$ & @{thm (rhs) code.simps(7)}\\
\mbox{\phantom{XX}}\\
\end{tabular}
\end{tabular}
\end{center}
\noindent
As can be seen, this coding is ``lossy'' in the sense that we do not
record explicitly character values and also not sequence values (for
them we just append two bitsequences). However, the
different alternatives for @{text Left}, respectively @{text Right}, are recorded as @{text Z} and
@{text S} followed by some bitsequence. Similarly, we use @{text Z} to indicate
if there is still a value coming in the list of @{text Stars}, whereas @{text S}
indicates the end of the list. The lossiness makes the process of
decoding a bit more involved, but the point is that if we have a
regular expression \emph{and} a bitsequence of a corresponding value,
then we can always decode the value accurately. The decoding can be
defined by using two functions called $\textit{decode}'$ and
\textit{decode}:
\begin{center}
\begin{tabular}{@ {}l@ {\hspace{1mm}}c@ {\hspace{1mm}}l@ {}}
$\textit{decode}'\,bs\,(\ONE)$ & $\dn$ & $(\Empty, bs)$\\
$\textit{decode}'\,bs\,(c)$ & $\dn$ & $(\Char\,c, bs)$\\
$\textit{decode}'\,(\Z\!::\!bs)\;(r_1 + r_2)$ & $\dn$ &
$\textit{let}\,(v, bs_1) = \textit{decode}'\,bs\,r_1\;\textit{in}\;
(\Left\,v, bs_1)$\\
$\textit{decode}'\,(\S\!::\!bs)\;(r_1 + r_2)$ & $\dn$ &
$\textit{let}\,(v, bs_1) = \textit{decode}'\,bs\,r_2\;\textit{in}\;
(\Right\,v, bs_1)$\\
$\textit{decode}'\,bs\;(r_1\cdot r_2)$ & $\dn$ &
$\textit{let}\,(v_1, bs_1) = \textit{decode}'\,bs\,r_1\;\textit{in}$\\
& & $\textit{let}\,(v_2, bs_2) = \textit{decode}'\,bs_1\,r_2$
\hspace{2mm}$\textit{in}\;(\Seq\,v_1\,v_2, bs_2)$\\
$\textit{decode}'\,(\Z\!::\!bs)\,(r^*)$ & $\dn$ & $(\Stars\,[], bs)$\\
$\textit{decode}'\,(\S\!::\!bs)\,(r^*)$ & $\dn$ &
$\textit{let}\,(v, bs_1) = \textit{decode}'\,bs\,r\;\textit{in}$\\
& & $\textit{let}\,(\Stars\,vs, bs_2) = \textit{decode}'\,bs_1\,r^*$
\hspace{2mm}$\textit{in}\;(\Stars\,v\!::\!vs, bs_2)$\bigskip\\
$\textit{decode}\,bs\,r$ & $\dn$ &
$\textit{let}\,(v, bs') = \textit{decode}'\,bs\,r\;\textit{in}$\\
& & \hspace{7mm}$\textit{if}\;bs' = []\;\textit{then}\;\textit{Some}\,v\;
\textit{else}\;\textit{None}$
\end{tabular}
\end{center}
\noindent
The function \textit{decode} checks whether all of the bitsequence is
consumed and returns the corresponding value as @{term "Some v"}; otherwise
it fails with @{text "None"}. We can establish that for a value $v$
inhabited by a regular expression $r$, the decoding of its
bitsequence never fails.
\begin{lemma}\label{codedecode}\it
If $\;\vdash v : r$ then
$\;\textit{decode}\,(\textit{code}\, v)\,r = \textit{Some}\, v$.
\end{lemma}
\begin{proof}
This follows from the property that
$\textit{decode}'\,((\textit{code}\,v) \,@\, bs)\,r = (v, bs)$ holds
for any bit-sequence $bs$ and $\vdash v : r$. This property can be
easily proved by induction on $\vdash v : r$.
\end{proof}
Sulzmann and Lu define the function \emph{internalise}
in order to transform (standard) regular expressions into annotated
regular expressions. We write this operation as $r^\uparrow$.
This internalisation uses the following
\emph{fuse} function.
\begin{center}
\begin{tabular}{lcl}
$\textit{fuse}\,bs\,(\textit{ZERO})$ & $\dn$ & $\textit{ZERO}$\\
$\textit{fuse}\,bs\,(\textit{ONE}\,bs')$ & $\dn$ &
$\textit{ONE}\,(bs\,@\,bs')$\\
$\textit{fuse}\,bs\,(\textit{CHAR}\,bs'\,c)$ & $\dn$ &
$\textit{CHAR}\,(bs\,@\,bs')\,c$\\
$\textit{fuse}\,bs\,(\textit{ALTs}\,bs'\,rs)$ & $\dn$ &
$\textit{ALTs}\,(bs\,@\,bs')\,rs$\\
$\textit{fuse}\,bs\,(\textit{SEQ}\,bs'\,r_1\,r_2)$ & $\dn$ &
$\textit{SEQ}\,(bs\,@\,bs')\,r_1\,r_2$\\
$\textit{fuse}\,bs\,(\textit{STAR}\,bs'\,r)$ & $\dn$ &
$\textit{STAR}\,(bs\,@\,bs')\,r$
\end{tabular}
\end{center}
\noindent
A regular expression can then be \emph{internalised} into a bitcoded
regular expression as follows:
\begin{center}
\begin{tabular}{lcl}
$(\ZERO)^\uparrow$ & $\dn$ & $\textit{ZERO}$\\
$(\ONE)^\uparrow$ & $\dn$ & $\textit{ONE}\,[]$\\
$(c)^\uparrow$ & $\dn$ & $\textit{CHAR}\,[]\,c$\\
$(r_1 + r_2)^\uparrow$ & $\dn$ &
$\textit{ALT}\;[]\,(\textit{fuse}\,[\Z]\,r_1^\uparrow)\,
(\textit{fuse}\,[\S]\,r_2^\uparrow)$\\
$(r_1\cdot r_2)^\uparrow$ & $\dn$ &
$\textit{SEQ}\;[]\,r_1^\uparrow\,r_2^\uparrow$\\
$(r^*)^\uparrow$ & $\dn$ &
$\textit{STAR}\;[]\,r^\uparrow$\\
\end{tabular}
\end{center}
\noindent
There is also an \emph{erase}-function, written $r^\downarrow$, which
transforms a bitcoded regular expression into a (standard) regular
expression by just erasing the annotated bitsequences. We omit the
straightforward definition. For defining the algorithm, we also need
the functions \textit{bnullable} and \textit{bmkeps}(\textit{s}), which are the
``lifted'' versions of \textit{nullable} and \textit{mkeps} acting on
bitcoded regular expressions.
