syntax: don't treat '#' as a comment inside [[ ]] tests

The lexer treated '#' as the start of a comment whenever the parser
quote state was not unquotedWordCont, which meant that mid-word '#'
inside [[ ]] tests was wrongly consumed as a comment. This broke
expressions like [[ -n $foo#bar ]] and zsh glob qualifiers like
[[ -n ./(../)#(.jj)(#qN/) ]], where parsing fell off the end of the
test clause.

Allow '#' as a word continuation in testExpr too when adjacent to the
previous token, since '#' has no comment meaning inside [[ ]].

Fixes #1326.

Author: mvdan

syntax/filetests_test.go

@@ -4021,6 +4021,14 @@ var fileTests = []fileTestCase{
 			X:  litWord("a"),
 		}}, LangBash),
 	),
+	fileTest(
+		// '#' inside [[ ]] is part of a word, not a comment.
+		[]string{"[[ -n $foo#bar ]]"},
+		langFile(&TestClause{X: &UnaryTest{
+			Op: TsNempStr,
+			X:  word(litParamExp("foo"), lit("#bar")),
+		}}, LangBash|LangMirBSDKorn|LangZsh),
+	),
 	fileTest(
 		[]string{"[[ a =~ b ]]", "[[ a =~ b ]];"},
 		langFile(&TestClause{X: &BinaryTest{

syntax/lexer.go

@@ -294,7 +294,8 @@ skipSpace:
 		case '#':
 			// If we're parsing $foo#bar, ${foo}#bar, 'foo'#bar, or "foo"#bar,
 			// #bar is a continuation of the same word, not a comment.
-			if p.quote == unquotedWordCont && !p.spaced {
+			// The same applies inside [[ ]] tests, where '#' has no comment meaning.
+			if !p.spaced && (p.quote == unquotedWordCont || p.quote == testExpr) {
 				p.advanceLitNone(r)
 				return
 			}