parse.go

// Copyright 2011 The Go Authors.  All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package syntax

import (
	"os"
	"sort"
	"strings"
	"unicode"
	"utf8"
)

// An Error describes a failure to parse a regular expression
// and gives the offending expression.
type Error struct {
	Code ErrorCode
	Expr string
}

func (e *Error) String() string {
	return "error parsing regexp: " + e.Code.String() + ": `" + e.Expr + "`"
}

// An ErrorCode describes a failure to parse a regular expression.
type ErrorCode string

const (
	// Unexpected error
	ErrInternalError ErrorCode = "regexp/syntax: internal error"

	// Parse errors
	ErrInvalidCharClass      ErrorCode = "invalid character class"
	ErrInvalidCharRange      ErrorCode = "invalid character class range"
	ErrInvalidEscape         ErrorCode = "invalid escape sequence"
	ErrInvalidNamedCapture   ErrorCode = "invalid named capture"
	ErrInvalidPerlOp         ErrorCode = "invalid or unsupported Perl syntax"
	ErrInvalidRepeatOp       ErrorCode = "invalid nested repetition operator"
	ErrInvalidRepeatSize     ErrorCode = "invalid repeat count"
	ErrInvalidUTF8           ErrorCode = "invalid UTF-8"
	ErrMissingBracket        ErrorCode = "missing closing ]"
	ErrMissingParen          ErrorCode = "missing closing )"
	ErrMissingRepeatArgument ErrorCode = "missing argument to repetition operator"
	ErrTrailingBackslash     ErrorCode = "trailing backslash at end of expression"
)

func (e ErrorCode) String() string {
	return string(e)
}

// Flags control the behavior of the parser and record information about regexp context.
type Flags uint16

const (
	FoldCase      Flags = 1 << iota // case-insensitive match
	Literal                         // treat pattern as literal string
	ClassNL                         // allow character classes like [^a-z] and [[:space:]] to match newline
	DotNL                           // allow . to match newline
	OneLine                         // treat ^ and $ as only matching at beginning and end of text
	NonGreedy                       // make repetition operators default to non-greedy
	PerlX                           // allow Perl extensions
	UnicodeGroups                   // allow \p{Han}, \P{Han} for Unicode group and negation
	WasDollar                       // regexp OpEndText was $, not \z
	Simple                          // regexp contains no counted repetition

	MatchNL = ClassNL | DotNL

	Perl        = ClassNL | OneLine | PerlX | UnicodeGroups // as close to Perl as possible
	POSIX Flags = 0                                         // POSIX syntax
)

// Pseudo-ops for parsing stack.
const (
	opLeftParen = opPseudo + iota
	opVerticalBar
)

type parser struct {
	flags       Flags     // parse mode flags
	stack       []*Regexp // stack of parsed expressions
	free        *Regexp
	numCap      int // number of capturing groups seen
	wholeRegexp string
	tmpClass    []int // temporary char class work space
}

func (p *parser) newRegexp(op Op) *Regexp {
	re := p.free
	if re != nil {
		p.free = re.Sub0[0]
		*re = Regexp{}
	} else {
		re = new(Regexp)
	}
	re.Op = op
	return re
}

func (p *parser) reuse(re *Regexp) {
	re.Sub0[0] = p.free
	p.free = re
}

// Parse stack manipulation.

// push pushes the regexp re onto the parse stack and returns the regexp.
func (p *parser) push(re *Regexp) *Regexp {
	// TODO: compute simple

	if re.Op == OpCharClass && len(re.Rune) == 2 && re.Rune[0] == re.Rune[1] {
		// Single rune.
		if p.maybeConcat(re.Rune[0], p.flags&^FoldCase) {
			return nil
		}
		re.Op = OpLiteral
		re.Rune = re.Rune[:1]
		re.Flags = p.flags &^ FoldCase
	} else if re.Op == OpCharClass && len(re.Rune) == 4 &&
		re.Rune[0] == re.Rune[1] && re.Rune[2] == re.Rune[3] &&
		unicode.SimpleFold(re.Rune[0]) == re.Rune[2] &&
		unicode.SimpleFold(re.Rune[2]) == re.Rune[0] ||
		re.Op == OpCharClass && len(re.Rune) == 2 &&
			re.Rune[0]+1 == re.Rune[1] &&
			unicode.SimpleFold(re.Rune[0]) == re.Rune[1] &&
			unicode.SimpleFold(re.Rune[1]) == re.Rune[0] {
		// Case-insensitive rune like [Aa] or [Δδ].
		if p.maybeConcat(re.Rune[0], p.flags|FoldCase) {
			return nil
		}