%
\begin{center}
\begin{tabular}{@ {}c@ {}c@ {}}
\begin{tabular}{@ {}l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
$\textit{bnullable}\,(\textit{ZERO})$ & $\dn$ & $\textit{False}$\\
$\textit{bnullable}\,(\textit{ONE}\,bs)$ & $\dn$ & $\textit{True}$\\
$\textit{bnullable}\,(\textit{CHAR}\,bs\,c)$ & $\dn$ & $\textit{False}$\\
$\textit{bnullable}\,(\textit{ALTs}\,bs\,\rs)$ & $\dn$ &
$\exists\, r \in \rs. \,\textit{bnullable}\,r$\\
$\textit{bnullable}\,(\textit{SEQ}\,bs\,r_1\,r_2)$ & $\dn$ &
$\textit{bnullable}\,r_1\wedge \textit{bnullable}\,r_2$\\
$\textit{bnullable}\,(\textit{STAR}\,bs\,r)$ & $\dn$ &
$\textit{True}$
\end{tabular}
&
\begin{tabular}{@ {}l@ {\hspace{1mm}}c@ {\hspace{1mm}}l@ {}}
$\textit{bmkeps}\,(\textit{ONE}\,bs)$ & $\dn$ & $bs$\\
$\textit{bmkeps}\,(\textit{ALTs}\,bs\,\rs)$ & $\dn$ &
$bs\,@\,\textit{bmkepss}\,\rs$\\
$\textit{bmkeps}\,(\textit{SEQ}\,bs\,r_1\,r_2)$ & $\dn$ &\\
\multicolumn{3}{r}{$bs \,@\,\textit{bmkeps}\,r_1\,@\, \textit{bmkeps}\,r_2$}\\
$\textit{bmkeps}\,(\textit{STAR}\,bs\,r)$ & $\dn$ &
$bs \,@\, [\S]$\\
$\textit{bmkepss}\,(r\!::\!\rs)$ & $\dn$ &
$\textit{if}\;\textit{bnullable}\,r$\\
& &$\textit{then}\;\textit{bmkeps}\,r$\\
& &$\textit{else}\;\textit{bmkepss}\,\rs$
\end{tabular}
\end{tabular}
\end{center}
\noindent
The key function in the bitcoded algorithm is the derivative of a
bitcoded regular expression. This derivative function calculates the
derivative but at the same time also the incremental part of the bitsequences
that contribute to constructing a POSIX value.
\begin{center}
\begin{tabular}{@ {}lcl@ {}}
$(\textit{ZERO})\backslash c$ & $\dn$ & $\textit{ZERO}$ \\
$(\textit{ONE}\;bs)\backslash c$ & $\dn$ & $\textit{ZERO}$\\
$(\textit{CHAR}\;bs\,d)\backslash c$ & $\dn$ &
$\textit{if}\;c=d\; \;\textit{then}\;
\textit{ONE}\;bs\;\textit{else}\;\textit{ZERO}$\\
$(\textit{ALTs}\;bs\,\rs)\backslash c$ & $\dn$ &
$\textit{ALTs}\,bs\,(\mathit{map}\,(\_\backslash c)\,\rs)$\\
$(\textit{SEQ}\;bs\,r_1\,r_2)\backslash c$ & $\dn$ &
$\textit{if}\;\textit{bnullable}\,r_1$\\
& &$\textit{then}\;\textit{ALT}\,bs\,(\textit{SEQ}\,[]\,(r_1\backslash c)\,r_2)$\\
& &$\phantom{\textit{then}\;\textit{ALT}\,bs\,}(\textit{fuse}\,(\textit{bmkeps}\,r_1)\,(r_2\backslash c))$\\
& &$\textit{else}\;\textit{SEQ}\,bs\,(r_1\backslash c)\,r_2$\\
$(\textit{STAR}\,bs\,r)\backslash c$ & $\dn$ &
$\textit{SEQ}\;bs\,(\textit{fuse}\, [\Z] (r\backslash c))\,
(\textit{STAR}\,[]\,r)$
\end{tabular}
\end{center}
\noindent
This function can also be extended to strings, written $r\backslash s$,
just like the standard derivative. We omit the details. Finally we
can define Sulzmann and Lu's bitcoded lexer, which we call \textit{blexer}:
\begin{center}
\begin{tabular}{lcl}
$\textit{blexer}\;r\,s$ & $\dn$ &
$\textit{let}\;r_{der} = (r^\uparrow)\backslash s\;\textit{in}$\\
& & $\;\;\;\;\textit{if}\; \textit{bnullable}(r_{der}) \;\;\textit{then}\;\textit{decode}\,(\textit{bmkeps}\,r_{der})\,r
\;\;\textit{else}\;\textit{None}$
\end{tabular}
\end{center}
\noindent
This bitcoded lexer first internalises the regular expression $r$ and then
builds the bitcoded derivative according to $s$. If the derivative is
(b)nullable the string is in the language of $r$ and it extracts the bitsequence using the
$\textit{bmkeps}$ function. Finally it decodes the bitsequence into a value. If
the derivative is \emph{not} nullable, then $\textit{None}$ is
returned. We can show that this way of calculating a value
generates the same result as \textit{lexer}.
