Binary file hws/hw07.pdf has changed
--- a/hws/hw07.tex Thu Nov 10 23:49:29 2022 +0000
+++ b/hws/hw07.tex Tue Nov 15 11:34:33 2022 +0000
@@ -13,9 +13,9 @@
\begin{center}
\begin{tabular}{lcl}
-$S$ & $\rightarrow$ & $bSAA\;|\; \epsilon$\\
-$A$ & $\rightarrow$ & $a$\\
-$bA$ & $\rightarrow$ & $Ab$\\
+$S$ & $::=$ & $bSAA\;|\; \epsilon$\\
+$A$ & $::=$ & $a$\\
+$bA$ & $::=$ & $Ab$\\
\end{tabular}
\end{center}
@@ -50,8 +50,8 @@
\begin{center}
\begin{tabular}{lcl}
-$A$ & $\rightarrow$ & $0A1 \;|\; BB$\\
-$B$ & $\rightarrow$ & $\epsilon \;|\; 2B$
+$A$ & $::=$ & $0A1 \;|\; BB$\\
+$B$ & $::=$ & $\epsilon \;|\; 2B$
\end{tabular}
\end{center}
@@ -62,14 +62,14 @@
\begin{center}
\begin{tabular}{l}
-$S \rightarrow \texttt{if0} \cdot E \cdot \texttt{then} \cdot S$\\
-$S \rightarrow \texttt{print} \cdot S$\\
-$S \rightarrow \texttt{begin} \cdot B\cdot \texttt{end}$\\
-$B \rightarrow S\cdot \texttt{;}$\\
-$B \rightarrow S\cdot \texttt{;} \cdot B$\\
-$S \rightarrow num$\\
-$E \rightarrow num$\\
-$B \rightarrow num$
+$S ::= \texttt{if0} \cdot E \cdot \texttt{then} \cdot S$\\
+$S ::= \texttt{print} \cdot S$\\
+$S ::= \texttt{begin} \cdot B\cdot \texttt{end}$\\
+$B ::= S\cdot \texttt{;}$\\
+$B ::= S\cdot \texttt{;} \cdot B$\\
+$S ::= num$\\
+$E ::= num$\\
+$B ::= num$
\end{tabular}
\end{center}
@@ -82,14 +82,37 @@
\begin{center}
\begin{tabular}{rl}
-(i) & $S \rightarrow \texttt{if0} \cdot E\cdot \texttt{then} \cdot S\cdot \texttt{else} \cdot S$\\
-(ii) & $B \rightarrow B \cdot B$\\
-(iii) & $E \rightarrow ( \cdot E \cdot )$\\
-(iv) & $E \rightarrow E \cdot + \cdot E$
+(i) & $S ::= \texttt{if0} \cdot E\cdot \texttt{then} \cdot S\cdot \texttt{else} \cdot S$\\
+(ii) & $B ::= B \cdot B$\\
+(iii) & $E ::= ( \cdot E \cdot )$\\
+(iv) & $E ::= E \cdot + \cdot E$
\end{tabular}
\end{center}
+\item Suppose the string $``9-5+2''$. Give all ASTs that
+ the following two grammars generate for this string.
+
+Grammar 1, where List is the starting symbol:
+
+\begin{center}
+\begin{tabular}{lcl}
+$List$ & $::=$ & $List + Digit \mid List - Digit \mid Digit$\\
+$Digit$ & $::=$ & $0 \mid 1 \mid 2 \mid 3 \mid 4 \mid 5 \mid 6 \mid 7 \mid 8 \mid 9$
+\end{tabular}
+\end{center}
+
+Grammar 2, where String is the starting symbol:
+
+\begin{center}
+\begin{tabular}{@{}lcl@{}}
+ $String$ & $::=$ & $String + String \mid String - String \mid$\\
+ & & $0 \mid 1 \mid 2 \mid 3 \mid 4 \mid 5 \mid 6 \mid 7 \mid 8 \mid 9$
+\end{tabular}
+\end{center}
+
+
+
%\item {\bf (Optional)} The task is to match strings where the letters are in alphabetical order---for example,
%\texttt{abcfjz} would pass, but \texttt{acb} would not. Whitespace should be ignored---for example
%\texttt{ab c d} should pass. The point is to try to get the regular expression as short as possible!