		// Rewrite as (case-insensitive) literal.
		re.Op = OpLiteral
		re.Rune = re.Rune[:1]
		re.Flags = p.flags | FoldCase
	} else {
		// Incremental concatenation.
		p.maybeConcat(-1, 0)
	}

	p.stack = append(p.stack, re)
	return re
}

// maybeConcat implements incremental concatenation
// of literal runes into string nodes.  The parser calls this
// before each push, so only the top fragment of the stack
// might need processing.  Since this is called before a push,
// the topmost literal is no longer subject to operators like *
// (Otherwise ab* would turn into (ab)*.)
// If r >= 0 and there's a node left over, maybeConcat uses it
// to push r with the given flags.
// maybeConcat reports whether r was pushed.
func (p *parser) maybeConcat(r int, flags Flags) bool {
	n := len(p.stack)
	if n < 2 {
		return false
	}

	re1 := p.stack[n-1]
	re2 := p.stack[n-2]
	if re1.Op != OpLiteral || re2.Op != OpLiteral || re1.Flags&FoldCase != re2.Flags&FoldCase {
		return false
	}

	// Push re1 into re2.
	re2.Rune = append(re2.Rune, re1.Rune...)

	// Reuse re1 if possible.
	if r >= 0 {
		re1.Rune = re1.Rune0[:1]
		re1.Rune[0] = r
		re1.Flags = flags
		return true
	}

	p.stack = p.stack[:n-1]
	p.reuse(re1)
	return false // did not push r
}

// newLiteral returns a new OpLiteral Regexp with the given flags
func (p *parser) newLiteral(r int, flags Flags) *Regexp {
	re := p.newRegexp(OpLiteral)
	re.Flags = flags
	re.Rune0[0] = r
	re.Rune = re.Rune0[:1]
	return re
}

// literal pushes a literal regexp for the rune r on the stack
// and returns that regexp.
func (p *parser) literal(r int) {
	p.push(p.newLiteral(r, p.flags))
}

// op pushes a regexp with the given op onto the stack
// and returns that regexp.
func (p *parser) op(op Op) *Regexp {
	re := p.newRegexp(op)
	re.Flags = p.flags
	return p.push(re)
}

// repeat replaces the top stack element with itself repeated
// according to op.
func (p *parser) repeat(op Op, min, max int, opstr, t, lastRepeat string) (string, os.Error) {
	flags := p.flags
	if p.flags&PerlX != 0 {
		if len(t) > 0 && t[0] == '?' {
			t = t[1:]
			flags ^= NonGreedy
		}
		if lastRepeat != "" {
			// In Perl it is not allowed to stack repetition operators:
			// a** is a syntax error, not a doubled star, and a++ means
			// something else entirely, which we don't support!
			return "", &Error{ErrInvalidRepeatOp, lastRepeat[:len(lastRepeat)-len(t)]}
		}
	}
	n := len(p.stack)
	if n == 0 {
		return "", &Error{ErrMissingRepeatArgument, opstr}
	}
	sub := p.stack[n-1]
	re := p.newRegexp(op)
	re.Min = min
	re.Max = max
	re.Flags = flags
	re.Sub = re.Sub0[:1]
	re.Sub[0] = sub
	p.stack[n-1] = re
	return t, nil
}

// concat replaces the top of the stack (above the topmost '|' or '(') with its concatenation.
func (p *parser) concat() *Regexp {
	p.maybeConcat(-1, 0)

	// Scan down to find pseudo-operator | or (.
	i := len(p.stack)
	for i > 0 && p.stack[i-1].Op < opPseudo {
		i--
	}
	subs := p.stack[i:]
	p.stack = p.stack[:i]