Before we can proceed we need to define a helper function, called
\textit{retrieve}, which Sulzmann and Lu introduced for the correctness proof.
\begin{center}
\begin{tabular}{lcl}
@{thm (lhs) retrieve.simps(1)} & $\dn$ & @{thm (rhs) retrieve.simps(1)}\\
@{thm (lhs) retrieve.simps(2)} & $\dn$ & @{thm (rhs) retrieve.simps(2)}\\
@{thm (lhs) retrieve.simps(3)} & $\dn$ & @{thm (rhs) retrieve.simps(3)}\\
@{thm (lhs) better_retrieve(1)} & $\dn$ & @{thm (rhs) better_retrieve(1)}\\
@{thm (lhs) better_retrieve(2)} & $\dn$ & @{thm (rhs) better_retrieve(2)}\\
@{thm (lhs) retrieve.simps(6)[of _ "r\<^sub>1" "r\<^sub>2" "v\<^sub>1" "v\<^sub>2"]}
& $\dn$ & @{thm (rhs) retrieve.simps(6)[of _ "r\<^sub>1" "r\<^sub>2" "v\<^sub>1" "v\<^sub>2"]}\\
@{thm (lhs) retrieve.simps(7)} & $\dn$ & @{thm (rhs) retrieve.simps(7)}\\
@{thm (lhs) retrieve.simps(8)} & $\dn$ & @{thm (rhs) retrieve.simps(8)}
\end{tabular}
\end{center}
\noindent
The idea behind this function is to retrieve a possibly partial
bitsequence from a bitcoded regular expression, where the retrieval is
guided by a value. For example if the value is $\Left$ then we
descend into the left-hand side of an alternative in order to
assemble the bitcode. Similarly for
$\Right$. The property we can show is that for a given $v$ and $r$
with $\vdash v : r$, the retrieved bitsequence from the internalised
regular expression is equal to the bitcoded version of $v$.
\begin{lemma}\label{retrievecode}
If $\vdash v : r$ then $\textit{code}\, v = \textit{retrieve}\,(r^\uparrow)\,v$.
\end{lemma}
\noindent
We also need some auxiliary facts about how the bitcoded operations
relate to the ``standard'' operations on regular expressions. For
example if we build a bitcoded derivative and erase the result, this
is the same as if we first erase the bitcoded regular expression and
then perform the ``standard'' derivative operation.
\begin{lemma}\label{bnullable}\mbox{}\smallskip\\
\begin{tabular}{ll}
\textit{(1)} & $(r\backslash s)^\downarrow = (r^\downarrow)\backslash s$\\
\textit{(2)} & $\textit{bnullable}(r)$ iff $\textit{nullable}(r^\downarrow)$\\
\textit{(3)} & $\textit{bmkeps}(r) = \textit{retrieve}\,r\,(\textit{mkeps}\,(r^\downarrow))$ provided $\textit{nullable}(r^\downarrow)$.
\end{tabular}
\end{lemma}
\begin{proof}
All properties are by induction on annotated regular expressions. There are no
interesting cases.
\end{proof}
\noindent
The only difficulty left for the correctness proof is that the bitcoded algorithm
has only a ``forward phase'' where POSIX values are generated incrementally.
We can achieve the same effect with @{text lexer} (which has two phases) by stacking up injection
functions during the forward phase. An auxiliary function, called $\textit{flex}$,
allows us to recast the rules of $\lexer$ in terms of a single
phase and stacked up injection functions.
\begin{center}
\begin{tabular}{lcl}
$\textit{flex}\;r\,f\,[]$ & $\dn$ & $f$\\
$\textit{flex}\;r\,f\,(c\!::\!s)$ & $\dn$ &
$\textit{flex}\,(r\backslash c)\,(\lambda v.\,f\,(\inj\,r\,c\,v))\,s$\\
\end{tabular}
\end{center}
\noindent
The point of this function is that when
reaching the end of the string, we just need to apply the stacked up
injection functions to the value generated by @{text mkeps}.
Using this function we can recast the success case in @{text lexer}
as follows:
\begin{proposition}\label{flex}
If @{text "lexer r s = Some v"} \;then\; @{text "v = "}$\,\textit{flex}\,r\,id\,s\,
(\mkeps (r\backslash s))$.
\end{proposition}
\noindent
Note we did not redefine \textit{lexer}, we just established that the
value generated by \textit{lexer} can also be obtained by a different
method. While this different method is not efficient (we essentially
need to traverse the string $s$ twice, once for building the
derivative $r\backslash s$ and another time for stacking up injection
functions using \textit{flex}), it helps us with proving
that incrementally building up values in @{text blexer} generates the same result.