--- a/progs/parser-combinators/comb1.sc Thu Nov 10 23:49:29 2022 +0000
+++ b/progs/parser-combinators/comb1.sc Tue Nov 15 11:34:33 2022 +0000
@@ -50,6 +50,7 @@
if (in != "" && in.head == c) Set((c, in.tail)) else Set()
}
+CharParser('c').parse("abc")
// an atomic parser for parsing strings according to a regex
import scala.util.matching.Regex
@@ -65,23 +66,27 @@
val NumParser = RegexParser("[0-9]+".r)
def StrParser(s: String) = RegexParser(Regex.quote(s).r)
+NumParser.parse("a123a123bc")
+StrParser("else").parse("eelsethen")
// NumParserInt transforms a "string integer" into a propper Int
// (needs "new" because MapParser is not a case class)
val NumParserInt = new MapParser(NumParser, (s: String) => s.toInt)
-
+NumParserInt.parse("123abc")
// the following string interpolation allows us to write
// StrParser(_some_string_) more conveniently as
//
// p"<_some_string_>"
+
implicit def parser_interpolation(sc: StringContext) = new {
def p(args: Any*) = StrParser(sc.s(args:_*))
}
-
+
+(p"else").parse("elsethen")
// more convenient syntax for parser combinators
implicit def ParserOps[I : IsSeq, T](p: Parser[I, T]) = new {
@@ -90,6 +95,10 @@
def map[S](f: => T => S) = new MapParser[I, T, S](p, f)
}
+def toU(s: String) = s.map(_.toUpper)
+
+(p"ELSE").map(toU(_)).parse("ELSEifthen")
+
// these implicits allow us to use an infix notation for
// sequences and alternatives; we also can write the usual
// map for a MapParser
@@ -101,9 +110,11 @@
val NumParserInt2 = NumParser.map(_.toInt)
+
+
// A parser for palindromes (just returns them as string)
lazy val Pal : Parser[String, String] = {
- ((p"a" ~ Pal) ~ p"a").map{ case ((x, y), z) => s"$x$y$z" } ||
+ (p"a" ~ Pal ~ p"a").map{ case ((x, y), z) => s"$x$y$z" } ||
(p"b" ~ Pal ~ p"b").map{ case ((x, y), z) => s"$x$y$z" } ||
p"a" || p"b" || p""
}
@@ -130,7 +141,7 @@
println(P.parse_all("()"))
// A parser for arithmetic expressions (Terms and Factors)
-
+{
lazy val E: Parser[String, Int] = {
(T ~ p"+" ~ E).map{ case ((x, _), z) => x + z } ||
(T ~ p"-" ~ E).map{ case ((x, _), z) => x - z } || T }
@@ -138,7 +149,7 @@
(F ~ p"*" ~ T).map{ case ((x, _), z) => x * z } || F }
lazy val F: Parser[String, Int] = {
(p"(" ~ E ~ p")").map{ case ((_, y), _) => y } || NumParserInt }
-
+}
println(E.parse_all("1+3+4"))
println(E.parse("1+3+4"))
println(E.parse_all("4*2+3"))
--- a/solutions/cw2/lexer.sc Thu Nov 10 23:49:29 2022 +0000
+++ b/solutions/cw2/lexer.sc Tue Nov 15 11:34:33 2022 +0000
@@ -196,6 +196,11 @@
}
}
+def ders_simp(cs: List[Char], r: Rexp) : Rexp = cs match {
+ case Nil => r
+ case c::cs => ders_simp(cs, simp(der(c, r))._1)
+}
+
def lexing_simp(r: Rexp, s: String) = env(lex_simp(r, s.toList))
// Language specific code
@@ -221,7 +226,7 @@
("w" $ WHITESPACE) |
("i" $ ID) |
("n" $ NUM) |
- ("c" $ COMMENT)).%
+ ("c" $ COMMENT)).%
def esc(raw: String): String = {
import scala.reflect.runtime.universe._
@@ -346,7 +351,10 @@
case ALT(r1, r2) => 1 + size(r1) + size (r2)
case SEQ(r1, r2) => 1 + size(r1) + size (r2)
case STAR(r1) => 1 + size(r1)
+ case PLUS(r1) => 1 + size(r1)
case NTIMES(r1, n) => 1 + size(r1)
+ case RECD(_, r1) => 1 + size(r1)
+ case RANGE(_) => 1
}
val reg = SEQ(STAR(STAR(CHAR('a'))), CHAR('b'))
@@ -355,3 +363,10 @@
size(simp(der('a', reg))._1) // 8
size(simp(der('a', der('a', reg)))._1) // 8
+
+
+// python tests
+println(size(WHILE_REGS))
+println(size(ders_simp("r".toList, WHILE_REGS)))
+println(size(ID))
+println(size(ders_simp("read".toList, ID)))