	// Empty concatenation is special case.
	if len(subs) == 0 {
		return p.push(p.newRegexp(OpEmptyMatch))
	}

	return p.collapse(subs, OpConcat)
}

// alternate replaces the top of the stack (above the topmost '(') with its alternation.
func (p *parser) alternate() *Regexp {
	// Scan down to find pseudo-operator (.
	// There are no | above (.
	i := len(p.stack)
	for i > 0 && p.stack[i-1].Op < opPseudo {
		i--
	}
	subs := p.stack[i:]
	p.stack = p.stack[:i]

	// Make sure top class is clean.
	// All the others already are (see swapVerticalBar).
	if len(subs) > 0 {
		cleanAlt(subs[len(subs)-1])
	}

	// Empty alternate is special case
	// (shouldn't happen but easy to handle).
	if len(subs) == 0 {
		return p.push(p.newRegexp(OpNoMatch))
	}

	return p.collapse(subs, OpAlternate)
}

// cleanAlt cleans re for eventual inclusion in an alternation.
func cleanAlt(re *Regexp) {
	switch re.Op {
	case OpCharClass:
		re.Rune = cleanClass(&re.Rune)
		if len(re.Rune) == 2 && re.Rune[0] == 0 && re.Rune[1] == unicode.MaxRune {
			re.Rune = nil
			re.Op = OpAnyChar
			return
		}
		if len(re.Rune) == 4 && re.Rune[0] == 0 && re.Rune[1] == '\n'-1 && re.Rune[2] == '\n'+1 && re.Rune[3] == unicode.MaxRune {
			re.Rune = nil
			re.Op = OpAnyCharNotNL
			return
		}
		if cap(re.Rune)-len(re.Rune) > 100 {
			// re.Rune will not grow any more.
			// Make a copy or inline to reclaim storage.
			re.Rune = append(re.Rune0[:0], re.Rune...)
		}
	}
}

// collapse pushes the result of applying op to sub
// onto the stack.  If sub contains op nodes, they all
// get flattened into a single node.
// sub points into p.stack so it cannot be kept.
func (p *parser) collapse(subs []*Regexp, op Op) *Regexp {
	if len(subs) == 1 {
		return p.push(subs[0])
	}
	re := p.newRegexp(op)
	re.Sub = re.Sub0[:0]
	for _, sub := range subs {
		if sub.Op == op {
			re.Sub = append(re.Sub, sub.Sub...)
			p.reuse(sub)
		} else {
			re.Sub = append(re.Sub, sub)
		}
	}
	return p.push(re)
}

func literalRegexp(s string, flags Flags) *Regexp {
	re := &Regexp{Op: OpLiteral}
	re.Flags = flags
	re.Rune = re.Rune0[:0] // use local storage for small strings
	for _, c := range s {
		if len(re.Rune) >= cap(re.Rune) {
			// string is too long to fit in Rune0.  let Go handle it
			re.Rune = []int(s)
			break
		}
		re.Rune = append(re.Rune, c)
	}
	return re
}

// Parsing.

func Parse(s string, flags Flags) (*Regexp, os.Error) {
	if flags&Literal != 0 {
		// Trivial parser for literal string.
		if err := checkUTF8(s); err != nil {
			return nil, err
		}
		return literalRegexp(s, flags), nil
	}

	// Otherwise, must do real work.
	var (
		p          parser
		err        os.Error
		c          int
		op         Op
		lastRepeat string
		min, max   int
	)
	p.flags = flags
	p.wholeRegexp = s
	t := s
	for t != "" {
		repeat := ""
	BigSwitch:
		switch t[0] {
		default:
			if c, t, err = nextRune(t); err != nil {
				return nil, err
			}
			p.literal(c)