This brings us to our main lemma in this section: if we calculate a
derivative, say $r\backslash s$, and have a value, say $v$, inhabited
by this derivative, then we can produce the result @{text lexer} generates
by applying this value to the stacked-up injection functions
that $\textit{flex}$ assembles. The lemma establishes that this is the same
value as if we build the annotated derivative $r^\uparrow\backslash s$
and then retrieve the corresponding bitcoded version, followed by a
decoding step.
\begin{lemma}[Main Lemma]\label{mainlemma}\it
If $\vdash v : r\backslash s$ then
\[\textit{Some}\,(\textit{flex}\,r\,\textit{id}\,s\,v) =
\textit{decode}(\textit{retrieve}\,(r^\uparrow \backslash s)\,v)\,r\]
\end{lemma}
\begin{proof}
This can be proved by induction on $s$ and generalising over
$v$. The interesting point is that we need to prove this in the
reverse direction for $s$. This means instead of cases $[]$ and
$c\!::\!s$, we have cases $[]$ and $s\,@\,[c]$ where we unravel the
string from the back.\footnote{Isabelle/HOL provides an induction principle
for this way of performing the induction.}
The case for $[]$ is routine using Lemmas~\ref{codedecode}
and~\ref{retrievecode}. In the case $s\,@\,[c]$, we can infer from
the assumption that $\vdash v : (r\backslash s)\backslash c$
holds. Hence by Prop.~\ref{Posix2} we know that
(*) $\vdash \inj\,(r\backslash s)\,c\,v : r\backslash s$ holds too.
By definition of $\textit{flex}$ we can unfold the left-hand side
to be
\[
\textit{Some}\,(\textit{flex}\;r\,\textit{id}\,(s\,@\,[c])\,v) =
\textit{Some}\,(\textit{flex}\;r\,\textit{id}\,s\,(\inj\,(r\backslash s)\,c\,v))
\]
\noindent
By induction hypothesis and (*) we can rewrite the right-hand side to
%
\[
\textit{decode}\,(\textit{retrieve}\,(r^\uparrow\backslash s)\;
(\inj\,(r\backslash s)\,c\,\,v))\,r
\]
\noindent
which is equal to
$\textit{decode}\,(\textit{retrieve}\, (r^\uparrow\backslash
(s\,@\,[c]))\,v)\,r$ as required. The last rewrite step is possible
because we generalised over $v$ in our induction.
\end{proof}
\noindent
With this lemma in place, we can prove the correctness of \textit{blexer}---it indeed
produces the same result as \textit{lexer}.
\begin{theorem}\label{thmone}
$\textit{lexer}\,r\,s = \textit{blexer}\,r\,s$
\end{theorem}
\begin{proof}
We can first expand both sides using Prop.~\ref{flex} and the
definition of \textit{blexer}. This gives us two
\textit{if}-statements, which we need to show to be equal. By
Lemma~\ref{bnullable}\textit{(2)} we know the \textit{if}-tests coincide:
\[
\textit{bnullable}(r^\uparrow\backslash s) \;\textit{iff}\;
\nullable(r\backslash s)
\]
\noindent
For the \textit{if}-branch suppose $r_d \dn r^\uparrow\backslash s$ and
$d \dn r\backslash s$. We have (*) @{text "nullable d"}. We can then show
by Lemma~\ref{bnullable}\textit{(3)} that
%
\[
\textit{decode}(\textit{bmkeps}\:r_d)\,r =
\textit{decode}(\textit{retrieve}\,r_d\,(\textit{mkeps}\,d))\,r
\]
\noindent
where the right-hand side is equal to
$\textit{Some}\,(\textit{flex}\,r\,\textit{id}\,s\,(\textit{mkeps}\,
d))$ by Lemma~\ref{mainlemma} (we know
$\vdash \textit{mkeps}\,d : d$ by (*)). This shows the
\textit{if}-branches return the same value. In the
\textit{else}-branches both \textit{lexer} and \textit{blexer} return
\textit{None}. Therefore we can conclude the proof.
\end{proof}
\noindent This establishes that the bitcoded algorithm by Sulzmann and
Lu \emph{without} simplification produces correct results. This was
only conjectured by Sulzmann and Lu in their paper
\cite{Sulzmann2014}. The next step is to add simplifications.
*}
section {* Simplification *}
text {*
Derivatives as calculated by Brzozowski’s method are usually more
complex regular expressions than the initial one; the result is
that derivative-based matching and lexing algorithms are
often abysmally slow if the ``growth problem'' is not addressed. As Sulzmann and Lu wrote, various
optimisations are possible, such as the simplifications
$\ZERO{}\,r \Rightarrow \ZERO$, $\ONE{}\,r \Rightarrow r$,
$\ZERO{} + r \Rightarrow r$ and $r + r \Rightarrow r$. While these
simplifications can considerably speed up the two algorithms in many
cases, they do not solve fundamentally the growth problem with
derivatives. To see this let us return to the example from the
Introduction that shows the derivatives for \mbox{@{text "(a + aa)\<^sup>*"}}.
If we delete in the 3rd step all $\ZERO{}s$ and $\ONE$s according to
the simplification rules shown above we obtain
%
\def\xll{\xrightarrow{\_\backslash{} [a, a, a]}}%%
%
\begin{equation}\label{derivex}
(a + aa)^* \quad\xll\quad
\underbrace{\mbox{$(\ONE + a) \cdot (a + aa)^*$}}_{r} \;+\;
((a + aa)^* + \underbrace{\mbox{$(\ONE + a) \cdot (a + aa)^*$}}_{r})
\end{equation}
\noindent This is a simpler derivative, but unfortunately we
cannot make any further simplifications. This is a problem because
the outermost alternatives contains two copies of the same
regular expression (underlined with $r$). These copies will
spawn new copies in later derivative steps and they in turn even more copies. This
destroys any hope of taming the size of the derivatives. But the
second copy of $r$ in \eqref{derivex} will never contribute to a
value, because POSIX lexing will always prefer matching a string
with the first copy. So it could be safely removed without affecting the correctness of the algorithm.
The dilemma with the simple-minded
simplification rules above is that the rule $r + r \Rightarrow r$
will never be applicable because as can be seen in this example the
regular expressions are not next to each other but separated by another regular expression.
