Nex

Nex is a lexer similar to Lex/Flex that:

  • generates Go code instead of C code

  • integrates with Go’s yacc instead of YACC/Bison

  • supports UTF-8

  • supports nested structural regular expressions.

See Structural Regular Expressions by Rob Pike. I wrote this code to get acquainted with Go and also to explore some of the ideas in the paper. Also, I’ve always been meaning to implement algorithms I learned from a compilers course I took many years ago. Back then, we never coded them; merely understanding the theory was enough to pass the exam.

Go has a less general scanner package, but it is especially suited for tokenizing Go code.

Installation

$ export GOPATH=/tmp/go
$ go get github.com/blynn/nex

Example

One simple example in the Flex manual is a scanner that counts characters and lines. The program is similar in Nex:

/\n/{ nLines++; nChars++ }
/./{ nChars++ }
//
package main
import ("fmt";"os")
func main() {
  var nLines, nChars int
  NN_FUN(NewLexer(os.Stdin))
  fmt.Printf("%d %d\n", nLines, nChars)
}

The syntax resembles Awk more than Flex: each regex must be delimited. An empty regex terminates the rules section and signifies the presence of user code, which is printed on standard output with NN_FUN replaced by the generated scanner.

Name the above example lc.nex. Then compile and run it by typing:

$ nex -r -s lc.nex

The program runs on standard input and output. For example:

$ nex -r -s lc.nex < /usr/share/dict/words
99171 938587

To generate Go code for a scanner without compiling and running it, type:

$ nex -s < lc.nex  # Prints code on standard output.

or:

$ nex -s lc.nex  # Writes code to lc.nn.go

The NN_FUN macro is primitive, but I was unable to think of another way to achieve an Awk-esque feel. Purists unable to tolerate text substitution will need more code:

/\n/{ lval.l++; lval.c++ }
/./{ lval.c++ }
//
package main
import ("fmt";"os")
type yySymType struct { l, c int }
func main() {
  v := new(yySymType)
  NewLexer(os.Stdin).Lex(v)
  fmt.Printf("%d %d\n", v.l, v.c)
}

and must run nex without the -s option:

$ nex lc.nex

We could avoid defining a struct by using globals instead, but even then we need a throwaway definition of yySymType.

Toy Pascal

The Flex manual also exhibits a scanner for a toy Pascal-like language, though last I checked, its comment regex was a little buggy. Here is a modified Nex version, without string-to-number conversions:

/[0-9]+/          { println("An integer:", txt()) }
/[0-9]+\.[0-9]*/  { println("A float:", txt()) }
/if|then|begin|end|procedure|function/
                  { println( "A keyword:", txt()) }
/[a-z][a-z0-9]*/  { println("An identifier:", txt()) }
/\+|-|\*|\//      { println("An operator:", txt()) }
/[ \t\n]+/        { /* eat up whitespace */ }
/./               { println("Unrecognized character:", txt()) }
/{[^\{\}\n]*}/    { /* eat up one-line comments */ }
//
package main
import "os"
func main() {
  lex := NewLexer(os.Stdin)
  txt := func() string { return lex.Text() }
  NN_FUN(lex)
}

Enough simple examples! Let us see what nesting can do.

Peter into silicon

In “Structural Regular Expressions”, Pike imagines a newline-agnostic Awk that operates on matched text, rather than on the whole line containing a match, and writes code converting an input array of characters into descriptions of rectangles. For example, given an input such as:

    #######
   #########
  ####  #####
 ####    ####   #
 ####      #####
####        ###
########   #####
#### #########
#### #  # ####
## #  ###   ##
###    #  ###
###    ##
 ##   #
  #   ####
  # #
##   #   ##

we wish to produce something like:

rect 5 12 1 2
rect 4 13 2 3
rect 3 7 3 4
rect 9 14 3 4
...
rect 10 12 16 17

With Nex, we don’t have to imagine: such programs are real. Below are practical Nex programs that strongly resemble their theoretical counterparts. The one-character-at-a-time variant:

/ /{ x++ }
/#/{ println("rect", x, x+1, y, y+1); x++ }
/\n/{ x=1; y++ }
//
package main
import "os"
func main() {
  x, y := 1, 1
  NN_FUN(NewLexer(os.Stdin))
}

The one-run-at-a-time variant:

/ +/{ x+=len(txt()) }
/#+/{ println("rect", x, x+len(txt()), y, y+1); x+=len(txt()) }
/\n/{ x=1; y++ }
//
package main
import "os"
func main() {
  x, y := 1, 1
  lex := NewLexer(os.Stdin)
  txt := func() string { return lex.Text() }
  NN_FUN(lex)
}

The programs are more verbose than Awk because Go is the backend.

Rob but not robot

Pike demonstrates how nesting structural expressions leads to a few simple text editor commands to print all lines containing "rob" but not "robot". Though Nex fails to separate looping from matching, a corresponding program is bearable:

/[^\n]*\n/ < { isrobot = false; isrob = false }
  /robot/    { isrobot = true }
  /rob/      { isrob = true }
>            { if isrob && !isrobot { fmt.print(lex.Text()) } }
//
package main
import ("fmt";"os")
func main() {
  var isrobot, isrob bool
  lex := NewLexer(os.Stdin)
  NN_FUN(lex)
}

The "<" and ">" delimit nested expressions, and work as follows. On reading a line, we find it matches the first regex, so we execute the code immediately following the opening "<".

Then it’s as if we run Nex again, except we focus only on the patterns and actions up to the closing ">", with the matched line as the entire input. Thus we look for occurrences of "rob" and "robot" in just the matched line and set flags accordingly.

After the line ends, we execute the code following the closing ">" and return to our original state, scanning for more lines.

Word count

We can simultaneously count lines, words, and characters with Nex thanks to nesting:

/[^\n]*\n/ < {}
  /[^ \t\r\n]*/ < {}
    /./  { nChars++ }
  >      { nWords++ }
  /./    { nChars++ }
>        { nLines++ }
//
package main
import ("fmt";"os")
func main() {
  var nLines, nWords, nChars int
  NN_FUN(NewLexer(os.Stdin))
  fmt.Printf("%d %d %d\n", nLines, nWords, nChars)
}

The first regex matches entire lines: each line is passed to the first level of nested regexes. Within this level, the first regex matches words in the line: each word is passed to the second level of nested regexes. Within the second level, a regex causes every character of the word to be counted.

Lastly, we also count whitespace characters, a task performed by the second regex of the first level of nested regexes. We could remove this statement to count only non-whitespace characters.

UTF-8

The following Nex program converts Eastern Arabic numerals to the digits used in the Western world, and also Chinese phrases for numbers (the analog of something like "one-hundred and fifty-three") into digits.

/[零一二三四五六七八九十百千]+/ { fmt.Print(zhToInt(txt())) }
/[٠-٩]/ {
  // The above character class might show up right-to-left in a browser.
  // The equivalent of 0 should be on the left, and the equivalent of 9 should
  // be on the right.
  //
  // The Eastern Arabic numerals are ٠١٢٣٤٥٦٧٨٩.
  fmt.Print([]rune(txt())[0] - rune('٠'))
}
/./ { fmt.Print(txt()) }
//
package main
import ("fmt";"os")
func zhToInt(s string) int {
  n := 0
  prev := 0
  f := func(m int) {
    if 0 == prev { prev = 1 }
    n += m * prev
    prev = 0
  }
  for _, c := range s {
    for m, v := range []rune("一二三四五六七八九") {
      if v == c {
        prev = m+1
        goto continue2
      }
    }
    switch c {
    case '零':
    case '十': f(10)
    case '百': f(100)
    case '千': f(1000)
    }
continue2:
  }
  n += prev
  return n
}
func main() {
  lex := NewLexer(os.Stdin)
  txt := func() string { return lex.Text() }
  NN_FUN(lex)
}

nex and Go’s yacc

The parser generated by go tool yacc exports so little that it’s easiest to keep the lexer and the parser in the same package.