		case '(':
			if p.flags&PerlX != 0 && len(t) >= 2 && t[1] == '?' {
				// Flag changes and non-capturing groups.
				if t, err = p.parsePerlFlags(t); err != nil {
					return nil, err
				}
				break
			}
			p.numCap++
			p.op(opLeftParen).Cap = p.numCap
			t = t[1:]
		case '|':
			if err = p.parseVerticalBar(); err != nil {
				return nil, err
			}
			t = t[1:]
		case ')':
			if err = p.parseRightParen(); err != nil {
				return nil, err
			}
			t = t[1:]
		case '^':
			if p.flags&OneLine != 0 {
				p.op(OpBeginText)
			} else {
				p.op(OpBeginLine)
			}
			t = t[1:]
		case '$':
			if p.flags&OneLine != 0 {
				p.op(OpEndText).Flags |= WasDollar
			} else {
				p.op(OpEndLine)
			}
			t = t[1:]
		case '.':
			if p.flags&DotNL != 0 {
				p.op(OpAnyChar)
			} else {
				p.op(OpAnyCharNotNL)
			}
			t = t[1:]
		case '[':
			if t, err = p.parseClass(t); err != nil {
				return nil, err
			}
		case '*', '+', '?':
			switch t[0] {
			case '*':
				op = OpStar
			case '+':
				op = OpPlus
			case '?':
				op = OpQuest
			}
			if t, err = p.repeat(op, min, max, t[:1], t[1:], lastRepeat); err != nil {
				return nil, err
			}
		case '{':
			op = OpRepeat
			min, max, tt, ok := p.parseRepeat(t)
			if !ok {
				// If the repeat cannot be parsed, { is a literal.
				p.literal('{')
				t = t[1:]
				break
			}
			if t, err = p.repeat(op, min, max, t[:len(t)-len(tt)], tt, lastRepeat); err != nil {
				return nil, err
			}
		case '\\':
			if p.flags&PerlX != 0 && len(t) >= 2 {
				switch t[1] {
				case 'A':
					p.op(OpBeginText)
					t = t[2:]
					break BigSwitch
				case 'b':
					p.op(OpWordBoundary)
					t = t[2:]
					break BigSwitch
				case 'B':
					p.op(OpNoWordBoundary)
					t = t[2:]
					break BigSwitch
				case 'C':
					// any byte; not supported
					return nil, &Error{ErrInvalidEscape, t[:2]}
				case 'Q':
					// \Q ... \E: the ... is always literals
					var lit string
					if i := strings.Index(t, `\E`); i < 0 {
						lit = t[2:]
						t = ""
					} else {
						lit = t[2:i]
						t = t[i+2:]
					}
					p.push(literalRegexp(lit, p.flags))
					break BigSwitch
				case 'z':
					p.op(OpEndText)
					t = t[2:]
					break BigSwitch
				}
			}

			re := p.newRegexp(OpCharClass)
			re.Flags = p.flags

			// Look for Unicode character group like \p{Han}
			if len(t) >= 2 && (t[1] == 'p' || t[1] == 'P') {
				r, rest, err := p.parseUnicodeClass(t, re.Rune0[:0])
				if err != nil {
					return nil, err
				}
				if r != nil {
					re.Rune = r
					t = rest
					p.push(re)
					break BigSwitch
				}
			}

			// Perl character class escape.
			if r, rest := p.parsePerlClassEscape(t, re.Rune0[:0]); r != nil {
				re.Rune = r
				t = rest
				p.push(re)
				break BigSwitch
			}
			p.reuse(re)

			// Ordinary single-character escape.
			if c, t, err = p.parseEscape(t); err != nil {
				return nil, err
			}
			p.literal(c)
		}
		lastRepeat = repeat
	}

	p.concat()
	if p.swapVerticalBar() {
		// pop vertical bar
		p.stack = p.stack[:len(p.stack)-1]
	}
	p.alternate()

	n := len(p.stack)
	if n != 1 {
		return nil, &Error{ErrMissingParen, s}
	}
	return p.stack[0], nil
}

// parseRepeat parses {min} (max=min) or {min,} (max=-1) or {min,max}.
// If s is not of that form, it returns ok == false.
func (p *parser) parseRepeat(s string) (min, max int, rest string, ok bool) {
	if s == "" || s[0] != '{' {
		return
	}
	s = s[1:]
	if min, s, ok = p.parseInt(s); !ok {
		return
	}
	if s == "" {
		return
	}
	if s[0] != ',' {
		max = min
	} else {
		s = s[1:]
		if s == "" {
			return
		}
		if s[0] == '}' {
			max = -1
		} else if max, s, ok = p.parseInt(s); !ok {
			return
		}
	}
	if s == "" || s[0] != '}' {
		return
	}
	rest = s[1:]
	ok = true
	return
}