But here is where Sulzmann and Lu's representation of generalised
alternatives in the bitcoded algorithm shines: in @{term
"ALTs bs rs"} we can define a more aggressive simplification by
recursively simplifying all regular expressions in @{text rs} and
then analyse the resulting list and remove any duplicates.
Another advantage with the bitsequences in bitcoded regular
expressions is that they can be easily modified such that simplification does not
interfere with the value constructions. For example we can ``flatten'', or
de-nest, @{text ALTs} as follows
%
\[
@{term "ALTs bs\<^sub>1 ((ALTs bs\<^sub>2 rs\<^sub>2) # rs\<^sub>1)"}
\quad\xrightarrow{bsimp}\quad
@{term "ALTs bs\<^sub>1 ((map (fuse bs\<^sub>2) rs\<^sub>2) # rs\<^sub>1)"}
\]
\noindent
where we just need to fuse the bitsequence that has accumulated in @{text "bs\<^sub>2"}
to the alternatives in @{text "rs\<^sub>2"}. As we shall show below this will
ensure that the correct value corresponding to the original (unsimplified)
regular expression can still be extracted. %In this way the value construction
%is not affected by simplification.
However there is one problem with the definition for the more
aggressive simplification rules described by Sulzmann and Lu. Recasting
their definition with our syntax they define the step of removing
duplicates as
%
\[ @{text "bsimp (ALTs bs rs)"} \dn @{text "ALTs
bs (nub (map bsimp rs))"}
\]
\noindent where they first recursively simplify the regular
expressions in @{text rs} (using @{text map}) and then use
Haskell's @{text nub}-function to remove potential
duplicates. While this makes sense when considering the example
shown in \eqref{derivex}, @{text nub} is the inappropriate
function in the case of bitcoded regular expressions. The reason
is that in general the elements in @{text rs} will have a
different annotated bitsequence and in this way @{text nub}
will never find a duplicate to be removed. One correct way to
handle this situation is to first \emph{erase} the regular
expressions when comparing potential duplicates. This is inspired
by Scala's list functions of the form \mbox{@{text "distinctBy rs f
acc"}} where a function is applied first before two elements
are compared. We define this function in Isabelle/HOL as
\begin{center}
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) distinctBy.simps(1)} & $\dn$ & @{thm (rhs) distinctBy.simps(1)}\\
@{thm (lhs) distinctBy.simps(2)} & $\dn$ & @{thm (rhs) distinctBy.simps(2)}
\end{tabular}
\end{center}
\noindent where we scan the list from left to right (because we
have to remove later copies). In @{text distinctBy}, @{text f} is a
function and @{text acc} is an accumulator for regular
expressions---essentially a set of regular expressions that we have already seen
while scanning the list. Therefore we delete an element, say @{text x},
from the list provided @{text "f x"} is already in the accumulator;
otherwise we keep @{text x} and scan the rest of the list but
add @{text "f x"} as another ``seen'' element to @{text acc}. We will use
@{term distinctBy} where @{text f} is the erase function, @{term "erase (DUMMY)"},
that deletes bitsequences from bitcoded regular expressions.
This is clearly a computationally more expensive operation than @{text nub},
but is needed in order to make the removal of unnecessary copies
to work properly.
Our simplification function depends on three helper functions, one is called
@{text flts} and analyses lists of regular expressions coming from alternatives.
It is defined as follows:
\begin{center}
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) flts.simps(1)} & $\dn$ & @{thm (rhs) flts.simps(1)}\\
@{thm (lhs) flts.simps(2)} & $\dn$ & @{thm (rhs) flts.simps(2)}\\
@{thm (lhs) flts.simps(3)[of "bs'" "rs'"]} & $\dn$ & @{thm (rhs) flts.simps(3)[of "bs'" "rs'"]}\\
\end{tabular}
\end{center}
\noindent
The second clause of @{text flts} removes all instances of @{text ZERO} in alternatives and
the third ``spills'' out nested alternatives (but retaining the
bitsequence @{text "bs'"} accumulated in the inner alternative). There are
some corner cases to be considered when the resulting list inside an alternative is
empty or a singleton list. We take care of those cases in the
@{text "bsimpALTs"} function; similarly we define a helper function that simplifies
sequences according to the usual rules about @{text ZERO}s and @{text ONE}s:
\begin{center}
\begin{tabular}{c@ {\hspace{5mm}}c}
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{text "bsimpALTs bs []"} & $\dn$ & @{text "ZERO"}\\
@{text "bsimpALTs bs [r]"} & $\dn$ & @{text "fuse bs r"}\\
@{text "bsimpALTs bs rs"} & $\dn$ & @{text "ALTs bs rs"}\\
\mbox{}\\
\end{tabular}
&
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{text "bsimpSEQ bs _ ZERO"} & $\dn$ & @{text "ZERO"}\\
@{text "bsimpSEQ bs ZERO _"} & $\dn$ & @{text "ZERO"}\\
@{text "bsimpSEQ bs\<^sub>1 (ONE bs\<^sub>2) r\<^sub>2"}
& $\dn$ & @{text "fuse (bs\<^sub>1 @ bs\<^sub>2) r\<^sub>2"}\\
@{text "bsimpSEQ bs r\<^sub>1 r\<^sub>2"} & $\dn$ & @{text "SEQ bs r\<^sub>1 r\<^sub>2"}
\end{tabular}
\end{tabular}
\end{center}
\noindent
With this in place we can define our simplification function as
\begin{center}
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) bsimp.simps(1)[of "bs" "r\<^sub>1" "r\<^sub>2"]} & $\dn$ &
@{thm (rhs) bsimp.simps(1)[of "bs" "r\<^sub>1" "r\<^sub>2"]}\\
@{thm (lhs) bsimp.simps(2)[of "bs" _]} & $\dn$ & @{thm (rhs) bsimp.simps(2)[of "bs" _]}\\
@{text "bsimp r"} & $\dn$ & @{text r}
\end{tabular}
\end{center}
\noindent
As far as we can see, our recursive function @{term bsimp} simplifies regular
expressions as intended by Sulzmann and Lu. There is no point in applying the
@{text bsimp} function repeatedly (like the simplification in their paper which needs to be
applied until a fixpoint is reached) because we can show that @{term bsimp} is idempotent,
that is
\begin{proposition}
@{term "bsimp (bsimp r) = bsimp r"}
\end{proposition}
\noindent
This can be proved by induction on @{text r} but requires a detailed analysis
that the de-nesting of alternatives always results in a flat list of regular
expressions. We omit the details since it does not concern the correctness proof.