%{
package main
import "fmt"
%}

%union {
  n int
}

%token NUM
%%
input:    /* empty */
       | input line
;

line:     '\n'
       | exp '\n'      { fmt.Println($1.n); }
;

exp:     NUM           { $$.n = $1.n;        }
       | exp exp '+'   { $$.n = $1.n + $2.n; }
       | exp exp '-'   { $$.n = $1.n - $2.n; }
       | exp exp '*'   { $$.n = $1.n * $2.n; }
       | exp exp '/'   { $$.n = $1.n / $2.n; }
        /* Unary minus    */
       | exp 'n'       { $$.n = -$1.n;       }
;
%%

We must import fmt even if we don’t use it, since code generated by yacc needs it. Also, the %union is mandatory; it generates yySymType.

Call the above rp.y. Then a suitable lexer, say rp.nex, might be:

/[ \t]/  { /* Skip blanks and tabs. */ }
/[0-9]*/ { lval.n,_ = strconv.Atoi(yylex.Text()); return NUM }
/./ { return int(yylex.Text()[0]) }
//
package main
import ("os";"strconv")
func main() {
  yyParse(NewLexer(os.Stdin))
}

Compile the two with:

$ nex rp.nex && go tool yacc rp.y && go build y.go rp.nn.go

For brevity, we work in the main package. In a larger project we might want to write a package that exports a function wrapped around yyParse(). This is fine, provided the parser and the lexer are both in the same package.

Alternatively, we could use yacc’s -p option to change the prefix from yy to one that begins with an uppercase letter.

Matching the beginning and end of input

We can simulate awk’s BEGIN and END blocks with a regex that matches the entire input:

/.*/ < { println("BEGIN") }
  /a/  { println("a") }
>      { println("END") }
//
package main
import "os"
func main() {
  NN_FUN(NewLexer(os.Stdin))
}

However, this causes Nex to read the entire input into memory. To solve this problem, Nex supports the following syntax:

<      { println("BEGIN") }
  /a/  { println("a") }
>      { println("END") }
package main
import "os"
func main() {
  NN_FUN(NewLexer(os.Stdin))
}

In other words, if a bare < appears as the first pattern, then its action is executed before reading the input. The last pattern must be a bare >, and its action is executed on end of input.

Additionally, no empty regex is needed to mark the beginning of the Go program. (Fortunately, an empty regex is also a Go comment, so there’s no harm done if present.)

Matching Nuances

Among rules in the same scope, the longest matching pattern takes precedence. In event of a tie, the first pattern wins.

Unanchored patterns never match the empty string. For example,

/(foo)*/ {}

matches "foo" and "foofoo", but not "".

Anchored patterns can match the empty string at most once; after the match, the start or end null strings are "used up" so will not match again.

Internally, this is implemented by omitting the very first check to see if the current state is accepted when running the DFA corresponding to the regex. An alternative would be to simply ignore matches of length 0, but I chose to allow anchored empty matches just in case there turn out to be applications for them. I’m open to changing this behaviour.

Testing

Run:

$ go build nex.go
$ cd test
$ go test

Reference

func NewLexer(in io.Reader) *Lexer
// NewLexerWithInit creates a new Lexer object, runs the given callback on it,
// then returns it.
func NewLexerWithInit(in io.Reader, initFun func(*Lexer)) *Lexer
// Lex runs the lexer. Always returns 0.
// When the -s option is given, this function is not generated;
// instead, the NN_FUN macro runs the lexer.
      func (yylex *Lexer) Lex(lval *yySymType) int
// Text returns the matched text.
func (yylex *Lexer) Text() string
// Line returns the current line number.
// The first line is 0.
func (yylex *Lexer) Line() int
// Column returns the current column number.
// The first column is 0.
func (yylex *Lexer) Column() int

Ben Lynn blynn@cs.stanford.edu