// parsePerlFlags parses a Perl flag setting or non-capturing group or both,
// like (?i) or (?: or (?i:.  It removes the prefix from s and updates the parse state.
// The caller must have ensured that s begins with "(?".
func (p *parser) parsePerlFlags(s string) (rest string, err os.Error) {
	t := s

	// Check for named captures, first introduced in Python's regexp library.
	// As usual, there are three slightly different syntaxes:
	//
	//   (?P<name>expr)   the original, introduced by Python
	//   (?<name>expr)    the .NET alteration, adopted by Perl 5.10
	//   (?'name'expr)    another .NET alteration, adopted by Perl 5.10
	//
	// Perl 5.10 gave in and implemented the Python version too,
	// but they claim that the last two are the preferred forms.
	// PCRE and languages based on it (specifically, PHP and Ruby)
	// support all three as well.  EcmaScript 4 uses only the Python form.
	//
	// In both the open source world (via Code Search) and the
	// Google source tree, (?P<expr>name) is the dominant form,
	// so that's the one we implement.  One is enough.
	if len(t) > 4 && t[2] == 'P' && t[3] == '<' {
		// Pull out name.
		end := strings.IndexRune(t, '>')
		if end < 0 {
			if err = checkUTF8(t); err != nil {
				return "", err
			}
			return "", &Error{ErrInvalidNamedCapture, s}
		}

		capture := t[:end+1] // "(?P<name>"
		name := t[4:end]     // "name"
		if err = checkUTF8(name); err != nil {
			return "", err
		}
		if !isValidCaptureName(name) {
			return "", &Error{ErrInvalidNamedCapture, capture}
		}

		// Like ordinary capture, but named.
		p.numCap++
		re := p.op(opLeftParen)
		re.Cap = p.numCap
		re.Name = name
		return t[end+1:], nil
	}

	// Non-capturing group.  Might also twiddle Perl flags.
	var c int
	t = t[2:] // skip (?
	flags := p.flags
	sign := +1
	sawFlag := false
Loop:
	for t != "" {
		if c, t, err = nextRune(t); err != nil {
			return "", err
		}
		switch c {
		default:
			break Loop

		// Flags.
		case 'i':
			flags |= FoldCase
			sawFlag = true
		case 'm':
			flags &^= OneLine
			sawFlag = true
		case 's':
			flags |= DotNL
			sawFlag = true
		case 'U':
			flags |= NonGreedy
			sawFlag = true

		// Switch to negation.
		case '-':
			if sign < 0 {
				break Loop
			}
			sign = -1
			// Invert flags so that | above turn into &^ and vice versa.
			// We'll invert flags again before using it below.
			flags = ^flags
			sawFlag = false

		// End of flags, starting group or not.
		case ':', ')':
			if sign < 0 {
				if !sawFlag {
					break Loop
				}
				flags = ^flags
			}
			if c == ':' {
				// Open new group
				p.op(opLeftParen)
			}
			p.flags = flags
			return t, nil
		}
	}

	return "", &Error{ErrInvalidPerlOp, s[:len(s)-len(t)]}
}

// isValidCaptureName reports whether name
// is a valid capture name: [A-Za-z0-9_]+.
// PCRE limits names to 32 bytes.
// Python rejects names starting with digits.
// We don't enforce either of those.
func isValidCaptureName(name string) bool {
	if name == "" {
		return false
	}
	for _, c := range name {
		if c != '_' && !isalnum(c) {
			return false
		}
	}
	return true
}

// parseInt parses a decimal integer.
func (p *parser) parseInt(s string) (n int, rest string, ok bool) {
	if s == "" || s[0] < '0' || '9' < s[0] {
		return
	}
	// Disallow leading zeros.
	if len(s) >= 2 && s[0] == '0' && '0' <= s[1] && s[1] <= '9' {
		return
	}
	for s != "" && '0' <= s[0] && s[0] <= '9' {
		// Avoid overflow.
		if n >= 1e8 {
			return
		}
		n = n*10 + int(s[0]) - '0'
		s = s[1:]
	}
	rest = s
	ok = true
	return
}