Next we can include simplification after each derivative step leading to the
following notion of bitcoded derivatives:
\begin{center}
\begin{tabular}{cc}
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) bders_simp.simps(1)} & $\dn$ & @{thm (rhs) bders_simp.simps(1)}
\end{tabular}
&
\begin{tabular}{l@ {\hspace{1mm}}c@ {\hspace{1mm}}l}
@{thm (lhs) bders_simp.simps(2)} & $\dn$ & @{thm (rhs) bders_simp.simps(2)}
\end{tabular}
\end{tabular}
\end{center}
\noindent
and use it in the improved lexing algorithm defined as
\begin{center}
\begin{tabular}{lcl}
$\textit{blexer}^+\;r\,s$ & $\dn$ &
$\textit{let}\;r_{der} = (r^\uparrow)\backslash_{bsimp}\, s\;\textit{in}$\\
& & $\;\;\;\;\textit{if}\; \textit{bnullable}(r_{der}) \;\;\textit{then}\;\textit{decode}\,(\textit{bmkeps}\,r_{der})\,r
\;\;\textit{else}\;\textit{None}$
\end{tabular}
\end{center}
\noindent The remaining task is to show that @{term blexer} and
@{term "blexer_simp"} generate the same answers.
When we first
attempted this proof we encountered a problem with the idea
in Sulzmann and Lu's paper where the argument seems to be to appeal
again to the @{text retrieve}-function defined for the unsimplified version
of the algorithm. But
this does not work, because desirable properties such as
%
\[
@{text "retrieve r v = retrieve (bsimp r) v"}
\]
\noindent do not hold under simplification---this property
essentially purports that we can retrieve the same value from a
simplified version of the regular expression. To start with @{text retrieve}
depends on the fact that the value @{text v} corresponds to the
structure of the regular expression @{text r}---but the whole point of simplification
is to ``destroy'' this structure by making the regular expression simpler.
To see this consider the regular expression @{text "r = r' + 0"} and a corresponding
value @{text "v = Left v'"}. If we annotate bitcodes to @{text "r"}, then
we can use @{text retrieve} with @{text r} and @{text v} in order to extract a corresponding
bitsequence. The reason that this works is that @{text r} is an alternative
regular expression and @{text v} a corresponding @{text "Left"}-value. However, if we simplify
@{text r}, then @{text v} does not correspond to the shape of the regular
expression anymore. So unless one can somehow
synchronise the change in the simplified regular expressions with
the original POSIX value, there is no hope of appealing to @{text retrieve} in the
correctness argument for @{term blexer_simp}.
We found it more helpful to introduce the rewriting systems shown in
Figure~\ref{SimpRewrites}. The idea is to generate
simplified regular expressions in small steps (unlike the @{text bsimp}-function which
does the same in a big step), and show that each of
the small steps preserves the bitcodes that lead to the final POSIX value.
The rewrite system is organised such that $\leadsto$ is for bitcoded regular
expressions and $\stackrel{s}{\leadsto}$ for lists of bitcoded regular
expressions. The former essentially implements the simplifications of
@{text "bsimpSEQ"} and @{text flts}; while the latter implements the
simplifications in @{text "bsimpALTs"}. We can show that any bitcoded
regular expression reduces in zero or more steps to the simplified
regular expression generated by @{text bsimp}:
\begin{lemma}\label{lemone}
@{thm[mode=IfThen] rewrites_to_bsimp}
\end{lemma}
\begin{proof}
By induction on @{text r}. For this we can use the properties
@{thm fltsfrewrites} and @{thm ss6_stronger}. The latter uses
repeated applications of the $LD$ rule which allows the removal
of duplicates that can recognise the same strings.
\end{proof}
\noindent
We can show that this rewrite system preserves @{term bnullable}, that
is simplification, essentially, does not affect nullability:
\begin{lemma}
@{thm[mode=IfThen] bnullable0(1)[of "r\<^sub>1" "r\<^sub>2"]}
\end{lemma}
\begin{proof}
Straightforward mutual induction on the definition of $\leadsto$ and $\stackrel{s}{\leadsto}$.
The only interesting case is the rule $LD$ where the property holds since by the side-conditions of that rule the empty string will
be in both @{text "L (rs\<^sub>a @ [r\<^sub>1] @ rs\<^sub>b @ [r\<^sub>2] @ rs\<^sub>c)"} and
@{text "L (rs\<^sub>a @ [r\<^sub>1] @ rs\<^sub>b @ rs\<^sub>c)"}.
\end{proof}
\noindent
From this, we can show that @{text bmkeps} will produce the same bitsequence
as long as one of the bitcoded regular expressions in $\leadsto$ is nullable (this lemma
establishes the missing fact we were not able to establish using @{text retrieve}, as suggested
in the paper by Sulzmannn and Lu).
\begin{lemma}\label{lemthree}
@{thm[mode=IfThen] rewrite_bmkeps_aux(1)[of "r\<^sub>1" "r\<^sub>2"]}
\end{lemma}
\begin{proof}
By straightforward mutual induction on the definition of $\leadsto$ and $\stackrel{s}{\leadsto}$.
Again the only interesting case is the rule $LD$ where we need to ensure that
\[
@{text "bmkeps (rs\<^sub>a @ [r\<^sub>1] @ rs\<^sub>b @ [r\<^sub>2] @ rs\<^sub>c) =
bmkeps (rs\<^sub>a @ [r\<^sub>1] @ rs\<^sub>b @ rs\<^sub>c)"}
\]
\noindent holds. This is indeed the case because according to the POSIX rules the
generated bitsequence is determined by the first alternative that can match the
string (in this case being nullable).
\end{proof}
\noindent
Crucial is also the fact that derivative steps and simplification steps can be interleaved,
which is shown by the fact that $\leadsto$ is preserved under derivatives.
\begin{lemma}
@{thm[mode=IfThen] rewrite_preserves_bder(1)[of "r\<^sub>1" "r\<^sub>2"]}
\end{lemma}
\begin{proof}
By straightforward mutual induction on the definition of $\leadsto$ and $\stackrel{s}{\leadsto}$.
The case for $LD$ holds because @{term "L (erase (bder c r\<^sub>2)) \<subseteq> L (erase (bder c r\<^sub>1))"}
if and only if @{term "L (erase (r\<^sub>2)) \<subseteq> L (erase (r\<^sub>1))"}.
\end{proof}
\noindent
Using this fact together with Lemma~\ref{lemone} allows us to prove the central lemma
that the unsimplified
derivative (with a string @{term s}) reduces to the simplified derivative (with the same string).
\begin{lemma}\label{lemtwo}
@{thm[mode=IfThen] central}
\end{lemma}
\begin{proof}
By reverse induction on @{term s} generalising over @{text r}.
\end{proof}
\noindent
With these lemmas in place we can finally establish that @{term "blexer_simp"} and @{term "blexer"}
generate the same value, and using Theorem~\ref{thmone} from the previous section that this value
is indeed the POSIX value.
\begin{theorem}
@{thm[mode=IfThen] main_blexer_simp}
\end{theorem}
\begin{proof}
By unfolding the definitions and using Lemmas~\ref{lemtwo} and \ref{lemthree}.
\end{proof}
\noindent
This completes the correctness proof for the second POSIX lexing algorithm by Sulzmann and Lu.
The interesting point of this algorithm is that the sizes of derivatives do not grow arbitrarily, which
we shall show next.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
@{thm[mode=Axiom] bs1[of _ "r\<^sub>2"]}$S\ZERO{}_l$\qquad
@{thm[mode=Axiom] bs2[of _ "r\<^sub>1"]}$S\ZERO{}_r$\\
@{thm[mode=Axiom] bs3[of "bs\<^sub>1" "bs\<^sub>2"]}$S\ONE$\\
@{thm[mode=Rule] bs4[of "r\<^sub>1" "r\<^sub>2" _ "r\<^sub>3"]}SL\qquad
@{thm[mode=Rule] bs5[of "r\<^sub>3" "r\<^sub>4" _ "r\<^sub>1"]}SR\\
@{thm[mode=Axiom] bs6}$A0$\qquad
@{thm[mode=Axiom] bs7}$A1$\\
@{thm[mode=Rule] bs8[of "rs\<^sub>1" "rs\<^sub>2"]}$AL$\\
@{thm[mode=Rule] ss2[of "rs\<^sub>1" "rs\<^sub>2"]}$LT$\qquad
@{thm[mode=Rule] ss3[of "r\<^sub>1" "r\<^sub>2"]}$LH$\\
@{thm[mode=Axiom] ss4}$L\ZERO$\qquad
@{thm[mode=Axiom] ss5[of "bs" "rs\<^sub>1" "rs\<^sub>2"]}$LS$\medskip\\
@{thm[mode=Rule] ss6[of "r\<^sub>2" "r\<^sub>1" "rs\<^sub>1" "rs\<^sub>2" "rs\<^sub>3"]}$LD$\\
\end{tabular}
\end{center}
\caption{The rewrite rules that generate simplified regular expressions
in small steps: @{term "rrewrite r\<^sub>1 r\<^sub>2"} is for bitcoded regular
expressions and @{term "srewrite rs\<^sub>1 rs\<^sub>2"} for \emph{lists} of bitcoded
regular expressions. Interesting is the $LD$ rule that allows copies of regular
expressions to be removed provided a regular expression earlier in the list can
match the same strings.}\label{SimpRewrites}
\end{figure}
*}
section {* Finiteness of Derivatives *}
text {*
In this section let us sketch our argument for why the size of the simplified
derivatives with the aggressive simplification function is finite. Suppose
we have a size function for bitcoded regular expressions, written
@{text "|r|"}, which counts the number of nodes if we regard $r$ as a tree
(we omit the precise definition). For this we show that for every $r$
there exists a bound $N$
such that
\begin{center}
$\forall s. \; |@{term "bders_simp r s"}| < N$
\end{center}
\noindent
We prove this by induction on $r$. The base cases for @{term AZERO},
@{term "AONE bs"} and @{term "ACHAR bs c"} are straightforward. The interesting case is
for sequences of the form @{term "ASEQ bs r\<^sub>1 r\<^sub>2"}. In this case our induction
hypotheses state $\forall s. \; |@{term "bders_simp r\<^sub>1 s"}| < N_1$ and
$\forall s. \; |@{term "bders_simp r\<^sub>2 s"}| < N_2$. We can reason as follows
\begin{center}
\begin{tabular}{lcll}
& & $ |@{term "bders_simp (ASEQ bs r\<^sub>1 r\<^sub>2) s"}|$\\
& $ = $ & $|bsimp(ALTs\;bs\;((@{term "bders_simp r\<^sub>1 s"}) \cdot r_2) ::
[@{term "bders_simp r\<^sub>2 s'"} \;|\; s' \in Suf\!fix(s)])| $ & (1) \\
& $\leq$ &
$|distinctBy\,(flts\,((@{term "bders_simp r\<^sub>1 s "}) \cdot r_2) ::
[@{term "bders_simp r\<^sub>2 s'"} \;|\; s' \in Suf\!fix(s)])| + 1 $ & (2) \\
& $\leq$ & $|(@{term "bders_simp r\<^sub>1 s"}) \cdot r_2| +
|distinctBy\,(flts\, [@{term "bders_simp r\<^sub>2 s'"} \;|\; s' \in Suf\!fix(s)])| + 1 $ & (3) \\
& $\leq$ & $N_1 + |r_2| + 2 + |distinctBy\,(flts\, [@{term "bders_simp r\<^sub>2 s'"} \;|\; s' \in Suf\!fix(s)])|$ & (4)\\