// can this be represented as a character class?
// single-rune literal string, char class, ., and .|\n.
func isCharClass(re *Regexp) bool {
	return re.Op == OpLiteral && len(re.Rune) == 1 ||
		re.Op == OpCharClass ||
		re.Op == OpAnyCharNotNL ||
		re.Op == OpAnyChar
}

// does re match r?
func matchRune(re *Regexp, r int) bool {
	switch re.Op {
	case OpLiteral:
		return len(re.Rune) == 1 && re.Rune[0] == r
	case OpCharClass:
		for i := 0; i < len(re.Rune); i += 2 {
			if re.Rune[i] <= r && r <= re.Rune[i+1] {
				return true
			}
		}
		return false
	case OpAnyCharNotNL:
		return r != '\n'
	case OpAnyChar:
		return true
	}
	return false
}

// parseVerticalBar handles a | in the input.
func (p *parser) parseVerticalBar() os.Error {
	p.concat()

	// The concatenation we just parsed is on top of the stack.
	// If it sits above an opVerticalBar, swap it below
	// (things below an opVerticalBar become an alternation).
	// Otherwise, push a new vertical bar.
	if !p.swapVerticalBar() {
		p.op(opVerticalBar)
	}

	return nil
}

// If the top of the stack is an element followed by an opVerticalBar
// swapVerticalBar swaps the two and returns true.
// Otherwise it returns false.
func (p *parser) swapVerticalBar() bool {
	// If above and below vertical bar are literal or char class,
	// can merge into a single char class.
	n := len(p.stack)
	if n >= 3 && p.stack[n-2].Op == opVerticalBar && isCharClass(p.stack[n-1]) && isCharClass(p.stack[n-3]) {
		re1 := p.stack[n-1]
		re3 := p.stack[n-3]
		// Make re3 the more complex of the two.
		if re1.Op > re3.Op {
			re1, re3 = re3, re1
			p.stack[n-3] = re3
		}
		switch re3.Op {
		case OpAnyChar:
			// re1 doesn't add anything.
		case OpAnyCharNotNL:
			// re1 might add \n
			if matchRune(re1, '\n') {
				re3.Op = OpAnyChar
			}
		case OpCharClass:
			// re1 is simpler, so either literal or char class
			if re1.Op == OpLiteral {
				re3.Rune = appendRange(re3.Rune, re1.Rune[0], re1.Rune[0])
			} else {
				re3.Rune = appendClass(re3.Rune, re1.Rune)
			}
		case OpLiteral:
			// both literal
			if re1.Rune[0] == re3.Rune[0] {
				break
			}
			re3.Op = OpCharClass
			re3.Rune = append(re3.Rune, re3.Rune[0])
			re3.Rune = appendRange(re3.Rune, re1.Rune[0], re1.Rune[0])
		}
		p.reuse(re1)
		p.stack = p.stack[:n-1]
		return true
	}

	if n >= 2 {
		re1 := p.stack[n-1]
		re2 := p.stack[n-2]
		if re2.Op == opVerticalBar {
			if n >= 3 {
				// Now out of reach.
				// Clean opportunistically.
				cleanAlt(p.stack[n-3])
			}
			p.stack[n-2] = re1
			p.stack[n-1] = re2
			return true
		}
	}
	return false
}

// parseRightParen handles a ) in the input.
func (p *parser) parseRightParen() os.Error {
	p.concat()
	if p.swapVerticalBar() {
		// pop vertical bar
		p.stack = p.stack[:len(p.stack)-1]
	}
	p.alternate()

	n := len(p.stack)
	if n < 2 {
		return &Error{ErrInternalError, ""}
	}
	re1 := p.stack[n-1]
	re2 := p.stack[n-2]
	p.stack = p.stack[:n-2]
	if re2.Op != opLeftParen {
		return &Error{ErrMissingParen, p.wholeRegexp}
	}
	if re2.Cap == 0 {
		// Just for grouping.
		p.push(re1)
	} else {
		re2.Op = OpCapture
		re2.Sub = re2.Sub0[:1]
		re2.Sub[0] = re1
		p.push(re2)
	}
	return nil
}