& $\leq$ & $N_1 + |r_2| + 2 + l_{N_{2}} * N_{2}$ & (5)
\end{tabular}
\end{center}
% tell Chengsong about Indian paper of closed forms of derivatives
\noindent
where in (1) the $Suf\!fix(s')$ are the suffixes where @{term "bders_simp r\<^sub>1 s''"} is nullable for
@{text "s = s'' @ s'"}. In (3) we know that $|(@{term "bders_simp r\<^sub>1 s"}) \cdot r_2|$ is
bounded by $N_1 + |r_2|$. In (5) we know the list comprehension contains only regular expressions of size smaller
than $N_2$. The list length after @{text distinctBy} is bounded by a number, which we call $l_{N_2}$. It stands
for the number of distinct regular expressions with a maximum size $N_2$ (there can only be finitely many of them).
We reason similarly in the @{text Star}-case.\medskip
\noindent
Clearly we give in this finiteness argument (Step (5)) a very loose bound that is
far from the actual bound we can expect. We can do better than this, but this does not improve
the finiteness property we are proving. If we are interested in a polynomial bound,
one would hope to obtain a similar tight bound as for partial
derivatives introduced by Antimirov \cite{Antimirov95}. After all the idea with
@{text distinctBy} is to maintain a ``set'' of alternatives (like the sets in
partial derivatives). Unfortunately to obtain the exact same bound would mean
we need to introduce simplifications such as
%
\[ (r_1 + r_2) \cdot r_3 \longrightarrow (r_1 \cdot r_3) + (r_2 \cdot r_3)
\]
\noindent
which exist for partial derivatives. However, if we introduce them in our
setting we would lose the POSIX property of our calculated values. We leave better
bounds for future work.
*}
section {* Conclusion *}
text {*
We set out in this work to prove in Isabelle/HOL the correctness of
the second POSIX lexing algorithm by Sulzmann and Lu
\cite{Sulzmann2014}. This follows earlier work where we established
the correctness of the first algorithm
\cite{AusafDyckhoffUrban2016}. In the earlier work we needed to
introduce our own specification about what POSIX values are,
because the informal definition given by Sulzmann and Lu did not
stand up to a formal proof. Also for the second algorithm we needed
to introduce our own definitions and proof ideas in order to establish the
correctness. Our interest in the second algorithm
lies in the fact that by using bitcoded regular expressions and an aggressive
simplification method there is a chance that the derivatives
can be kept universally small (we established in this paper that
they can be kept finite for any string). This is important if one is after
an efficient POSIX lexing algorithm based on derivatives.
Having proved the correctness of the POSIX lexing algorithm, which
lessons have we learned? Well, we feel this is a very good example
where formal proofs give further insight into the matter at
hand. For example it is very hard to see a problem with @{text nub}
vs @{text distinctBy} with only experimental data---one would still
see the correct result but find that simplification does not simplify in well-chosen, but not
obscure, examples. We found that from an implementation
point-of-view it is really important to have the formal proofs of
the corresponding properties at hand.
We have also developed a
healthy suspicion when experimental data is used to back up
efficiency claims. For example Sulzmann and Lu write about their
equivalent of @{term blexer_simp} ``...we can incrementally compute
bitcoded parse trees in linear time in the size of the input''
\cite[Page 14]{Sulzmann2014}.
Given the growth of the
derivatives in some cases even after aggressive simplification, this
is a hard to believe claim. A similar claim about a theoretical runtime
of @{text "O(n\<^sup>2)"} is made for the Verbatim lexer, which calculates
tokens according to POSIX rules~\cite{verbatim}. For this Verbatim uses Brzozowski's
derivatives like in our work.
The authors write: ``The results of our empirical tests [..] confirm that Verbatim has
@{text "O(n\<^sup>2)"} time complexity.'' \cite[Section~VII]{verbatim}.
While their correctness proof for Verbatim is formalised in Coq, the claim about
the runtime complexity is only supported by some emperical evidence obtained
by using the code extraction facilities of Coq.
Given our observation with the ``growth problem'' of derivatives,
we
tried out their extracted OCaml code with the example
\mbox{@{text "(a + aa)\<^sup>*"}} as a single lexing rule, and it took for us around 5 minutes to tokenise a
string of 40 $a$'s and that increased to approximately 19 minutes when the
string is 50 $a$'s long. Taking into account that derivatives are not simplified in the Verbatim
lexer, such numbers are not surprising.
Clearly our result of having finite
derivatives might sound rather weak in this context but we think such effeciency claims
really require further scrutiny.\medskip
\noindent
Our Isabelle/HOL code is available under \url{https://github.com/urbanchr/posix}.
%%\bibliographystyle{plain}
\bibliography{root}
*}
(*<*)
end
(*>*)