// parseEscape parses an escape sequence at the beginning of s
// and returns the rune.
func (p *parser) parseEscape(s string) (r int, rest string, err os.Error) {
	t := s[1:]
	if t == "" {
		return 0, "", &Error{ErrTrailingBackslash, ""}
	}
	c, t, err := nextRune(t)
	if err != nil {
		return 0, "", err
	}

Switch:
	switch c {
	default:
		if c < utf8.RuneSelf && !isalnum(c) {
			// Escaped non-word characters are always themselves.
			// PCRE is not quite so rigorous: it accepts things like
			// \q, but we don't.  We once rejected \_, but too many
			// programs and people insist on using it, so allow \_.
			return c, t, nil
		}

	// Octal escapes.
	case '1', '2', '3', '4', '5', '6', '7':
		// Single non-zero digit is a backreference; not supported
		if t == "" || t[0] < '0' || t[0] > '7' {
			break
		}
		fallthrough
	case '0':
		// Consume up to three octal digits; already have one.
		r = c - '0'
		for i := 1; i < 3; i++ {
			if t == "" || t[0] < '0' || t[0] > '7' {
				break
			}
			r = r*8 + int(t[0]) - '0'
			t = t[1:]
		}
		return r, t, nil

	// Hexadecimal escapes.
	case 'x':
		if t == "" {
			break
		}
		if c, t, err = nextRune(t); err != nil {
			return 0, "", err
		}
		if c == '{' {
			// Any number of digits in braces.
			// Perl accepts any text at all; it ignores all text
			// after the first non-hex digit.  We require only hex digits,
			// and at least one.
			nhex := 0
			r = 0
			for {
				if t == "" {
					break Switch
				}
				if c, t, err = nextRune(t); err != nil {
					return 0, "", err
				}
				if c == '}' {
					break
				}
				v := unhex(c)
				if v < 0 {
					break Switch
				}
				r = r*16 + v
				if r > unicode.MaxRune {
					break Switch
				}
				nhex++
			}
			if nhex == 0 {
				break Switch
			}
			return r, t, nil
		}

		// Easy case: two hex digits.
		x := unhex(c)
		if c, t, err = nextRune(t); err != nil {
			return 0, "", err
		}
		y := unhex(c)
		if x < 0 || y < 0 {
			break
		}
		return x*16 + y, t, nil

	// C escapes.  There is no case 'b', to avoid misparsing
	// the Perl word-boundary \b as the C backspace \b
	// when in POSIX mode.  In Perl, /\b/ means word-boundary
	// but /[\b]/ means backspace.  We don't support that.
	// If you want a backspace, embed a literal backspace
	// character or use \x08.
	case 'a':
		return '\a', t, err
	case 'f':
		return '\f', t, err
	case 'n':
		return '\n', t, err
	case 'r':
		return '\r', t, err
	case 't':
		return '\t', t, err
	case 'v':
		return '\v', t, err
	}
	return 0, "", &Error{ErrInvalidEscape, s[:len(s)-len(t)]}
}

// parseClassChar parses a character class character at the beginning of s
// and returns it.
func (p *parser) parseClassChar(s, wholeClass string) (r int, rest string, err os.Error) {
	if s == "" {
		return 0, "", &Error{Code: ErrMissingBracket, Expr: wholeClass}
	}

	// Allow regular escape sequences even though
	// many need not be escaped in this context.
	if s[0] == '\\' {
		return p.parseEscape(s)
	}

	return nextRune(s)
}

type charGroup struct {
	sign  int
	class []int
}

// parsePerlClassEscape parses a leading Perl character class escape like \d
// from the beginning of s.  If one is present, it appends the characters to r
// and returns the new slice r and the remainder of the string.
func (p *parser) parsePerlClassEscape(s string, r []int) (out []int, rest string) {
	if p.flags&PerlX == 0 || len(s) < 2 || s[0] != '\\' {
		return
	}
	g := perlGroup[s[0:2]]
	if g.sign == 0 {
		return
	}
	return p.appendGroup(r, g), s[2:]
}

// parseNamedClass parses a leading POSIX named character class like [:alnum:]
// from the beginning of s.  If one is present, it appends the characters to r
// and returns the new slice r and the remainder of the string.